Symposium F.MT07—Data Science and Automation to Accelerate Materials Development and Discovery

Machine leaning (ML) [1,2] has become popular method in materials design, and came to be used in various fields nowadays, especially in metal material [3,4] and organic material [5]. Actually, so far it was reported that ML have the effect of shorting the lead time of product development. By contrast, at the development site, there are multiple required characteristics. Because there is a trade-off relationship for multiple characteristics, it is difficult to achieve all of them in same time. Therefore, researchers will conduct experiments on the combinations of various ingredients and their compound ratios, and search for the optimal synthesis method by trial and error. However, it is not easy to find out optimum conditions breaking the trade-off from an enormous number of combinations, even if you use existing ML.
In this study, we verified the effectiveness of machine learning in developing thermosetting transparent flexible films. The thin films were obtained by cross-linking the pre-polymers synthesized by using the mixture of three diol ingredients and a single isocyanate ingredient. Here, we explored the optimum experimental conditions for maximizing transparency, elongation, and elastic modulus.
The predictive model was trained using 27 data obtained by the experimentalist. The hydroxyl values and Extended-connectivity fingerprints (ECFP) of pre-polymers were employed as the explanatory variables X [6,7], and the film properties were used as the objective variables y. The ECFP of each pre-polymers was calculated by summing up the ECFPs of all ingredients with weighted by mixing ratio.
We compared the number of trials required for improving the performance of thin films between that performed by the experimentalist and that by machine learning. As a result, the predictive model successfully found the high-performance films by a lower number of trials compared to the experimentalist.
Additionally, we also report the results by the above model applying Bayesian optimization.

＊This work was supported by a grant from the New Energy and Industrial Technology Development Organization (NEDO) of Japan (JPNP16010).

[1] K. Rajan. Materials Today, 8, 38 (2005)
[2] A. Jain, et al., APL Materials, 1, 011002 (2013)
[3] A. Agrawal, et al., Integr. Mater. Innovation 3, 1-19 (2014)
[4] Frang Ren, et al., Science Advances 4 4 1566 (2018)
[5] R. Gomez-Bombarelli. et al., Nature Materials, 15, 1120-1127 (2016)
[6] T. Minami, et al., MRS Advances, 3(49), 2975 (2018)
[7] T. Minami, et al., MRS Advances, 4(19), 1125 (2019)

Thermal conductivity is a fundamental material property but challenging to predict, with less than 5% out of about 10⁵ synthesized inorganic materials being documented. In this work, we extract the structural chemistry that governs lattice thermal conductivity, by combining graph neural networks and random forest approaches. We show that both mean and variation of unit-cell configurational properties, such as atomic volume and bond length, are the most important features, followed by mass and elemental electronegativity. We chart the structural chemistry of lattice thermal conductivity into extended van-Arkel triangles, and predict the thermal conductivity of all known inorganic materials in the Inorganic Crystal Structure Database (stoichiometric). For the latter, we develop a transfer learning framework extendable for other applications.

Effective data visualization accelerates the analysis and communication of data-driven science by enabling humans to interpret, comprehend, and form hypotheses around complex datasets. A large number of high-quality data visualization tools have emerged in recent years, however the adoption of proprietary formats or stand-alone software can hinder interoperability in the pursuit of FAIR data principles. We have developed a visualization authoring tool for the NanoMine knowledge graph (KG) database that combines the query language for RDF data (SPARQL) with a high-level specification for transforming raw data into interactive charts (Vega-Lite). This combination of open-source tools provides a glimpse into the broad collection of experimental materials data curated from the polymer nanocomposites literature into the NanoMine KG. Charts created using this tool are saved to the KG along with provenance information and made accessible via a Gallery on the NanoMine website. The Gallery allows visitors and new users to assess the available data in NanoMine, while advanced users can use the source code (SPARQL query, Vega-Lite spec) behind these charts as inspiration for their own contributions to the Gallery. The reusability of charts created with this approach allows a community of users to learn, build upon examples, and realize the rich expressiveness of SPARQL and Vega-Lite. Together, this data visualization authoring tool provides an interoperable, reusable, and extensible mechanism for users of NanoMine to inspect data quality, perform their own analyses, and pursue new research questions.

Recent advancements in computing technologies coupled with the need to make sense of large amounts of raw data have renewed much interest in data-driven materials properties modeling and prediction. Traditional materials science research relies heavily on experimental data to gauge the properties of materials. However, this paradigm is purely based on trial and error and ongoing research can take decades to discover new materials. Data-driven modeling tools such as machine learning and its proven libraries can help speed up the materials discovery process through the implementation of powerful algorithms on readily available materials datasets mined from the ever-increasing privately and government-funded materials databases. In this work, we applied various machine learning models on tens of hundreds of thermoelectric compounds obtained from Density Functional Theory (DFT) calculation results. In our preliminary analysis, we made use of pymatgen and the powerful materials science library matminer to add and explore key materials features that have the propensity to accurately predict our achievable target output. We evaluated the accuracy and performance of our models with K-Fold cross-validation estimators and identified the most important descriptors for our materials. Finally, we reviewed the current state-of-the-art in data-driven materials research, its prospects, and current challenges.

Colossal permittivity dielectric materials exhibit great potential as essential building blocks for future electronic devices and energy storage applications. Recent research has demonstrated that a giant dielectric constant could be achieved for certain materials with perovskite-like crystal structure (e.g. CaCu3Ti4O12) or by co-doping the transition metal oxides (e.g. In and Nb doped in TiO2). This opens up tremendous opportunities for inventing and tailoring materials with desired dielectric properties, while it also comes with significant challenges including a gigantic high-dimensional materials design space and unclear colossal permittivity mechanisms. Here, we leverage first-principles calculations and deep learning via the crystal graph convolutional neural network (CGCNN) method to predict and understand dielectric properties of a diverse array of materials. In doing so, we greatly accelerate the exploration and screening of candidate materials for colossal permittivity, and we clarify some of the complex relationships between dielectric properties and material structure and composition.

Data-driven material research can radically shorten time to discover new materials compared to traditional methods. However, collecting and classifying materials data is time consuming. A number of researchers focus on automation methods, however automated analysis of 2-dimensional X-ray diffraction (2D-XRD), which contains various information on structure and is beneficial to thin film research, is not well researched compared to conventional 1D-XRD. We consider automated cluster analysis of 2D-XRD with non-negative matrix factorization (NMF) and also visualization method with variational autoencoder (VAE). In this work, we analyzed 2D-XRD of (Ga_xIn_1-x)₂O₃ composition spread film fabricated on various crystal growth conditions.
For X-ray diffraction (XRD) data, which is the basic analysis method for structural analysis, identification of each peak in a XRD data and classifying each XRD data into some group are essential to understand the correlation between crystal structures and physical properties, but tough work. NMF is already reported to separate XRD dataset into clusters and, consequently, identifying few XRD patterns which are characteristic to each cluster is enough to understand the cluster analysis result [1]. However, the reported XRD data is from one composition spread film, so the cluster analysis result is based on the effects of composition differences on the crystal structure. Effects of crystal growth conditions such as substrates, growth temperatures, and reactive gases on cluster analysis result are still unknown. Also, for thin films, properties such as high oriented, mosaicity, and grained structures affect physical properties, which can be identified by 2D-XRD. Therefore, we have applied NMF to 2D-XRD, and as a case study, the 2D-XRD data of (Ga_xIn_1-x)₂O₃ composition spread samples fabricated by pulsed laser deposition varying composition ratio, substrate, oxygen partial pressure, and substrate temperature were analyzed. We confirmed that NMF can automatically separate the 2D-XRD data into some clusters corresponding to single crystalline, oriented polycrystalline, grained structures, etc. In addition, we applied VAE for converting differences among 2D-XRD data into a graphical image according to physical properties and growth conditions. VAE can convert input data into latent variables following some probability distribution, so differences in the input data is represented as the different positions in the distribution. Analyzing relationships among probability distribution of converted 2D-XRD data, cluster analysis result from NMF, and physical condition will guide us to improved fabrication process based on the obtained graphical image.

[1] C. J. Long, D. Bunker, X. Li, V. L. Karen, and I. Takeuchi, Review of Scientific Instruments 80, 103902 (2009)

In order to make accurate predictions of crystalline materials properties, current machine-learning approaches generally require large amounts of data. However, in practice, problems that could benefit from machine learning are intrinsically slow and thus small in data size. In this work, a novel all-round framework is presented which relies on a feedforward neural network and the selection of a limited set of physically-meaningful and property-relevant descriptors. Moreover, it enables the prediction of multiple properties, such as temperature or pressure curves, by using a novel tree-like architecture. This does slightly improve accuracy by taking advantage of information stored across different properties, known as joint transfer learning.
Next to being faster in terms of training time, this approach is shown to outperform current graph-network models on small datasets.
Furthermore, excellent accuracy is found when using the composition only, very helpful for identifying experimental candidates. In particular, the vibrational entropy at 305K of crystals is predicted with a mean absolute error of 0.01 meV/K/atom (four times lower than previous studies).
Finally, next to improving prediction performances, the selection algorithm highlights the most important features and thus helps understanding the underlying physics.

New embedded atom method (EAM) potentials have been developed to describe the interatomic interaction for face-centered cubic (FCC) gold represented by all-atom and coarse-grained (CG) models using particle swarm optimization (PSO) algorithm. These EAM potentials show good physical properties such as lattice constant, bulk modulus, and elastic constant, in agreement with the experimental and computational data. The reliability of each EAM potential is assessed using Gaussian Process (GP) regression and Bayesian statistical framework. The GP model predicted quantities of interest (QOI) such as surface energies, density, and vacancy formation energy, show good consistency with QOI obtained from molecular dynamics simulation, which indicates the GP model is good to reduce the computational cost significantly. Further, the GP model is used to develop 10,000 data points for the Bayesian framework. The Bayesian analysis shows that this EAM can predicate the physical properties with a high confidence level. The above optimization framework will be extended to develop the interatomic potentials for other metals and their alloys in a more efficient way.

Nanocomposites are at the forefront of materials research as they could meet the challenges and growing needs of diverse strategic fields from energy storage to transportation to advanced manufacturing. The investigation of the mechanical properties of such materials requires the combination of novel experimental microscale testing combined with labor/time-intensive finite element computations. In this talk, we will present a new approach for ceramic nanocomposite materials characterization based on the latest developments in deep learning. Specifically, we will discuss how multi-fidelity deep learning can significantly accelerate fracture toughness characterizations at small scales, without the need for large finite element-generated datasets. The predictive capabilities of this method will be accessed by direct comparisons with single-fidelity machine learning.

Machine Learning (ML) has been more and more widely used in accelerating the design of new materials due to its close accuracy and less computational cost in predicting material properties compared to simulation methods. However, existing ML methods for designing crystalline materials are usually not flexible to handle different crystal types or hard to interpret because of the arbitrary size. Although there are more generalized models such as Crystal Graph Convolutional Neural Networks (CGCNN) that utilize graph CNN to encode the structure of a crystal, they are over-localized and impossible to learn beyond the bond connectivity. Here we show a stronger and more adaptive framework, based on spectral convolutional neural network, which is not only capable for arbitrary crystal structure but also learns a unique graph representation based on both global and local topology of each crystal.

First a crystal graph is produced base on the cif file of the crystal similar to CGCNN, and each node is encoded with its atomic properties such as electronegativity, covalent radius, etc. The connectivity is determined based on the distance between each atom. The graph Laplacian L is then obtained by performing graph Fourier transformation on the crystal graph. The transformation process can be decomposed by a complete set of eigenvectors U and the eigenvalue matrix LAMBDA(λ), which represents the topology of the graph. The spectral filter g_θ(λ) generates a customized convolution kernel on the graph in space which can be formulated by:

g_θ(λ) =(from k=0 to k=k-1)∑θ_kλ^k
Where k can be interpreted as the number of convolutions in space.

However, this graph convolution cannot exploit the topological property of the graph. It is possible that the disconnected nodes have larger correlation than those connected nodes. To achieve the capability to learn from the geometric structure, we adapted a new spectral filter from Li et al.¹ which takes a non-Euclean metric so that it learns a residual graph apart from the intrinsic crystal graph by making a correction on the original graph Laplacian L:

L_new = L + αL_res

The overall intrinsic plus residue graph is trained with the spectral graph convoluted neural network to predict chemical properties of crystal and results are compared with various existing models. The proposed method has achieved better prediction accuracy especially in larger crystals with more than 4 elements.

Further work has been done by encoding the momentum space feature of the crystal with a graph of reciprocal lattice. Both graphs in real space and momentum space are trained with spectral graph convoluted neural network, but with their own parameters. The outcomes are concatenated and used to make predictions with a neural network. The results are further improved than the model with only the real space graph trained.

[1] Li, R., Wang, S., Zhu, F. and Huang, J., 2018, April. Adaptive graph convolutional neural networks. In Thirty-second AAAI conference on artificial intelligence.

Many next-generation technologies rely on increasing the capacity of batteries. Although lithium-ion batteries have attracted the most attention, batteries that shuttle anions like fluoride remain relatively unexplored. Anion shuttle batteries may offer high capacity to enable next-generation technologies and may be made from abundant materials. The promise of anion shuttle batteries warrants a comprehensive investigation of phase-space to discover new materials that can intercalate anions. In the past, materials discovery was done exclusively through experiment, often tedious and time-consuming work. Within the past 20 years, many methods have been developed to explore phase-space more efficiently by computationally predicting structures and calculating properties. With the goal of predicting new materials for anion intercalation, we have developed a computational method of rapid stable and metastable structure prediction using known chemical intuition rules and well-established structure prediction methods. This method uses an evolutionary algorithm to predict complete phase diagrams combined with density functional theory calculations to optimize each predicted structure and obtain an accurate energy. Stable and metastable phases of similar chemical systems can then be predicted based on the calculated phase diagram, with similarity of systems quantified using a previously published measure of atomic similarity. Using this method, we can rapidly predict novel, synthesizable structures that have desirable properties for various applications, including next-generation batteries.

When applying deep learning in material characterization tasks, generating and manually labeling experimental data can be expensive and time-consuming [1]. Especially, this is not an option when large amounts of data are necessary to obtain accurate results from a machine learning model. For this reason, the idea of training deep learning models with simulated data is appealing to tackle data scarcity, particularly when the simulation data can be generated easily and cheaply. A recent study about XRD pattern classification [2] found that the simulated XRD patterns provide some but a limited improvement on accuracy due to the differences between simulated data and experimental data. To close this gap, we explore a Generative Adversarial Network (GAN) [3] to augment the simulated XRD patterns with “realistic” features from experimental data, so that the training of the highly-accurate classifier can be completed with the simulation data only. The GAN consists of a refiner network and a discriminative network with an adversarial loss, where the refiner network reintroduces those realistic features, and the discriminative network penalizes large changes between the original simulated data and the refined data. To avoid the instability of two competing networks in the refined XRD patterns, the discriminative network in our study uses a history of the refined patterns rather than only ones from the current batch. By doing this, our GAN model has a longer-term memory to prevent the refiner network from reintroducing errors. First, we are building and testing this GAN model for data augmentation with 2000 simulated data with labels and 88 unlabeled experimental data, generated following the protocol established in Ref. [2]. Second, we will compare the prediction accuracy of classifying the space groups of the XRD patterns with the model training model using three types of training data: the original simulation data only, the original simulation and experimental data, and the refined simulation data. Third, following the protocol established in Ref. [2], we will adopt the class activation map (CAM) method to identify and understand the root causes of classification error when trained by different training datasets. We envision the method is generalizable to tackle the data scarcity in materials research, where large labeled experimental datasets are not readily available, and further accelerate many characterization tasks with machine learning.

References:
[1] Christopher Bowles et al. Gan augmentation: Augmenting training data using generative adversarial networks. ArXiv, abs/1810.10863, 2018.

[2] Oviedo, F., Ren, Z., Sun, S. et al. Fast and interpretable classification of small X-ray diffraction datasets using data augmentation and deep neural networks. npj Comput Mater 5, 60 (2019). https://doi.org/10.1038/s41524-019-0196-x

[3] A. Shrivastava, T. Pfister, O. Tuzel, J. Susskind, W. Wang, and R. Webb. Learning from simulated and unsupervised images through adversarial training. arXiv preprint arXiv:1612.07828, 2016.

Monte Carlo simulation is a standard tool for the investigation of the thermodynamic properties of materials. However, the usage of these methods in a high-throughput framework is limited by the efficiency of the current algorithms. The problem becomes even more challenging when considering the difficulty of modeling the multi-component systems at the forefront of many materials disciplines. Recent advances in autoregressive generative models have introduced new sampling methods to optimize the simulation of spin systems through Neural Markov Chain Monte Carlo (NMCMC) and Neural Importance Sampling (NIS). We will demonstrate the applicability of these methods to the study of lattice models in materials science. In particular, we will present the performance of these approaches on benchmark systems, including Copper-Gold and Nickel-Gold. Furthermore, we will discuss results when extending the model’s possible state-space beyond two components. Finally, we will evaluate the potential for using these machine learning led simulations in high-throughput screening platforms for perovskites.

The first reports of the application of the Boltzmann Sigmoidal model on pitch-dependency in patterned nanowires (NWs) of dilute nitride GaAsSbN on p-Si (111) substrates by self-catalyzed plasma-assisted molecular beam epitaxy are presented in this study. Sizeable bandgap tuning of ~75 meV, as ascertained from 4K photoluminescence (PL), over a pitch length variation of 200 nm to 1200 nm has been demonstrated. Axial and radial growth rates reveal a logistic sigmoidal growth trend with respect to pitch, which differs from those commonly observed in other patterned non-nitride III-V NWs. Sigmoidal modeling offers additional insight into the shift of the PL spectral arising from differences in Sb and N incorporation from pitch induced variation in secondary fluxes. Results signal that sigmoidal fitting can be a valuable tool to ascertain optimal pitch length for patterned NW arrays of dilute nitrides, and other highly mismatched alloys.

This work is financially supported by National Science Foundation (Award No. 1649517) and through the Title III HGBI PhD Fellowship. This work was performed at Joint School of Nanoscience and Nanoengineering, a member of the South eastern Nanotechnology Infrastructure Corridor (SENIC). The authors acknowledge the use of the Shared Materials Instrumentation Facility (SMIF) at Duke University.

High-throughput data generation and analysis via automation and applications of machine learning offer incredible potential for accelerating materials discovery and development, especially if used in conjunction with well-managed computational workflows. The open-source signac data management framework* (signac.io) enables researchers to maintain well-formed, reusable data spaces from early exploration through automated production runs and analysis on platforms ranging from laptops to supercomputers. This is achieved through a transparent data and workflow model combined with a simple and unobtrusive programmatic interface. The framework is application-agnostic, and has been applied in molecular simulations, quantum chemistry, photonics, computational fluid dynamics, machine learning, graph mining, and even organizing experimental data. Recently, the framework has been extended significantly with the addition of FlowGroups and aggregation to signac’s workflow model. We have added the ability to group operations on the data space for larger throughput on computer clusters and multiple independent resource specifications allowing for different resource requests on different computing platforms. In addition, we generalized operations from working with individual data points to operating on arbitrary subsets of an arbitrarily large data space. In this presentation, we show examples of applications of these new additions to signac that demonstrate the efficacy and versatility of signac in materials discovery and development.

*The signac framework is a NumFOCUS affiliated project.

Scientists have used deep learning algorithms to accelerate drug screening and design in the past few years. For example, transforming the organic compound structure into SMILES notation and leveraging the n-gram model have been used to compute the embedding for the downstream task. Most of these models are based on heavily hand-craft rules to preprocess the training data, and it remains challenging to encode the molecular property. In order to solve these problems, we propose a self-supervised method to learn the molecular representation from the molecular graph directly. We develop a novel approach to encode the information of bond rather than the node and use this embedding for membrane penetrating predictions. The self-supervised graph representation learning method is used to learn the molecular representation both hierarchically and as a whole, to understand the influence of the molecular size. Our model is evaluated with membrane-penetrating peptide database and our results show that this model can successfully predict the cell-penetrating peptides with less human heuristics.

We study the quasiparticle bandstructure and excitonic properties of 52 materials, which were recently shortlisted for their potential as a catalyst for CO photo-reduction through rigorous firstprinciples computation-based screening strategy. Many body perturbation theory within GW approximation has been used to explore the electronic structure of these materials. To inspect the excitonic effects we use state-of-the-art Bethe-Salpeter formalism. A high-throughput computational workflow using the “atomate” package was used to perform the GW-BSE calculations and analyze the results. We validate our results with 10 previously studied (using both theoretical and experimental methods) materials found in literature and report the results for 42 promising unexplored materials. In addition to providing a more accurate quasiparticle description of the electronic structure, our study further investigates the suitability of these materials in applications such as CO photo-reduction, efficient solar cells, etc. by examining their absorption spectra and excitonic properties.

The density of states (DOS) is a diagram to describe the electronic structure of the material and provides direct information on chemical bonding and materials properties. For example, electronic and photonic properties are mainly correlated to the DOS profiles around bandgap, and catalytic activity is also known to be correlated to the DOS profile of d-band[1]. The DOS is usually obtained by a density functional theory (DFT) simulations which are often time-consuming, especially in a large system. Machine learning (ML) methodologies have been used to create alternative models to predict some DFT outputs for overcoming this problem[2]. Besides, while a bond formation is a fundamental phenomenon in materials and molecules, changes in the orbital interactions and DOS profiles accompanying the bond formation are often complicated.
In this study, we have predicted the changes in DOS using ML techniques. Artificial neural network models have been developed to predict DOS of the atom combined system using DOS of the corresponding isolated system as inputs. We obtained successful results including prediction of gradual change of DOS with respect to bond distance.
First, the DOS database has been created by DFT computations on isolated atoms, diatomic molecules, and single-atom-adsorbed graphene and ethylene with various bond distances using various elements from H to Kr. Second, the ML models based on the artificial neural network has been developed and trained. Finally, the prediction of DOS has been demonstrated for unseen test species on each database.
All the first-principles DFT calculations are performed with the Vienna Ab-initio Simulation Package (VASP)[3] under the generalized gradient approximation (GGA) of Perdew-Burke-Ernzerhof (PBE). Ion cores are modeled with projector augmented wave (PAW) potentials. Semi-core orbital is included in valence. Cut-off energy of 500 eV is used. Spin polarization is considered. SCAN + rVV10[4] is used for considering van der Waals forces. All the structures and the cells are fixed and only electronic relaxation is performed.
For each system, a feedforward neural network model has been developed for DOS prediction. DOS of isolated atoms or adatom is used as input data and DOS of corresponding diatomic molecules or adsorbed systems is used as output data. the DOS of diatomic molecules is also used as input data only for adsorbed systems. The size of the dataset is 14750, 5694, and 2670 sets for diatomic molecules, adsorbed graphene, and adatom-ethylene system respectively. Each dataset is separated into 8:2 for train data and test data. Note that species used in the train data is not included in the test data. Mean squared error (MSE) is used as a loss function.
The DOS prediction for diatomic molecules which are the unseen test species have shown good agreement with calculated DOS. The prediction model reproduces change in shapes with respect to the bond distance. For instance, in the case of O-O molecule, characteristic changes in the DOS profile depending on the bond length would be predicted sufficiently. Main peaks appear around -14 eV in the most stable model with the bond distance of 1.2Å which is shifted to -10 eV in models with larger bond distance, and then it becomes a double-peak feature in the 4.0Å model. On the other hand, the accurate predictions for the adsorbed graphene and ethylene have achieved only when the DOS of the diatomic molecule (calculated) is served as the additional inputs for the prediction model. We are going to present the details in my presentation.
[Reference]
[1] B. Hammer and J.K.Nørskov, Adv. Catal. 45, 71 (2000).
[2] A. Chandrasekaran, D. Kamal, R. Batra et al., npj Comput. Mater. 5, 22 (2019).
[3] G. Kresse and J. Furthmüller, Comput. Mater. Sci. 6, 15 (1996).
[4] H. Peng et al., Phys. Rev. X 6, 041005 (2016).

For decades, fiber-reinforced composite materials have been integral to the aerospace, automotive, and military industries due to their lightweight properties. Fiber-reinforced composite manufacturing process involves the curing of the matrix material, which is typically a resin, polymer, or ceramic. The curing process is dependent on several factors, including humidity, temperature, and cycle time, altogether referred to as the curing environment. Curing environments are known to have a significant impact on the mechanical properties, such as modulus and strength, of the final product. While many studies have focused on predicting the mechanical properties of composites, the curing environment is usually not considered. In this work, a data-driven method is applied to various uni-directional carbon fiber laminates to investigate the effects of curing environments on mechanical properties such as longitudinal/transversal strength and modulus. We have conducted statistical and exploratory data analyses to identify trends, using data from the National Center for Advanced Materials Performance (NCAMP). Results show that high curing temperatures can yield stronger composites, and that the variability in material property values may be resistant to temperature for some resins but not all. Additionally, using machine learning techniques, we develop and compare regressive models considering varying curing environments to predict the strength and modulus of these materials in both longitudinal and transverse directions. This work establishes a statistical framework to analyze complex empirical data for both inference and insight for optimal designs.

Singlet fission has emerged as one of the most exciting phenomena known to improve the efficiencies of different types of solar cells and has found uses in diverse optoelectronic applications. The number of discovered singlet fission molecules is however limited, since molecules have to satisfy certain energy conditions in order to make the phenomenon possible. Recent advances in drug and molecular design have made use of state of the art Deep Learning models for inverse design that have enabled the prediction of drugs by optimizing targeted properties of interest.
These techniques have been particularly helpful in uncovering information from large datasets and using the learned information to generate new structures. However, unlike the field of drug discovery, we seldom encounter large datasets linking structures to properties in the field of materials science.
In our work, we tackle this scarcity of materials data with a transfer learning approach and demonstrate the utility of the inverse design approach to the design of singlet fission molecules.
Our approach is based on a deep generative model (Variational Autoencoder (VAE)) which has been inspired from Gomez-Bombarelli et al.’s work on de novo drug design. We attempt to extend the utility of their approach to Materials Science datasets (singlet fission in our case) that are often small in size.
We initially pre-train the VAE on an unsupervised task of SMILES string reconstruction on a large database (ChEMBL) of drug molecules. This pre-training allows the model to learn basic SMILES string grammar and the resulting trained weights can serve as a good initialization for the subsequent fine-tuning task on the singlet fission dataset. This can be thought of as similar to the common approach used in Computer Vision of pre-training models on the large ImageNet dataset and subsequently fine-tuning on various downstream tasks. This pre-training approach has been shown to remarkably improve the performance on downstream tasks, especially when the downstream datasets are small.
An important condition for the Singlet Fission phenomenon to occur is E(S1) - 2 x E(T1) >= 0. This condition ensures that Singlet Fission can proceed without the need for external thermal activation. So, we use E(S1) - 2 x E(T1) as the targeted property to optimize in our fine-tuning task. We fine-tune the pre-trained VAE model and an MLP (Multilayer Perceptron) regression head jointly on SMILES string reconstruction and property (E(S1) - 2 x E(T1)) regression tasks, on the Singlet Fission dataset. This guides the model to learn a latent (encoded) representation that segregates molecules based on the value of E(S1) - 2 x E(T1). We were then able to predict potential optimized molecules by performing targeted sampling from this learned latent space.
A second important condition for Singlet Fission to occur is E(T2) - 2 x E(T1) >= 0. This is to ensure that the generated triplet excitons do not annihilate each other immediately after their formation. We model this condition by training an MLP on the prediction of E(T2) - 2 x E(T1). We apply this trained model to the previously sampled molecules and eliminate the ones with negative valued predictions. Finally, we were left with four molecules that could potentially show Singlet Fission properties.
Our work demonstrates the formulation of a transfer learning based inverse design framework that can significantly minimize the efforts involved in the computational screening of molecules by allowing targeted traversal of a boundless search space. This approach can easily be extended to many classes of materials though we demonstrate it’s utility only to Singlet Fission molecules.

Physical unclonable function (PUF) is a cornerstone of anticounterfeiting. However, conventional PUF key-based secure tags encounter several bottlenecks, such as complicated manufacturing, specialized and tedious readout, long authentication time, and insufficient stability. To address these challenges, we utilize various two-dimensional materials (2DMs), including Ti₃C₂T_x MXene and graphene oxide, to construct multigenerational microstructures as PUF anticounterfeiting patterns. Two intermediate treatments, cation intercalation and moisture-induced lubrication, are introduced in between sequential substrate contractions to in situ engineer the multiscale patterns in a transfer-free and scalable fashion. To accelerate the authentication, a deep learning (DL)-facilitated software is developed to pre-categorize the hierarchical topographies with classifiable features. Afterwards, the search-and-compare is conducted within a smaller database to shorten the overall authentication time. The synergy between 2DM tags and DL-facilitated authentication software enables a reliable and environmentally stable anticounterfeiting technology, DeepKey, showing superior encoding capacity (>10^144,494) and short authentication time (~3.5 minutes). Our 2DM anticounterfeiting tag is finally integrated with QR codes to provide two-layer information security.

Prediction of the highest or lowest interfacial thermal conductance (ITC) across dissimilar materials is of high technological importance for effective heat management in nuclear and fusion reactors [1], thermal barrier coatings, nanoelectronics and thermoelectric devices [2]. It might take several decades to discover new materials using traditional methods such as physics-based modeling and testing of the pilot samples. In contrast, newly emerging and dynamically developing machine learning algorithms allow to accelerate this process by minimizing the trial-failure routine thanks to an availability of large number of measurement and simulation data generated over the years of active materials research. Today’s industry requires much higher rates of materials improvement in thermal and mechanical properties. In this regard, the application of machine learning algorithms allows to discover high-performance materials and interfaces with sufficient accuracy at the lowest consumption of materials and time to discovery.
In this work we present an application of Gaussian Process Regression (GPR) method applied to earlier published experimental reports of ITCs measured across various metal-nonmetal and nonmetal-nonmetal two-layer interfaces. The following physical descriptors of both, the film and the substrate, were considered: thermal conductivity k, Debye temperature θ_D, longitudinal sound velocity v_l, melting temperature T_m, mass density ρ, Young’s modulus E, molar volume V_µ. These parameters of the corresponding films and substrates were found to be substantially affecting the resulting ITC across them [3, 4].
Pearson correlation coefficient map was built to evaluate the linear correlation of different physical descriptors with the corresponding ITC value. The analysis based on the given data set showed that such film parameters as v_l, E and θ_D have the strongest correlation with the ITC.
The values of G of two-layered structures fell in the range of 1.5 to 1500 W/m²K. The tendency of the parameter’s ratio for the 5 cases of G values were compared. The results show that the ITC is maximum if the ratios of the film and the substrate descriptors are in the range of 0.7 - 1.7. Moreover, we see that matching of v_l, k , T_m and V_µ between films and substrates is of critical importance to maximize ITC values.
Acknowledgements
Funding support is acknowledged from grant AP05130446 and state-targeted program # BR05236454 by Kazakhstan Ministry of Education & Science and FDCR grant # 110119FD4501 by Nazarbayev University.

References
[1] M. Khafizov, V. Chauhan, Y. Wang, F. Riyad, N. Hang, D.H. Hurley, Investigation of thermal transport in composites and ion beam irradiated materials for nuclear energy applications, J. Mater. Res. 32 (1), 204–216 (2016).
[2] T. Zhan, L. Fang, and Y. Xu, Prediction of thermal boundary resistance by the machine learning method, Nature Scientific Reports 7, 7109 (2017).
[3] Y.-J. Wu, T. Zhan, Zh. Hou, L. Fang, and Y. Xu, Physical and chemical descriptors for predicting interfacial thermal resistance, Nature Scientific Data 7, 36 (2020).
[4] Ruxandra M. Costescu, Marcel A. Wall, and David G. Cahill, Thermal conductance of epitaxial interfaces, Phys. Rev. B 67, 054302 (2003).

Machine-learning-inspired computational approaches have accelerated the chemical composition space exploration hence motivating experimental laboratories to adopt this strategy for novel materials discovery. In this work, we build an extensive dataset based on density functional theory calculations using 2x2x2 supercells of halide and chalcogenide mixed cation Perovskites () with varying ratios of A/A′-cations mixture to account for the local distortions in the BX₆ octahedra. An automated workflow is employed to handle the DFT calculation, followed by the extraction of relevant physical properties and target features form the converged and verified density functional calculations (DFT) calculations. Subsequently, the dataset is built, curated and prepared using the chemical and structural features of the perovskites. This dataset is divided into two datasets sampled randomly: training and test dataset. Subsequently, this procedure is followed by a series of machine learning training based on a 5-fold cross-validation for each hyper-parameter combination to select the best hyper-parameters set until the model is optimized using the training dataset.
Here, we report the details of building a highly accurate predictor for halide and chalcogenide mixed cation Perovskites () bandgaps that takes into account the subtle structural and octahedral deformation a perovskite structure might experience upon mixing the organic cations. A high level of accuracy could be achieved with a careful investigation of the input features [1]. Our analysis emphasizes how important the feature selection stands for constructing the predictive model as we challenge the robustness of our machine learning predictor in a lab validation setup.
A model-to-lab validation is undertaken by synthesizing mixed-cation MA_xDMA_(1−x)PbI₃ perovskite thin films while varying the DMA concentration over a broad range, namely, for (1 − x)=0.0, 0.005, 0.01, 0.02, 0.05, 0.1, and 0.15. We find that the incorporation of DMA at concentrations of 5, 10, and 15 mol % results in non-perovskites in agreement with earlier findings. This indicates that beyond 2 mol % the local strain of DMA on MAPbI₃, due to its size, is too large to be accommodated within the perovskite structure. The experimental bandgaps of MA_xDMA_{(1 x)}PbI₃ thin films, as measured by UV–vis spectrophotometer analysis feature systematic bandgap increase with the DMA concentration, with modest variations of the order of 0.02 eV, which are expected given the small concentration range explored. Hence, the observed bandgap increase upon DMA doping is related to small local distortions of the PbI₆ octahedra and not to volumetric effects. Reverting back to our perovskite band gap predictor that have trained and tested over a limited number of DFT calculations for cation concentrations requiring moderate-sized supercells, the model enables to to map the bandgap of mixed-cation hybrid perovskites over a broad range of chemical and structural compositions not accessible by DFT. In this presentation, we will discuss not only the performances of our model but also the machine-learning training procedure to obtain a robust machine learning model.
Acknowledgements:
This work is sponsored by the Qatar National Research Fund (QNRF) through the National Priorities Research Program (NPRP8-090-2-047 & NPRP12S-0209-190063) and by the Qatar Environment and Energy Research Institute (FE). Computational resources have been provided by the research computing group at Texas A&M University at Qatar. We are grateful to QEERI core labs for the XRD and SEM characterizations.

[1] Park et al. “Importance of Structural Deformation Features in the Prediction of Hybrid Perovskite Bandgaps” Comput. Mater. Sci., 2020, 184, 109858 https://doi.org/10.1016/j.commatsci.2020.109858

Predicting the critical temperature Tc of superconductors is a notoriously difficult task, even for electron-phonon systems. We build on earlier efforts by McMillan and Allen and Dynes [1] to model Tc from various measures of the phonon spectrum and the electron-phonon interaction by using machine learning algorithms. Specifically, we use symbolic regression implemented in the Sure Independence and Sparsifying Operator (SISSO) method [2] to identify a new, physically interpretable equation for Tc as a function of a small number of physical quantities. We show that our model [3], trained using the relatively small Tc < 10K data tested by Allen and Dynes, improves upon the Allen-Dynes fit and can reasonably generalize to superconducting materials with higher Tc such as H3S. We find that recently-discovered superconducting systems with a more two-dimensional electron-phonon coupling are outliers in both the Allen-Dynes equation and our analytical expression, suggesting a need for a descriptor that also characterizes Fermi surface properties. Furthermore, we explore the use of variational autoencoders and Eliashberg theory to augment the data. By incorporating physical insights and constraints into a data-driven approach, we demonstrate that machine-learning methods can identify the relevant physical quantities and obtain predictive equations using small but high-quality datasets.

[1] P. B. Allen and R. C. Dynes, Phys. Rev. B12, 905 (1975).
[2] R. Ouyang, S. Curtarolo, E. Ahmetcik, M. Scheffler, and L. M. Ghiringhelli, Phys. Rev. Materials2, 083802(2018).
[3] S. R. Xie, G. R. Stewart, J. J. Hamlin, P. J. Hirschfeld, and R. G. Hennig, Phys. Rev. B100, 174513 (2019).

Novel approaches of Artificial Intelligence (AI) can find patterns and correlations in data that cannot be obtained from individual calculations or experiments and not even from high-throughput studies. In fact, data-driven research is adding a new research paradigm to the scientific landscape. For a real breakthrough, Open Data and sharing, as well as an efficient data infrastructure is key [1]. In other words, for shaping this forth paradigm and contributing to the development or discovery of improved and novel materials, data must be what is now called FAIR - Findable, Accessible, Interoperable, and Reusable [2].

The NOMAD Laboratory [3,4] is a living example for such infrastructure in computational materials science, comprising the NOMAD Repository (raw data) and its Archive (normalized, i.e. code-independent data). The NOMAD Encyclopedia is a web-based public platform that visualizes the results of this vast amount of calculations. The NOMAD Analytics Toolkit provides a collection of examples and tools to demonstrate how materials data can be turned into knowledge by AI approaches (e.g. [5,6]). I will give a guided tour through this data lab and discuss successes and challenges [7] ahead of us for exploiting the whole wealth of materials data – including experiment and sample synthesis.

[1] C. Draxl and M. Scheffler, Big-Data-Driven Materials Science and its FAIR Data Infrastructure, Invited Perspective in Handbook Andreoni W., Yip S. (eds) Handbook of Materials Modeling. Springer, Cham (2019).
[2] M. D. Wilkinson et al., The FAIR Guiding Principles for scientific data management and stewardship, Sci Data 3, 160018 (2016).
[3] The NOMAD Laboratory, https://nomad-coe.eu/, including metadata and various software for parsing, analysis, and visualization.
[4] C. Draxl and M. Scheffler, NOMAD: The FAIR Concept for Big-Data-Driven Materials Science, MRS Bulletin 43, 676 (2018).
[5] L. M. Ghiringhelli, J. Vybiral, S. V. Levchenko, C. Draxl, and M. Scheffler, Big Data of Materials Science - Critical Role of the Descriptor, Phys. Rev. Lett. 114, 105503 (2015).
[6] L. M. Ghiringhelli, J. Vybiral, E. Ahmetcik, R. Ouyang, S. V. Levchenko, C. Draxl, and M. Scheffler, Learning physical descriptors for materials science by compressed sensing, New J. Phys. 19, 023017 (2017).
[7] C. Draxl and M. Scheffler, The NOMAD Laboratory: From Data Sharing to Artificial Intelligence, J. Phys. Mater. 2, 036001 (2019).

Lead halide perovskites have in the last decade gained a large interest as a promising solar cell material and the number of published papers has now surpassed 15000. The sheer amount of papers makes it increasingly difficult to keep track of the field, and the lack of standards for describing, reporting, and sharing device data is something that seriously limit the pace of discovery. To deal with those problems, we have initiated the Perovskite Database Project where we aim at collecting all past and future perovskite device data in one place in a standardised machine-readable format. The project includes a web app, interactive graphics enabling anyone to explore and learn from the data, open source code, and mechanisms for collecting new data. The project is scheduled to go online during the autumn of 2020. In this talk we will discuss challenges and insights gained during this project including: how to motivate volunteers to gather previously published data, how do we get a sufficiently large part of a research community to share all their new data, how to organise data in a way that is simple to collect and robust to reporting errors, how do we ensure that the data is useful for as many as possible, and what can we really do with the data once we have it?

The materials science community has established that data-driven modeling, based on chemical composition, generates accurate and actionable results. Tools for performing this task have evolved from simple statistical methods to large endeavors containing complex neural-networks with millions of parameters. With each successive work, we have seen a marked improvement in property predictions with a trend towards raw model accuracy. However, model performance often includes additional considerations such as uncertainty quantification or interpretability.

In this talk, I will present our new model architecture, the Compositionally-Restricted Attention-Based Network (CrabNet). CrabNet generates high-fidelity predictions based on the self-attention mechanism, a fundamental component of the transformer architecture which revolutionized natural language processing. The transformer encoder uses self-attention to encode the context-dependent behavior for the components within a system. In physical environments, elements contribute differently to a material's property based on the materials system itself. For example, boron behaving as an electrical dopant in one system while behaving as a mechanical strengthening bond modification in another. CrabNet's ability to potentially capture this type of context-dependent behavior allows for highly accurate model predictions.

CrabNet does not stop there. Our modeling approach uses a robust loss function and model ensembling techniques to provide uncertainty values along with the property predictions. Importantly, we achieve this with a model architecture that generates simple and inspectable self-attention maps. These attention maps govern the learned material property by representing element importance and interactions. The visualization and analysis of these attention maps are available during training and inference periods.

As a result, CrabNet provides property prediction, uncertainty estimation, and the governing element interactions for every prediction. We expect that the ideas and capabilities of CrabNet will encourage additional research and discussion around the uses and benefits of our community's maturing data-driven approaches.

Stimulated by the Materials Genome Initiative (MGI), more databases with materials and materials properties are becoming publicly available. Among them, the databases are the Materials Project and the Harvard Clean Energy Project. Databases at the next scale - microstructure - are relatively less explored and consist of various application-specific and sparse datasets. This is because, among others, microstructural datasets are relatively expensive to obtain due to the high cost of sample preparation, limitations of imaging instruments, simulation cost, and individual labs' capabilities to generate a large volume of data. This limits machine learning's applicability and necessitates some form of data consolidation between various labs with the following challenges:
(i) data consolidation is hindered by the inherently diverse nature of materials data (as most data repositories focus on a specific subset of materials data) but also lab effects (e.g., different protocols of data processing or sample preparation);
(ii) microstructural data is sparse and noisy data due to incomplete observations or imprecise measurements.
In this talk, we present the microstructural database build on the principles of probabilistic and incomplete databases. Unlike a classic database, the probabilistic database can deal with uncertain, incomplete information-missing values or uncertain information (value of an attribute represented as probability density function over the set of possible value instead of a single value). They inherently handle problems that emerge from data integration from various sources that are diverse and often incompatible. We show two frameworks: vizier and mimir that are used to integrate and store the datasets, respectively.
To demonstrate our database's effectiveness, we showcase the integration of three data sets generated through three projects aimed to improve the efficiency of organic solar cells. In each project, the microstructures were analyzed to predict the solar cell efficiencies. However, the microstructures were annotated with a different set of descriptors. Moreover, different types of microstructures were analyzed (i.e., dendritic, spinodal, and wavy). We illustrate how the integrated, although incomplete, data is used to impute the missing data in the respective datasets. Finally, we showcase the framework's capabilities to mark database entry or imputed values in conflict with other entries or violating the physics-based constraints.

Data-driven R&D, especially machine learning (ML) and artificial intelligence (AI), are receiving lots of attention as avenues to accelerate new product development and to avoid/supplement costly lab experimentation. The pre-requisite for all these approaches is of course the availability of reliable high-quality data. However in real-life AI/ML projects an excess of time is spent aggregating and curating data, leading inevitably to dirty datasets and poor prediction results. Sadly, the IT infrastructure required to automate the provision of clean and properly contextualized data is often an afterthought.

This presentation gives an overview on our recent progress on an R&D IT infrastructure that solves the data provisioning problem in process-driven R&D. We will focus on three key areas: the automation of instrument data acquisition and their mass storage, combined querying across semantic and highly structured data, and the application of graph searching across chemistry experiments for reaction pathway mapping.

Managing instrument data is challenging not only because of their volume, but also because of their variability in data formats. Commonly, R&D IT systems capture incomplete data, e.g. by only saving processed files, or by placing them in inaccessible data silos where they can only provide limited use to data science efforts. We will discuss our approach that retains the full information content of all instrument data generated, while at the same time providing a framework to automate data processing and streamline lab operations.

Scientific data are notorious difficult to deal with in data modeling contexts. This is because of the combination of highly structured data, e.g. structure-property relationships, and unstructured semantic data that is generated as part of analyses, observations, conclusions, etc. Here, we will discuss an approach to embed semantic technologies into our scientific data platform to solve this complex problem.

Finally, we will discuss an application of graph technologies to mining reaction data that allows analysis of reaction pathways that span a multitude of experiments. This functionality is useful to compile information for patent submissions, or to optimize and discover synthetic routes based on existing IP.

With the increasing use of data-driven methodologies, concerns around data discovery, data access, and data interoperability have come to the forefront. Beginning in August 2019, communities have convened to work towards convergence of three complementary visions: (1) Digital Object Architecture, (2) Linked Data and Semantic Web, (3) FAIR Data Principles. This convergence has established the FAIR Digital Object Framework. This talk will overview these developments, summarize work within the NIST Material Measurement Laboratory to support adoption of the FAIR Digital Object Framework within the materials science and engineering community, and provide practical examples of how researchers can leverage these developments.

X-ray powder diffraction (powder XRD) is an important experimental characterization technique used to determine crystal structure. Current powder XRD analysis involves several time-consuming and error-prone human intervention steps including peak identification and indexing. In this work, we use one-dimensional Convolutional Neural Networks (1D CNNs) to perform the task of lattice parameter prediction, a critical step in the analysis pipeline.

We train independent regression models for each of the six crystal systems in order to take advantage of specific symmetry constraints on the lattice parameters. We perform simulations based on allowed Bragg angles for each crystal system in order to quantify best-case performance under conditions of perfect information. We show that, with sufficient data, it is possible to meet arbitrarily low mean percentage error (MPE) thresholds. Furthermore, the amount of data needed for perfect prediction scales directly with the number of independent lattice parameters in each crystal system. We also present results for simulated data based on the Cambridge Structural Database (CSD), containing over one million unique entries which span a very wide range of lattice parameter space. We achieve an MPE value of less than 10% for the cubic crystal system and MPE values of less than 20% for all other crystal systems. Validation of these results is performed in GSASII using standard refinement procedures.

This analysis builds on previous works focusing on classification tasks such as crystal system and space group prediction from powder XRD data. We hope that our results will further the eventual goal of high-throughput, real-time crystal structure analysis.

We have investigated structural transformations in the InGaO₃(ZnO)_m (IGZO) homologous series using optical and X-ray techniques. Following thermal processing via lateral gradient Laser Spike Annealing (lgLSA), structural transformation kinetics were observed as a function of composition (1<m<12), processing time (250 us - 10 ms), and peak temperature (500-1400°C). The high dimensionality of this composition and processing space dictates the use of automation in both sample generation and data analysis. Two complementary high-throughput and low-cost spatially resolved optical techniques were initially used to identify optical transitions that correlate strongly with structural transformations. As these samples are thin transparent films, the optical techniques are sensitive to extremely small changes in the optical path length due to physical thickness and optical index (phase) variations. This ability to efficiently prescreen the experimental phase space enabled automated and accelerated determination of structural transformations (if any) associated with these optically determined boundaries using X-ray diffraction. We propose a model for structural evolution along this IGZO pseudo-binary based on the ordering of In-O and (Ga,Zn)-O layers.

Instruments of scientific discovery (e.g., electron microscopes, scanning probe microscopes, and others) acquire vast collections of images that contain physical insight. Humans, however, struggle to search and draw correlations from enormous databases of images; thus, only a small fraction of the data collected is translated into knowledge. Researchers need an analytical toolbox that can create an image similarity latent manifold and allow visual interaction in an informative, comprehensive, and intuitive way. Here, we develop machine learning algorithms to create image similarity projections using a dataset of 25,000 piezoresponse force microscopy images. We implement a variety of pre-trained convolutional neural networks optimized on the ImageNet dataset to extract latent features of size [1x1x4096]. We extend this concept by updating the pre-trained model by classifying wallpaper group symmetries to create physics-informed latent representations. We use manifold learning techniques [e.g., t-Distributed Stochastic Neighbor Embedding (t-SNE) and Uniform Manifold Approximation and Projection (UMAP)] to generate interpretable 2D projections of these features. We demonstrate how this tool can be used for recursive image similarity searching and exploration accelerating discovery of trends in materials microscopy. Our algorithms are capable of identifying microstructure in diverse microscopy images across length scales. We will discuss how this concept can be integrated with structured scientific databases containing images and metadata as well as interactive graphical user interfaces to enable the rapid extraction of actionable information from materials microscopy data. While we demonstrate the aptitude of this approach on piezoresponse force microscopy images, this approach is amenable to other forms of microscopy and imaging techniques.

Voltammetry is a ubiquitous electroanalytical technique used for qualitative and quantitative characterization of redox-active species for a range of purposes (e.g., materials development, detection).^1,2 For dilute electrolyte solutions, known fundamental relationships, based on the mathematical treatment of reaction-transport processes, can be used to discern key physical and electrochemical properties of deterministically prepared electrolytes.^1,3 Further, voltammetry may be readily performed in-situ or in-operando, making it a minimally invasive and rapid diagnostic technique. However, many studies seek to track the chemical evolution of redox-active species in non-deterministically generated electrolytes, often by identifying the chemical by-products of this evolution.^4,5 Voltammetric methods alone are inadequate to uniquely identify compounds, as multiple species may have similar electrochemical properties; as a result, electrolyte characterization also includes suites of ex-situ techniques (e.g., mass spectrometry, UV-Vis spectroscopy). Advanced labeling protocols – potentially involving machine learning (ML) methods – may enable voltammetry to uniquely identify more compounds, which, in turn, may streamline characterization workflows. Indeed, ML methods have already been shown to enhance voltammetric capabilities,^6,7 but, to our knowledge, no studies have combined ML methods and electrochemical descriptors to identify compounds.

In this presentation, we describe a framework to identify solution-phase redox couples using experimental voltammetry, electrochemical descriptors from physics-based simulations, and Bayesian inference. This protocol references a library of candidate compounds comprised of molecular property sets (e.g., diffusion coefficient, redox potential) and compares it to experimental data to characterize the electrolyte being examined. The compound identities and the associated probabilities are then reported, enabling targeted follow-up experiments. We discuss the development and the validation of this protocol using a model set of redox-active molecules, demonstrating viability and motivating further development. We also explore possible expansions of this Bayesian framework, and we conclude by discussing potential applications (e.g., electrolytes for energy storage technologies). As a result, this protocol may extend the capabilities of voltammetric analysis, thereby accelerating the materials development pipeline for systems involving redox-active compounds.

Acknowledgements: This work was funded by the National Science Foundation under Award Number 1805566.

References:
1. Kwabi, D. G. et al. Alkaline Quinone Flow Battery with Long Lifetime at pH 12. Joule 2, 1894–1906 (2018).
2. Komorsky-Lovrić, Š., Vukašinović, N. & Penovski, R. Voltammetric determination of microparticles of some local anesthetics and antithusics immobilized on the graphite electrode. Electroanalysis 15, 544–547 (2003).
3. Northrop, P. W. C. & Cole, J. V. A Pulse Voltammetry Analysis Toolkit for Battery and Fuel Cell Material. ECS Trans. 85, 23–42 (2018).
4. Goulet, M. & Aziz, M. J. Flow Battery Molecular Reactant Stability Determined by Symmetric Cell Cycling Methods. J. Electrochem. Soc. 165, 1466–1477 (2018).
5. Day, R. P. et al. Differential thermal analysis of Li-Ion cells as an effective probe of liquid electrolyte evolution during aging. J. Electrochem. Soc. 162, A2577–A2581 (2015).
6. Ye, J. J., Lin, C. H. & Huang, X. J. Analyzing the anodic stripping square wave voltammetry of heavy metal ions via machine learning: Information beyond a single voltammetric peak. J. Electroanal. Chem. 113934 (2020).
7. Li, J., Kennedy, G. F., Gundry, L., Bond, A. M. & Zhang, J. Application of Bayesian Inference in Fourier-Transformed Alternating Current Voltammetry for Electrode Kinetic Mechanism Distinction. Anal. Chem. 91, 5303–5309 (2019).

Scientists are increasingly turning to the power of artificial intelligence (AI) to increase the rate of knowledge extraction, provide insight into new materials and discover unexpected correlations. In some cases, robust AIs have even been coupled with experimental (or theoretical) automation to generate autonomous scientific platforms capable of planning and executing experimental designs without human intervention. With the excitement around AI rapidly increasing it is time to begin thinking about the bursting of the AI bubble and what are the actions we can take as a community to mitigate its severity. Here scientific-AI and the opportunities and paths forward for creating a sustainable and scalable paradigm within materials science will be discussed. Of particular note will be issues with the reproducibility of AI model results, the use of archival data to build models, and the common usage of single scalar labels to summarize data sets. We will show building the same model using the same data set using a different implementation of the same algorithm does not result in statistically identical results. Next we will discuss the importance of preserving a record of primary data underlying archival labels, by highlighting the impact of a using a fuzzy metric to classify the state of a material. Finally, we will discuss the issue of labeling uncertainty with the idea that an AI will not always be presented with pristine data for evaluation and make recommendations for possible methods to capture disagreement between experts in the literature.

Core-loss spectroscopy observed by electron, electron energy loss spectroscopy (EELS), and X-ray, X-ray adsorption spectroscopy (XAS), are analytical techniques used for wide variety of materials, including battery, catalysis, and functional oxides. In particular, electron energy loss near edge structure (ELNES) and x-ray adsorption near edge structure (XANES) correspond to the electronic transition from a core orbital to conduction bands and contain various information on local atomic structures and electronic states.
In general, ELNES/XANES spectra of target samples are interpreted through comparison with reference spectra which are often obtained with time-consuming theoretical calculations. However, the interpretation of these spectra is very difficult because of the ambiguous relationship between the spectral features and the properties of target materials. Moreover, up to tens of spectra can be obtained in an experiment due to time-resolved and/or space resolved observation by high performance experimental equipment. For these reasons, individual interpretation of the spectra by the conventional approaches is becoming unrealistic.
In recent years, machine learning (ML) approaches, which can overcome obstacles to handling multidimensional big-data, have spread in materials science. ML has been applied to spectroscopic data including ELNES/XANES spectra and succeeded in predicting local chemical environment and properties including bond length, bond angle and Mulliken charge [1]. However, few studies have attempted to utilize ML on carbon-K edge ELNES/XANES spectra of functional organic molecules for investigating local chemical properties.
In this research, ML was applied to carbon-K edge dataset generated by theoretical calculations. A neural network model was adopted to extract information about properties of organic molecules from carbon-K edges. Moreover, we added some data about molecules as input data to improve the accuracy of prediction.

Input data of neural network was composed of 240-dimensional vector data of spectral intensities, which were referred to range of 24 eV in 0.1 eV steps. For output data, several molecular properties were selected. Hyperparameters, including the number of hidden layers, the number of neurons and the weight decay parameter, were tuned through cross validation.
To construct the carbon-K edge dataset, 7,941 structural data of organic molecules were extracted from the QM9 database [2,3]. Carbon-K edges were calculated from these data by a first principle pseudopotential method using CASTEP code [4].

Several properties such as HOMO-LUMO gap were successfully predicted, whereas the regression accuracy of internal energy remained at a low level. Careful investigation suggested that it was difficult to accurately predict the extensive properties when only spectral data was used as the input data. We found that the prediction accuracy of these extensive properties could be drastically improved using additional input data about molecules.

In summary, we clarified that carbon-K edges contained sufficient information on organic molecular properties. Our neural network model enebles fast, accurate, and direct prediction of molecular properties from ELNES/XANES, which contributes to a new way to investigate materials properties beyond the conventional comparison approach.

[1] S. Kiyohara et al., J. Phys. Mater., 2 (2019) 024003,
[2] L. Ruddigkeit et al., J. Chem. Inf. Model., 52 (2012), 2864-2875,
[3] R. Ramakrishnan et al., Scientific Data, 1 (2012), 140022,
[4] M. D. Segall et al., J. Phys. Condens. Matter., 24 (2002), 2717-2744.

Materials scientists are trying to accelerate the discovery of new materials, for example, by using high-throughput computations and machine learning. Chemical heuristics such as the Pauling rules^[1] have long served as the only guide for experimentalists to discover and understand new materials. The Pauling rules promised to be a structure-stability relationship that could accelerate the discovery of stable materials. They relate the coordination environments and their connections to the stability of materials. Despite being a corner-stone of solid-state chemistry and materials science, the Pauling rules have not been systematically and statistically evaluated. Here, we will present such a systematic assessment using modern informatics tools and a recently developed code which automatically detects coordination environments and their connections.^[2] Using a database of 5000 known oxides, we will report on the quality of the different Pauling rules and provide a general assessment of their predictive power.^[3] Our work shows how traditional empirical rules can now be tested and challenged using modern tools of data analysis and available structural databases.

References:
[1] L. Pauling, J. Am. Chem. Soc. 1929, 51, 1010.
[2] D. Waroquiers, J. George, M. Horton, S. Schenk, K. Persson, G.-M. Rignanese, X. Gonze, G. Hautier, ChemRxiv Preprint 2019, DOI 10.26434/chemrxiv.11294480.v1.
[3] J. George, D. Waroquiers, D. Di Stefano, G. Petretto, G. Rignanese, G. Hautier, Angew. Chem. Int. Ed. 2020, 59, 7569–7575.

Acknowledgements:
J.G. acknowledges funding from the European Union's Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement No 837910.

The identification of single-crystal elastic constants (SEC) of a high strength near-β titanium alloy, used in aeronautical applications, has been investigated with a Bayesian approach. The literature on the identified SEC of the cubic β phase in the Ti-alloys shows no substantial agreement, with anisotropic index (Zener ratio) extending from 1.6 to 8 [1]. The directional elastic moduli (DEM) from the X-ray diffraction (XRD) analysis are often used to identify the SEC following the method of Kröner (inverse identification) [2,3]. However, several studies [3,4] have reported that the DEM for Ti-alloy can be insensitive to the bulk modulus, thus resulting in several possible SECs that satisfy the lattice strain evolution. This is addressed by evaluating the SEC at a fixed bulk modulus. In the current study, the high energy XRD and elastic self-consistent modeling [5,6] is revisited with a Bayesian framework and applied to a near-β Ti–10V–2Fe–3Al alloy in its as-forged state (dual phase: 15% of primary hexagonal α_p nodules embedded in a cubic β matrix). The effect of each phase elastic anisotropy was systematically investigated along with different parameters of the micro-mechanical model and also the biases introduced by the XRD data on the identification of the SEC of the β phase. The grain aspect ratio in the self-consistent model, which is often simplified in the literature was found to be the most significant parameter affecting the identification of SEC. The grain shape should therefore be taken into account in the micro-mechanical model when identifying the SEC from diffraction data.

Reference:
1. S. Lhadi, S. Berbenni, N. Gey, T. Richteton, L. Germain, Micromechanical modeling of the effect of elastic and plastic anisotropies on the mechanical behavior of β-Ti alloys, Int. J. Plast., Vol. 109 (2018), 88-107
2. Matthies, S., H. G. Priesmeyer, and M. R. Daymond, On the diffractive determination of single-crystal elastic constants using polycrystalline samples. J. Appl. Cryst 34.5 (2001): 585-601.
3. Cho, J. R., D. Dye, K. T. Conlon, M. R. Daymond, and R. C. Reed. "Intergranular strain accumulation in a near-alpha titanium alloy during plastic deformation." Acta materialia 50, no. 19 (2002): 4847-4864.
4. Obbard, E. G., Y. L. Hao, R. J. Talling, S. J. Li, Y. W. Zhang, D. Dye, and R. Yang. The effect of oxygen on α ″martensite and superelasticity in Ti–24Nb–4Zr–8Sn. Acta materialia 59, no. 1 (2011): 112-125.
5. Meng, Q, Qing L, Shun G, Yongqi Z, and Xinqing Z. Effect of thermo-mechanical treatment on mechanical and elastic properties of Ti–36Nb–5Zr alloy. Progress in Natural Science: Materials International 25, no. 3 (2015): 229-235.
6. Faurie D., O. Castelnau, R. Brenner, P.O. Renault, E. Le Bourhis, P. Goudeau, In situ diffraction strain analysis of elastically deformed polycrystalline thin films, and micromechanical interpretation, J. Appl. Cryst., 42 (2009), p.1073-1084.
7. C. Mareau, S. Berbenni, An affine formulation for the self-consistent modeling of elasto-viscoplastic heterogeneous materials based on the translated field method, Int.J. Plast., 64 (2015) 134-150

Reaction networks, graph-based representations of the reactions occurring between chemical species, have long served as tools for the identification of chemical pathways and for the study of chemical dynamics. In recent years, high-throughput first-principles calculations have been used to generate computational reaction networks (CRNs) consisting of hundreds of molecules and thousands of chemical reactions. However, the systems thus far considered in CRN literature have been limited by insufficient automation. Specifically, automating calculations involving complex chemistries (ions, open-shell molecules, metal-coordinated molecules, solvated molecules, etc.) remains a challenge.

Here, we present a computational framework for the automatic generation of massive CRNs based on density functional theory (DFT) calculations. Using a robust and accurate level of theory and automated error handling, we have achieved a 95% success rate for DFT calculations of electrochemically relevant molecules. Using graph theory, we can generate CRNs consisting of all possible elementary reactions between the chemical species that we have studied without the use of complex pre-defined mechanisms. Our framework can also automatically identify optimal reaction pathways in massive CRNs based on reaction thermodynamics or energy barriers. As a demonstration of these automated tools, we present a CRN for the formation of the solid-electrolyte interphase in Li-ion batteries consisting of over 8,000 species and over 100,000 chemical reactions, the largest CRN based on reaction thermodynamics yet reported. Using shortest path algorithms, we have recovered critical pathways for electrolyte decomposition in a fully automated fashion, indicating the power of unbiased pathway detection for enhancing chemical understanding.

Discovery, optimization, and design of novel materials is the primary challenge facing civilization in the XXI century. The transition from simple functionalities in structural, optical, and semiconductor materials to complex intertwined functionalities in batteries, fuel cells, information technology and sensor devices renders the classical paradigm of serendipitous discovery and sequential optimization impractical and necessitates alternative pathways for materials discovery and optimization. One, actively explored at present, paradigm is automated experimentation and guided or combinatorial synthesis combined with rapid throughput characterization. However, a second pathway can be developed based on modern imaging tools including scanning probes and electron microscopy that allow us to not only visualize and probe functionalities of materials on the nanometer and atomic scales, but also to introduce controllable modifications in materials. Effectively, we aim to discover the defined mesoscopic and atomic structures and probe their functional properties, establish the pathways towards controllable creation of these structures, and ideally extend these towards macroscopic assemblies. Here, we discuss possible strategies towards autonomous experimentation (AE) in STEM and SPM. We demonstrate the examples of Bayesian optimization-based workflows for AE SPM and STEM using prior data sets, and discuss the development of the acquisition and pathfinder functions. We further discuss opportunities for the development of self-training AE workflows in microscopy, where the features of interest are unknown a-priori and are discovered during the experiment. The experimental realization of these concepts is demonstrated for ferroelectric domain manipulation and electron-beam induced chemical reactions.

The realization of autonomous materials discovery depends critically on the ability to predict whether a given material will be synthesizable. This problem is generally addressed in silico by calculating the thermodynamic stability of candidate materials using density functional theory (DFT). There have been many recent claims that this step can be greatly accelerated using machine learning instead of DFT to predict the formation energy of inorganic solids. However, thermodynamic stability is ultimately dictated by relative formation energies between chemically similar competing phases, and not by the absolute formation energy of any single compound. It is therefore essential to assess how well these machine learning predictions of intrinsic thermodynamic quantities generalize to the problem of relative materials stability. In this work, we systematically probe this question by testing a diverse set of materials representations and learning algorithms on a variety of problems that simulate real applications of materials discovery. Our findings emphasize the value of including structure in materials representations and suggest that error cancellation plays a critical role in the prediction of thermodynamic stability.

Defect transport is a key process in materials science and catalysis, but as migration mechanisms are often too complex to enumerate a priori, calculation of transport tensors typically have no measure of convergence and require significant end user intervention. These two bottlenecks prevent high-throughput implementations essential to propagate model-form uncertainty from interatomic interactions to predictive simulations. In order to address these issues, we extend a massively parallel accelerated sampling scheme[1], autonomously controlled by Bayesian estimators of statewise sampling completeness, to build atomistic kinetic Monte Carlo models on a state space irreducible under exchange and space group symmetries. Focusing on isolated defects, we derive analytic expressions for defect transport tensors and provide a convergence metric by calculating the Kullback-Leiber divergence across the ensemble of diffusion processes consistent with the sampling uncertainty. The autonomy and efficacy of the method is demonstrated on surface trimers in tungsten and hexa-interstitials in magnesium oxide, both of which exhibit complex, correlated migration mechanisms[2].
[1] Swinburne & Perez, Phys. Rev. Mat, 2018
[2] Swinburne & Perez, arXiv:2003.07752, 2020

Over the past decade, the family of 2D crystals increased dramatically, with 2D materials now covering a huge range of properties: metallic, semiconducting, ferromagnetic, superconducting… As the field attempts to engineer materials properties by design, an outstanding challenge is to develop high-accuracy predictive models that reversibly connect material structure with their properties. Herein, we adapt the recently-proposed Fourier Transform of Crystal Properties (FTCP) framework [1] to 2D materials, toward generating an invertible crystallographic representation of 2D materials suitable for materials search and inverse design tasks. When we train on ~2000 2D materials in the publicly available 2DMatPedia database, we can predict simple ground-state properties such as formation energy with mean average errors of 281 meV/atom. When a transfer-learning approach is employed, whereby ~51,000 3D materials from MaterialsProject are used, the mean average error of formation energy predictions improves to 209 meV/atom. We believe this framework presents a promising approach to predict certain ground-state properties of 2D materials, and may potentially extend to other properties in the near future.

[1] D. Ren et al., https://arxiv.org/abs/2005.07609

We propose an autonomous AI planner that recommends the next experiments based on the outcomes of all past experiments with the goal of materials discovery in accelerated paces. We are particularly interested in accelerating an experimental campaign for mapping out an unknown relation of experimental conditions to the experimental outcomes of interest, e.g. resulting material properties and kinetic rates of reaction. The relation often has discontinuities in that there are abrupt changes in the experimental outcomes for a small change in the experimental conditions over some boundary regions such as boundaries of reaction equilibrium or chemical regime changes. In such cases, many conventional experimental planners, exploiting the continuity and smoothness of the relation, cannot be applicable. We propose a jump planner to optimize the experimental planning under discontinuities of the underlying relation to discover. Its application will be illustrated with carbon nanotube growth experiments with various catalysts.

The development of automated computational tools is required to accelerate the discovery of novel battery materials. In this work, we design and implement a workflow, based on Density Functional Theory calculations, to autonomously identify materials to be used as intercalation electrodes in batteries. The materials selection is based on thermodynamic and kinetic descriptors, such as adsorption energies and diffusion barriers. A substantial acceleration for the calculations of the kinetic properties is obtained thanks to the recent implementation of the Nudged Elastic Bands (NEB) method, which takes into consideration the symmetries of the system to reduce the number of images to calculate. We have applied this workflow to discover new cathode materials for Mg batteries, where two of these materials display a threefold increase in the potential of the Chevrel phase, the state of the art cathode in commercial prototype Mg batteries. In addition to the discovery of new materials, the dataset produced with this workflow has been used to train a machine-learning algorithm to accelerate and generalize the discovery of electrodes, beyond Mg-ion batteries.

A large database of known molecular semiconductors is used to define a plausible physical limit to the charge carrier mobility achievable within this materials class. From a detailed study of the desirable properties in a large dataset, it is possible to establish whether such properties can be optimized independently and what would be a reasonably achievable optimum for each of them. All relevant parameters are computed from a set of almost five thousand known molecular semiconductors, finding that the best known materials are not ideal with respect to all properties. These parameters in decreasing order of importance are the molecular area, the nonlocal electron-phonon coupling, the 2D nature of transport, the local electron-phonon coupling, and the highest transfer integral. It is also found that the key properties related to the charge transport are either uncorrelated or “constructively” correlated (i.e., they improve together) concluding that a tenfold increase in mobility is within reach in a statistical sense, on the basis of the available data. It is demonstrated that high throughput screenings, when coupled with physical models of transport produce both specific target materials and a more general physical understanding of the materials space.

Predicting the synthesizability of hypothetical crystals has proven to be challenging due to the wide range of parameters that govern the crystalline materials synthesis. In this work, we convert the atomic structures of known synthesized or observed crystals in databases into three-dimensional pixel-wise images color-coded by chemical species and electronegativity and use them to train a deep-network convolutional autoencoder(CAE). We extract the latent features of synthesizability hidden in structural and chemical arrangements of crystalline materials embedded in the auto-encoder. The accurate classification of materials into synthesizable crystals vs. crystal anomalies based on these features across a broad range of crystal structure types and chemical compositions confirms the validity of our model.

Photodynamic therapy is arising as a non-invasive treatment modality for cancer and other diseases. One of the key factors to determine the therapeutic function is the efficiency of photosensitizers (PSs). According to the latest studies, reducing the energy gap (ΔE_ST) between the lowest excited singlet state (S₁) and lowest excited triplet state (T₁) was an effective approach to yield highly efficient PSs, and ΔE_ST could be fine-tuned through precise molecular donor-acceptor engineering. There is a growing literature where small families of small-ΔE_ST PS compounds are explored experimentally, or purely computationally at a slightly broader scale. Here, we report a large-scale data-driven search for novel PSs, with a special focus on red PS. This work comprehensively integrates computational quantum chemistry, machine learning, organic synthesis, and nanoagent fabrication and testing.

By providing the list of commonly used donors, acceptors and bridges, there is a chemical design space of more than 7 million possible PSs. Such a vast search space can hardly be explored by chemical intuitions or Edison’s approaches. Even performing quantum calculations of this vast space can take infeasibly long time. A data-driven approach by screening the ΔE_ST of 7 million PSs with a well-trained deep learning model provides a possible solution. Herein, we have created a high-throughput quantum calculation toolset for the construction of a novel dataset of 10k PSs with ΔE_ST information. This training dataset contains representative donors, acceptors and bridges with both Donor-Acceptor and Donor-Acceptor-Donor structures. A molecular graph convolutional neural network with molecular structure as features is trained on this dataset. In combination with a high-throughput PS generator, we could efficiently predict the ΔE_ST of possible PSs within the chemically intuited design framework. In order to further improve the performance of the predictive model, we have also employed a transfer learning approach. In addition to this physically informed design model, we utilize a motif-based hierarchical encoder-decoder (hgraph2graph) as another deep learning method for the direct generation of possible PS candidates. These two approaches will be compared with results discussions. Furthermore, the experimental validation of chosen PSs with low ΔE_ST is conducted and presented. Finally, the potentials and limitations of this work will be discussed.

The rapid development of deep learning enabled artificial intelligence (AI) has brought new opportunities to accelerate conventional experimental design in materials science. The discovery of high-dimensional synthesis recipes that yield desired material properties used to be costly and time-consuming, while advances in machine learning (ML) driven high-throughput experiments are able to rapidly achieve optimal conditions to produce the target materials. I will introduce our recently developed online active learning approaches that can effectively guide experiments and achieve full-map understanding of the design space.

The first case study demonstrates a two-step active learning framework that combines Bayesian Optimization (BO) and Deep Neural Network (DNN) in a loop with a high-throughput microfluidic platform, to optimize the synthesis of silver nanoparticles with the desired absorbance spectrum. In Step I, the optimization process is accelerated by the offline introduction of DNN after a few runs of targeted sampling through BO with a Gaussian Process. In step II, the DNN is allowed to sample the parameter space using grid optimization to test its regression function. The fully trained DNN can then predict absorbance spectra around the target and possibly optimize the synthesis towards a new target.

The second case study further extends the active learning method to a hierarchical AI framework, which is tested for strain sensor material synthesis. The challenge to obtain customized sensing devices with designated working windows comes from the complex and unclear structure-composition-property relationships that are hard to predict and pre-design. A feasibility classifier using support vector machine (SVM) is first constructed as the screening layer towards the material design. An active machine learning loop is then designed to perform full-map exploring and learning by sampling the “most informative” data points to perform explorations. Besides, comprehensive algorithm optimization is performed for better AI prediction performance. The framework enables a full-map understanding to the system and provides new insights and principles for sensor design.

Moreover, we have developed a series of interpretation and visualization approaches to understand "black-box" ML models and make them more favoured by domain experts to extract knowledge of complex systems that embed a huge amount of hidden information. Although promising opportunities are identified, many challenges exist this highly interdisciplinary field, such as construction of valuable and open datasets, in-depth understandings of descriptors, lack of standard algorithms workflow and fully autonomous experimental platforms, which will be briefly discussed as future directions.

Discovery and development of polymer materials is strongly driven by experimental data acquisition. Polymer materials have to be prepared and characterized in order to be patentable. The immediate outcome of the polymerization reaction can be transformed into multiple materials with strikingly different properties by means of formulation and processing. Potential for IP development increases sharply as the polymerization product evolves via formulation and processing into the final material. Experimental data acquisition, therefore, unfolds under conditions of delayed rewards on incredibly rich landscapes shaped by access to multiple experimental degrees of freedom that can be both continuous (concentration, temperature, radiation, time) and categorical (monomers, catalysts, initiators, solvents) [1,2].

In the presented contribution, we report results of the ongoing effort in the development of an end-to-end reinforcement learning (RL) approach to the experimental data acquisition in the polymer material domain. The application is the development of a simple spin-on-glass. It starts with the acquisition of initial experimental data, proceeds to training a RL agent to search for the experimental settings (generating experimental hypothesis) that produce materials with desired properties, then to applying a trained RL agent to design the experimental plan, and finally to executing the experimental plan using the existing robotic platform. We identify and discuss the following factors that we are systematically addressing:

- Direct access to the lab equipment during the training phase is impractical; given the complexity of the polymer synthesis and processing, therefore, the most viable option is to prepare a surrogate model of the system and use it to set up the RL environment. The optimal choice of the surrogate as well as handling the uncertainties deserves systematic investigation.

- The task of materials discovery often comprises a search for multiple promising solutions, both search and enumeration, but not necessarily global optimization. The search task can be learned in the narrow context of the specific experimental project (cf. navigation of one specific maze); a significantly stronger result is to train a RL agent to navigate experimental landscapes as such (cf. learning general strategies of maze navigation). This brings meta-learning to the top of the priority list [3].

- RL is notoriously data-greedy; meta-learning further amplifies the need for abundant and diverse training data. Therefore, data augmentation is necessary, particularly around creating families of the surrogate models. The specifics of the data augmentation, that leads to optimal performance, needs careful observation.

- As is generally the case with RL, the reward system plays a big role. One possible (and informative) approach is to develop reward systems that mimic effective “rewards” and “penalties” for human researchers solving similar data acquisition tasks.

On the artificial intelligence angle, we describe the specifics of the implementation (OpenAI Gym based environment, and RL algorithms) and discuss obtained performance metrics and approaches to the diagnostics of the RL agent behavior.

1. Li, H. et al. “Tuning the Molecular Weight Distribution from Atom Transfer Radical Polymerization Using Deep Reinforcement Learning” Mol. Syst. Des. Eng., 2018.
2. Zhou, Z. et al. “Optimizing Chemical Reactions with Deep Reinforcement Learning” ACS Cent. Sci. 2017.
3. Kobbe, K. et al. “Quantifying Generalization in Reinforcement Learning” PMLR 97, 2019.

Scalable synthesis of two-dimensional (2D) materials is a major bottleneck to more widespread adoption of layered material-based devices. Chemical vapor deposition (CVD) has emerged as a viable method for large-scale synthesis of 2D materials. However, neither experiment nor theory has been able to decipher mechanisms and selection rules for different growth scenarios or make predictions of optimized growth parameters. Here, we have designed a workflow using model-based reinforcement learning (RL) and neural autoregressive density estimator (NADE-CVD) to generate the optimal synthesis condition for the prototypical two-dimensional semiconductor, monolayer MoS₂ with high 2H-phase fraction in minimum time from MoO₃ crystal. Here, the NADE-CVD models the probability density function of CVD-synthesis and is trained using 15,000 reactive molecular dynamics (RMD) simulation trajectories each for 20 ns of MoS₂. After training, NADE-CVD serves as a dynamics model of CVD-synthesis and interacts with the RL agent. Using the reward singles from the dynamics model, the RL agent is able to propose synthesis conditions (such as temperature profile and S₂, H₂ and H₂S partial pressures) and their dependencies among each other that generates MoS₂ crystal with maximum 2H phase fraction quickly, thus paving the way toward rational design of CVD synthesis techniques for MoS₂ and other layered materials.

This work was supported as part of the Computational Materials Sciences Program funded by the U.S. Department of Energy, Office of Science, Basic Energy Sciences, under Award Number DE-SC0014607. Simulations were performed at the Argonne Leadership Computing Facility under the DOE INCITE program and Aurora Early Science programs and at the Center for High Performance Computing of the University of Southern California.

Humanity is accumulating magnitudes of data for many discrete components — but how do we best utilize these data to improve complex systems? The world of “big data” begins to intersect “systems engineering.” In this contribution, we contemplate a model system — conjoined photovoltaic and thermoelectric generators (PV+TEG) — which has tens of design variables and corresponding response curves [1]. However, with such voluminous design and response data (in machine-learning parlance, “vast state space”), it is unclear how we can inversely map the response data back to our design variables to improve system energy yield under a given set of operating conditions. Therefore, to leverage the supply of big data coming from our simulated and experimental systems such as PV+TEG, we propose a framework for system optimization via supervised surrogate models (e.g., neural networks) and model-free optimization (Bayesian optimization, BO), which combines the advantages of embedded physics domain knowledge and adaptability to complex action-responses outside the model.

Conducting system optimization using this framework promotes high-throughput iterative design cycles for faster rates of learning [2]. There is great importance in optimizing new or existing systems using high-throughput iterative design because by informing a new design based on learning data trends rather than experimentally building a new system each iteration, significant savings are seen in time, energy, and material costs [2]. In the example of solar energy, photovoltaics and thermoelectric modules are both well-established technologies individually, however, by drilling down to the design components of each device, the thermoelectric module can be better optimized to improve the energy yield of a photovoltaic device, in turn, requiring a smaller area to generate the same power. The working principle of this system improvement framework is that: (1) physics-driven models establish the fundamental relationship between PV+TEG design variables and response outputs; (2) experimental response curves provide ground-truth relationships for a narrow band of design variables and outputs; (3) physics-fusion neural networks combine simulated and experimental response data to build robust inferences of feasible PV+TEG designs [3]; (4) the manifold of PV+TEG design variables is searched by BO to generate operational system designs that maximize energy yield. We demonstrate that with large volumes of operational system data, neural networks and Bayesian optimization are used to accelerate the convergence on optimal photovoltaic and thermoelectric generator designs by 10x relative to the state of the art Multiphysics simulation while allowing more parameter customization [4]. By reducing design time, fiber laser-powered gallium arsenide PV+TEGs will be explored in a gamut of small-scale applications to maximize energy transmission efficiency.

References:
[1] H. Hashim, J. J. Bomphrey, and G. Min, “Model for geometry optimisation of thermoelectric devices in a hybrid PV/TE system,” Renew. Energy, vol. 87, pp. 458–463, 2016, doi: 10.1016/j.renene.2015.10.029.
[2] A. T. D. Perera, P. U. Wickramasinghe, V. M. Nik, and J. L. Scartezzini, “Machine learning methods to assist energy system optimization,” Appl. Energy, vol. 243, pp. 191–205, 2019, doi: 10.1016/j.apenergy.2019.03.202.
[3] Y. Ba, G. Zhao, and A. Kadambi, “Blending Diverse Physical Priors with Neural Networks,” no. 1, pp. 1–15, 2019.
[4] Z. Ren et al., “Embedding physics domain knowledge into a Bayesian network enables layer-by-layer process innovation for photovoltaics,” npj Comput. Mater., vol. 6, no. 1, pp. 1–9, 2020, doi: 10.1038/s41524-020-0277-x.

Reticular frameworks are crystalline porous materials that form via the self-assembly of molecular building blocks (i.e., nodes and linkers) in different topologies. Many of them have high internal surface areas and other desirable properties for gas storage, separation, and other applications. The notable variety of the possible building blocks and the diverse ways they can be assembled endow reticular frameworks with a near-infinite combinatorial design space, making reticular chemistry both promising and challenging for prospective materials design. Here, we propose an automated nanoporous materials discovery platform powered by a supramolecular variational autoencoder (SmVAE) for the generative design of reticular materials with desired functions. We demonstrate the automated design process with a class of metal-organic framework (MOF) structures and the goal of separating CO₂ from natural gas or flue gas. Our model exhibits high fidelity in capturing structural features and reconstructing MOF structures. We show that the autoencoder has a promising optimization capability when jointly trained with multiple top adsorbent candidates identified for superior gas separation. MOFs discovered here are strongly competitive against some of the best-performing MOFs/zeolites ever reported. This platform lays the groundwork for the design of reticular frameworks for desired applications.

Combinatorial experimentation has become a well-established technique to explore new compounds with enhanced physical properties in a variety of functional materials fields. High-throughput experiments are particularly effective when designs and layouts of combinatorial libraries are guided by predictions. With the advent of high-throughput computational screening, it is now possible to come up with a large list of possible new compounds quickly. In many areas of functional and quantum materials, however, it remains a formidable challenge to make accurate predictions through computational means. When properly curated experimental databases are available, it is also possible to mine such databases in order to identify trends among known materials compositions and make predictions of new materials. As an example, we have previously developed machine learning models of superconductors based on the SuperCon database within MatNavi (NIMS, Japan) which contains over 29,000 experimental entries. Although SuperCon lacks structural property information, random forest regression was able to build models, which can decipher key distinctions between different families of superconductors. In many materials areas, such an approach is unfortunately hampered by the lack of sufficiently large experimental databases with very few exceptions. I will discuss our recent work where we have combined structural property predictions using ICSD with compositional information in order to identify trends in materials discoveries and make predictions based on clustering of known compounds through neural network. This work is carried out in collaboration with Valentin Stanev, Haotong Liang, and A. Gilad Kusne, and is funded by NIST.

1. Introduction
Designing “tailored” material that possesses desired characteristics is a long-awaited technology. Especially molecular design is important across broad industrial domains. It is to solve a complex combinatorial problem in terms of molecule’s atomistic configuration patterns, that significantly exceeds 10⁶⁰. In this extra-vast parameter space, human SMEs (Subject-Matter-Experts) can only explore around a tiny space by trial-and-error relying on their experience and knowledge. To accelerate their design speed and expand the structural variety of molecules, AI can play a key role. Several AI molecular design tools and algorithms exist [1,2], but following issues in practical use are unaddressed; (1) deep learning - based methods rely on extremely large amount of material data, (2) models are black box, therefore hard to tune for chemists, (3) customization for specific material domains is not supported, (4) a certain level of IT skill is required to use. To address those issues, we developed a platform for small organic molecules [3]. In this paper, we exhibit customization of the tool for polymer domains and demonstrate to design new acrylic polymers having specific Tg (glass transition temperature).
2. Method
The fundamental workflow of our tool consists of sequential 5 steps; (1) data input, (2) feature encoding, (3) property prediction, (4) solution search, and (5) structure generation. First, a data set listing pairs of chemical structures and target properties is input as a training dataset. Structures are encoded to a set of feature vectors, by which a regression model to predict target properties is built. After the model is built, a user inputs target property values. The system performs the solution search to identify candidate feature vectors that satisfy the target property values. Finally, the feature vectors are decoded to concrete molecular structures by improved Mckay’s Canonical Costruction Path algorithm, that builds up a molecular graph by connecting atoms and substructures indicated in the feature vector. That process is called “inverse-design”; to solve a model inversely starting from target properties.
We further customized the tool for polymer domain; first, the tool identifies monomer’s main chain and extract only side chain structures as a part to design. Second, structure generation algorithm is reinforced by structural constraints to avoid chemically unrealistic structures.
3. Results
To demonstrate inverse-design of polymers, from PolyInfo we extracted acrylic polymer data consisting of 378 pairs of SMILES and Tg. Side chain structures are extracted from the acrylic main chain, and encoded to about 50 dimensional feature vectors, in which each element represents a number of substructures. Regressions were carried out on the feature vectors to predict Tg. The best regression model exhibited determination coefficient R²~ 0.68 corresponding to RMSE~ 30 C°. Inverse-design was carried out on the model targeting Tg ~ 150 C°. Running several hours, more than 100 structures were generated, that is significant acceleration in design speed compared with typical SMEs. The generated structures were reviewed by a polymer SME for screening, and some of the candidates were successfully synthesized in the lab.
Platform in Service
The above platform is running on the cloud of IBM. They are provided as-a-service to our client companies for research purpose. Two interfaces; Python-based command line interface and GUI-based web application are provided to cover broad range of user’s IT-skills. In the presentation, we will also demonstrate the web application.
Reference
[1] R.G.Bombarelli, et al. “Automatic chemical design using a data-driven continuous representation of molecules”, 2018
[2] G.B.Goh, et al. “Using Rule-Based Labels for Weak Supervised Learning: A ChemNet for Transferable Chemical Property Prediction”, 2018
[3] S.Takeda, et al. “Molecular Inverse-Design Platform for Material Industries”, 2020

Rapid experimentation and discovery of materials has typically centered around high throughput screening of materials in high-dimensional composition settings. Machine learning assisted selection of composition spaces is expected to significantly accelerate the experimental materials discovery process. To achieve this, frameworks that enable learning across composition spaces are necessary. A central aspect of such a framework is digital representation of compositions that are tuned for the composition-property mapping task of interest. While generic descriptors of elements exist we show that a framework that learns these representations for a specific task significantly outperforms generic representations.

Specifically, in this talk we present methodologies to learn composable digital representations of the chemical elements that enable learning about composition-property relationships across composition spaces. We demonstrate these methodologies by predicting formation energy, band gap in compositional spaces that cannot be interpolated from the training composition spaces. We also demonstrate that digital representations learned from simulation data can be transferred to predict experimental data.

Further, we demonstrate large scale predictions of formation energy by combining digital representation of elements, physical constraints into a Gaussian process framework. We demonstrate that using physical constraints and clustering based on covariances we can address the scalability limitations of Gaussian processes, allowing their deployment for large scale discovery applications. Finally, this work provides a framework to accelerate extrapolative exploration of composition spaces for materials discovery.

Homogeneous catalysis using transition metal complexes is ubiquitously used for organic synthesis, as well as technologically relevant in applications such as water splitting and CO₂ reduction. Automatic selection and optimization of catalysts also plays a very important role for automated and autonomous experimental synthesis platforms. The key steps underlying homogeneous catalysis require a specific combination of electronic and steric effects from the ligands bound to the metal center. Finding the optimal combination of catalysts and ligands is a challenging task due to the exceedingly large number of possibilities and the non-trivial ligand–ligand interactions. We show how quantum mechanical calculations and machine learning (ML) methods can be combined to enable the prediction of reactivity within large chemical spaces containing thousands of complexes. We illustrate our approach using the classic examples of the Vaska's complex [1] as well as a gold catalyzed cyclization reaction. We compare multiple different machine learning methods based on their accuracy, interpretability as well as their ability to learn from small amounts of data. In contrast to experiments and calculations requiring several days to be completed, the ML models were trained and used on a laptop on a time-scale of minutes.

[1] Pascal Friederich, Gabriel dos Passos Gomes, Riccardo De Bin, Alán Aspuru-Guzik and David Balcells, Chemical Science, 11, 4584-4601, 2020. DOI: 10.1039/D0SC00445F

High throughput combinatorial methods have accelerated searches for new materials, but most remain constrained by experimental limitations and human decisions. We are developing fully autonomous methods that enable discovery of functional materials, especially non-equilibrium materials that have yet to be synthesized by traditional techniques. To achieve this goal, we use a hierarchy of closed-loop processing/characterization agents (systems) to identify and execute an experimental strategy that optimizes the deployment of experimental resources (time, money, and access). This results in extremely efficient exploration of broad composition and processing spaces leading to the rapid and autonomous discovery of new metastable phases.

At the experimental level, we employ lateral-gradient Laser Spike Annealing to process amorphous thin film composition-spread libraries prepared on standard Si wafers. A variety of optical and X-diffraction (XRD) techniques are used to identify regions in the high-dimensional space {composition fractions {xi}, temperature T, dwell/quench time, t} that are associated with distinct crystal structures and/or microstructures. While optical characterization alone cannot label crystal structures or crystallographic texture, these low cost optical techniques are very sensitive and can rapidly identify transitions that provide valuable prior knowledge for structural XRD studies, either asynchronously through subsequent follow-up autonomous investigations at synchrotron sources, or synchronously in a fully coupled autonomous experimental loop.
Intelligent exploration of the high-dimensional space requires effective active learning—using the results of initial and accumulated characterization of the phase space to decide the next [{xi},T,t] condition to be explored. This involves real-time interaction of the experiments with a community of AI-based agents that must: (a) interpret optical or XRD patterns to create a model of the evolving “processing phase diagram,” (b) identify the conditions that, when characterized, are most likely to decrease uncertainty in the boundaries of this map, (c) direct automated equipment to execute the chosen processing, and (d) complete the loop by managing and controlling optical or XRD characterization. To tackle this challenge, we employ various Gaussian Process (GP) models with carefully designed kernels that capture subtleties of the physics associated with the transformations.

This approach enables us to identify regions with distinct structures. To determine the particular crystal structure(s) associated with each region, we apply novel computational techniques that we have developed and that have proven to be highly effective at automated crystal structure phase mapping. Example results from several systems will be used to demonstrate the current state of these agents and the overall process.

As these methods are fully realized, we will achieve the key goals of:
* developing a methodology for real-time closed-loop autonomous synthesis/process/characterization experiments using active learning
* identifying opportunities and demonstrating examples where state-of-the-art AI methods offer unique advantages for scientific experimentation
* creating new materials with superior functional properties, and advancing our understanding of the fundamentals of phase transformation in rapid-thermal-processed materials.

Data is a resource to drive discovery, fueled by multiple factors such as new funding, open source, and AI. Inspired by demands from the materials industry, an inverse design solver accompanied with an end-to-end AI-based tool for small organic molecule design is developed. Our published studies show good correspondence between predicted properties and calculated values [1]. In this report, we extend the scope of our tool, ensuring that the newly generated ligands form optimized Ir(III) complex structures, and finally perform validation through DFT simulations confirming energy gaps large enough to blue light LED material candidates.

Cyclometalated Ir(III) complexes are a well-established class of metal-organic compounds with attractive photophysical properties [2]. The design of the ligands around the Ir ion dominates the performance. By using open data sets and reliable models, a vast number of ligand candidates can be generated. However, considering the complexity of the three-fold symmetric structure, it is still difficult to effectively calculate representative characteristics even when subject matter expert insight is included. An automatic, high throughput workflow is discussed in the followings.

Initial searches extracted 1400 samples with CIE 1931 color coordinates and SMILES information. For deep blue emission the target values are (0.15, 0.12). This corresponds to a calculated energy gap larger than 2.5 eV according to experience. Multiple fused rings in the ligands leads to encode features with atom count, ring count, substructures, and fingerprints. In the prediction step, the hyper-parameters are further optimized by exhaustive grid search algorithm. The model with highest cross validation score is input to estimate new features and ligand generation.

In parallel, another framework is established to inherit the generated results. Newly generated ligands are converted to 3D structures and optimized by force field. Although the tris-heteroleptic Ir(III) complex is the more general use case, the detail of dealing with tris-homoleptic ones is described here. The two coordinating ions (A and B) to Ir ion are identified by a rule-based algorithm with more than 90 % accuracy. The three ligands coordinates are put in a 3D space without interference and Ir ion is located at the origin. The atom and bonding information are updated and do 2nd structure optimization. Because the ligand structures are optimized prior to the attachment, the degrees of freedom can be reduced to only three for the purpose of finding energy in global minimum. Three ligands are rotated simultaneously along the vector sum of Ir-A and Ir-B. By using genetic algorithm with many enough iterations, the angle combination with the lowest total energy is obtained. Finally, the hydrogen atoms are attached and ready for DFT simulation.

The workflow is validated by reproducing the numbers in [3]. There are 100 samples with tabulated values of LUMO, HOMO, dipole moment, S1, and T1 wavelength. Different software and parameters (G98 vs. GAMESS) are used. The results show good correspondence with R-square scores: 0.98, 0.98, 0.74, 0.71, and 0.89. With this framework, we can validate the newly generated ligands also have large enough band gap energy such as c1cnc2ccn3cc(C4CCCCC4)nc3c2c1 (4.20 eV), c1ccc2c(c1)ncn1c3ccccc3nc21 (3.64 eV).

The framework is capable of forming facial and meridional coordination geometries by adjusting the fitness in the optimization algorithm. Multiple dockerized packages provide such workflow automation. The potential of the developed tool is still increasing by adapting the customization for other functional materials.

[1] S. Takeda, et al. “Molecular Inverse-Design Platform for Material Industries”, Proc. of KDD (2020)
[2] Z. Wu, et al, “Novel Design of Iridium Phosphors with Pyridinylphosphinate Ligands for High-Efficiency Blue Organic Light-emitting Diodes”, Scientific Reports (2016)
[3] D. B. Knowles, U.S. Patent 2008/0297033A1

Metal-insulator transition (MIT) compounds are both of high technological interest, with potential applications to low-power computation and transistors, glass coatings and of broad general scientific interest, as the metal-insulator transition is due to a rich interplay of complex electronic and atomic degrees of freedom. This interplay makes it difficult for high-throughput, first-principles approaches to find potential MIT candidates, while machine-learning approaches can be difficult to implement and understand due to the limited number of known MIT compounds. In this work we show that, using a gradient boosted decision tree, we can nonetheless obtain a surprisingly strong model to predict which materials may or may not exhibit thermally driven MITs. By comparing a model in which the most important features of the materials are extracted automatically, to one in which they are added gradually and physically informed using domain knowledge, we extract novel insights into the physics of metal-insulator transition materials, as well as into how best to search for improved MIT material candidates.

Acknowledgement: This work was supported in part by the National Science Foundation (NSF) under award number DMR-1729303. The information, data, or work presented herein was also funded in part by the Advanced Research Projects Agency-Energy (ARPA-E), U.S. Department of Energy, under Award Number DE-AR0001209. The views and opinions of authors expressed herein do not necessarily state or reflect those of the United States Government or any agency thereof.

To date, most applications of machine learning (ML) in materials science have focused on process optimization and improving engineering parameters of merit. Inferring physical laws directly from data has a great potential to accelerate materials research by allowing us to extract generalizable scientific information from the data. In this study, we explore the use of ML algorithms to study perovskite degradation, with the aim of gaining knowledge of the underlying degradation mechanisms. The environmental instability of halide perovskite materials hinders the commercialization of perovskite solar cells. The degradation mechanisms are still not fully understood due to the complexity of degradation phenomena. The functional space of feasible physics equations capturing these mechanisms is enormous [1]. Furthermore, experimental data is noisy and often sparse, making inference of quantitative analytic descriptions even more challenging.

We show a case study of perovskite degradation to infer physical laws from data and we interpret them in the context of proposed theoretical degradation pathways for perovskites [1]. We analyze simulated datasets for the degradation of perovskites under heat and continue to experimental data: Methylammonium lead iodide (MAPI) films subjected to degradation in our in-house environmental chamber at various temperatures, low humidity, and 0.15 sun illumination. We monitor their degradation using an optical camera extracting color versus time, capturing the transition from black MAPI to yellow lead iodide (PbI₂). Utilizing experimental and modeled data, we conduct a systematic study modeling degradation processes using a (1) traditional regression approach, (2) fitting parameters to partial differential equations (PDEs), and finally shifting to (3) symbolic regression of PDEs [2, 3, 4].

We find that the regression approaches extract variables of interest with relative errors as low as 2%, even with 5% Gaussian noise superimposed upon the simulated datasets. However, this requires knowledge of the equation form. We demonstrate that sparse regression [2] identifies the underlying equation from noiseless data with errors less than 10%. We have PDE extraction calculations underway, with preliminary results, governing physical laws for MAPI degradation using a genetic programming algorithm for symbolic regression [3]. We examine three data-driven algorithms for learning differential equations and find that they are generally challenged to fit exponential equations (common to many materials-science problems) and handle noise. We conclude that the study of materials degradation can be rendered more quantitative, by including inference of physical laws of degradation, using regression and potentially ML-based PDE extraction techniques. We also opine about the gap that exists between ML-based tools today and common materials science challenges.

[1] Smecca, Emanuele, et al. "Stability of solution-processed MAPbI 3 and FAPbI 3 layers." Physical Chemistry Chemical Physics 18.19 (2016): 13413-13422

[2] Rudy, Samuel H., et al. "Data-driven discovery of partial differential equations." Science Advances
3.4 (2017): e1602614.

[3] Atkinson, Steven, et al. "Data-driven discovery of free-form governing differential equations." arXiv preprint arXiv:1910.05117 (2019).

[4] Willard, Jared, et al. "Integrating physics-based modeling with machine learning: A survey." arXiv preprint arXiv:2003.04919 (2020).

Autonomous laboratory systems continually learn by adaptively planning and executing campaigns of physical and/or in silico experiments to achieve a scientific or engineering goal without requiring direct human intervention. The current research community is demonstrating new opportunities in accelerated materials synthesis, evaluation, and optimization, but major challenges remain for application of robotic systems for broader scientific discovery in materials and chemistry. Foremost, current AI systems used in robotic science applications make and test predictions, but lack the mechanistic modeling components needed for formulating and evaluating scientific hypotheses that explain these predictions in terms of generalizable physical principles. Furthermore, learning to reliably synthesize materials, and reconciling multiple structure and property data streams make it difficult to comprehensively explore complex and multiscale structures that determine the performance of engineering materials.

This talk will focus on technical and methodological aspects of building AI agents that drive closed-loop synthesis and characterization platforms and deploying them for scientific discovery. We explore these issues using an autonomous scanning droplet cell (ASDC) platform designed for on-demand alloy electrodeposition and real-time electrochemical characterization. Our initial studies seek to jointly optimize electrochemical deposition of binary ZnNi alloys while maximizing their corrosion resistance properties. We present exploratory research incorporating mechanistic transport models into the Gaussian Process modeling framework we use for surrogate modeling and active selection of experiments. We also discuss the challenge of disentangling the effects of chemistry, crystal structure, and microstructure on material properties, focusing on corrosion resistance within a ternary NbTiTa refractory alloy system that we have characterized with a multimodal X-ray diffraction and corrosion study.

As the materials community surmounts the challenge of incorporating mechanistic modeling components into our machine learning systems, automated materials synthesis and characterization platforms offer much more than black-box optimization of engineering materials: they offer the potential to develop and deploy new algorithms for generating and testing scientific hypotheses.

Representations play a crucial role in building predictive models using data driven approaches. In this work we present a representation based approach to high-throughput experimental data analysis of X-ray diffraction (XRD) and Cyclic Voltammetry (CV)--where the input is systematically swept and the response is collected for all the materials screened. A common data representation in above cases is to use the output as an n-dimensional vector with discrete inputs as corresponding dimensions/components. In this talk, we present a couple of other representations namely a metric space and a function space representation with a goal to encode known physical knowledge. We evaluate the representations in two different case studies with applications involving material discovery and cyclic voltammetry. Our first case study involves using a metric space representation where topology of the sampled data is captured using a physics based metric. Metric space representation presented in this talk is an automated way of computing distance measure and achieves comparable accuracy to computationally complex, expert defined distance measures. A second case study involves a function space representation tailored for CV data analytics. We use covariance functions of Gaussian process (GP) as basis functions to encode the two dimensional CV curve onto a vector space represented by the basis functions. We show that the GP based methodology presented here is analogous to using images with each pixel represented using a Gaussian distribution thus allowing us to use well developed image analysis techniques for downstream analysis. We present two case studies of function space representations 1) iterative design of experiments to identify materials with fast kinetics in a electro-chemical environment with application to catalyst discovery of oxygen reduction and evolution reaction 2) iterative identification of an optimal mechanism from a set of postulated mechanisms by combining experiments with a physics based model. Approaches presented here will also highlight the importance of building a physics based representation of data for downstream analysis.

Efficient discovery of novel functional materials strongly relies on development of the exploration strategies in the chemical space. In this contribution we are describing an approach that helps to identify lacunae in the existing data and complete the missing pieces in a targeted manner. The proposed approach consists of 2 main steps.

Step 1: We perform topological data analysis on the set of molecules. Molecular set is treated as a point cloud with pair-wise distances defined using standard chemoinformatic approaches. We apply the Mapper algorithm [1] to produce a simplified description of the data in a form of a graph G=(C,E), where C is a set of clusters represented as nodes, and E is the set of all edges. Each node in the graph represents a cluster of molecules, and an edge between 2 clusters indicates that the clusters overlap. Mapper graph directly visualizes such aspects of the data shape as holes and branches (loops and flares on the Mapper graph, respectively), which indicate that there is missing data. Our task, therefore, is to generate new molecules that fill out the loops (elimination of the branches can be done in a similar fashion). First, we search for the scaffolds in the dataset that are most promising for the molecule generation. Operationally, we prioritize scaffolds found in small clusters along big loops on the Mapper graph. For each cluster we identify the smallest (in terms of hops) loop l using the Dijkstra's algorithm [2], then we calculate the length of the cycle w_l, which is the sum of the length of each edge in the loop. For each molecule in each cluster we calculate the Bemis-Murcko scaffolds S={s₁,s₂,...s_n}. For each scaffold s we introduce the generative potential g_s = Σ_cs mean(|C|) w_lc /|c_i| where c_s are all the clusters where the scaffold appears, and the w_lc is the length of the cycle for that cluster. A high g_s value indicates that the scaffold appears in small clusters with large cycle length, meaning they are part of bigger loops, and have a higher potential for generating novel molecules. We normalize the g_s values between 0 and 1.

Step 2: To generate novel molecules that complete loops on Mapper graph we adapt an existing graph-generative model for scaffold-based molecular design, using variational autoencoders (VAE). Here, the input scaffold is extended by sequentially adding atoms and bonds, i.e., the generation is conditioned on the existing scaffolds, making sure that all generated molecules contain the input scaffold. We modify the standard VAE loss function to take in consideration the generative potential of the scaffolds: L = (1 - g_s) (L_r + L_KM + α (g_s - g_sn)), where α is a hyperparameter between 0 and 1, g_sn is the generative potential of the scaffold of the newly generated molecule, L_r is the reconstruction error of the input s and generated scaffold s_n and L_KL is the Kullback-Leibler divergence between the prior and approximate posterior distribution. With this loss function we reduce the influence of the scaffolds with low generative potential and penalize the model when it generates molecules with low generative potential g_sn. To calculate the g_sn of a new scaffold, we include the newly generated molecule in the Mapper analysis after each iteration.

We will discuss the application and performance of the described approach to the exploration of the space of photo-acid generators.

[1] Gurjeet, S., Mémoli, F., & Carlsson, G. E. "Topological methods for the analysis of high dimensional data sets and 3d object recognition." SPBG (2007).
[2] Dijkstra, Edsger W. "A note on two problems in connexion with graphs." Numerische mathematik 1.1 (1959): 269-271.
[3] Lim, J., Hwang, S. Y., Moon, S., Kim, S., & Kim, W. Y. "Scaffold-based molecular design with a graph generative model." Chemical Science (2020)

Much progress in molecular property prediction has been made by generating and curating datasets, and developing machine-learning (ML) models. However, to increase predictive accuracy of machine-learning-driven research, combining these two with improved data representations (which capture the structural and chemical features relevant for property prediction) remains an outstanding challenge. This is especially true for large, non-rigid molecules containing more than 30 atoms. Proper data representation maximizes the probability that an ML algorithm will successfully infer correlations between molecular structure and chemistry and output properties of interest.

There are a multitude of successful molecule representations in literature [1-2], but a remaining limitation especially in understanding large molecules is the ability to efficiently capture both chemistry and three-dimensional (3D) structure. Molecular fingerprint sets can be designed as a combination of features reflecting the 3D structure and chemistry of the molecule but may have a limited generalizability. SMILES strings provide a straightforward method for mapping the molecular structure into 1D, and graph representations develop the idea further to capture bonds between the molecules, albeit using a dimensionless representation.

We investigate molecule representations from the viewpoint of capturing both chemistry and 3D structure. We show a case study of predicting antibiotic potential of large (20–100 carbon atoms) antibiotic molecule candidates via an experimentally measured indicator, minimum inhibitory concentration. Our consistent investigation involves comparison of different representations ranging from chemistry only to ones attempting to integrate 1D – 3D structure (chemistry-only and structure-capturing features downselected automatically from a set of 5000 potential features, SMILES, and molecular bond graph with combinations of the previous ones), compared in traditional and complex models (including random forest regression and directed message passing neural network [3]).

We demonstrate the power of improving model predictions by focusing on optimum representation, show that representation can be tuned utilizing models with faster ramp up time, and find that structure-capturing representation improves predictive accuracy especially when the data consists of large, non-rigid molecules. Our systematic case study highlights the importance of data representation when optimizing for predictive accuracy.

References:
[1] Elton, Daniel C., Zois Boukouvalas, Mark D. Fuge, and Peter W. Chung. "Deep learning for molecular design - a review of the state of the art." Molecular Systems Design & Engineering 4, no. 4 (2019): 828-849.
[2] Schwalbe-Koda, Daniel, and Rafael Gómez-Bombarelli. "Generative models for automatic chemical design." In Machine Learning Meets Quantum Physics, pp. 445-467. Springer, Cham, 2020.
[3] Yang, Kevin, Kyle Swanson, Wengong Jin, Connor Coley, Philipp Eiden, Hua Gao, Angel Guzman-Perez et al. "Analyzing learned molecular representations for property prediction." Journal of chemical information and modeling 59, no. 8 (2019): 3370-3388.

The development of high performance, stable solar cell materials poses a multi-objective optimization problem in a considerable multidimensional parameter space. Solving it requires reproducible, user independent laboratory work and intelligent preselection of conducted experiments. However, research on organic photovoltaics (OPV) is a field where manual routines are still predominant, although other domains like pharmacy have long used robotics and automation.
To better utilize automation in materials science, we introduce AMANDA (Autonomous Materials and Device Application Platform), a generic software framework for distributed materials research and its research line LineOne (L1), a robotic facility for the production and characterization of solution processed solar cells.
We integrated pipetting and transport robots with spin coating and evaporation units as well as characterization tools like spectrometers and electrical measurements into LineOne. With this setup, the AMANDA Platform automatically produces and measures OPV cells without human interaction while it is capable of rapidly screening large materials and processing parameter spaces in a combinatorial approach at high precision. The current hardware configuration is fully scalable and allows the high throughput screening of up to 270 variations per day on LineOne. The output comprises the complete fabrication of solar cells, full characterization and comprehensive documentation of all applied process steps as well as compiled results for each individual solar cell. We demonstrate the capabilities of AMANDA by the process development for OPV solar cells based on PM6:Y6 with 13.7% efficiency when processed in air and through the detailed investigations of the impact of subtle process parameters, like solution concentration and volume of solution deposited for spin coating, on film quality and performance.

Formulation chemistry is big business for a wide variety of industries, where formulated fluids cover a market worth hundreds of billions of dollars per year. Formulations for printing applications represent one of many components of this market and include a diverse range of products from consumer inks to advanced electronic materials. Industrial ink formulations can be extremely complex with a wide range of specifically selected components to impart multifunctionality on the final dry film. Film materials have important requirements for optical, electrical, and mechanical properties as well as longer-term chemical, temperature, and UV stability. In addition, formulations must meet specific surface tension and viscosity criteria to be printable. Meeting these complex requirements requires optimization over many dimensions. In our experience working with various specialty chemical companies, we have observed that, most often, these formulation optimization experiments are still performed using slow, labor-intensive workflows. This dramatically limits what can be explored given realistic budget constraints.

In recent years, advances in machine learning together with greater accessibility of robotic automation tools have enabled a new paradigm of accelerated materials development using autonomous research systems. In the printed electronics space, several exemplary autonomous research systems have been recently created.^1,2 However, these systems have been designed to create thin films by spin-coating or drop-casting, which are less applicable to large-area, high-throughput industrial manufacturing processes. Previous systems have also used robotic pipetting to measure out different formulations in a slow batch-based process.

To tackle real industry problems with our clients, we have developed a modular automated system for rapid flow-based microfluidic formulation, fluid characterization, inkjet printing, and film characterization. As opposed to the methods used in previous autonomous systems, inkjet printing is a key enabling additive manufacturing technology that is mature and widely used for high volume manufacturing. The system includes multiple characterization modes along the workflow, including shear-rate-dependent viscosity, complex jetting behavior, and final film quality. Where possible, we have also been developing low-cost hardware modules to help alleviate the sticker shock that often occurs when new automated systems are proposed to business units with tight R&D budgets.

In our first case study, the system was able to improve upon an existing manual process used to develop polymer films for display applications. The legacy process took around 35 person-hours per recipe, including slow ink formulation, printing, characterization, and cleaning. The new system reduced that worker time to almost zero and features automated cleaning, freeing up skilled researchers’ time to work on other tasks. The system now takes about 15 minutes to investigate each recipe with minimal human interaction and is also able to rapidly map out the multidimensional printability window. One of the key enablers of this was the inkjet stroboscopic imaging setup. There is a wealth of information in the dynamic shape of jetted droplets which correlates strongly with the viscoelasticity, surface chemistry, and microstructure of the fluid. We are currently implementing an experiment planning and scheduling system that can execute various strategies for hierarchical optimization. With this in place, the system will not simply maximize sample throughput but also focus on performing experiments where the likelihood of knowledge gain is the highest.

1. Langner, S. et al. Beyond Ternary OPV: High-Throughput Experimentation and Self-Driving Laboratories Optimize Multicomponent Systems. Adv. Mater. 32, 1907801 (2020).
2. MacLeod, B. P. et al. Self-driving laboratory for accelerated discovery of thin-film materials. Sci. Adv. 6, eaaz8867 (2020).

We present our work on the development and application of a mesofluidic electrochemical deposition and analysis platform. Our scanning droplet cell (SDC), coupled with a bank of syringe pumps is capable of serial electrodeposition and electrochemical corrosion measurements on millimeter scale areas. We have also demonstrated integration of our platform into the 6-BM end-station at NSLS-II to add in situ x-ray fluorescence and diffraction capabilities. In addition to describing this platform in detail, we will present two validation studies: In the first we use this system for serial corrosion measurements of sputtered combinatorial thin-film metallic glass and high-entropy alloy libraries. Second, we will discuss our on-going efforts to combine synthesis and characterization, controlled by an autonomous AI agent. In this work we look at the effect of electrodeposition processing parameters for Ni-Zn plating, such as solution concentration and bias, on alloy corrosion resistance, composition, and structure. Our goal here is to experimentally generate a machine learning based model to describe optimal deposition conditions for an arbitrary alloy composition in its most corrosion resistant state and describe the corrosion behavior of that alloy. In this way, we hope that this system can transcend simple optimization and can begin to decouple the interplay of deposition, structure, and corrosion resistance in order to generate new knowledge.

With ever increasing demand of the modern world for the materials with advanced properties, we are facing the pressure to accelerate the discovery through a paradigmatic shift. One of the major practical bottlenecks – actionability bottleneck - is majority of the computationally generated hypotheses fail to transition to the experimental phase. In this case, the target properties are expected to meet project requirements, but the candidates fail to satisfy broader project-specific constraints. The most problematic constraints represent institutional knowledge, broadly defined. Institutional knowledge is not systematically captured in publications, but rather carried by subject matter experts (SMEs). This explains fundamental difficulty of capturing institutional knowledge, introducing it into the training phase of statistical modeling, and propagating into the hypothesis generation phase.
This contribution addresses resolution of the actionability bottleneck via capture and instantiation of SME’s creativity and institutional knowledge in the form of AI models. This is accomplished using a two-tiered expert-in-the-loop (EITL) approach [1,2]. Specifically, we considered monomer design for acrylic polymers. The first tier of EITL AI included training a discriminator that was reproducing SME’s decision-making in the adjudication of the candidate monomers. The candidate monomers were generated using recently developed system for molecular inverse-design [3]. The SME – a polymer chemist responsible for the polymerization experiments – accepted or rejected the candidates as they answered the question: “Is it practical to synthesize this monomer?” The question, as discussed with the SME, implied subjective assessment of the synthetic accessibility, expected behavior during polymerization, and novelty. The second tier of EITL AI leveraged the adjudication results and the trained discriminator to train a generative model producing monomer candidates that were likely to be accepted by the SME. We trained 2 sequential neural network models and 3 generative adversarial networks (GANs); generative models were refined via separate adjudication of their output until acceptable performance was reached. At this point the generative form of EITL AI combined implicit models of SME’s knowledge and creativity as both these components informed the adjudication processes.
The trained generative component of EITL AI was queried by the SME to identify candidates for the experimental phase. The generated structures were evaluated on the basis of novelty and synthetic complexity. Selected structures were additionally assessed based on the following criteria:

- the ease of synthesis or commercial availability for the monomer precursor
- the number of transformations, their difficulty, commercial availability and price of starting materials and reagents
- the ease of polymerizable handle installation on the monomer precursor
- the ability of the obtained monomer to be polymerized in a free radical polymerization process

Five monomers were selected to produce acrylic polymers. Performance metrics of the AI models, experimental procedures, and polymer characterization results will be discussed along with existing challenges and potential extensions.

We present the development of high-throughput, data-enabled strategies for understanding the crystallization and chemical transformations of metal halide perovskites in the form of macroscopic single crystals and nanocrystals. These approaches combine robotic synthesis workflows developed at the Molecular Foundry with data-driven approaches that can predict crystallization outcomes and elucidate reaction networks. We will first discuss our Robot-Accelerated Perovskite Investigation and Discovery (RAPID) workflow, which we used to rapidly identify and optimize synthesis conditions for the formation of high quality single crystals via inverse temperature crystallization (ITC). A total of 8172 metal halide perovskite synthesis reactions were performed using 45 organic ammonium cations. Robotic screening increased the number of metal halide perovskite materials accessible by an ITC synthesis route by more than five-fold and resulted in the formation of two new phases. The resulting dataset enabled the construction and evaluation of machine learning models that were able to predict the formation of a diverse range of crystal compositions and morphologies with >85% accuracy after only 100 reactions. We used the same robotic apparatus to systematically explore the reaction networks that produce cesium lead bromide nanocrystals. Deploying automated spectroscopy and a custom deconvolution algorithm to extract the relative populations of each reaction product (i.e. molecular clusters, nanoplatelets, and nanocubes), we generated phase maps of the high-dimensional Cs-Pb-Br synthetic space. Based on these maps, we attempted chemical transformations from each of the as-synthesized products to all other species. The resulting transformation data revealed a dynamic equilibrium between species and allowed us to construct chemical reaction network that provides a holistic view of this weakly reversible system. This work demonstrates the concept that combining high throughput experimentation, machine learning, and data-enabled approaches accelerates and enhances the study of metal halide perovskite crystallization, and that this approach is generalizable across different synthetic routes, materials, and their length scales.

The acceleration of high-throughput experimental discovery involves several recurring challenges such as data ingestion, visualization, analysis, and recommendation. In this talk, we describe how the ACE (AI-assisted Catalyst Experimentation) framework addresses these challenges and enables scientists to improve decision-making in high-throughput experimentation.

We highlight the following aspects of our framework:

1. Data Ingestion: We describe an experimental database schema that has minimal built-in assumptions, follows FAIR principles, and is expected to generalize to a wide variety of scientific applications using one of the largest experimental datasets as an example.¹

2. Data Visualization: Scientists can easily navigate the database by using a domain specific language as part of the UI. These queries are bolstered with visual aids about metadata. In addition, the UI automatically provides a set of useful visualizations for dataset statistics, trends, and details of individual samples and processes.

3. Analysis: Analysis of materials characterization data remains one of the critical bottlenecks for high-throughput experimental materials discovery. We present specific examples for AI-assisted analysis of X-ray spectroscopy to demonstrate the capabilities of our framework to address this bottleneck.

4. Recommendation: The current recommendation capabilities of our framework include a) sequential recommendation for accelerated discovery within a composition space, b) recommendation of composition spaces for experiments using representation learning. We demonstrate these capabilities for learning formation energy, band gap, and catalytic properties. We also show human and machine benchmarks for sequential learning within a specific composition space.

These components that make up ACE are designed to be modular and generalizable; and aim to provide foundational infrastructure for accelerated adoption of machine learning for diverse high-throughput experimental applications.

References:
1. Soedarmadji, E. et al. Tracking materials science data lineage to manage millions of materials experiments and analyses. npj Comput Mater 5, 79 (2019).

High-throughput autonomous research has recently emerged as the new frontier of accelerated materials discovery. This success has sparked efforts to improve not only lab automation infrastructure but also optimization algorithms. In materials optimization campaigns, sequential learning approaches such as Bayesian Optimization (BO) have gained great popularity due to their ability to improve parameters of merit with fewer number of experiment cycles in comparison to brute-force approaches. Previous attempts to quantitatively evaluate the acceleration of research through sequential learning have been limited to specific materials systems [1]. Therefore, two open questions that remain are: (a) how does performance of BO in sequential learning approaches vary across multiple experimental domains and manifold complexities. (b) how we could improve upon current optimization algorithms and attend to the needs of a broader materials science community.

In this work, we benchmark the performance of sequential learning approaches within the BO framework across four experimental domains. Each focusing on a different materials system, four experimental datasets with varying sizes, dimensionalities and dataset manifold complexities are used to build representations of ground truths, where we draw from these to simulate optimization campaigns. We compare the performance of BO algorithms employing one of three surrogate models and five myopic acquisition functions. The performance of each surrogate model and acquisition function pair is benchmarked against random acquisition, and further quantified via learning rate metrics designed to measure different research objectives.

It is observed that BO consistently reduces the number of experiment cycles necessary to reach superior materials response by factors of 5 to 15 across all experimental systems. Random forest surrogate model paired with a lower confidence bound acquisition function of balanced mean and standard deviation shows higher-ranking performance for research tasks such as finding any or all of the candidates with top 5% response. Interestingly, in the case of surrogate models, random forests outclass gaussian processes in optimization tasks when using the same acquisition functions across all datasets, likely due to their nature of making fewer structural assumptions about manifold landscapes. In addition, we further explore the benefits of utilizing look-ahead decision policies as an alternative to the traditional myopic acquisition functions in the BO framework. Complementary to our findings, our benchmarking effort has yielded many useful insights on how experimentalists interested in novel materials systems could select the most suitable BO methods prior to embarking on an optimization campaign. These can be utilized by a broader materials science community in further progressing the field of autonomous research.


[1] Rohr, B., Stein, H.S., Guevarra, D., Wang, Y., Haber, J.A., Aykol, M., Suram, S.K. and Gregoire, J.M., 2020. Benchmarking the acceleration of materials discovery by sequential learning. Chemical Science, 11(10), pp.2696-2706.

Developing novel scalable fabrication processes for perovskite PV devices involves optimizing many process variables, which is challenging for human intuition alone to solve the problem efficiently. Autonomous high-throughput experimentation has recently demonstrated a very encouraging outcome to aid rapid materials development^1–3. Preliminary success has also been demonstrated by introducing machine learning (ML) in the loop to guide the experiments⁴ for developing a fabrication process. However, the black-box nature of machine learning is a key limitation that hinders wide adoption in physical science research⁵. Although ML is a powerful statistical tool to navigate and optimize in a high-dimensional space, rigorously-trained materials scientists are not able to easily understand the decisions made by ML algorithms. Due to this lack of model interpretability, researchers usually find it difficult to intellectually engage with ML-driven optimization.
In this work, we investigate how we can work cooperatively with an ML algorithm and maximize the acceleration of the optimization process. To do so, we revisit a historical dataset of perovskite devices with open-air spray deposition, and then adopt commonly used sequential learning methods to navigate the process optimization of perovskite solar cells. Based on this case study, we find three generalizable points to better utilize machine learning tools for materials research. First, initial sampling is crucial to ensure rapid optimization. We compare different initial sampling methods, e.g., chronological, random, and Latin hypercube sequences, and find the initialization method plays an important role to facilitate the subsequent optimization. Without a good initial design of experiments guided by domain knowledge, the acceleration benefits with an ML algorithm is very limited. Second, the acquisition and objective functions need to be tuned for different research objectives of optimizing device performance or building a predictive regression model. By assigning different weights to “exploitation” and “exploration”, we evaluate the optimization outcomes for various acquisition functions with different regression models. Echoing previous findings by Rohr et al.⁶, we also observe that acceleration in building a predictive model is not necessarily easy with sequential learning algorithms, but it could significantly accelerate the process optimization of device efficiency (where up to 2x acceleration could be found). Third, the algorithm interpretability allows researchers to engage with the ML model and avoid mistakes. At each step, our framework helps visualize how sampling decisions are made by the ML algorithm: whether an acquisition decision is influenced more by the predicted mean or the predicted uncertainty, whether it favors exploration in the unsampled regions or exploitation in the known regions, or which process variable has the highest impact on the decision. We envision the impact of this work as a human-in-loop machine-learning platform where researchers can understand the framework of ML and utilize ML algorithms to accelerate the research of novel materials.

Reference:
1. Stein, H. S. & Gregoire, J. M. Progress and prospects for accelerating materials science with automated and autonomous workflows. Chem. Sci. 10, 9640–9649 (2019).
2. MacLeod, B. P. et al. Self-driving laboratory for accelerated discovery of thin-film materials. Sci. Adv. 6, eaaz8867 (2020).
3. Li, Z. et al. Robot-Accelerated Perovskite Investigation and Discovery (RAPID): 1. Inverse Temperature Crystallization. (2019) chemrxiv.10013090.v1.
4. Attia, P. M. et al. Closed-loop optimization of fast-charging protocols for batteries with machine learning. Nature 578, 397–402 (2020).
5. Towards trustable machine learning. Nat. Biomed. Eng. 2, 709–710 (2018).
6. Rohr, B. et al. Benchmarking the acceleration of materials discovery by sequential learning. Chem. Sci. 11, 2696–2706 (2020).

Metal halide perovskite quantum dots (QDs) have recently emerged as an exciting class of semiconducting materials, outperforming conventional II-VI, IV-VI, and III-V semiconductor nanocrystals in QD-based optoelectronic devices.¹ Despite the substantial improvement in performance of perovskite QD-based optoelectronic devices over the past five years, two major obstacles hinder their development: (i) Edisonian (trial and error)-based QD synthesis, discovery, and optimization methods, and (ii) poor size dispersity of batch-synthesized colloidal nanocrystals, resulting in surface trap states. Existing material development strategies very often fail to overcome the demands of colloidal QDs’ vast colloidal synthesis and processing universe, resulting in time- and cost-intensive QD development efforts.
Recent advances in machine learning (ML),² including deep neural networks (DNNs) and reinforcement learning algorithms, provide an exciting opportunity for reshaping the synthesis and development of perovskite QDs, through the machine-based direction of a high-throughput QD synthesis bot. Capitalizing on the recent progress of ML-based optimization algorithms, smart QD manufacturing strategies relying on DNNs trained on experimentally measured QD properties can be envisioned to accelerate the discovery, synthesis, and manufacturing of metal halide perovskite QDs with application-guided optoelectronic properties.
In this talk, I will present an Artificial Chemist,³ a modular intelligent fluidic microprocessor⁴ capable of autonomous synthesis, optimization, and end-to-end manufacturing of colloidal QDs for direct utilization in high-performance and low-cost optoelectronic devices. The Artificial Chemist can rapidly and efficiently (i) explore the massive parameter space of colloidal perovskite QDs, (ii) learn their synthesis and processing schemes, (iii) identify the composition and relevant synthesis route(s) of perovskite QDs to achieve specific optoelectronic properties, and (iv) continuously manufacture the rapidly optimized perovskite QDs at a fraction of time/cost of batch techniques. The reconfigurable Artificial Chemist technology utilizes a multimodal in-situ material diagnostic probe (absorption/photoluminescence (PL) spectroscopy) in conjunction with a real-time, ensemble DNN adaptive sampling algorithm to enable simultaneous optimization of PL quantum yield and emission linewidth of colloidal QDs for any desired peak emission energy. The Artificial Chemist technology can be readily adapted for accelerated development and autonomous end-to-end manufacturing of other solution-processed nanomaterials.

References
1. (a) Hazarika, A., et al., ACS Nano 2018, 12 (10), 10327-10337; (b) Sanehira, E. M., et al., Sci. Adv. 2017, 3 (10); (c) Swarnkar, A., et al., Science 2016, 354 (6308), 92-95; (d) Yoon, H. C., et al., ACS Applied Materials & Interfaces 2016, 8 (28), 18189-18200; (e) Lu, M., et al., Advanced Functional Materials 0 (0), 1902008.
2. (a) LeCun, Y., et al., Nature 2015, 521, 436; (b) Le, T. C.; Winkler, D. A., Chem. Rev. 2016, 116 (10), 6107-6132.
3. Epps, R. W., et al., Advanced Materials 2020, (in press), 2001626.
https://doi.org/10.1002/adma.202001626
4. (a) Abdel-Latif, K., et al., Advanced Functional Materials 2019, 29 (23), 1900712; (b) Epps, R. W., et al., Lab Chip 2017, 17 (23), 4040-4047.

Phase diagrams are a fundamental tool in materials design, but thorough experimental determination is challenging, expensive, and time consuming. Phase diagrams calculated entirely from first-principles may reduce time and expense, providing information at the prediction stage. Our previous work demonstrated a methodology to obtain a first-principles only CALPHAD-type solid phase diagram reproducing all major features, with little or no prior knowledge of the system [1]. This can guide reduced experiments needed for database validation.

Considering the quantified uncertainty of the first-principles phase diagram using PDUQ [2] and ESPEI [3], a sequential learning approach is taken to systematically add thermodynamic and phase boundary data in regions of highest uncertainty. This simulates how the first-principles only phase diagram can most efficiently guide experimental investigation towards a thorough understanding of a materials system. This convergence towards a target phase diagram will be quantified and compared to the chronological introduction of experimental data.

[1] T. Davey et al., CALPHAD XLVIII, June 2019.
[2] N. Paulson et al., Acta Mater. 174 (2019) 9–15.
[3] B. Bocklund et al., MRS Commun. (2019) 1-10.

A number of competing frameworks for learning interatomic potentials have been proposed in the last decade, including Gaussian Approximation Potentials (GAP), Spectral Neighbor Analysis Potentials (SNAP), and the Atomic Cluster Expansion (ACE). We here present a Bayesian approach that unites these three frameworks, allowing uncertainty estimates on many-body energies, forces, and stresses to be rigorously computed. We first show how a multi-species generalization of the 3-body ACE descriptor can be used to define an efficient many-body kernel similar to the Smooth Overlap of Atomic Positions kernel used in GAP models. We use this kernel to construct Gaussian process (GP) models of the potential energy surface, and show that uncertainties derived from the predictive posterior distribution of the GP correlate with true model error on independent test sets. Mean predictions of the GP are shown to be linear in an integer power of the descriptor, eliminating the expensive loop over training points required in GP regression and establishing a connection to the SNAP framework. The resulting linear models are implemented in the molecular dynamics program LAMMPS, and shown to achieve competitive performance for a range of molecular and extended systems.

The predictive uncertainties of our Gaussian process models open the door to automated training during molecular dynamics simulations, allowing a wide range of material compositions to be explored in a closed-loop fashion. In this approach, training structures are selected with active learning by making calls to density functional theory only when the predictive uncertainty on a force component exceeds a threshold value. We apply this on-the-fly molecular dynamics method to the shape memory alloy nitinol, and show that we accurately capture the martensite/austenite phase transition that occurs in this material at ambient pressure. Guided by model uncertainties, we systematically expand the training set to examine the effects of dopants, defects, and nonequiatomic compositions on the martensitic transition temperature.

Automating materials calculations to populate a searchable and curated web-accessible database is non-trivial problem. Moreover, we are interested in enabling the model creation, initialisation of simulations and control of the workflow (as well as the content of the database) from a webpage interface which unfortunately leads to an additional conflict between increasing the desirable accessibility/functionality whilst maintaining security to prevent malicious cyber-attacks. Over a period of 10 years we have developed and implemented a combined hardware (housing the webserver, database and dedicated compute node) and software solution to handle such workflows. To date, we have applied our approach to assisting two different communities: those interested in research based on the atomic structure of clusters, or small sized nanoparticles; and those interested in simulations of surfaces and their reactivity with probe molecules. The former is part of the WASP@N project, whereas the latter is part of the UK’s Materials Chemistry HEC Consortium’s SAINT project. Both offer access to a database of their respective models, which users can also contribute, and include useful tools for analysis via their web-interfaces.

WASP@N (Web-Assisted Structure Prediction at the Nanoscale) was developed to support the dissemination and comparison of atomic configurations found from global optimisation studies of energy landscapes, which can subsequently be used as models for further studies (their electronic and physical properties or their interaction with other media). All uploaded structures are automatically optimised using the FHI-aims electronic structure code and hyperlinks – based on either composition, energy, and connectivity – that users can follow are created between both the original and refined structures and other entries in the database (all uploaded structures are also linked via DOI to their original publication).

SAINT (The Surface and Interface Toolkit) was developed to support researchers’ day-to-day surface simulations, by providing tools for generating surface models and enabling automatic calculation of surface energies using the CRYSTAL electronic structure code. SAINT offers fully relaxed surface models for a range of metal oxides and maps of their reactivity to three probe molecules; the ability to generate surfaces and then download input files for a range of different DFT and IP codes; and a method of dissemination of published surface models.

Here we will first discuss our insights into hardware and software solutions for such automation and database population along with the need to balance security of the data with ease of accessibility and how this can be applied to other databases and automation. Next, we will demonstrate a case-study using SAINT to produce reactivity maps of non-polar surfaces for a range of compounds with respect to three probe molecules, and how this information can streamline simulations of surfaces with more complex reaction species.

Optimizing multicomponent chemical processes is a challenge which lends itself well to automation due to the large number of experiments usually required. In addition to classic high throughput and Design of Experiment methodologies, the application of online analytical tools provides real-time feedback and therefore acute control over the chemistry in question. This approach enables fundamental insights to be translated into practical application at an unprecedented pace, particularly when coupled to machine learning capabilities. Herein we present two examples of automated process analytical technology designed to resolve bottlenecks in materials R&D and chemical manufacturing.

The supply of specific, high-performance molecular conductors is critical to the development of organic photovoltaic devices. Unfortunately, the gold-standard molecule used almost ubiquitously (Spiro-OMeTAD) is notoriously expensive as its synthesis remains underdeveloped. To address this challenge, we have developed a range of automated tools capable of formulating reactions, taking samples, and performing online HPLC analysis. Our versatile platforms are also compatible with air sensitive chemistry occurring in an inert atmosphere. We use these tools to generate temporal reaction profiles and rapidly optimize the synthesis of key materials such as Spiro-OMeTAD.

As a second case study, we apply advanced crystallization methodologies inspired by the pharmaceutical industry to access battery-grade lithium carbonate from waste brines. The dynamic nature of crystallization requires the continuous assessment of sample purity and the corresponding manipulation of crystallization parameters. Our suite of process analytical technology provides comprehensive data on sample composition (IC, FTIR), turbidity (computer vision), and crystal growth (microscopy) to enable feedback control and consistently isolate high-purity product.

Due to its computational efficiency compared to other quantum mechanical methods, semi-local density functional theory (DFT) is a popular tool for simulating large, complex systems like surfaces, interfaces, point defects, and battery cathode materials. However, the complex electronic structure of these systems cannot be described accurately by the approximate exchange-correlation (XC) functionals used in semi-local DFT. More accurate methods, such as exact exchange and post-Hartree Fock, are prohibitively expensive for large systems. To remedy this problem, we introduce nonlocal exchange density functionals that are fit via machine learning (ML) and can be computed more efficiently than the exact exchange energy. We build on work in both analytical and ML functionals to perform feature engineering that enables the use of lightweight, smooth, and transferable machine learning models. We explore how Gaussian Processes can be used to construct uncertainty estimates of the exchange energy and assess the performance of different kernels. To construct large training databases, we also develop code to evaluate the exchange energy density on real-space grids. To benchmark this new class of functionals, we implement the functional derivatives to find the exchange potential and use it in self-consistent field calculations on a diverse set of small molecules.

The completely automated nature of high-throughput density functional theory (HT-DFT) creates unique challenges, such as how to choose DFT parameters and post-processing techniques that work across all materials classes, and strike an effective balance between computational cost and accuracy. These challenges have been tackled in different ways by the several large HT-DFT databases that have emerged so far. However, the effect of the different approaches on the calculated materials properties is as yet unclear. Here, we quantify the uncertainty in the HT-DFT properties resulting from such different-yet-reasonable parameter choices, by consolidating three large HT-DFT databases: Automatic-FLOW (AFLOW), Materials Project, and the Open Quantum Materials Database (OQMD), onto a common data platform and analyzing comparable records. Overall, we find the variance in reported properties between HT-DFT databases to be comparable to that between DFT and experiment. We identify the underlying physical factors responsible for some of the larger discrepancies, and outline paths toward further standardization of HT-DFT.

Aldehydes, in particular formaldehyde (CH₂O), are key precursors of many important chemical compounds. Cu-based catalysts are attractive for its production since they allow anhydrous alcohol dehydrogenation that does not need expensive distillation and produces a valuable byproduct, H₂.

However, the fundamental mechanism of alcohol dehydrogenation on Cu is not yet fully understood. Static DFT calculations fail to predict the nonlinear dependency of the dehydrogenation reaction rates on the size of the alkoxy molecule [1]. To address this problem, we use molecular dynamics to directly model the dehydrogenation process on the Cu(110) surface, which can capture the dynamical effects that are missed in static DFT calculations.

But due to the slow reaction rates, it is prohibitive to directly compute these reactions with ab initio molecular dynamics. Therefore, we use a two-step approach to accelerate molecular dynamics simulations. First, DFT force calculations are replaced by machine learning force fields. We train two different models, Tensor-Field neural networks and mapped Gaussian processes[2]. Both models can reach near DFT level accuracy at much lower computational cost. Second, we use nonlinear dimensionality reduction to learn reaction coordinates and employ them to explore the free energy landscape around the reaction path, speeding up sampling of rare transition events and obtaining reaction rates from free energy barriers.

[1] Chen, W.; Cubuk, E. D.; Montemore, M. M.; Reece, C.; Madix, R. J.; Friend, C. M.; Kaxiras, E. A Comparative Ab Initio Study of Anhydrous Dehydrogenation of Linear-Chain Alcohols on Cu(110). J. Phys. Chem. C 2018, 122 (14), 7806–7815.
[2] Vandermause, J.; Torrisi, S. B.; Batzner, S.; Xie, Y.; Sun, L.; Kolpak, A. M.; Kozinsky, B. On-the-Fly Active Learning of Interpretable Bayesian Force Fields for Atomistic Rare Events. npj Computational Materials 2020, 6 (1), 1–11.

Molecular dynamics simulations use statistical mechanics at the atomistic scale to enable both the elucidation of fundamental mechanisms and the engineering of matter for desired tasks. The behavior of molecular systems at the microscale is typically simulated with differential equations parameterized by a Hamiltonian, or energy function. The Hamiltonian describes the state of the system and its interactions with the environment. In order to derive predictive microscopic models, one wishes to infer a molecular Hamiltonian that agrees with observed macroscopic quantities. From the perspective of engineering, one wishes to control the Hamiltonian to achieve desired simulation outcomes and structures, as in self-assembly and optical control, to then realize systems with the desired Hamiltonian in the lab. In both cases, the goal is to modify the Hamiltonian such that emergent properties of the simulated system match a given target. We demonstrate how this can be achieved using differentiable simulations where bulk target observables and simulation outcomes can be analytically differentiated with respect to Hamiltonians, opening up new routes for parameterizing Hamiltonians to infer macroscopic models and develop control protocols.

Materials chemists develop chemical compounds to meet often conflicting demands of industrial applications. This process may not be properly modeled by black-box optimization because the target property is not well defined in some cases. Herein, we propose a new algorithm for automated materials discovery called BoundLess Objective-free eXploration (BLOX) that uses a novel criterion based on kernel-based Stein discrepancy in the property space. Unlike other objective-free exploration methods, a boundary for the materials properties is not needed; hence, BLOX is suitable for open-ended scientific endeavors. We demonstrate the effectiveness of BLOX by finding light-absorbing molecules from a drug database. Our goal is to minimize the number of density functional theory calculations required to discover out-of-trend compounds in the intensity–wavelength property space. Using absorption spectroscopy, we experimentally verified that eight compounds identified as outstanding exhibit the expected optical properties. Our results show that BLOX is useful for chemical repurposing, and we expect this search method to have numerous applications in various scientific disciplines.

Pseudo-magnetic fields (PMFs) in graphene are an intriguing strain-emergent phenomenon that lead to exciting physics such as Landau level quantization and zero-field quantum Hall effect. PMFs have been demonstrated experimentally first on small nanostructures such as nanobubbles [1]. Experimentally, generating uniform PMF of sufficiently large magnitude is of interest. However, as the strength of the generated PMF is inversely related to the size of the strained region, only small dimensions of graphene have been explored for this purpose. Of more technological interest are larger structures with engineered PMFs [2].

In this work, we are interested in the strain-induced uniform PMF and PMF-zero edges in various geometries in the scale of micrometers. From finite element simulations of strained graphene shapes with varying PMF distributions, we utilize a variational autoencoder that learns the pseudo-magnetic field emergent from geometric images. Our aim is to discover hitherto novel geometries by learning over the geometry-PMF possibility space (a generative approach). We further analyze geometries that will have optimal, robust PMF-zero edges under uniaxial strain via advanced graphing theory algorithms. Finally, we train an auxiliary neural network to rapidly classify large batch, simulated images of strained graphene to arrive at optimally strained graphene geometries that would produce well-defined pseudo-Landau levels for use in advanced electronic devices, such as transistors.

[1] N. Levy et al., “Strain-induced pseudo-magnetic fields greater than 300 Tesla in Graphene nanobubbles,” Science, vol. 329, no. July, pp. 544–547, 2010, doi: 10.1126/science.1191700.
[2] S. Zhu, J. A. Stroscio, and T. Li, “Programmable Extreme Pseudomagnetic Fields in Graphene by a Uniaxial Stretch,” Phys. Rev. Lett., vol. 115, no. 24, p. 245501, Dec. 2015, doi: 10.1103/PhysRevLett.115.245501.

Neural network potentials (NNPs) are gaining much attention as they enable fast atomic simulations for a wide range of systems while maintaining the accuracy of density functional theory calculations. Since NNP is constructed by machine learning on training data, its prediction uncertainty increases drastically as atomic environments deviate from training points. Therefore, it is essential to monitor the uncertainty level during the simulations to judge the soundness of the results. In this presentation, we present an uncertainty estimator based on the replica ensemble in which NNPs are trained over atomic energies of a reference NNP that drives the target simulations. The replica ensemble is trained efficiently and its standard deviation provides atomic-resolution uncertainties. We apply this method to a silicidation process of Ni deposited on Si(001) and confirm that the replica ensemble can spatially and temporally trace simulation errors at the atomic resolution, which in turn guides on augmenting the training set. The refined NNP completes a 3.6-ns molecular dynamics simulation without any noticeable defects. The efficient and atomic-resolution uncertainty indicator suggested in this presentation will contribute to achieving reliable NNP simulations.

An essential component of materials research is the use of simulations based on density functional theory (DFT), which imposes severe limitations on the size of the system under study. A promising development in recent years is the use of machine learning (ML) methodologies to train surrogate models with DFT data to predict quantum-accurate results for larger systems. Many successful ML models have been created to predict higher-level DFT results such as the total potential energy and atomic forces, and initial steps have been taken to create deep-learning based ML methodologies that can predict fundamental DFT outputs such as the charge density, wave functions and corresponding energy levels [1]. Here, we explore the applicability of this latter methodology using convolutional and recurrent neural networks to learn and predict the electronic charge density and the density of states of carbon, for a large variety of allotropes spanning from metallic to insulating behavior. Further improvements to the speed, accuracy and versatility of this DFT-emulation methodology will also be presented.

[1] A. Chandrasekaran, D. Kamal, R. Batra, C. Kim, L. Chen, and R. Ramprasad, Npj Comput. Mater. 5, 22 (2019)

For understanding and design of functional materials, density of states (DOS) is an important property. DOS hosts an essential information on various materials properties such as stability, thermal property, magnetism and catalytic activities [1]. The DOS projected on orbitals, so-called projected DOS (proDOS), provides site-specific properties and is closely related to the local physical property and chemical activity. The DOS is also used for understanding experimental data such as core-loss spectroscopy and X-ray photoelectron spectroscopy. Generally, the DOS and proDOS is obtained by a first-principles calculation, which requires a high computational cost. If a lower-cost computational method can alternate the first-principles calculation, search for new functional materials and extraction of information from huge experimental data can be accelerated.

In recent years, Materials Informatics, a combination of materials science and informatics methods, has been attracting great attention as a way for realizing the search for substances that exhibit excellent physical properties and the extraction of information from experimental data. In principle, the DOS is a quantity that is uniquely determined by an atomic arrangement, although its distribution drastically changes in various ways by a splitting and a shift of states due to a formation of bonded/anti-bonded states according to the atomic arrangement. Regarding the application of informatics on DOS, predictions on DOS of bulk crystals and surface using principal component analysis have been reported [2], which focuses on the composition dependence of alloys with a certain specific structure and the plane orientation dependence of surface DOS. from the viewpoint of atomic coordinates, however, the machine learning prediction on DOS has not been investigated well so far. In this research, we have tried to predict the DOS of metal nanoclusters from their atomic arrangements using the machine learning.

The nanocluster structures composed of several metal atoms were randomly sampled with a constraint on an interatomic distance. The first-principles calculation was performed using VASP code [3] for the generated atomic arrangements without a structural relaxation by PAW method with a planewave cutoff energy of 500 eV and a generalized gradient approximation with considering a spin polarization. Each ProDOS data for the analysis was extracted as two 64-dimensional vectors for each spin up and down, 128-dimensional vectors in total. In order to predict ProDOS, we first trained a variational auto-encoder (VAE) to extract features of ProDOS as latent vectors.
The trained VAE reproduces ProDOS to the extent that the physical property values, such as a d-band center, are roughly reproduced. Interestingly, we found that latent variables are one-dimensionally distributed as a function of interatomic distance in the latent variable space. We then trained another neural network for regression of the encoded latent vectors by atomic configurations. The predicted ProDOS obtained by decoding predicted latent vector from atomic configurations with the trained VAE shows good agreement of the calculated ProDOS.

In summary, we constructed the combination of the VAE and the neural network in order to predict the ProDOS of the metal nanoclusters. In the presentation, we will show examples of the ProDOS prediction performance, the model details, the distribution of the VAE latent variables, the generalization performance, and the examples of the diatomic molecules composed of different atoms.

Reference
[1] B. Hammer and J.K. Nørskov, Adv. Catal 45 (2000) 71,
[2] B. C. Yeo, D. Kim , C. Kim & S. S. Han, Sci. Rep. 9 (2019) 5879,
[3] G. Kresse and J. Furthmüller, Comput. Mater. Sci. 6 (1996) 15.

Many current autonomous materials platforms employ Bayesian optimization or active learning techniques, in which synthesis and processing parameters are sequentially considered in a search for the growth conditions that optimize some material response. While these techniques optimize within a closed-loop framework, each experiment done considers the input parameter settings as fixed and treats the synthesis and characterization of the material as a black-box, with input parameters ultimately yielding characterized output properties. We can, however, take advantage of the ability to perform in-situ characterization to provide yet another level of control in the search for optimized parameters by allowing strategically selected changes to real-time-controllable parameters such as temperature or gas flow rates within an experiment. In this talk, we describe the modeling framework and decision-making policies needed to take advantage of real-time control and characterization currently available in many autonomous materials platforms. We provide examples of how such a system can be used to dynamically control and optimize synthesis pathways in chemical kinetics, and show how to overcome hurdles in characterization resolution associated to in-situ measurements.

By combining automated experimentation and active learning, autonomous experimentation systems have accelerated the pace of research in fields such as materials science, chemistry, biology, and mechanics. While advancements in automated experimentation such as robotic platforms can increase the throughput of experiments to further accelerate the pace of research, an alternative path to accelerating learning is to utilize multiple information sources, with varying predictive power and cost, to evaluate the property of interest in the active learning loop. For example, a property of interest can be evaluated via a computational model or experimental measurement. This is particularly salient in the field of mechanics where both computational tools such as finite element analysis (FEA) and experimental characterization such as mechanical testing offer measurements of mechanical properties with varying fidelity. While imperfect and limited to mechanical properties that can be reasonably predicted, FEA encompasses useful information that can be utilized to explore vast design spaces for optimal designs. Conversely, for properties that cannot be reliably simulated, such as toughness, autonomous experimentation systems have emerged to efficiently search the design space for high-performing structures.¹ An open question that remains, however, is how to accelerate learning by utilizing multiple information sources available to an autonomous experimentation system.

In this work, we present a multi-information source optimization (MISO) framework for autonomous experimentation systems by incorporating simulated mechanical responses from FEA in the learning campaign. In this way, the learning campaigns become physics-informed instead of proceeding as a strictly black-box approach. We evaluate this MISO approach by simulating physics-informed learning campaigns and compare the resulting performance to that of black-box learning campaigns. From this study, we observed that including physics-based simulations, in our case FEA, leads to a decrease in both the number of experiments and time necessary to find optimal mechanical designs. Additionally, we evaluate the utility of these approaches in autonomous experimental campaigns. The development of this MISO framework by combining machine learning, physics-based simulations, and physical experimentation presents an approach for incorporating and leveraging numerous information sources available to accelerate research.

1. A. E. Gongora, B. Xu, W. Perry, C. Okoye, P. Riley, K. G. Reyes, E. F. Morgan, K. A. Brown “A Bayesian Experimental Autonomous Researcher for Mechanical Design” Sci. Adv. 2020

Developing novel glasses with new, improved properties and functionalities is key to address some of the Grand Challenges facing our society. Although machine learning offers a unique opportunity to accelerate the discovery of novel glasses with exotic functionalities, it faces the “curse of dimensionality.” This is especially significant for glass discovery, since glasses can virtually accommodate the entire periodic table, which, in turn, results in highly dimensional models. Such high dimensionality must be matched with high numbers of glass property data, which is presently lacking. Here, to address this issue, we present a new machine learning pipeline that simultaneously leverages experimental and simulation-based (synthetic) data by means of multi-fidelity Gaussian process regression (GPR) models. We show that our hybrid model systematically outperforms models relying solely relying on experimental data.