Symposium DS01—Accelerating Experimental Materials Research with Machine Learning

A high level introduction to the basic concepts of Machine Learning and central challenges in applying these concepts to materials data.

A high level introduction to the basics of Gaussian Processes (GP), their applications to regression and classification tasks, and their relationship to deep learning and linear regression. We will also have a brief hands-on demonstration of GPs applied to materials data. For attendees interested in following along, please bring a laptop and have a Google account.

A brief introduction to artificial neural networks (Neural Network Definition and Elements, Custom layers, Activation Functions, Loss functions) as well as to Deep Neural Networks (CNN, RNN) will be given. This is complemented by examples that focus on materials science applications, such as semantic segmentation of images.

In this module, we shall discuss sequential experimental design, and the Bayesian modeling of experimental responses. Key to this is the idea of a decision-making policy. We shall give many examples of such policies for various design tasks. Topics discussed will include Gaussian Processes, Multi-armed Bandits and Bayesian Optimization.

A hands-on demonstration of active learning in action. This includes a competition where attendees will compete in using Bayesian Optimization to identify the global extrema of a hidden function. Attendees who would like to participate in the competition should bring a laptop and have a Google account. Prizes will be awarded.

Access to machine learning methods has grown significantly over the past decade. Robust, open-source tools are readily available allowing materials scientists to explore the benefits of a modern materials informatics approach. Here, we present a selection of recent case studies covering a range of problems intended to test the application of machine learning to accelerate materials research at GE. New alloy and coating development projects, component life prediction, and part inspection are all discussed. Successes are highlighted where machine learning has directly led to new insights, but we also address challenges in applying these methods to complex industrially relevant problems. We cover the perennial lack-of-data issue and ongoing efforts at GE to tackle this problem through a deliberate data strategy, integrating both internal and external sources of information to improve model performance. The role of high-throughput experiments is also discussed as a critical enabler for machine learning. While many successes have been published in the academic literature, we cover some of the unique challenges in applying a modern materials informatics framework in industry and ongoing work at GE to modernize the way we conduct materials research and development.

Despite that graphene synthesis via chemical vapor deposition (CVD) was first demonstrated in 2009, today economically-viable manufacturing of continuous single crystals of graphene is still elusive. Hundreds of coupled synthesis parameters and complex kinetic pathways impede the development of recipes that, for instance, are reproducible, environmentally sustainable and compatible with electronics fabrication. Building an accurate model for synthesis is challenging due to the coupled chemistry-multi-physics, multi-time and -length scale nature of the growth process. To overcome these challenges, we have adopted a data driven approach to address the complexities of CVD synthesis of graphene. First, we will present a data project called Gr-ResQ (Graphene Recipes for Synthesis of High-Quality Materials). Gr-ResQ is a platform enabling the sharing and use of synthesis data towards building a more accurate predictive model for graphene synthesis. At the core of Gr-ResQ is a crowd-sourced database of CVD recipes and characterization. A suite of associated tools enable fast, automated, and standardized processing of Raman spectra and scanning electron microscopy images. To facilitate community-based efforts, Gr-ResQ provides tools for cyber-physical collaborations among research groups, allowing experiments to be designed, executed, and analyzed by different teams. Second, we present a data set of synthesis experiments where the selected parameters have a reasonably trackable influence on the outcome. Finally, we apply a Bayesian optimization technique to exploit the results of these experiments. This data-driven approach could facilitate more efficient discovery of complex synthesis recipes, where experimentation is time or capital intensive. This technique can guide the selection of the next experiment and iterate through the process till a predictive model is prepared.

The ability to deliver innovative products to the marketplace is at the core of many chemicals and materials companies. One of the main obstacles in doing so are high R&D cost and long development timelines. As a promising solution, data-driven R&D has made its way on executive agendas. At the same time, the underlying IT infrastructure required to provision R&D data to these data-modeling initiatives is often overlooked and/or under resourced. Consequently, many initiatives that aim at implementing data-driven R&D are facing data accessibility challenges, including insufficient data quality, lack of data contextualization or incomplete data sets because of poor system adoption.

The solution to this problem is a uniform platform approach that can directly ingest data from their point of creation and serve up these data to lab managers, scientists, or directly to data modeling tools. This approach is a departure from traditional R&D IT infrastructure designs, as it avoids compartmentalizing data in silos e.g., in LIMS, ELN, or SDMS databases. This is especially important for materials innovation, as workflows are highly variable and generated by scientists from multiple disciplines, leading to data aggregation and curation issues across data silos.

Our unified, data-driven platform that allows for flexible configuration of workflows and roles that naturally contextualizes R&D data. This is paired with a set of scientific data intelligence capabilities, that spans data querying across heterogeneous data sources, data analytics, and advanced scientific data visualization applications. The platform can operate in conjunction with other IT systems and data sources or as a standalone system that includes lab digitalization. This presentation reviews the general aspects of this approach and discusses various customer scenarios of digitalization efforts in the chemicals and materials space.

In previous work¹, it was established that the use of Artificial Intelligence (AI) and Computer Vision (CV) can augment the current approaches used for thin film analysis. By applying cutting edge Machine Learning techniques to Atomic Force Microscopy (AFM) images, fast, accurate and reproducible analysis can be done at much larger scales and levels of detail than previously feasible. The analysis pipeline uses the AFM height data to calculate metrics that are currently collected manually – such as linear density – as well as providing entirely new information not available from any machine or manual technique, such as the positions, orientations, and lengths of each individual carbon nanotubes in a device.
However, the main drawback of my previous work¹ is that an experienced Machine Learning practitioner is required to properly operate the software and algorithms. In practice, there are many parameters and configurations possible for each algorithm, and processing the data requires a very specific software environment as well as access to the source code. In this work, the previous approaches are implemented in a graphical interface, allowing chemists and materials scientists to leverage the cutting edges techniques without any technical support from a Machine Learning researcher. Various parameters can be tuned, configurations explored, and the results can be computed and visualized in near real-time.

In the analysis software, there are several steps to the data analysis procedure. First, the AFM height data is ingested in the form of a color or black-and-white image. Next, there are several pre-processing steps available; the image can be cropped, converted to black-and-white and even remove outliers (pixels that are particularly bright or dark). Once the image has been preprocessed, a scientist can apply various segmentation algorithms to the AFM data, this segmentation will produce a bitmap which represents the presence or absence of a nanoscale object at a given position in the microscope image. The segmentation processes consist of several steps which are designed to identify if a given point in a film belongs to a nanoscale object or not. In the previous work¹ this is accomplished by using adaptive thresholding techniques such as Otsu’s method or Yen thresholding. However, improved algorithms have been added to augment the capabilities of m Thresholding techniques and more sophisticated algorithms such as Deep Neural Networks. The final output of this segmentation process is a matrix of true and false values, indicating if a given pixel belongs to a nanotube (true) or if belongs to the substrate that the nanotube was deposited on (false). Once the segmentation is complete, a variety of solvers can be applied to extract quantitative measures of the nanoscale objects (or film morphologies in the case of thin film devices). We even propose the use of an algorithm in which a scientist can select automatic segmentation or manually specify points of interest, which are then used to segment and extract the polygonal boundary of crystal grains.
Using our graphical interface, materials scientist will be able to quickly and easily apply a wide range of cutting-edge Machine Learning techniques to extract detailed quantitative descriptors of their thin film devices or nanoscale objects without the need for any special hardware or Machine Learning expertise.

1. Mirka, B. et al. ACS Nano 15, 8252–8266 (2021)

The use of artificial intelligence (AI) or machine learning (ML) in the physical sciences has exploded over the past 5 years - 10 years. In that time frame several truly remarkable discoveries have been made including the discovery of new phase change materials, amorphous alloys, and catalysts. The continued success of these methods relies upon the availability of open data, meta-data and scientific code conforming to the findable, accessible, interoperable and reusable (F.A.I.R.) guidelines. Here I will briefly discuss successes and failures at NIST in collaboration with NREL, SLAC and the University of Maryland to create the first multi-institution combinatorial dataset and code repository to comply with F.A.I.R. guidelines. But even scientifically sound AI models built from open data sets can only be as trusted as the labels and values contained within them. The second part of my talk will focus on the tenuousness of ground truth and the need for an honest and open accounting for experimental uncertainties within our data sets. An example to be discussed will be of phase attribution from x-ray diffraction data. This will drive home the difficulties in forming unanimous expert consensus and how a consensus with variance impacts the perceived performance of ML model attributions.

We introduce an approach for the prediction of capacity for over 100,000 spinel compounds relevant for battery materials, from which we propose the few most promising candidate materials. In the design of batteries, selecting the proper material is difficult because there are so many metrics to consider, including capacity which is a fundamental engineering property. Using reported experimental data as our starting point, we demonstrate how we can build a dataset that provides a guide for the selection of battery materials. Although we focus on capacity of Li based spinel structures for electrode materials relevant for usage in batteries, the methodology developed and demonstrated here can be adapted to other properties, structures, and site occupancies. This work integrates non-linear manifold learning, prediction approaches accounting for co-linearity in the data, and uncertainty assessments. This addresses important issues in design of materials, where often the underlying physics is complex and difficult to visualize, while also accounting for challenges with limited data. Our methodology provides a reasonably accurate computational representation of experimental data while accounting for model uncertainty.

The future of research and manufacturing is autonomous and data-driven. Well-curated data is the fuel for this change, but research and high-mix low-volume manufacturing tend to produce highly dynamic and complex data. To date, most solutions have been limited to single pieces of equipment in highly controlled environments, thus limiting the applicability of solutions to the originating environments. Despite this limit, standardized portions of the data can be added to communal material data repositories like NOMAD or Materials Cloud. The tools to interact with the data in these cloud repositories are growing quickly, but there is very limited support for the process data, non-standardized data, and integration of the repository into closed-loop control of research or manufacturing automation. The authors propose that the ROSEDA (Repository of Operations & Subject Expertise for Decision Autonomy) framework may close these gaps, enabling autonomous learning, not just as a lab, but as a national manufacturing/research system. The ROSEDA framework includes an agile cyber-physical framework involving automation, dynamic database engineering, cloud/edge computing, human-machine interfacing, and artificial intelligence. As a first demonstration, the authors are attempting to give a ROSEDA system an "intuition" about new inks similar to the intuition possessed by expert print technicians. This presentation will cover the structure of the ROSEDA framework and current progress toward this demonstration.

In the past decade, the Materials Genome Initiative (MGI) effort has significantly facilitated and accelerated materials discovery and design by deploying large-scale ab-initio computation and building computational databases of structure-property relationships, thereby significantly accelerating the design of novel compounds. In contrast, important experimental properties aspects of materials research are only available in manually curated databases, or are not collected at all as large data sets. In this presentation, we focus on inorganic synthesis recipes, as a comprehensive large-scale collection of synthesis formulations would assist in the development of more quantitative and AI-driven approaches to synthesis routes for novel compounds.

Scientific publications represent the largest repository of knowledge about material synthesis and could be used as a reliable source of data. However, human-written descriptions of syntheses require additional levels of interpretation in order to be converted into a codified, machine-operable format. In this work, we built an extraction pipeline consisting of algorithms and models which use various advanced machine learning and natural language processing techniques to extract essential solution-based inorganic materials synthesis information from the scientific literature. By applying the extraction pipeline, we extracted 35,675 codified solution-based inorganic materials synthesis “recipes” from over 4 million papers. Each recipe includes precursors, targets, the quantities of materials used, and synthesis actions with the corresponding attributes. Information about targets and precursors was used to build the reaction formula for every synthesis procedure. This work has created the first large-scale, freely available dataset of solution-based inorganic materials synthesis recipes. The dataset not only paves the way for future data-driven approaches to investigate inorganic materials synthesis and synthesizability, but also allows researchers to design optimized synthesis procedures for automated experimentation.

For crude oil separation, polymer membranes consume as little as 10 percent of the total energy as required by thermal distillation processes, hence showing potential to surpass traditional distillation processes in terms of carbon emissions and energy efficiency. To build a polymer membrane, it is of utmost importance to investigate thermodynamic properties like solubility, miscibility, diffusivity, absorption, etc. The Flory-Huggins Interaction Parameter (χ) is often used to describe the polymer-solvent behavior in terms of these thermodynamic quantities.

This work focuses on building a temperature-dependent machine learning model for the prediction of χ for a given polymer-solvent pair, by making use of existing experimental data from numerous literature sources.

The experimental data was collected from over 130 research articles, to build a database of temperature-dependent χ, for over 50 polymers and 130 solvents. This dataset was also augmented with a dataset of Hildebrand solubility parameters for polymers and solvents. Both datasets were used to build a Gaussian Process Regression (GPR) based model in the “multi-task” learning mode. To further boost the performance of the model, feature engineering using Recursive Feature Elimination (RFE) and LASSO reduction was employed. Multi-task learning, which involves dataset fusion, has recently proved powerful in treating situations with data sparsity. We show that the resultant model is superior in performance to the conventional approach of estimating χ using empirical relationships.

This presentation will cover learnings around the manufacturing scalability of experimental research through automatic mining of published text for key phrases and facet similarity across multiple disciplines. In this way we inform at least three principal consequences of materials selection: (1) the availability, scaling capacity, and price volatility of the chosen materials’ constituents, (2) the manufacturing processes needed to integrate the chosen materials towards manufacturable devices, and (3) the performance that may be practically achieved with the chosen materials and processes. While each of these factors is, in isolation, a pivotal determinant of manufacturing scalability, we show that consideration and optimization of their collective effects and trade-offs is necessary to more completely chart a scalable pathway towards experimental research that moves more rapidly towards manufacturability. Successful scale up and integration of these novel materials is conditioned on a multitude of factors that are difficult to simultaneously optimize, particularly given the wide range of chemical and structural archetypes currently under investigation. Here we discuss key materials and processing considerations that could impact the scale up to production.

Computational methods can guide materials discovery by screening relevant compositions and structures prior to experimentation. Despite the acceleration provided by these approaches, synthesis efforts still rely on trial-and-error to realize computationally designed materials. Here, we combine high-throughput simulations, automated literature extraction, human-computer interaction, and experimentation to control materials synthesis from first principles. We focus on zeolite structures, whose challenging synthesis hinders the discovery of new catalysts. We start developing algorithms that accelerate zeolite simulations by two orders of magnitude when compared to traditional methods. Then, we compute more than half a million zeolite-molecule pairs, thus simulating relevant synthesis conditions for these materials. The calculations explain synthesis outcomes from more than one thousand articles extracted from the literature. In addition, computational predictions enable the design of organic structure-directing agents (OSDAs) with tailored selectivity towards the materials of interest. To visualize the generated data, we developed a web-based platform that provides a visual interface for data exploration by human experts. This human-computer partnership enables researchers to compare templates, obtain information on prior art and literature for zeolite-OSDA pairs, and downselect molecules based on their selectivity, synthetic accessibility, or physical descriptors prior to attempting the synthesis. Experimental results show that the proposed synthesis routes are effective in crystallizing the target materials and additionally enhance the catalytic properties of the obtained frameworks. This work offers a design platform for zeolite synthesis and can be extended to other classes of materials.

Zeolites are porous, crystalline aluminosilicate materials that have a wide range of industrial uses. These materials exhibit a variety of pore sizes and shapes, which enable them to selectively and efficiently catalyze specific chemical reactions. To template the formation of target frameworks, organic structure-directing agents (OSDAs) are incorporated into the zeolite synthesis route. However, selecting appropriate OSDAs to synthesize a desired zeolite often requires exhaustive experimentation and trial-and-error. While atomistic simulations can suggest favorable OSDAs for zeolites, their computational cost hinders a vast exploration of the combinatorial host-guest space. Here, we developed a machine learning approach to address this combinatorial design challenge. We started by creating an interpretable representation to capture OSDA shapes and sizes. Then, we trained machine learning models on more than 500,000 simulated zeolite-OSDA pairs to predict binding energies at a fraction of the computational cost from simulations. To thoroughly explore the chemical space of OSDAs, we created a library of candidates using a computational synthesis approach. The molecules were combinatorially generated from known precursors taken from vendors and ranked by their final cost. Finally, we evaluated the potential of millions of molecules as selective OSDAs for hundreds of zeolites using the ML-predicted energies. Our library covers a wide portion of the OSDA molecular-shape space and can drastically accelerate zeolite discovery.

Concrete—which is by far the most manufactured material in the world—is responsible for 5-to-10% of human CO₂ emissions. Here, we present an uncertainty-aware, machine-learning-enabled optimization scheme that aims to accelerate the discovery of new optimized concrete mixes featuring minimum embodied CO₂ while meeting target performance and manufacturing constraints. We first develop an unprecedented dataset comprising more than 10,000 concretes with varying mixture proportions, together with their measured properties (compressive strength, slump, etc.). The dataset is used to train a series of Gaussian Process regression (GPR) models that accurately map concretes’ mixture proportions to their properties, and uncertainty thereof. We then adopt Bayesian optimization to guide the discovery of new concrete mixtures featuring minimum embodied CO₂ while presenting required performance metrics (e.g., with a strength meeting or exceeding a given target) and obeying manufacturing constraints (e.g., with compliant slump). This pipeline leads to the discovery of several new concrete mixtures presenting a >50% decrease in embodied CO₂.

Given the natural graph structure in crystalline materials, numerous graph neural network (GNN) architectures have been developed to learn robust and generalizable representations that enable fast property prediction and accelerate the design of crystalline materials. However, with a vast array of material properties of interest, many of which differ from one another in fundamental ways, a one-size-fits-all approach to designing machine learning (ML) models is ill-suited to property prediction tasks. In this work, we develop the Integrated Graph Neural Network (I-GNN) framework to systematically explore the architecture design space for GNNs over a wide range of material property prediction tasks, ranging from mechanical properties, thermal properties to electronic properties. We construct an expansive design space of GNNs to enable methodical exploration of design choices. We show how the design space can be searched to find a task-specific optimal architecture by performing Bayesian optimizations over the model architecture itself. We demonstrate that the I-GNN framework is able to find architectures that outperform existing GNN architectures, such as CGCNN, MEGNet and GATGNN, on a wide range of properties, and the optimal architecture is indeed property-dependent. To provide insights into individual GNN designs for predicting specific material properties, we analyze the correlation between property similarities and the associated optimal architecture similarities, and study the transferability of the best GNN design across different properties. In summary, the I-GNN framework offers a methodical and scalable approach to explore the GNN design space, identifies the relationship between the best GNN designs and the target material properties, and provides insights into selecting the appropriate GNN to train given the target prediction task.

Perovskite solar cells have attracted significant attention in the PV community, because >25% power conversion efficiency was recently demonstrated in lab-scale small-area devices. However, manufacturing scaleup of perovskite solar cells is one of the critical areas requiring further innovations [1]. Photonic curing process is an industrially compatible technique that could facilitate crystallization of large-area perovskite thin films in milliseconds without heating the substrate [2]. This key feature enables perovskite device manufacturing on a flexible substrate that cannot sustain high temperature at high web speed. Due to the vast number of possible experimental combinations, process optimization typically requires a systematic effort of experimental planning to find an optimal window of process conditions.

In this work, we demonstrate rapid optimization of photonic curing with four process variables (namely, MAPbI₃ precursor concentration, additive concentration, exposure time and radiant energy during photonic curing) and aim at achieving the power conversion efficiency improvement from the current baseline of 10% on 100 µm thick Willow glass substrates [3]. Machine-learning-based Bayesian optimization framework is deployed to search the optimal process conditions in a high-dimensional experimental space. Using this framework, we can quickly establish the numerical regression model between process variables and device efficiency, and suggest new experimental planning based on evaluating an acquisition function that balances between exploitation and exploration. Especially, to increase human-machine interactions and reduce the possibility of potential mistakes in the model, we provide human-understandable 2D visualization (e.g., contour plot, tSNE plot) in every iteration and produce real-time feature importance ranking by the Shapley Additive Explanation (SHAP) method. The Bayesian optimization framework is shown to be efficient in process optimization and could accelerate the transition from the conventional thermal annealing process to photonic curing. Furthermore, we explore the potential of transferring the knowledge of fabricating the perovskite absorber layer to the hole and electron transport layers. We envision machine learning could be a useful tool to accelerate the development of manufacturing scale-up in perovskite PV devices.

Reference:
[1] Z. Li, T. R. Klein, D. H. Kim, M. Yang, J. J. Berry, M. F. A. M. Van Hest, and K. Zhu, “Scalable fabrication of perovskite solar cells,” Nature Reviews Materials, vol. 3, pp. 1–20, 2018.
[2] Weijie Xu, Trey B. Daunis, Robert T. Piper, Julia W.P. Hsu. Effects of Photonic Curing Processing Conditions on MAPbI3 Film Properties and Solar Cell Performance. ACS Applied Energy Materials 2020, 3 (9) , 8636-8645.
[3] R. T. Piper, T. B. Daunis, W. Xu, K. A. Schroder, and J. W. P. Hsu, “Photonic Curing of Nickel Oxide Transport Layer and Perovskite Active Layer for Flexible Perovskite Solar Cells: A Path towards High-throughput Manufacturing,” Frontiers in Energy Research 9, 640960 (2021)

Polymer film capacitors are favored by engineers because of their graceful failure mechanisms and low cost. However, certain applications, such as electrified aircraft, require capacitors that can withstand temperatures over 250 C, which is at least 100 C higher than the operating temperature of state-of-the-art polymer film capacitors. We combined machine learning with a genetic algorithm to predict hypothetical polymers with bandgaps above 5 eV, glass transition temperatures above 500 K, and dielectric constants above 4 at 100 Hz. These are useful properties, as a large bandgap can be used as a proxy for dielectric breakdown field strength, a high glass transition temperature indicates the polymer is thermally stable at high temperatures, and a high dielectric constant improves energy density per volume.

Basic genetic algorithms, however, suffer from low chemical diversity. To solve this challenge, we implemented a migration-assisted genetic algorithm where polymers evolved on separate islands, with each island enforcing a different design criterion, and then migrated. This increased the chemical space explored and improved the speed at which polymers achieving target properties were discovered. This is a robust design that could be used for other design applications as well. After predicting polymers that achieved target properties, a retrosynthetic algorithm was used to screen them for ease of synthesizability and likely reactants. A list of synthesis-friendly polymers achieving target properties and their likely reactants is shown.

Polymer electrolytes (PEs) have the potential to improve safety and enable higher energy densities in energy storage devices. However, PEs suffer from significantly lower ionic conductivity than liquid electrolytes¹, preventing their adoption in functional devices. Additionally, because PE synthesis and characterization are time consuming, relatively few PE formulations have been experimentally tested despite decades of study. In this work we leverage advances in machine learning (ML) and high throughput experimentation² to accelerate the design and discovery of PEs. The ML model we’ve developed incorporates Arrhenius behavior into a state-of-the-art message passing neural network³ to predict ionic conductivity of PEs. To train the model, we’ve compiled the largest known PE ionic conductivity dataset with nearly 15,000 values, which were gathered from experimental PE characterization publications and hand checked for quality. Ionic conductivity values were predicted for several thousand previously untested PE formulations, from which promising candidates—subject to synthesizability and processability requirements—will be selected for high throughput testing. The resulting data from high throughput experiments will be used to retrain the ML model, improving model predictions in an active learning fashion. The ML model shows marked improvement with increasing training data, indicating that data scarcity is a key limitation in machine learning PE ionic conductivity that our high throughput experimental setup addresses. This work offers an integrated computational-experimental platform to accelerate materials discovery for energy storage applications.


References
1. Xue, Z., He, D. & Xie, X. Poly(ethylene oxide)-based electrolytes for lithium-ion batteries. Journal of Materials Chemistry A vol. 3 19218–19253 (2015).
2. Ren, F. et al. Accelerated discovery of metallic glasses through iteration of machine learning and high-throughput experiments. Sci. Adv. 4, eaaq1566 (2018).
3. Yang, K. et al. Analyzing Learned Molecular Representations for Property Prediction. J. Chem. Inf. Model. 59, 3370–3388 (2019).

Polymers, due to advantages such as low-cost processing, chemical stability, low density and tunable design, have emerged as a powerhouse class of materials for a wide range of applications, including dielectrics. In certain applications, the performance of dielectrics is limited by low electric breakdown strength. Using this real-world application as a technology driver, we describe a novel artificial intelligence (AI) based approach for the design of polymers. We call this polyG2G. The key concept underlying polyG2G is graph-to-graph translation. In graph-to-graph translation, the distribution of subtle chemical differences between high- and low-performing polymers is learned and then sampled to generate novel, high-performing, hypothetical materials. Our approach, with respect to a host of presently adopted design methods, exhibits a favorable trade-off between generation of chemically valid materials and available chemical search space. polyG2G finds thousands of potentially high-value targets (in terms of glass-transition temperature, band gap, and electron injection barrier) from an otherwise intractable search space. Density functional theory simulations of band gap and electron injection barrier confirm that a large fraction of the polymers designed by polyG2G are indeed high-value. Finally, we find that polyG2G is able to learn established structure-property relationships.

Polymer blends is one of the most rapidly growing fields in the plastics industry. These multicomponent systems can satisfy complex material requirements by blending known polymers. There is a huge demand for property prediction models to find superior materials in the vast chemical space of polymer blends. In this study, we train a multi-task deep neural network to predict the glass transition, melting, and degradation temperature of homopolymers, copolymers, and polymer blends in a single model. For this, we propose a universal fingerprinting scheme applicable to all three polymer types that shows excellent performance. Our models were built on a joint dataset of more than 20,000 data points containing experimental values of the thermal properties of homopolymers, copolymers, and polymer blends. Our developed models are accurate, fast, flexible, and scalable, and aim to supersede previously developed models for only homopolymers and copolymers.

High performance polymers which can exhibit controlled degradation to monomer feedstock at end-of-use are essential to solve the global polymeric waste problem. This work focuses on polymers which undergo ring-opening-polymerization (ROP), as these polymers have shown promise within the context of high performance, depolymerizable polymers. We attempt to accelerate the development of such ROP depolymerizable polymers using machine learning to predict properties indicative of whether or not a polymer can undergo stable depolymerization. These properties, which include ceiling temperature (T_c), enthalpy of polymerization (ΔH_p), and entropy of polymerization (ΔS_p), are reported in literature for only a small subset of ROP polymers. In order to overcome this problem of data scarcity, classical molecular dynamics and density functional theory based simulations were utilized to compute ΔH_p for several hundreds of ROP polymers spanning a much larger chemical diversity than studied thus far. The computational dataset (treated as a “low-fidelity” version of the true measured ΔH_p) was combined with the much smaller dataset of experimentally measured ΔH_p (treated as the “high-fidelity” data) to develop predictive models of ΔH_p at the high-fidelity level. A variety of multi-fidelity learning approaches, including co-kriging and multi-task learning, were adopted. Finally, the computed ΔH_p, measured ΔH_p, and T_c datasets were utilized together to develop predictive models of T_c, the temperature above which depolymerization will be thermodynamically favored. A critical analysis of the performance of the models, challenges that remain in terms of data sparsity and diversity, the true benefits and limitations of the information fusion approaches, and guidelines on chemical factors that control ΔH_p and T_c will also be presented.

Despite the intriguing properties and widespread applications of colloidal nanomaterials in energy and chemical technologies, their discovery, synthesis, and manufacturing are still based on Edisonian trial-and-error-based techniques. Existing nanomaterials development strategies using resource-intensive batch reactors with irreproducible and uncontrollable heat/mass transport rates very often fail to overcome the demands of the vast synthesis and processing universe of colloidal nanomaterials, resulting in a slow and expensive discovery and development timeframe (8-10 years). Recent advances in lab automation and machine learning (ML)-guided modeling/decision-making strategies provide an exciting opportunity to reshape the discovery and manufacturing of colloidal nanomaterials.¹ In this talk, I will present an intelligent closed-loop nanomaterial synthesizer for on-demand development, formulation optimization, and manufacturing of colloidal nanomaterials through the convergence of robo-fluidic experimentation and in-situ nanomaterial characterization with machine learning.² I will discuss how modularization of different stages of colloidal nanomaterials synthesis and processing in tandem with a modular neural network ensemble modeling and decision-making under uncertainty can enable resource-efficient navigation through an experimentally accessible high dimensional space. An application of the developed modular robotic experimentation workflow for the autonomous synthesis of colloidal metal halide perovskite nanocrystals will be presented.

References
1. (a) Volk, A. A.; Abolhasani, M., Trends in Chemistry 2021, Advance Article, DOI: 10.1016/j.trechm.2021.04.001; (b) Volk, A. A., et al., Advanced Materials 2020, 33 (4), 2004495.
2. (a) Abdel-Latif, K., et al., Advanced Intelligent Systems 2021, 3 (2), 2000245; (b) Epps, R. W., et al., Chemical Science 2021, 12 (17), 6025-6036.

Semiconducting colloidal quantum dots hold promise for and are starting to be applied in a number of applications ranging from optoelectronics and photovoltaics to bioimaging. Their properties can be tuned by controlling their composition, structure, size, and size distribution. Given the vastness of synthesis parameter space, identifying optimal compositions and structures as well associated optimal synthesis recipes for them, however, remains a major hurdle for accelerated screening and discovery. The implementation of modular flow reactors, integrated with real-time reaction monitoring, automation and data science, for resource-efficient synthesis has the potential to accelerate this material discovery and synthesis screening process.
This presentation will highlight the development and application of a modular continuous flow reactor platform for synthesis of heavy-metal-free colloidal quantum dots with improved size-uniformity and enhanced optical properties. For example, our efforts, in collaboration with experts from industry and academia, have led to multiple promising insights into the synthesis of InP-based quantum dots. Second, the presentation will feature the development of a flow reactor platform that integrates inline spectroscopy with artificial intelligence-based feedback algorithms for efficient exploration of the multidimensional chemical synthesis parameter space. This platform autonomously learns the parameter space from a small number of experiments (<100), develops a predictive model, and synthesizes InP quantum dots of targeted bandgap and polydispersity (with no-prior-knowledge of the reaction space) all through self-driven experiments in less than 30 hours. Our results underscore the promise and critical role of data-science-assisted experimentation for accelerating the screening and discovery of semiconducting colloidal nanocrystals with desired properties, but also for maximizing synthesis insights across multidimensional parameter space for different classes of colloidal materials.

Perovskite solar cell technology has achieved rapid, unprecedented development in the research lab in terms of its record power conversion efficiency (PCE); however, successful commercialization requires the development of low-cost, scalable, and high-throughput manufacturing techniques. One of the key challenges to develop a new fabrication technique is the high-dimensional process optimization. Sequential learning, an iterative framework of machine learning for experimental planning, has been adopted for high-dimensional optimization tasks in many fields of physical science. This framework has been adopted to achieve research autonomy for optimization tasks, which is often combined with robotic research equipment, and in some research studies, sequential learning method can overperform the experimental plan made by human experts.

The current reports of sequential learning studies for process optimization have some common drawbacks: e.g., disregarding knowledge from related background studies, or neglecting qualitative insights from human experts. For example, during manual optimization of perovskite thin films, a visual assessment of film thickness, color, and large asperities has been used as an effective screening method. Based on this idea, we introduce an ML manufacturing process optimization framework that combines the output of a regression-based model (learning by doing) with probabilistic knowledge constraints (physical intuition and previously collected data) in sequential learning.

We optimized six parameters in the rapid spray plasma processing (RSPP) technique — substrate temperature, linear process speed, spray flow rate, plasma gas flow rate, plasma height, and plasma duty cycle. Process optimization enables 19% efficiency, which is on par with the highest performance reported to date for open-air spray coating processing technique. The resulting acceleration factor to achieve this highest efficiency is > 3X compared to conventional pseudo-random or factorial design of experiments sampling methods. Additionally, the model has an enhancement factor of greater than 30X in finding the top-one-percentile process conditions, identifying 10 new conditions that produce efficiencies comparable to previous top performing devices before implementing the machine learning approach.

In closing, we observe a disconnection between the machine learning community and photovoltaics research community, where machine learning methods tend to take human out of the loop. Herein, we developed this sequential learning framework to provide an alternative demonstration, where researchers’ inputs and background data can be easily incorporated with decision making during the iterative experimental planning.

The path from computational prediction to realization of new materials is clouded by processing parameters and kinetics. We present a new approach to unify computational and experimental data sets, wherein the computed material properties and processing parameters form a joint property+process search space for experimental efforts. Active learning over the manifold defined by computed properties utilizes their topology rather than their absolute values. This approach leverages similarities in computed and measured property manifolds while sidestepping problems arising from absolute errors in computed properties.

We demonstrate high-throughput optimization of composition and processing conditions for hybrid halide perovskites thin-films using a robotic platform guided by active learning on a joint property-process manifold. The property manifold is constructed using a limited set of density functional theory (DFT) calculations, augmented by trained materials graph networks (MEGNet [1]). The graph network densely interpolates the DFT-computed properties over the compositional search space, the manifold of which forms the “property space”. Process variables (annealing temperature, solvent, antisolvent, etc) form the “process space”. The joint property+process space constitutes our overall search space.
We search this property+process space using a high-throughput, automated thin-film deposition platform, Automated Materials by Solution Synthesis (AMaSS). AMaSS deposits solution-based thin films and characterizes them by darkfield, brightfield, and photoluminescence imaging and transmission and photoluminescence spectroscopy. These characterizations are processed to extract features describing film morphology, homogeneity, and optoelectronic behavior. AMaSS experiments are guided by Bayesian Optimization, in which the inputs are points in property+process space and the outputs are features extracted from the in-line characterization.

We will present several examples of validation of the approach, using optimizations within well understood search spaces, followed by results from discovery in new materials spaces. By simultaneously optimizing over property predictions and process variables, we bridge computational and experimental results. Our approach is broadly applicable to materials systems for which properties can be computed and offers a framework for translation of computed property manifolds into experimentally realized materials.

[1] Chen, Chi, Weike Ye, Yunxing Zuo, Chen Zheng, and Shyue Ping Ong. “Graph Networks as a Universal Machine Learning Framework for Molecules and Crystals.” Chemistry of Materials 31, no. 9 (2019): 3564–72.

With the widespread adoption of 3D printing processes, it is now a simple matter to build architected solids with complex geometries, such as cellular structures with non-uniform struts and non-periodic arrangements, to achieve desirable mechanical properties. However, finding the structures that optimize a set of mechanical properties, particularly nonlinear and/or failure properties, is challenging due to the large number of parameters that arise. To design more complex architected lattices, machine learning algorithms such as neural networks can help in the design process by generating all possible geometric patterns and finding candidates that match the desired properties. However, this approach is limiting, as it is only feasible for low dimensions and can only be experimentally validated for a few design candidates that have been identified. Here, we discuss the use of a framework based on Bayesian optimization which accelerates the convergence process. Bayesian optimization is a closed-loop design process based on Gaussian process regression which guides the choice of design variables for each iteration of experiments. We show the efficacy of the framework by first examining 3D-printed single edge notched bend bars with embedded cellular inclusions that act as defects. These defect regions can deflect cracks and lead to higher energy absorption during bending-induced fracture. Our Bayesian optimization framework explores the entire parameter space to find the geometry that leads to the globally optimal energy absorption, thereby identifying the parameters to use for the next experimental specimen. In a second application, we apply this framework to maximize energy absorption of cellular solids during in-plane compression. Groups of struts in the cellular solid are allowed to have non-uniform thickness, resulting in sixteen parameters. To accelerate the convergence process we use multiple data streams, including “high-fidelity” (but costly) experimental sources and “low-fidelity” (but inexpensive) finite element simulations using ABAQUS. We show that by including low-fidelity sources, the framework can converge more quickly than by using high-fidelity sources alone, thereby reducing the time and cost needed to discover an optimal design.

Investigating mechanical properties at small scales is a challenging endeavor. It requires sophisticated micro-/nano-scale experimental methods combined with laborious/time-intensive finite element computations. In this talk, a new framework for materials characterization at small scales using the latest developments in machine learning will be presented. This framework involves multi-fidelity deep learning and active learning methods limiting the need for finite element simulations. Its application for predicting the fracture toughness of microscale pentagonal cross-sectional ceramic cantilevers will be showcased, demonstrating that it can significantly accelerate fracture toughness characterizations at small scales.

Proteins can form biomaterials with exceptional mechanical properties equal or even superior to synthetic materials. Currently, using experimental atomic force microscopy or computational molecular dynamics to evaluate protein mechanical strength remains costly and time-consuming, limiting large-scale de novo protein investigations. Therefore, there exists a growing demand for fast and accurate prediction of protein mechanical strength. To address this challenge, here we present PRESTO, a rapid end-to-end deep learning model to predict protein resistance to pulling directly from sequences. By integrating a natural language processing model with simulation-based protein stretching data, we first demonstrate that PRESTO can accurately predict the maximal pulling force, F_max, for given protein sequences with unprecedented efficiency, bypassing the costly steps of conventional methods. Enabled by this rapid prediction power, we further find that PRESTO can successfully identify specific mutation locations that may greatly change protein strength in a biologically plausible manner, such as at the center of polyalanine regions. Finally, we apply PRESTO to design de novo protein sequences by randomly mixing two known sequences at varying ratios. Interestingly, PRESTO predicts that the strength of these mixed proteins follows up- or down-opening “banana curves”, breaking away from the general law of mixtures in conventional material systems. By discovering key insights and optimal sequences, we demonstrate the versatility of PRESTO as a tool in a rapid protein design pipeline. Our model may offer new pathways for protein material science that focus on analysis and testing of large-scale novel protein sets, as a discovery tool that can be complemented with other modeling methods, and ultimately, experimental synthesis and testing.

Direct ink writing (DIW) is capable of patterning viscoelastic structural, functional, and biological inks into planar and 3D architectures. During extrusion-based printing, these inks are fluidized under an applied stress enabling their flow through fine nozzles. However, upon exiting the nozzle, they rapidly recover a solid-like state allowing the extruded filaments to retain their shape. For identical printing conditions (i.e., nozzle size, printing pressure, and speed), modest changes in ink rheology can significantly impact the printed feature size and fidelity. However, the optimization of key printing parameters for a given ink is time consuming. Here, we describe a data-driven approach to estimate the critical rheological parameters from images of printed standardized patterns. Our model is trained on data generated with automated experimentation using multiple samples of several different inks that exhibit a wide range of flow behavior. Using oscillatory shear rheometry, we collect sets of data that serve as “truth” for a machine learning model. Our work represents a significant step towards the establishment of autonomous, material-aware 3D printing platforms.

AFM images were collected on a variety of polymer blend samples that differ in bulk mechanical properties and microstructure. A deep learning model based on a convolutional neural net (CNN) successfully classified various combinations of images pointing to real and meaningful differences in their microstructure. A separate regression-based CNN was built to correlate the AFM images with various bulk mechanical properties. The results observed from the deep learning model reveal a relationship between the microstructures as captured by the AFM images with the different bulk material properties.

A machine learning model is established to predict the influence of Hot isostatic pressing (HIP) parameters on tensile properties of Ti-6Al-4V parts prepared by selective laser melting (SLM). The database was established by collecting published reports on HIP treatment of SLM Ti6Al4V from 2010 to 2019. Using the established model, it is possible to prescribe HIP parameters and predict properties after HIP for SLM Ti-6Al-4V parts with high confidence. It is found that the YS and UTS are sensitive to the HIP parameters including temperature and holding time. The initial YS and UTS of as-printed parts are weaker influencing factors for YS and UTS of final parts. The model suggests that a HIP process with holding temperature lower than 970°C and holding time around 2 hours is desirable for SLMed Ti-6Al-4V. However, the prediction of strain to failure shows lower accuracy because of the variation of defects among different sample.

Zeolites are crystalline, nanoporous materials extensively studied for a variety of industrial applications including catalysis, absorption, and separations. Zeolites are typically synthesized hydrothermally using an organic structure-directing agent in combination with other inorganic precursors. While a number of studies have used simulation and data-driven methods to examine the effect of the organic component of zeolite studies, very little research has studied the inorganic aspects including the gel composition, reaction conditions, and precursor selection with similar techniques. Through data extraction and generative modeling, we examine and develop tools to improve these inorganic aspects of zeolite synthesis. We use natural language processing and human-in-the-loop data cleaning to optimize data extraction from the literature resulting in an exhaustive data set of approximately 20,000 zeolite synthesis routes. We then construct a conditional variational autoencoder trained to generate the inorganic synthesis components including gel composition, reaction conditions, and precursors given an OSDA and zeolite structure. This model is capable of generating synthesis conditions for newly predicted OSDA-zeolite pairs and can be paired with other models and simulations designed to optimize the organic aspect of zeolite synthesis. Additionally, we estimate basic kinetic information including the nucleation and growth rates by fitting crystallization curves found in the data set. We estimate synthesis systems missing intermediate crystallization points by fitting the parameters of existing curves as a function of the synthesis variables. We then examine the effect of various synthesis variables on these kinetic variables.

Conjugated polymers have garnered great interest as potential candidates for a wide variety of applications, ranging from electronics to biomedical usages. Doping conjugated polymers is an effective strategy for manipulating their electrical conductivity by generating charge carriers via addition of p-type or n-type dopants. However, identifying a suitable polymer:dopant combination generally relies on heuristic experimental approaches, and is exceptionally time- and labor-consuming due to the vastness of the chemical, configurational and morphological space. Here, we present high-performance surrogate models that are developed to predict the electrical conductivity of p-type and n-type doped polymers. Based on support vector machine (SVM) and Gaussian process regression (GPR) algorithms, the models are trained with available experimentally measured data, and are subsequently used to screen more than 800,000 polymer and dopant candidates to identify promising polymer:dopant combinations. Additionally, we extract new design guidelines that extend knowledge on important molecular fragments that correlate to high conductivity. Our work, thus, provides a roadmap towards robust and rapid material discoveries for doped conjugated polymers. Models are deployed at www.polymergenome.org for broader open-access community use.

Machine learning (ML) approaches have the potential to accelerate and transform materials discovery and property prediction. One challenge is the development of accurate material descriptors, i.e., accurate ways to describe the characteristics of a real material to a ML model. Several techniques to generate descriptors have already been developed, such as RooSt and Magpie, which convert a material's composition string to a long-form descriptor: a weighted graph or a matrix of attributes. In contrast, Crystal Graph Convolutional Neural Networks (CGCNN) ML framework develops a descriptor using a material's full crystal structure – its crystallographic information file (CIF) – creating a multigraph. Until now, no rigorous comparison of descriptors has been performed. We assess these descriptors and their performance in ML models for predicting band gap, formation energy, and crystal system, for a database of ~800,000 material composition strings and CIFs. Insights into relative performance of descriptors can help assess what level of materials knowledge is required for accurate prediction of specific materials properties.

Synthesis of materials with desired structures, e.g. metal-organic frameworks (MOFs), involves optimization of highly complex chemical and reaction spaces due to multiple choices of chemical elements and reaction parameters/routes. Traditionally, realizing such an aim requires rapid screening of these nonlinear spaces by experimental conduction under human intuition, which is quite inefficient and may cause errors or bias. In this work, we report a platform that integrates a synthesis robot with machine learning algorithms to accelerate the synthesis of MOFs. The robotic platform consists of an automatic direct laser writing apparatus, precursors injecting and Joule-heating components. It can automate the synthesis upon fed reaction parameters which are recommended by the genetic algorithm or the Bayesian optimization algorithm. Without any prior knowledge, this integrated platform continuously improves the crystallinity of ZIF-67, a demo MOF employed in this study, as the operation iterations of the two algorithms increase. This work represents a methodology enabled by a data-driven synthesis robot which achieves the goal of material synthesis with improved crystal structures, thus greatly shortening the reaction time and reducing energy consuming. It can be easily generalized to other material systems and paves a new route to the autonomous discovery of a variety of materials in a cost-effective way in future.

Bayesian optimization is a well-developed machine learning field for black box function optimization. In Bayesian optimization a surrogate predictive model, here a Gaussian process, is used to approximate the black box function. The estimated mean and uncertainty of the surrogate model are paired with an acquisition function to decide where to sample next. Bayesian optimization with Gaussian processes can be used for structure-property relationship optimization in materials systems with second order phase transitions or gradual disordering without a distinct phase transition. In this study we applied this technique to known ferromagnetic thin film materials such as ferromagnetic (Fe_100−y Ga_y)_1−xB_x (x=0–21 & y=9–17) and (Fe_100−y Ga_y)_1−xC_x (x=1–26 and y=2–18) to demonstrate optimization of structure-property relationships, specifically the dopant concentration or stoichiometry effect on magnetostriction and ferromagnetic resonance linewidth. These systems both experience gradual disordering without distinct phase transition with increase in boron and carbon concentration, thus are great case systems for Bayesian optimization with Gaussian processes. We have shown that using simulated Bayesian optimization methods to guide experiments, the number of samples required to statistically determine the optimum was reduced by 50 % compared with traditional research methods. Our results suggest that Bayesian optimization can be used to save time and resources to optimize ferromagnetic films. This method is transferrable to other ferromagnetic material structure-property relationships, providing an accessible implementation of machine learning to magnetic materials development.

High entropy alloys (HEAs) are multicomponent materials with nearly equal amount of four or more principal elements. Compared to conventional metal alloys, HEAs exhibit excellent mechanical, thermal, and electrochemical properties with immense potential for materials design due to a vast compositional space. HEAs are typically solid solutions with face centered cubic (FCC), body centered cubic (BCC), or hexagonally closed packed (HCP) phases. There is a long-standing in the application of machine learning (ML) methods to accurately classify the HEA phases as a function of chemistry and composition.

In this work, we develop an ML model with post hoc interpretability capability for HEA phase prediction. We accomplished our objective by combining an ensemble support vector machines (eSVM) algorithm with a variable attribution method. A dataset of 1,821 compositions was compiled from surveying the published literature. Each composition was represented by a high-dimensional space of 125 features that include atomic size, electronegativity, valence electron number, enthalpy of formation, and mixing entropy to name a few. The purpose of eSVM was to build a data-driven multi-class classification learning model that establishes a quantitative relationship between chemical composition and experimentally determined phase information. We also compared the performance of the eSVM model with a random forest model. Both models had an accuracy of ~85% on the independent test set. We then performed post hoc interpretability of the trained eSVM models to quantify the relative contribution of each feature to the ML prediction score with respect to each instance in the training data. We developed a new clustering algorithm that identifies the similarity between training instances based on the variable attribution scores. Our analysis uncovers previously unknown phase-specific correlations between key features and the HEA phases.

Abnormal grain growth, where a few grains grow faster at the expense of their neighbors, is hypothesized to occur because of anisotropic grain boundary (GB) character and unique GB networks. In a polycrystalline material, there are thousands of potential GB types and unlimited combinations of those GBs that makeup the 3D topology of a GB network. This combinatorial, high-complex space is too large for conventional data analysis methods to identify the physical descriptors that are necessary to build anisotropic grain growth models. Reinforcement learning is a machine learning tool that can capture the underlying behavior of an evolving Markov decision process and “teach” it to maximize the “rewards” regarding the agreement between prediction and simulation. Here, we applied a deep reinforcement model to emulate grain growth by conducting learning on Monte Carlo Potts grain growth simulations. Relative to other machine learning networks, reinforcement learning is a relatively interpretable model since the outputs of the models are tied directly to action. The developed reinforcement model can also be used to test different microstructural architectures and, thus, identify how GBs evolve over time. The accuracy of our short and long-term predictions for grain growth will be evaluated. Additionally, the potential and challenges associated with this method for exploring the cause of abnormal grain growth will be discussed.

Data driven methodology has become a key tool in predicting material properties computationally at low cost. We present a detailed analysis on predicting stability of Si electrodes for multivalent ion batteries using Support Vector Regression (SVR). Emphasis is on transferability of machine learning (ML) SVR model trained on a set of Si alloys to new Si electrodes for multivalent ion batteries. SVR has the ability to deal with non-linear higher dimensional data in feature space by using kernel trick. We utilize previously reported Si based alloys derived from database for training and testing. Hyperparameters are tuned using two different approaches namely, Grid search CV and Bayesian optimization. The performance of the two approaches is compared in predicting energies and packing fraction of the structures. Furthermore, dependence of these two output properties on the crystal structures calls for an additional analysis on structural representation to the ML model. Three different featurizers are used to represent material structures as input features for ML. Our results show that featurizer Orbital Field Matrix performed best on validation dataset against featurizers X-Ray Diffraction and Sine Coulomb Matrix, while it predicted poor energy values of test data. In contrast, Sine Coulomb Matrix data achieved the lowest error between actual energies and predicted energies of test structures. Our work highlights the potential of ML methodologies in accurately foretelling the stability of reaction intermediates in alloying electrodes for ion batteries beyond Li and guides the experimental studies in selecting appropriate battery materials.

Modern scanning transmission electron microscopes (STEM) can focus sub-angstrom electron probes to extract quantitative atomic structure and chemistry from elastic and inelastic scattering processes. The chemical composition of specimens is revealed by spectroscopic techniques produced by inelastic interactions in the form of energy dispersive X-rays (EDX) or electron energy loss of the transmitted electrons (EELS). Each atom’s chemical identity and coordination provides essential information about the performance of nanomaterials across a wide range of applications from clean energy, batteries, and optoelectronics, among many others. Unfortunately, the dose requirements for high-resolution chemical-spectroscopy (e.g. 10⁶ e/Ų) often far exceeds the dose limits of a specimen—chemical maps are noisy or missing entirely. More reliable interpretation of material structure is made in combination with elastically scattered electrons that can measure structure, but not chemistry, at high signal-to-noise ratios (SNR). These inelastic and elastic signals (i.e. modalities) are typically analyzed separately. However, inspired by machine learning and compressed sensing methods, we show high-SNR recovery of chemical maps down to atomic resolution can be achieved by linking information between modalities.

Here we introduce multi-modal electron microscopy, a technique that offers high SNR recovery (500% increase in favorable cases) of nanoscale material chemistry by leveraging correlated information encoded within both high-angle annular dark-field (HAADF) and EDX / EELS. We recover chemical maps by reformulating the inverse problem as an optimization that seeks a solution which surpasses traditional dose limits. First, we model the physics of elastic scattering process to the inelastic spectroscopic chemistries. Second, we ensure the recovered chemical maps maintain a high degree of agreement with the initial measurements using a maximum log-likelihood. Lastly, following recent achievements in compressed sensing, we ensure the recovered chemical maps are maximally sparse in the gradient domain.

We demonstrate chemical recovery at the nanoscale in commercial CoS₂ catalysts for oxygen-reduction and atomic length scales for Co_1.51Mn_1.49O₃ and ZnS/Cu_1.81S interfaces. In all cases, noise in the raw chemical maps is dramatically reduced, even at lower doses, while maintaining structure at and around interfaces. Moreover, our multi-modal approach also recovers the specimen’s chemical relative concentration, allowing researchers to measure local stoichiometry. Uncertainty can be further quantified using confidence intervals from Fisher Information maps. This data-driven approach opens new pathways for high-throughput chemical characterization and the investigation of radiation sensitive specimens.

As a result, multi-modal electron microscopy reveals atomic scale chemistry with enhanced clarity at significantly lower electron doses than traditionally allowed.

High-throughput experimentation is widely used to rapidly assess phase behavior across composition and processing space. Fast and accurate phase identification (on the timescales of active learning loops) from synchrotron-based X-ray diffraction data is a critical step in many of these experiments. In contrast to conventional nonnegative matrix factorization(NMF) algorithms or modern neural network(NN) methods, we have developed a probabilistic algorithm to search and label the specific phases existing in processed libraries. The algorithm builds from a database of potential phases, exploiting the crystallography space group to model physically possible peak shifts (e.g., due to alloying or strain). This makes this method training free, lightweight and flexible. Finally, by modeling the peak shifting, peak ratio deviation, and other variations in the spectrum using Bayesian statistics, we are able to give a probabilistic prediction of the phases that are presented in each diffraction pattern, including the probability of an uncatalogued phase. The algorithm has been benchmarked using a range of binary and ternary oxide systems, with analysis speeds that are commensurate with the lateral-gradient Laser Spike Annealing high throughput technique.

A key challenge in materials development is the ability to identify material phase in-situ during synthesis and processing. In this talk we present a phase mapping and phase determination method based on Reflection High Energy Electron Diffraction (RHEED) data. RHEED is an in-situ surface-structure characterization technique commonly used to observe the growth of thin films that produces diffraction patterns containing a wealth of static and dynamic information. However, the ability to extract this information is limited by the lack of versatile systems for automated analysis of the patterns. We used a Deep Learning-based tool for automating the analysis of RHEED diffraction patterns. Our method combines several supervised and unsupervised learning methods and permits the extraction of important qualitative and quantitative information characterizing the structure. As a test of the developed pipeline, we applied it to map the phase diagram of hematite structures grown under different conditions. With the increased robustness of the information extracted from the RHEED measurements, one would be able to attain quantitative insight as to the underlying structural phase as well as how the surface structure of a thin film is evolving during the deposition.

The current materials research process is slow and expensive; taking decades from invention to commercialization. The Air Force Research Laboratory pioneered ARES™, the first autonomous research systems for materials development. A rapidly growing number of researchers are now exploiting advances in artificial intelligence (AI), autonomy & robotics, along with modeling and simulation to create research robots capable of doing iterative experimentation orders of magnitude faster than today. We will discuss concepts and advances in autonomous experimentation in general, and associated hardware, software and autonomous methods.

We expect autonomous research to revolutionize the research process, and propose a “Moore’s Law for the Speed of Research,” where the rate of advancement increases exponentially, and the cost of research drops exponentially. We also consider a renaissance in “Citizen Science” where access to online research robots makes science widely available. This presentation will highlight advances in autonomous research and consider the implications of AI-driven experimentation on the materials landscape.

Many experiments in materials science and chemistry are executed within multi-stage experimental and computational pipelines. Examples of such pipelines include the sequential screening and filtering of candidate materials or drugs through a sequence of computer models of increasing fidelity, or multi-step materials synthesis and characterization. In this talk, we present some models and algorithms for executing multi-stage experimental and computational pipelines under the context of optimal experimental design and decision-making under uncertainty. In particular, we present techniques that capture the multi-stage nature of the experimental procedure, including both the need to model experimental decisions for intermediate stages and the ability to characterize the outputs of such stages. We demonstrate how such models and decision-making algorithms can lead to improvements over other methods that treat the experimental structure as a black box, or techniques that perform sequential screening without measuring and mitigating uncertainty.

Beamtime at neutron diffraction instruments is a valuable resource uniquely capable of characterizing the magnetic structure of materials but with very limited availability. Here we develop a Scientific AI (SciAI) that is able to autonomously discover the magnetic transition temperature, and thus make efficient use of the beamtime. Several physical principles of the problem are encoded into this Bayesian algorithm, such as the Poissonian-like detector physics and the relationship between diffraction intensity and material magnetization. Furthermore, this system demonstrates hypothesis testing wherein the algorithm identifies which physical principles govern the magnetic transition. This SciAI has been implemented on the BT-4 beamline at NCNR at NIST as well as at WAND² at the HB-2C beamline at HIFR at ORNL, and has been used to autonomously discover the magnetic transition temperature of both MnO and Fe_1.09Te, materials with different magnetic behaviors. Compared to an a priori scheduled experiments, this system reduced the number of measurements needed for the experiment by a factor of approximately 5.

Electron Backscatter Diffraction (EBSD) mapping is a powerful technique that can be used to study the crystal grain orientations in a material, but it can be time consuming to exhaustively collect diffraction data at every location in the sample. Here we present a Scientific AI (SciAI) that autonomously discovers the grain map, collecting data at the most informative locations. This SciAI encodes physics to predict the grain map such as: the crystal symmetry of the material under study, and the rule that crystal grains should be contiguous. Crucial to the success of the active learning scheme employed in this SciAI is the quantification of uncertainty in the assigned grain membership and propagation of that uncertainty to the prediction of the grain map. This SciAI is able to discover a grain map for ZnO with only ~10% of the data used in a traditional measurement.

Autonomous experimentation enabled by artificial intelligence (AI) offers a new paradigm for accelerating scientific discovery. Non-equilibrium materials synthesis is emblematic of complex, resource-intensive experimentation whose acceleration would be a watershed for materials discovery and development. The mapping of non-equilibrium synthesis phase diagrams has recently been accelerated via high throughput experimentation but still limits materials research because the parameter space is too vast to be exhaustively explored.

In this talk we will present our efforts to accelerate synthesis and exploration of metastable materials through hierarchical autonomous experimentation governed by the Scientific Autonomous Reasoning Agent (SARA). SARA integrates robotic materials synthesis and characterization along with a hierarchy of AI methods that efficiently reveal the structure of processing phase diagrams. SARA designs lateral gradient laser spike annealing (lg-LSA) experiments for parallel materials synthesis and employs optical spectroscopy to rapidly identify phase transitions. Efficient exploration of the multi-dimensional parameter space is achieved with nested active learning (AL) cycles built upon advanced machine learning models that incorporate the underlying physics of the experiments as well as end-to-end uncertainty quantification. With this, and the coordination of AL at multiple scales, SARA embodies AI harnessing of complex scientific tasks. We demonstrate its performance by autonomously mapping synthesis phase boundaries for the Bi₂O₃ system, leading to orders-of-magnitude acceleration in establishment of a synthesis phase diagram that includes conditions for kinetically stabilizing δ-Bi₂O₃ at room temperature, a critical development for electrochemical technologies such as solid oxide fuel cells.

We are incorporating active learning in screening of combinatorial libraries of functional materials. The array format with which samples of different compositions are laid out on combinatorial libraries is particularly conducive to active learning driven autonomous experimentation. For some physical properties, each characterization/measurement requires time/resources long/large enough that true "high"-throughput measurement is not possible. Examples include detection of martensitic transformation and superconducting transitions in thin film libraries. By incorporating active learning into the protocol of combinatorial characterization, we can streamline the measurement and the analysis process substantially. We have previously demonstrated discovery of a new phase change memory (PCM) material using the closed-loop autonomous materials exploration and optimization (CAMEO) strategy. The discovered PCM material has been tested in various scaled-up device formats and continues to exhibit superior performance to industrial standards. Recent efforts in developing synthesis – measurement closed loops on a combinatorial thin film platform will be discussed. This work is performed in collaboration with A. Gilad Kusne, V. Stanev, H. Yu, H. Liang, M. Li, E. Pop, and A. Mehta. This work is funded by SRC, ONR, AFOSR, and NIST.

Autonomous physical sciences is a rapidly developing field, where artificial intelligence controls experiment design, execution, and analysis in a closed-loop. The field promises to revolutionize the physical sciences, with the next generation of scientists using autonomous systems to address research challenges that are out of reach of Edisonian trial and error methods. However, as of yet, there are no education platforms for teaching the necessary skills to the next generation workforce. In this talk, we present a low-cost education platform which was used during an undergraduate course on machine learning for materials science at the University of Maryland. Through python code, students are taught the fundamentals of machine learning to identify trends from chemistry experiments. Students can also deploy active learning to guide sequential experiments to investigate chemical reactions and perform hypothesis testing.

Technology motivates the perpetual need for novel advanced materials. As a result, materials scientists are driven to investigate materials of increasing complexity. However, with each new synthesis or processing parameter, the materials search space grows exponentially, making human-driven exhaustive experimentation infeasible. In this talk we will compare the use of physics-informed machine learning to off-the-shelf machine learning in their ability to accelerate materials exploration, optimization, and discovery. We will explore different levels of physics integration in the updated Closed-Loop Autonomous Materials Exploration and Optimization 2.0 (CAMEO) algorithm and its impact on the coupled challenges of phase mapping and materials optimization. In particular, we will look at the use of probabilistic priors based on exogenous data, the integration of phase mapping and functional property physics and heuristics, and exploiting mutual information shared between phase mapping and materials optimization for physics-informed active learning. Performance is evaluated on combinatorial library data including the Fe-Ga-Pd benchmark dataset.

Technologies such as batteries, biomaterials, and heterogeneous catalysts have functions that are defined by mixtures of molecular and mesoscale components. As yet, this multi-length scale complexity cannot be fully captured by atomistic simulations, and the design of such materials from first principles is still rare. Likewise, experimental complexity scales exponentially with the number of variables, restricting most searches to narrow areas of materials space. Robots can assist in experimental searches but their widespread adoption in materials research is challenging because of the diversity of operations, instruments and measurements that is required. Here we use a mobile robot to search for improved photocatalysts for hydrogen production from water (Nature, 2020, 583, 237). The robot operated autonomously over 8 days, performing 688 experiments within a 10-variable experimental space that has more than 46 million points, driven by a batched Bayesian search algorithm. This autonomous search identified photocatalyst mixtures that were six times more active than the initial formulations, selecting beneficial components and deselecting negative ones. Our strategy uses a dexterous free-roaming robot, automating the researcher rather than the laboratory instruments, which are largely unmodified. This modular approach may have applications across a broad range of research problems. The mobility and non-linearity of the robot scientist offers the potential to unlock the power of artificial intelligence and machine learning – for example, to make complex, conditional decisions about what measurements and experiments to do next.

Natural porous materials such as nanoporous clays are recognized as green and low-cost adsorbents and catalysts. The industrial development of novel products requires properties optimization tailored to specific application. Machine learning-aided material design cut time and costs accelerating the experimental research. In our approach we apply predictive algorithms to diverse experimental datasets, obtained by a clay-manufacturer company, for multiobjective optimization. The methods were exploited for the development of (1) acid nanocatalyzers, (2) mycotoxin detoxifiers (3) and rheological additives for drilling fluid application.
In the (1) project, we applied tree-based algorithms trained on experimental datasets to accelerate the characterization of the morphology and the surface activity of processed clay-based materials. We proposed a design function finding optimal processing conditions for clays exploitable in acid catalysis. One of such identified materials was synthesized and tested in catalytic degradation of chlorophyll-a, achieving the 79% of removal.
In the (2) project, we explored promising mycotoxin detoxifiers which form bulky non-absorbable complexes with toxic molecules, reducing their bioavailability. The machine learning models were trained using an extensive set of in vitro adsorption experiments imitating the gastrointestinal tract of the animals. The powerfulness of the models allowed to identify high performing formulations for the regulated toxins, predict the detoxification of a wide set of yet-unregulated mycotoxins, and gain insights into the in vitro detoxification mode of action through model feature importance analysis. Furthermore, the machine learning algorithms were used to design the in vivo experiment for validating the performance of the selected composite material, i.e., the most challenging detoxification experiment. The biomarker detection-based in vivo detoxification of deoxynivalenol in broiler chickens confirmed the outcomes of our approach.
The (3) project was focused on the optimization of the thixotropic properties (apparent viscosity and yield point) of clay-based additives for drilling fluid process. In this case we generated experimental data by advanced space-filling design of experiments (latin hypercube sampling with multidimensional uniformity). The latter provided a powerful solution to generate the maximum information by the minimum number of needed experiments. The homogenous dataset allowed to build high predictive models enabling exploration of novel optimal formulations by multiobjective optimization based on pareto-front. Four optimal materials were prepared and tested finding correspondence with the virtual experiments.

Autonomous physical science has the ability to revolutionize how materials science is performed: algorithmically searching materials systems for optimal materials. This reduces the time and cost of the experiments by optimally collecting knowledge-rich data. Such an approach is becoming ever more popular with developments in active learning and scientific AI. We present a benchmarking study to compare the efficacy of popular off-the-shelf active learning methods and those designed with incorporated knowledge of the materials optimization challenge. A primary focus is the data fusion of structural and functional properties information. In this study we benchmark the active learning methods on 5 ternary composition materials datasets containing composition, structure, and functional property data and evaluate the impact of prior knowledge on materials optimization.

Electron microscopy has been on the forefront of materials discovery in numerous fields ranging from ceramics to biology [1-3]. Advancements in instrument controls and detectors have expanded the available pathways for uncovering physical phenomenon. However, despite the technological progression, the fundamental principle of scanning microscopies has remained largely the same, namely the sequential acquisition of data and the data collection workflow. The majority of workflows remain human operated, vastly limiting the data acquisition and feature identification through the latency time of a human operator. Thus, the integration of artificial intelligence and automated manipulation of microscope parameters opens new avenues for experiments with higher control, lower beam damage, and an overall decreased time per experiment. Here, we develop various approaches for automated experiment workflows in scanning transmission electron microscopy (STEM) based on building feature of interest descriptors using the raw image, image segmentation, deep convolutional neural networks (DCNN), and rotationally-invariant variational autoencoders (rVAE). The descriptor is used as a basis for the acquisition function which guides the Bayesian optimization algorithm. Each workflow is applied to three datasets, with a range of feature sizes, comprised of NiO pillars within a La: SrMnO₃ matrix, ferroelectric domains in BiFeO₃, and topological defects in graphene. This approach is compared with the simpler method based on the linear transform-based AE workflows. This work was performed at the Center for Nanophase Materials Sciences, a US Department of Energy Office of Science User Facility.

1. S. J. Pennycook, et al., in Advances in Imaging and Electron Physics, Vol 153, (2008), p. 327.
2. S. J. Pennycook, et al., Journal of Electron Microscopy 58 (2009), p.87.
3. Y. Jiang, et al., Nature 559 (2018) p.343.

Scaling electrochemical energy storage to TWh/a requires the proliferation of new post-Li ion battery chemistries to overcome materials supply constraints. Identification of lead materials for electrodes, electrolytes and the associated optimal electrochemical protocols do, however, remain elusive for many emergent chemistries like Na-, Mg-, Ca- or Al-ion batteries. This talk will present first results obtained from the platform for accelerated electrochemical energy storage research located at the Helmholtz Institute in Ulm as part of the Karlsruhe Institute of Technology. The talk will showcase transfer learning from high-throughput millimeter scale electrochemistry towards high-throughput coin and pouch cell manufacturing and beyond. The talk will also highlight bottlenecks in deploying a highly integrated laboratory and demonstrate how strict implementation of data management protocols and analysis servers fast track machine learning guided experimentation across the entire battery research value chain.

Active or sequential learning holds great promise for accelerating materials experiments, although full realization of this promise relies on the ability of machine learning models to be predictive in new composition spaces where no training data is available. Expert scientists make such predictions based on their materials chemistry knowledge, i.e. how the individual elements and interactions among them give rise to properties in composition spaces that have not been measured. Additionally, experts transfer knowledge about materials chemistry from other domains. To emulate these aspects of expert extrapolative predictions, we introduce the Hierarchical Correlation Learning for Multi-property Prediction (H-CLMP) framework that seamlessly integrates (i) prediction using only a material’s composition, (ii) learning and exploitation of correlations among target properties in multi-target regression, and (iii) leveraging training data from tangential domains via generative transfer learning. H-CLMP accurately predicts non-linear composition-property relationships in composition spaces for which no training data is available, which broadens the purview of machine learning to the discovery of materials with exceptional properties. This achievement results from the principled integration of latent embedding learning, property correlation learning, generative transfer learning, and attention models. The best performance is obtained using H-CLMP with Transfer learning (H-CLMP(T)) wherein a generative adversarial network is trained on computational density of states data and deployed in the target domain to augment prediction of optical absorption from composition. H-CLMP(T) aggregates multiple knowledge sources with a framework that is well-suited for multi-target regression across the physical sciences.

Enzyme catalysts are an integral part of green chemistry strategies towards a more sustainable and resource-efficient chemical synthesis. However, the retrosynthesis of given targets with biocatalysed reactions remains a significant challenge: the substrate specificity, the potential to catalyse unreported substrates, and the specific stereo- and regioselectivity properties are domain-specific knowledge factors that hinders the adoption of biocatalysis in daily laboratory works. [1,2]
Here, we use the molecular transformer architecture [3] to capture the latent knowledge about enzymatic activity from a large data set of publicly available enzymatic data, extending forward reaction and retrosynthetic pathway prediction to the domain of biocatalysis. We introduce class tokens based on the EC classification scheme that allows to capture catalysis patterns among different enzymes belonging to same hierarchical families. The forward prediction model achieves a top-5 accuracy of 62.7%, while the single step retrosynthetic model shows a top-1 round-trip accuracy of 39.6%. The enzymatic data and the trained models are available through the RXN for Chemistry platform (https://rxn.res.ibm.com and https://github.com/rxn4chemistry).

[1] Katrin Hecht et al., Catalysts, 2020, 10(12), 1420
[2] Shuke Wu et al., Angewandte Chemie, 2021, 60(1), 88-119
[3] Philippe Schwaller et al., ACS Central Science, 2019, 5(9), 1572–1583

Uncertainty quantification in AI-based predictions of material properties is of immense importance for the success and reliability of AI applications in material science. While confidence intervals are commonly reported for machine learning (ML) models, prediction intervals, i.e. the evaluation of the uncertainty on each prediction, are seldomly available. In this work we compare 3 different approaches to obtain such individual uncertainty, testing them on 12 ML-physical properties. Specifically, we investigated using the Quantile loss function, machine learning the prediction intervals directly, and using Gaussian Processes. We identify each approach’s advantages and disadvantages and compare their results. All data for training and testing were taken from the publicly available JARVIS-DFT database, and the codes developed for computing the prediction intervals are available through JARVIS-tools.

Over the years a tremendous amount of data on materials have been gathered whether by academic experimentalists building up databases, or during manufacturing process investigations. The shear amount of data makes the use of it intractable by humans at a global scale and is hence limited to very specific uses by specialists. We present a data-driven approach for formulations of novel materials via autoencoders-based models. Using deep-learning techniques, we are trying to find important correlations and patterns in the underlaying data and by doing so, improve existing products and design new ones¹, using the groundwork laid by Kingma² and Bombarelli.³ Starting from data that can consists of various inputs like compositions, processes or even aging conditions, and as outputs, product properties. We encoded the input into three latent representations utilizing encoder-decoder neural-network structures. From these trained latent space vector, the property can be predicted by a separate feed-forward neural network. Alternatively, one can optimize any given property with a given target value, by searching the latent space using a Gaussian process, to retrieve the corresponding compositions and processes. The scheme has been successfully applied to various material case studies in the past: ranging from polymers and epoxy resin to different alloys and it is general enough to be applied to much broader types of industrial problems. As an example of the broad potential use of such an algorithm, we present an application to the phase prediction of ceramic materials. We used commercially available databases consisting of binary and ternary diagrams to train an AI system able to predict properties such as the phases resulting from the mixture of compounds at any given temperatures and pressures conditions.


References:
1) T. Gaudin, O. Schilter, F. Zipoli and T. Laino, https://ercim-news.ercim.eu/en122/special/advanced-data-driven-manufacturing.
2) D. P. Kingma and M. Welling, Auto-encoding Variational Bayes, ICRL 2014, https://arxiv.org/abs/1312.6114
3) R. Gómez-Bombarelli, J. N. Wei, D. Duvenaud, J. M. Hernández-Lobato, B. Sánchez-Lengeling, D. Sheberla, J. Aguilera-Iparraguirre, T. D. Hirzel, R. P. Adams and A. Aspuru-Guzik, ACS Central Science, 2018, 4, 268–276.

Material discovery is bottlenecked by the validation of the vast number of promising hypothetical materials generated from high-throughput calculations. The main reason is that the traditional experimental approach relies on trial-and-error. Therefore, time is wasted in attempting to synthesize non-synthesizable material or using the wrong conditions or methods. In recent years, natural language processing (NLP) techniques have been utilized to extract material synthesis conditions from the literature. By leveraging text mined dataset (TMD), researchers have developed data-driven approaches for experiment planning and synthesizability prediction, such as predicting the synthesis outcomes¹ and choosing the correct precursors for solid state reaction (SSR)², in hopes to reduce the number of failed synthesis attempts and accelerate material discovery. High quality data is required for any data-driven approaches to avoid error propagation. While it is generally agreed that the accuracies of TMD are lowered than their manually collected counterparts³, there is no quantitative comparison between the two in the material domain.

In this study, SSR information of 4103 ternary metal oxides was manually collected from the literature. The information includes whether the oxide has been synthesized via SSR in the past and the SSR conditions. The accuracy of this dataset was quantitatively compared with reported performances of recent TMDs in materials, which highlights the limitation of current NLP to extract relatively complicated synthesis information. An example is the mixing/grinding conditions, where the manually collected dataset has a precision of 0.96 as opposed to 0.62-0.82 for a TMD⁴. The accuracy of this manually extracted dataset can serve as a milestone for future NLP applications in materials. A filter based on the distribution of heating temperature and binary oxide melting point of the manually collected dataset was created to identify over 100 erroneous entries in a subset of the TMD of solid state reaction⁴. Aside from common errors observed in the literature, some issues with the NLP set up and their implications to the TMD usages were found and discussed.

As a demonstration, this manually collected dataset was used to train a positive-unlabelled learning model to predict the solid state synthesizability of hypothetical ternary oxides. This approach can aid researchers to decide the suitability of using SSR to synthesize new materials.

1. Malik, S. A., Goodall, R. E. A. & Lee, A. A. Predicting the Outcomes of Material Syntheses with Deep Learning. Chem. Mater. (2021). doi:10.1021/acs.chemmater.0c03885
2. Kim, E. et al. Inorganic Materials Synthesis Planning with Literature-Trained Neural Networks. J. Chem. Inf. Model. 60, 1194–1201 (2020).
3. Kononova, O. et al. Opportunities and challenges of text mining in aterials research. iScience 24, 1–20 (2021).
4. Kononova, O. et al. Text-mined dataset of inorganic materials synthesis recipes. Sci. data 6, 203 (2019).

The water absorption and proton conductivity are crucial parameters for designing polymers for fuel cell applications. Proton conductivity can directly affect the performance of a proton exchange membrane fuel cell and is often discussed with water absorption as water prompts dissociation of acid groups and forms transport pathways. Experimental determination of these properties for new polymers is an expensive and time-intensive process. Here, we develop a machine learning model to predict the temperature and humidity-dependent water uptake and proton conductivity of polymers instantly and accurately. Our model is trained using a dataset of over 4000 experimentally measured water uptake and proton conductivity values, an advanced polymer fingerprinting scheme, and the multi-task Gaussian process regression algorithm. With this model, we demonstrate that we can accurately predict not only the two properties of complex copolymer systems but also those over a wide range of temperature and humidity. Moreover, new design principles are extracted that identify critical polymer segments or functional groups that correlate to high proton conductivity.

Over the last decades, electron and scanning probe microscopies have evolved as one of the primary tools to study systems in the domain of physical and life sciences on the atomic and mesoscale levels. These measurements give rise to highly reliable structural and spectral data containing a wealth of information on structures and functionalities of the systems. On the other hand, development and availability of more computational capabilities including accessible CPU/GPUs, efficient algorithms and corresponding implementations have significantly boosted the advancement of physical simulations. Physical models constructed using first-principle calculations to quantum Monte Carlo and finite-element methods, spanning over quantum-mechanical to continuum scales leads to abundance of insights on thermodynamic, and electronic properties of materials. There has also been a surge in applications of deep learning and data-driven techniques in each of these avenues in recent years. However, studies utilizing a combination of all three of these avenues to establish a bridge between such learnings is still in its infancy.
In this work, we show how deep learning can bridge together the knowledge learned from microscopic images and first-principles simulations to develop a comprehensive understanding of the physics of the materials of interest. Here we focus on how deep convolutional neural networks can be employed to identify atomic features (type and position) in graphene, use them to construct supercells perform density functional theory simulations to find optimized geometry of the structures followed by studying temperature-dependent dynamics of system evolutions with ad-atoms and defects. The results along with associated uncertainties in predictions at various levels as obtained utilizing this framework may be used to evaluate and modify experimental conditions and regions of interest. We aim to further expand this approach to incorporate edge-computing involving direct transfer of image-based data from microscopes via Jetson AGX Xavier and then analyze, train deep neural networks using a GPU-based platform followed by performing simulations using CPU-based high-performance computing resources and feed back to the human in the loop, altogether on-the-fly, to better guide experiments while learning from theoretical models.

This effort (machine learning) is based upon work supported by the U.S. Department of Energy (DOE), Office of Science, Office of Basic Energy Sciences Data, Artificial Intelligence and Machine Learning at DOE Scientific User Facilities (A.G., S.V.K.) and was also supported (STEM experiment) by the DOE, Office of Science, Basic Energy Sciences (BES), Materials Sciences and Engineering Division (O.D.), and was performed and partially supported (M.Z.) at the Oak Ridge National Laboratory’s Center for Nanophase Materials Sciences (CNMS), a
DOE Office of Science User Facility.

The stoichiometry of inorganic glasses controls their material property. Therefore, understanding the compositional dependence of glass properties is critical in developing novel glasses. Herein, we use a glass database (>450,000 glass compositions) with up to 232 glass components to train XGBoost (Extreme Gradient Boosting) models for 25 glass properties (including optical, physical, electrical, and mechanical properties). Further, we determine the role of each input glass component in controlling the glass property us SHAP (Shapely additive explanations) analysis. The SHAP analysis reveals a strong interdependence among the glass components for properties like liquidus temperature and glass transition, whereas no such interdependence for properties like density. While some of this interdependence can be explained as “boron anomaly” and “mixed modifier effect”, the others need further exploration. Thus, our work is critical in understanding the component–structure–property relationship of inorganic glasses and discovering novel inorganic glasses.

All-inorganic lead halide perovskite (LHP) quantum dots (QDs) have recently emerged as a promising class of semiconducting nanomaterials for a wide range of solution-processed photonic devices, including color-tunable light-emitting diodes (LEDs), solar cells, and lasers. Despite the outstanding optoelectronic properties of LHP QDs, their widespread adoption in chemical and energy technologies has been limited due to the health and environmental concerns associated with the high amount of lead ions (Pb²⁺) present in their crystalline network. Partial cation exchange of Pb²⁺ ions with less toxic impurity dopants (i.e., manganese, Mn²⁺) has been promoted as an effective strategy to simultaneously mitigate the toxicity concerns associated with Pb and incorporate new optical/magnetic properties into the pristine LHP QDs.
The synthesis, fundamental studies, and development of metal cation-doped LHP QDs, similar to other colloidal QDs, are conventionally conducted using time-, material- and labor-intensive flask-based techniques. In contrast, microfluidic synthesis strategies with their tunable and reproducible heat/mass transport rates as well as their reduced chemical consumption and waste generation compared to batch reactors (e.g., round-bottom flasks) have been demonstrated as a reliable time- and material-efficient tool for accelerated fundamental and applied studies of colloidal QDs. Despite the high-throughput nature of microfluidic QD synthesis strategies, the user-guided experiment selection restricts a comprehensive mapping of the multivariable chemical universe and complicates the search for the optimal formulation of metal cation-doped LHP QDs. Recently, the use of deep neural networks (DNNs) and reinforcement learning algorithms have enabled accelerated chemical space exploration in organic syntheses and discovery and optimization of advanced LHP QDs. In this work, we present a self-driven modular robotic QD synthesizer for closed-loop metal cation doping of LHP QDs with desired optoelectronic properties. The robotic QD synthesizer is equipped with (i) a formulation, (ii) multiple flow synthesis reactors, and an (iii) in-situ spectral monitoring module. Through convergence of automated robotic QD synthesis with machine learning (ML)-guided modeling (DNN ensemble) and decision-making under uncertainty, we demonstrate accelerated formulation discovery of high-quality Mn-doped LHP QDs. We study the effect of black-box vs modular ML modeling and decision-making on the overall experimental cost for multiple formulation optimization campaigns of Mn-doped LHP QDs. Additionally, we demonstrate the role of in house-generated prior knowledge within the same modular flow reactor on the performance of the autonomous QD synthesizer. The developed ML-guided material synthesis strategy provides a smart workflow for data-driven formulation optimization and continuous manufacturing of metal cation-doped LHP QDs.

In this research, SCAPS (Solar Cell Capacitance Simulator) was utilized to build and probe non-toxic Cs-based perovskite solar devices, and investigate modulations of key materials parameters on ultimate power conversion efficiency (PCE). The input materials parameters of the absorber Cs-perovskite layer were incrementally changed, and with the various resulting combinations, 63,500 unique devices were formed and probed to produce device PCE. Versatile and well-established machine learning algorithms were thereafter utilized to train, test and evaluate the output dataset with a focused goal to delineate and rank the input materials parameters for their impact on ultimate device performance and PCE. The most impactful parameters were then tuned to showcase unique ranges that would ultimately lead to higher device PCE values. As a validation step, the predicted results were confirmed against SCAPS simulated results as well, highlighting high accuracy and low error metrics. Overall, the results from this investigation provide much-needed insight and guidance for researchers at large, and experimentalists in particular, towards fabricating commercially viable non-toxic inorganic perovskite alternatives for the burgeoning solar industry.

The optimization space for battery materials is vast, requiring better instrumentation and orchestration for efficient exploration. Herein, we present a hierarchical experimental laboratory automation and orchestration (HELAO) framework that is capable of operating multiple distributed and resource sharing instruments at once. HELAO is a versatile, highly modular, user-friendly, lightweight framework that incorporates active learning, AI-assisted data analysis, and high-level hardware abstraction to support globally distributed laboratories. We will demonstrate the open source HELAO framework both through a distributed multi-instrument active learning run and through a combinatorial investigation of multi-cation electrodeposition optimization for battery electrodes using a scanning droplet cell. We will demonstrate the development and discovery of a Lithium and a Mg-cathode post-Li ion battery material and optimal charging protocols. These demonstrations will illustrate the processes required in partially autonomous workflows as envisioned for the Platform for accelerated electrochemical energy storage research (PLACES/R) at the Karlsruhe Institute of Technology and Helmholtz Institute Ulm.

Scanning probe microscopy (SPM) has become a mainstay of nanoscience, enabling nanoscale measurements of a vast array of different materials. SPM hosts an immense number of modalities of both imaging as well as spectroscopy, to enable functional property measurements to then be correlated with nanoscale (or atomic scale) features of the sample. Despite over three decades since the initial SPM developments, the method of spectroscopy acquisition has remained largely the same, restricted to either individual point measurements, or those on a uniformly spaced grid. This is problematic both due to high levels of redundancy (neighboring pixels are likely to contain the same behavior) as well as inefficiency (time taken per spectra can be prohibitive), thus resulting in tradeoffs between spectral and spatial resolution.

Here, we explored methods to improve the efficiency of spectroscopy in SPM via utilization of Bayesian optimization schemes, implemented via incorporation of edge hardware in the form of GPU servers connected to SPM instrumentation. We first collect hysteresis loops on a ferroelectric thin film from random locations, and then train a surrogate model via Gaussian process regression to predict specific spectral features at the unmeasured sites with quantified uncertainty. We then use an acquisition functions in the Bayesian optimization algorithm to determine the optimal locations to sample to maximize the target spectral feature of interest. We find efficiency gains on the order of 2-3x over conventional grid-based methods. We further incorporate prior knowledge into the algorithm through use of high-resolution imaging, via deep kernel learning, to further improve the predictions of the surrogate model. The tradeoffs between the cost of computation of the more advanced approaches and the cost of actual measurements are discussed. In sum, we show that it is possible to substantially reduce the quantity of spectra captured whilst maintaining highly accurate spectral reconstructions with sparse data on samples with a significant degree of spatial correlation. Notes on how to extend the framework to other spectroscopies will be discussed. This work was conducted at and supported by the Center for Nanophase Materials Sciences, a US DOE Office of Science User Facility.

Evaluating the potential of novel functional materials and devices for industrial viability is a multi-dimensional large parameter space exploration. Manual experimentation is extremely limited in throughput and reproducibility. Automated platforms for fabricating and characterizing complete functional devices can accelerate experimentation speed within tight processing parameter variations. In this talk I will introduce the concept for the automated research platform AMANDA Line 1 and outline the R&D challenges such a research platform is able to address by today in the broad scope of semiconducting materials
As a first test case I will demonstrate a multi-target evaluation of photovoltaic materials at full device level. About 100 processing variations and nearly 1000 devices have been tested for one photovoltaic absorber on AMANDA Line 1 and evaluated by a Gaussian process regression-based data evaluation. The unique data quality in combination with GPR allowed to discover hidden structure – property relationships within 70 hours only which otherwise would have taken months to establish. Similar acceleration factors are found for further material systems. Engineering efficient and stable perovskite devices was recently demonstrated as well, and the automated material research quest is now directed towards Pb free perovskites. The outlook projects the transition from high throughput engineering towards autonomous optimization. The design, synthesis, testing and optimization of hundreds of interface materials out of a library with more than a million candidates is discussed as a real world scenario for ML guided and autonomous material invention.

Materials discovery from infinite earth repository is always considered as the bottleneck of each revolutionary technological progress. The technique is hindered by the labor-intensive and time-consuming process. Although the collection of the machine learning techniques has shown excellent capability for speeding up materials discovery, the material feature representation, especially crystal structural information, is still a challenge.
This work focuses on developing an active learning-enabled adaptive materials exploration system for accelerating piezoelectric materials discovery. The system works in a closed-loop mode, which starts from the material’s stoichiometry and crystal structure representation. As the piezoelectric properties highly depend on the materials crystal structure, this work graphically encodes crystal structure by newly adopted adjacency matrix. Further, the system learned the materials’ properties by the implemented machine learning algorithm and well-established piezoelectric properties database. In this way, the materials exploration system could provide general guidance for piezoelectric properties optimization. Evaluated by the realistic experimental data, the system outcome will yield feedback within each cycle. This adaptive materials exploration system has demonstrated the efficiency of optimizing piezoelectric properties. The presented work mode can also be utilized when speeding up the material exploration system for properties other than piezoelectricity, such as photovoltaic, thermoelectricity, etc. This knowledge will have a broad interest in materials science and the machine learning community.

Ni-based superalloys are the preferred choice of materials for high temperature applications like turbines due to their exceptional properties at such elevated temperatures. However, the current generation of Ni-based components are operating at close to their melting point (1100 °C). Therefore, there has been an increase in the demand for metallic alloys that display superior mechanical properties at temperatures as high as 1600 °C. High-Entropy Alloys (HEAs) are suitable candidates as they show promising properties at these operational temperatures. However, very few HEAs have been discovered so far that surpass the performance of Ni based superalloys. Designing new HEAs that can meet these requirements is therefore a challenging task that not only requires domain knowledge, but also depends on fortuitous discovery from expensive experiments and computation.
Due to the complexity of the problem, generative deep learning is a prime candidate to accelerate the discovery of HEAs. Generative Adversarial Networks (GANs) are a promising method for learning complex distributions and have been demonstrated to great success in the field of computer vision. When trained on HEA compositions, these models can generate candidate materials rapidly as well as interpolate continuously between desirable structures. Our approach involves training a continuous conditional GAN (ccGAN) model on the HEA compositions with their corresponding figures of merit (such as high-temperature mechanical properties) as the conditioning vector. This allows us to generate novel HEAs which approach the desired figure of merit by querying the network for new samples with a given conditioning vector. We train the model on composition-property pairs found in the literature and evaluate the results using cross-validation and DFT calculations.

Machine learning (ML) has evolved to become a powerful cognitive assistant for optimizing processes and materials, as evidenced by numerous academic examples and industrial applications. However, the ability of ML to integrate and generate generalizable scientific knowledge generally still requires a human. In this talk, I’ll review the state of the art of ML-based knowledge representation in scientific research broadly. Then, I’ll discuss nascent examples applying knowledge representations to materials for sustainability, including elucidating the root cause(s) of perovskite degradation. In conclusion, I’ll summarize by sharing a personal perspective of where the field can evolve over the coming years, what classes of discovery may be enabled by further evolution of knowledge representations, and the ethical imperatives associated with this evolution.

In past several years, the material industry has paid much effort to accelerate material design by introducing data driven methods. Usually design problems attacked by the material industry are more challenging than those reported in the literature in some aspects. For example, in functional materials design, there are often more than three target properties (usually 5 to 10) whose experimental errors are not negligible, and this situation is much more complex than most data-driven material design studies reported in the literature. Moreover, in most cases, the number of possible experiments for functional materials design is extremely limited, because the evaluation of some target properties, such as the durability, requires extremely long-time experiments, and the speed of materials design is always important in order to win the competition among material manufacturers. Furthermore, efficient screening methods, such as virtual and real-world high throughput screenings, are still not available for most functional material design problems, due to the complexity and diversity of their experimental systems. Multi-objective Bayesian optimization is promising for such design problems, which can be applied to any multi-objective design problems requiring time-consuming experiments in order to accelerate their design processes. However, in contrast to single objective Bayesian optimization, which is rapidly spreading in the materials industry, the industrial application of multi-objective Bayesian optimization remains in its nascent stages, and trials and errors by experienced data scientists are required to design materials that satisfy multiple customer requirements using multi-objective Bayesian optimization.

In this study, we developed a straightforward graphical user interface (GUI) based system for multi-objective Bayesian optimization. This system was designed for a common situation of industrial functional materials design where there are multiple target properties with quantitative goals that are defined to satisfy customer requirements and material design efforts will be stopped after finding materials achieving quantitative goals for all target properties. The developed system employs the joint probability of the achievement of all goals (PA) as an acquisition function¹⁾. This allows users to easily control Bayesian optimization with many objectives and to efficiently achieve the goals only by inputting the target value for each objective. This system also provides users an interpretable visualization of the progress of multi-objective Bayesian optimization based on the PA, which allows material scientists not familiar with data science to see whether optimization efforts should be continued, whether the current choice of the design space is promising, and which material is the best solution. Furthermore, using this system, material scientists can easily feed experimental data obtained in past material design projects to the machine learning model driving Bayesian optimization, which further accelerate multi-objective Bayesian optimization. The development of the straightforward GUI-based system, which does not require experienced data scientists, will popularize multi-objective Bayesian optimization in the material industry. The developed system will be used for the actual functional material design of Showa Denko Materials Co., Ltd. from the beginning of 2022.

Hanaoka, K.; Bayesian Optimization for Goal-Oriented Multi-Objective Inverse Material Design, iScience 2021, in press.

Many products called functional materials are made by mixing multiple raw materials in order to achieve various properties. The design of such materials requires both tuning of the composition ratio of materials and selection of optimal combination of raw materials from many candidate materials. However, it is difficult to optimize both raw material combination and their composition ratio simultaneously due to their vast experimental space. Therefore, experimenter usually fix compositional or combinational freedom in order to optimize the other freedom. However, such experiment largely limits accessible experimental space even with acceleration by data-driven methods, and true global minimum in the vast experimental space is difficult to be found. It is expected that Bayesian optimization employing machine learning model that can learn both structural and compositional information of functional materials can be a potential solution to this problem, however, its efficiency remains unclear due to difficulty in data acquisition. Unlike other material systems, for functional materials, it is difficult to generate data by simulation, because, functional materials are extremely complex systems that are usually consisted of such as multiple polymeric raw materials and nano-particles. Moreover, for real experiments, it is difficult to acquire many data due to experimental costs and time issues. In this study, we demonstrate the efficiency of Bayesian optimization for tuning of both structural and compositional freedoms of functional materials by developing an experimental system specially designed for the rapid evaluation of the performances of optimization approaches. The developed experimental system is a multi-component polymeric film consisting of around 20 raw materials in maximum and was tuned to allow rapid evaluation of target properties. The target property chosen for the demonstration is the adhesiveness of the film which is expected to show non-linear responses to compositions of materials and is also expected to be difficult to optimize. As expected, by using machine learning model that can learn both structural and compositional freedoms of materials, we successfully found the optimal design within the large experimental space of the developed system. Additionally, we compared the performances of optimization strategies employing Bayesian optimization, classical design of experiment method, and experiments based on knowledge of materials experts, and we report both advantages and disadvantages of each strategy. From the early 2022, we plan to release the PA system as in-house plat-form for speeding up the actual functional material design.

Data-driven materials design methods are rapidly spreading in the functional materials industry. Among them, Bayesian optimization is especially helpful for inverse design problems of materials. Single-objective (single target property) Bayesian optimization is known to work well under common situations in the materials industry where the amount of available data is small and the number of experiments that can be conducted is also small. However, we actually need to overcome trade-off for multiple target properties in order to satisfy several customer demands. Currently, most applications of Bayesian optimization are limited to single-objective systems or computational problems, and the efficiency of multi-objective Bayesian optimization in industrial functional material design remains unclear.
In previous our study, we proposed a new Bayesian optimization method for tuning the multiple target properties with quantitative goals that satisfy customer demands¹⁾. The proposed method employs the joint probability of achievement (PA) of predefined quantitative target values for all objective properties as the acquisition function. We have represented the efficiency of Bayesian optimization with the PA for a virtual multi-objective inverse design of functional materials employing regression models obtained from the previous experimental study²⁾.
In this study, we actually demonstrated a real-world multi-objective inverse design for functional materials using Bayesian optimization with the PA. Only by inputting the quantitative goal values for each target properties, the PA efficiently navigated the multi-objective optimization trajectory in the space of multiple target properties. We therefore successfully designed the film-shaped polymeric functional materials that satisfy three desired properties within moderate number of experiments. As far as we know, this is the first achievement that validates multi-objective inverse design for the functional material targeting multiple quantitative goals by Bayesian optimization. We plan to release the PA system as the in-house plat-form in the early 2022 for speeding up the actual functional material design.

1) Hanaoka, K.; Bayesian Optimization for Goal-Oriented Multi-Objective Inverse Material Design, iScience 2021, in press.
2)Wang, B., Cai, J., Liu, C., Yang, J., and Ding, X.; Harnessing a Novel Machine-learning-assisted Evolutionary Algorithm to Co-optimize Three Characteristics of an Electrospun Oil Sorbent. ACS Appl. Mater. Interfaces 12, 2020, p42842–42849.

Lead-acid batteries are a type of rechargeable battery that uses lead, lead dioxide, and dilute sulfuric acid. They are used in batteries for automobiles, uninterruptible power supplies (UPS) for industrial use (for example, in data centers, in hospitals) and other infrastructure facilities, and main power supplies for electric vehicles such as battery-powered electric forklifts.
One of the important characteristics of lead-acid batteries and other storage batteries is their discharge characteristics, such as battery capacity and internal resistance. Discharge characteristics are evaluated from the transition of voltage values as time series data, which are obtained from measurements that a fully charged storage battery is discharged to the end-of-discharge voltage at a certain temperature and current. Since the battery capacity is evaluated from the above discharge data, the measurement is a time-consuming process.
For lithium-ion batteries, which are a type of rechargeable battery like lead-acid batteries and are widely used in the world, researches on time series analytics (for example, residual life prediction) has been actively conducted. However, there have been few reports of research on time series prediction for lead-acid batteries.
The purpose of this study was to shorten the measurement time of data required to evaluate the discharge characteristics of lead-acid batteries. Therefore, we focused on whether it is possible to predict the actual discharge characteristics from the first half of the time series data obtained in the discharge experiment.
In this study, we used about 11,000 discharge data obtained from in-house measurements. These data consist of a time series data of voltage values every second. A machine learning model for discharging time prediction was developed using support vector regression (SVR).
As a result, it was found that the SVR model can accurately predict the time from start to finish discharge by using both the first 30% or 50% of the time series data and the difference every 10 seconds as an explanatory variable.
This result suggests that the measurement time required to predict battery characteristics can be shortened and that the present method has a potential to speed up the evaluation and development process of lead-acid battery materials.

Regression-based optimization methods are known to be effective tools in the field of material developments. However, in practical applications, we often face methodological problems of the poor performance in the case of utilizing high dimensional and incomplete datasets. Bayesian Optimization (BO) is a well-known regression-based optimization method. Although BO works well to reach global minimum/maximum to objective functions in many cases, its efficiency decreases as the dimension of design spaces become higher, such as molecular descriptors. Furthermore, BO cannot deal with incomplete initial datasets without arbitrary treatments for missing values, such as imputation. Gaussian mixture regression-based optimization (GMRBO) is proposed as an efficient optimization method with direct inverse analysis, which works better than BO in high dimensional design space [1]. However, GMRBO also has the same drawback as BO, that it cannot deal with incomplete datasets.
In this study, we have developed an optimization method by extending GMRBO to deal with incomplete datasets. The performance of the extended GMRBO (exGMRBO) was investigated and compared with BO and GMRBO. For the investigation, we used the benchmark function ZDT1 with 30 design variables and two objective functions. The percentage of missing data in the initial dataset was maintained from 10% to 70%.
The results show that while the exGMRBO performance difference between GMRBO and BO was smaller for higher missing rates (above 60%), it achieved a higher performance difference for missing rates between 10 to 50%. However, the performance of exGMRBO outperformed that of GMRBO and that of BO in all cases. Therefore, it was verified that exGMRBO indicates a better performance than BO and GMRBO when the initial dataset contains missing values.
We have also applied the exGMRBO to find Pareto front of the conflict properties of polymers.

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

[1] H. Kaneko, Chemometrics and Intelligent Laboratory Systems, Volume 208, 104226 (2021)

1. Introduction
Meticulous design of molecules is a key of material development. However, molecular design processes heretofore has been driven by human expert’s trial-end-error cycles, therefore, a typical lead time to design new material takes more than 10 years. Over the past half a decade, data-driven approaches leveraging artificial intelligence (AI) have been introduced to this field, especially on the context of generative model. The primary requirement for molecular generative models is industrial practicality, including high speed design with large structural variety. Another important requirement is that a model should be accessible for researchers and engineers of experimental chemistry, who are not necessarily familiar with software-related works; writing Python scripts, using a command line interface, etc. Releasing a state-of-the-art molecular generative model to a wide variety of potential users is important.
In this paper, we will present on our web application (webapp) – based molecular design platform that has practical advantages in terms of speed and diversity of molecule generation, interpretability of model, no requirement of pre-training. The user interface (UI) is designed so to make the tool accessible for experimental chemists.

2. Method
The main workflow is composed of four steps. (1) Training a model: the system encodes input molecular structures to a set of feature vectors by graph kernel approaches. Those feature vectors are used to build a regression model to predict target properties. (2) Setting target properties: a user sets target properties values that should be satisfied by new molecules. (3) Structure generation: molecular graphs are built by repeatedly connecting graph resources, which consists of atoms, rings, and user-defined substructures, while avoiding isomorphic duplications. (4) Online evaluation : during running the graph generation algorithm, each generated structure is encoded to a feature vector and then evaluated on the regression model. If the predicted property satisfies the user-set target value, the structure is accepted and stored in the candidate list. Repeating the above process (3) and (4), molecular structures satisfying target properties are generated.

3. Results
First, we used a small subset (300 sample) of QM9 dataset to evaluate the performance. We trained a property prediction model on the energy of the highest occupied molecular orbital (HOMO); E_homo. Carrying out property prediction by changing models and sweeping hyper parameters, the system produced reasonable accuracy when using kernel ridge regression. We use this model for the following molecular generation process. In structure generation, we used three target properties; E_homo ~ -0.28 (Ha), -0.25 (Ha), and -0.20 (Ha). Depending on the target E_homo value, the generation speed (i.e. the number of molecules generated per second) ranges from 29 to 77 per sec. The speed of other state-of-the-art molecular generative models is typically 1 to 10 per sec when QM9 is used, that confirms our method’s significant acceleration in terms of speed.
For more practical use, we design new photoacid generator (PAG). We trained our model with ~1,300 PAG structures extracted from U.S. patents, and corresponding property data which we calculated by DFT simulation. We targeted five properties; LogW, LogP, T_bio, LD50, and lamda_max, satisfying specific ranges set by a PAG chemist. After running MolGX for 6 hours, more than 2,000 molecular structures were generated, that is more than 100 time of acceleration in design speed. The designed PAG structures were experimentally synthesized.

4. Web Application
We implemented the algorithms as a GUI-based web application for a wide variety of users especially experimental chemists. We designed the workflow from a viewpoint of user experience (UX), developed API set, and implemented it on a Kubernetes-based cloud environment. The system is today in-service on IBM Cloud.

Multicrystalline materials are often the most immediate choice for various applications including solar cells because of their affordable price and productivity. As for solar cells, the conversion efficiency based on multicrystalline materials is generally inferior to that based on monocrystalline ones. This is mainly explained by the presence of dislocation clusters, which is closely related to crystal structure components such as crystal orientations. We aim to control the formation of dislocation clusters by clarifying the generation mechanisms with the aid of an ingot-scale analysis of crystal orientations, grain boundaries, and dislocation clusters.
For this purpose, we have developed a machine learning-based method for estimating crystal orientations from multiple optical images of textured multicrystalline wafers and reported that the estimation accuracy can be improved by data augmentation. In this study, we report on further improvement of the prediction accuracy by optimizing training data through statistical analysis of the estimation results in different conditions.
In the first step, we perform alkaline texturing on the wafer surface. This shaves off some surface particles and leaves only {111} planes on the surface, creating a pyramid-like structure. Shining light on this pyramid will generate specific reflection patterns that inherit crystal orientation information. A house-made apparatus is used to generate and capture these patterns. The apparatus mainly consists of a camera, a rotating collimate light for illumination, and a stage. As the illumination light rotates around the sample, optical images are taken within every 5 degrees of rotation. We integrate these 72 images into a 72-dimensional signal intensity matrix which we call the “reflection profile.”
In the second step, we carry out the machine learning-based orientation estimation. An LSTM neural network and some connected layers constitute our model. The pre-obtained reflection profile is used as the input to the model, and quaternions as the output. Correct grain orientations are measured by the Laue scanner method as teacher data. Note that the training, validation, and test datasets consist of the reflection profiles of 3838, 960, and 6373 grains, respectively.
Successful training of the model resulted in an estimation error median of 8.61 degrees. Based on the assumption that a diverse reflection profile should enhance estimation accuracy, we increased the size and diversity of the training data. Diverse data was created by a variety of incident light angles. We set the elevation angle of the incident light to three different values; 30, 45, and 60 degrees. Using all three elevation data we improved accuracy by 15%. Comparing the results trained on each angle, we discovered that there is an optimal elevation angle. The estimation accuracy varied by ~30% depending on the elevation angle. To acquire the best results most efficiently, we then performed double angle training using two elevations to train the model. Considering results from single angle training, we selected the two angles with higher accuracy;30 and 45 degrees, and trained the model, expecting better results compared to other selections. However, this resulted in a poorly trained model, indicating that the combination of angles with good results alone does not always train the model effectively. Further evaluation of results revealed that the three elevations were capable of correctly estimating different orientations. For example, at 30 degrees, planes {100} were correctly estimated while {111}~{101} were poorly estimated. At 60 degrees, {111}~{100} were poorly estimated. By calculating the correlation coefficiency between each two elevations, we revealed that the selection of angles with smaller correlation resulted in a better-trained model. This indicates that the selection of angles with different estimation capabilities should train the model most effectively.

Material science research papers have a large amount of information about materials scattered in numerous images, tables, and text. Images and text are forms of unstructured data while tables are structured ones. To retrieve research papers related to a particular topic in a specialized materials science domain or get information from both forms of data are trivial tasks. Therefore, to streamline the information extraction from research papers, we present latent Dirichlet allocation (LDA) assisted topic labelling to obtain glass science papers based on their abstract. Further, we develop “Caption Cluster Plots” (CCP) to automate information extraction from figure captions. Using both LDA and CCP, we have also developed “Elemental Maps” which disseminate the information about which chemical elements are used in abstracts of which research papers and associated figure captions. Also, we use table extraction to create a dataset of compositions and properties of inorganic glasses. Hence, this pipeline will enable researchers to explore different material science domains and excavate the hidden information from the vast corpora of research articles.

Nanoparticles that block fundamental bacterial functions like quorum sensing could potentially replace conventional antibiotic therapies. Vesicles that bind interfacially to charged biomolecules could be used to block these quorum sensing pathways. Here, lipid-based vesicles consisting of dipalmitoyl-sn-glycero-3-phosphocholine lipids and polyamidoamine dendron-grafted amphiphiles (PDAs) are studied. PDAs have charged terminal groups that can electrostatically bind to oppositely charged biomolecules. Variation of the generation and relative concentration of PDAs could affect the functionality and stability of the vesicle. To identify the optimal control parameters, i.e., PDA generation and concentration, a large phase space of vesicles is studied as a function of the two parameters. For each model, a flat membrane is simulated under conditions that forces the membrane to bend its edges, eventually forming a spherical vesicle. This process, also known as membrane closing, entails several structural changes and hence there are numerous fluctuations in the system. This often leads to simulation failure. In this situation, a new set of conservative parameters, e.g. a smaller integration timestep, could be used to successfully proceed with the simulation. However, if the conservative parameter is used for the remaining duration of the simulation, the entire process will turn out to be computationally expensive and time consuming. In this study, a computational workflow is designed to divide the entire membrane closing process into multiple stages, each associated with a short simulation. Each simulation has a predetermined set of regular and conservative parameters. Analysis routines are weaved into the design of this workflow. On the basis of the performance of the previous simulation (success/failure), the analysis routines adaptively decide on an appropriate set of parameters (regular/conservative) for the next simulation. In this way, the membrane closing process is conducted efficiently without any human intervention. The workflow is generalized for biomolecular simulations and has been named PACE² (Pipeline for Automating Compliance Based Elimination and Extension). The PACE² framework lays out the process to investigate a large molecular parameter phase space to identify optimal designs for nanoparticles.

Given the increase in global population and the continuous depletion of fresh water sources, it is necessary to keep developing technologies to provide fresh water. Thus, seawater desalination is increasing in adoption and development. For this, membrane technology is capable of removing salts from water and providing fresh water, the most common membranes for this task are based on polyamide. Membrane desalination is characterized by permeate flux and salt rejection. Previously we have developed polyamide carbon nanotube composite membranes, with enhanced antifouling properties. In this work, we developed a machine learning (ML) model to predict permeate flux and salt rejection according to membrane synthesis conditions such as carbon nanotube concentration, m-phenylenediamine concentration, trimesoyl chloride concentration, and others. To construct the machine learning model we used feature selection and soft voting ensemble using data collected through two years of experimentation with salt rejections above 90%. The obtained ML model could predict permeate flux and salt rejection with R² of 0.8 and 0.4, respectively. Furthermore, we used Shapley additive explanations to extract feature importance that could give further insights of the ML model and membrane synthesis.

We recently developed a system that autonomously searches for the best conditions to fabricate thin-film materials. The system integrates a sputter deposition system, electrical resistance measurement system, and Bayesian optimization algorithm; the samples are transferred using robots to perform the closed-loop optimization. The system autonomously optimizes the deposition conditions to obtain the TiO₂ thin film with minimum electrical resistance. [1]
In this study, the autonomous system was used to search for new lithium solid electrolyte materials. Two target materials, Li₃PO₄ and Li₂CO₃. were co-sputtered onto a quartz substrate with comb-shaped Au electrodes. The sputtering power of each target was separately tuned to control the composition of films. The sputtering power and the flow rate of Ar/N₂ process gas were optimized using the Bayesian optimization algorithm. The ionic conductivity of thin films was evaluated using AC impedance spectroscopy.
The ionic conductivity of thin films improved to 2~3 times higher than the simple amorphous Li₃PO₄ or Li₂CO₃ solid electrolytes within twelve experimental cycles. Such an autonomous system will accelerate the development of new materials from an expanded chemical space.

[1] R. Shimizu, T. Hitosugi et al., APL Mater 2020, 8, 111110.