MRS Meetings and Events


DS04.04.01 2022 MRS Spring Meeting

Predicting and Understanding Perovskite Nanostructure Formation Through Machine Learning and Data-Driven Modelling of In Situ Spectroscopic Data

When and Where

May 10, 2022
9:15am - 9:30am

Hawai'i Convention Center, Level 3, 313B



Jakob Dahl1,2,Emory Chan2,A. Paul Alivisatos1

University of California, Berkeley1,Lawrence Berkeley National Laboratory2


Jakob Dahl1,2,Emory Chan2,A. Paul Alivisatos1

University of California, Berkeley1,Lawrence Berkeley National Laboratory2
Non-equilibrium materials synthesis is complex, poorly understood and mostly explored using researchers' intuition. As a result, this area has been a prime target for applying machine learning to materials sciences in recent years. There has been significant success in creating black box algorithms that predict synthesis outcomes. However, it has been difficult to integrate physical systems knowledge of synthesis kinetics into these models and glean further understanding of the process using more interpretable techniques such as physics informed neural networks (PINNs) or bayesian optimization of complex models. Part of this difficulty lies in the difficulty in obtaining the required time series data : most materials systems require elaborate and expensive measurement setups such as in-situ X-ray beamlines to observe the synthesis process, which do not lend themselves to the collection of many experiments at different conditions. Perovskite nanocrystals are an ideal study subject in this regard: They can be synthesized at room temperature and their entire formation process can be observed optically in a stopped-flow instrument in dilute solution within a few minutes. The optical spectra contain a wealth of information, including crystallite size, identity and concentration. Through collection of kinetic data at over 100 synthesis conditions as well as product endpoints at over 2000 synthesis conditions, we have obtained a dataset which can be used to train models that explicitly fit the formation process of these crystals. We employ both active learning of a generative physical model on a limited experimental subspace and neural network models for the entire experimental space. For the physics based model, it is possible to interpret the parameters and hyperparameters used as chemical kinetic rates, rate laws and other physical constants such as activation energies and exciton effective masses. In the case of neural networks, it is possible to interpolate the entire time series efficiently and predict a physically reasonable and differentiable formation result for unknown samples. We believe that using in-situ data is key to any realistic prediction of synthesis problems, not only to inform physical models of synthesis, but also to train machine learning algorithms and will be useful in generating more predictive and interpretable models in the future.


chemical synthesis | in situ

Symposium Organizers

Jeffrey Lopez, Northwestern University
Chibueze Amanchukwu, University of Chicago
Rajeev Surendran Assary, Argonne National Laboratory
Tian Xie, Massachusetts Institute of Technology

Symposium Support

Pacific Northwest National Laboratory

Publishing Alliance

MRS publishes with Springer Nature