Creating Automated Workflows for Functional Materials Discovery
When
Where
Dr. Thomas Purcell
Assistant Professor, Chemistry & Biochemistry
University of Arizona
Artificial intelligence (AI) creates models that can accelerate the discovery of functional materials. An open question is selecting the relevant materials features (descriptive parameters that characterize the material, that should be used to represent the material's function of interest, especially when there is a paucity of good-quality data. Here we present an approach that combines symbolic regression, and other regressors, with feature importance methods such as SHAP to select an optimal subset of primary features from a large pool of candidate inputs for a given dataset. We then test the approach on two different problems: thermal conductivity and electron mobility. For thermal conductivity we use a set of primary features related to different aspects of thermal conductivity and use the models plus importance metrics to learn the conditions needed for a computational funnel style workflow. We then supplement this dataset with information about the DFT calculated electronic bandstructure to learn the experimentally measured electron mobilities of 64 materials. For this example, the reduced dataset not only preserves the main signal found by SISSO, but also significantly enhances the model performance. Finally, we highlight how this approach can be used as a feature-selection criteria before learning a final model. The presented approach highlights how explainable AI techniques can not only act as a post hoc explanation generator for machine learning but also improve the training of models for smaller datasets.
Host: Dr. Mark Marley