Predictive Materials Design and Discovery with Causal Machine Learning

Applications of machine learning/deep learning (ML/DL) methods have become common in a variety of scientific disciplines. However, capabilities of such techniques are yet to be fully realized in fundamental physical disciplines due to limited access to labeled data, lack of extensibility, and robustness of traditional ML models. More importantly, the intrinsically in-built correlative nature of ML models does not capture the causal hypothesis-driven nature of physical sciences. The interpretability and explainability of large parameter-dependent deep networks also pose significant challenges for fashioning more informative and physically intuitive data-driven approaches. As a result, current ML approaches for materials design fail to provide quantitative and reliable predictions about structure/property relationships, useful for theory-based design workflows for targeted synthesis/characterization to accelerate time-to-solution and reduce expensive laboratory costs. This talk will focus on how causal ML models can be exploited in combination with generalized materials representation to solidify our understanding of governing fundamental atomistic mechanisms with direct ties to experimental observables. While the family of perovskite oxides remain at the center of this exploration, discussion on extending such frameworks to other material classes with real-life applications in energy, catalysis, photovoltaics, drug design, reaction mechanisms mapping, will also be incorporated in the presentation.<br/><br/>This research is sponsored by the Laboratory Directed Research and Development Program of Oak Ridge National Laboratory, managed by UT-Battelle, LLC, for the U. S. Department of Energy.