Science-informed AI for learning materials physics

Ayana Ghosh
Computational Sciences and Engineering Division
Oak Ridge National Laboratory, Oak Ridge, TN 37831, USA

Abstract: In recent years, artificial intelligence (AI) and machine learning (ML) methods have been rapidly adapted in the physical sciences to gain a comprehensive understanding of material structures, properties, evolution of systems over spatial-temporal resolution, and processes involving phase transitions across various time- and length-scales. Numerous studies exist that encode complex graphs, symbolic representations, invariances, and positional embeddings in these models for targeted design. However, the inherent correlative nature of ML models does not capture the causal, hypothesis-driven nature of the physical sciences. This presentation will focus on a few instances to demonstrate how causal ML models and hypotheses-driven active learning approaches can be exploited in combination with materials representation to extract fundamental atomistic mechanisms that are directly tied to experimental observables, especially for functional materials in the domain of perovskites and two-dimensional systems.