Voxelized Atomic Structure Framework for Developing Structure-Property Relationships from Density Functional Theory Calculations

The behavior of material systems is governed by complex physical phenomena taking place over a hierarchy of length scales, making it computationally infeasible to model the full material response with physics-driven simulation methods alone. Machine learning approaches offer new opportunities to efficiently learn the underlying structure-property relationships in material systems using low-computational cost surrogate models trained on expensive physics-based computations. In this talk, we present a comprehensive framework for formulating physics-based, high-fidelity reduced-order material structure-property relationships using density functional theory computations. In this framework, the charge density field is directly utilized as the definition of atomic structure. This provides a complete, purely physics-based definition of the atomic structure that does not require any <i>ad hoc</i> feature engineering or idealizations beyond those of the first-principles computations. The spatial features underlying the atomic structure that dictate the physics underlying the material response are then captured by the directionally resolved two-point spatial correlations of the charge density field, projected to a salient low-dimensional feature-space using principal component analysis and correlated to physical properties using Gaussian process regression (GPR). An active learning strategy based on Bayesian experiment design is implemented using the uncertainty quantification provided by GPR to minimize the number of computationally expensive physics simulations required for model training. We utilize this framework to elucidate physical insight into the relationship between chemical composition and bulk modulus in high entropy alloys.