Voxelized atomic structure framework for materials design and discovery

Abstract

We present a computational framework for developing physics-based, high-fidelity structure–property relationships with atomic systems. In this framework, atomic structure is quantified by directionally resolved two-point spatial correlations of the charge density field, projected to a salient low-dimensional feature space via principal component analysis (PCA), and correlated to physical properties by Gaussian process regression (GPR). The charge density field provides a complete, purely physics-based definition of the atomic structure that is independent of chemical species information and does not require additional feature engineering or idealizations beyond those of first-principles computations. The two-point spatial correlations capture the salient spatial features underlying the atomic structure that dictate the physics underlying the material response. Since the feature engineering approach explored in this work is universally applicable to all atomic structures independent of the chemical species present in the structure, it offers new avenues for efficiently exploring the space of atomic structures for desired property combinations. A further contribution of this work comes from utilizing the uncertainty quantification inherently provided by GPR to deploy a Bayesian experiment design strategy to minimize the number of computationally expensive physics simulations required to achieve the desired accuracy. In this work, we demonstrate the proposed framework to elucidate the relationship between the chemical composition and bulk modulus in AlNbTiZr high entropy alloys. It is shown that a highly accurate structure–property relationship with less than 2% average error can be established using a small training dataset of less than 30 samples.