The Voxelized Atomic Structure Machine Learning Framework for Modeling Structure-Property Relationships in High Entropy Alloys

Machine learning (ML) has emerged as a powerful tool for the discovery of novel materials that optimize a combination of desired properties. In this talk, we present the Voxelized Atomic Structure (VASt) ML framework for modeling structure-property relationships in atomic systems. In this framework, the VASt structure-property relationship connects the charge density field of an atomic system to one or more properties of interest using a Gaussian process regression (GPR) model. Because a VASt structure-property relationship depends only on the charge density field, it can model a highly complex and diverse space of atomic systems – consisting of many different element types — with high fidelity. Furthermore, using Bayesian optimization, highly accurate VASt structure-property relationships can be achieved using a small number of training samples. As a result, accurate VASt structure-property relationships can be developed for complex material properties that require computationally expensive Density Functional Theory (DFT) simulations.<br/><br/>We demonstrate the VASt framework for modeling structure-property relationships by developing a VASt structure-property relationship to optimize the composition of AlNbTiZr refractory high entropy alloys (RHEAs) for superior thermal properties (i.e., Gibbs free energy, constant pressure heat capacity, and thermal expansion coefficient) at high temperatures. Although these properties can be computed from DFT using the quasi-harmonic approximation, the number of atoms and lack of symmetry in the atomic structure of RHEAs makes computing their second-order force constants computationally expensive. As a result, it is computationally infeasible to compute these properties directly from DFT for the large number of possible RHEA compositions. By developing a VASt structure-property relationship, we can rapidly predict these properties with high accuracy over the entire composition space using only a small number of computationally expensive training samples.