Theory-guided machine learning to predict configurational energies of high distortion alloy systems

• Main Author: Huang, Xufa
• Other Authors: Kedar Hippalgaonkar
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Materials; Engineering::Computer science and engineering::Computing methodologies::Simulation and modeling
• Online Access: https://hdl.handle.net/10356/165985
• Institution: Nanyang Technological University

Cluster expansion (CE) is a popular surrogate model to density functional theory (DFT) for modeling the stability of alloy systems through configurational energies. However, since CE is a lattice-based model, its accuracy is often poor when applied to high-entropy alloys (HEAs) with significant structural distortion. State-of-the-art attempts at using CE with machine learning (ML) models like Lasso and Bayesian for selecting meaningful clusters show high prediction errors for these high distortion alloy systems, where the contributions of long-range effective cluster interactions (ECIs) to configurational energetics remain significant. Adopting only clusters as descriptors has proven insufficient for accurate and robust predictions. This paper presents the novel integration of feature generation from clusters in CE and over 3000 Matminer material descriptors, to comprehensively capture the behavior of complex high distortion systems. Matminer features have proved effective for predicting material properties such as bandgap, elastic constants, formation energies, adsorption energies, and ferromagnetic properties in the past. Using recursive feature elimination, optimized based on stable weight assignment of ridge regularization, we sieved out only important descriptors in a high dimensional framework where configurational energy labels vastly exceed the number of descriptors. The pipeline is applied to the ten constituent binary alloys of HEA Mo-Nb-V-Ti-Zr, which is known to have large structural distortions, and we discovered that the prediction accuracy significantly improved by an average of 56%, consistent across all ten binary alloy systems. More importantly, we found the four important classes of features—coordination number, XRD, dihedral-angle distribution function, and clusters—that our model consistently select across all ten binaries. Our results are robust, showing that the additional descriptors from Matminer can better capture the behavior of high-distortion alloy systems. These important classes of descriptors are also transferable to other complex systems, such as HEAs, that are currently poorly understood, and to give robust prediction of their properties, accelerating the discovery of these high-performance alloys.