Machine Learning Prediction of Material Properties Using Computational Physics Models

Abstract

Experimental synthesis and characterisation of materials are costly, time-consuming, and challenging, as the chemical and structural design space is enormous, and predictive methods for material properties are the centrepiece of computational materials research. Physically based descriptors and target labels are derived from computational physics methods, such as density functional theory (DFT), molecular dynamics (MD), finite element analysis (FEA), and Monte Carlo simulation; however, these methods can be costly to use for high-throughput screening. This article explores the potential of machine learning (ML) models to predict thermal conductivity, elastic modulus, and electronic band gap using computational physics and materials informatics descriptors. A representative set of records from the open materials databases (Materials Project, AFLOW, and Open Quantum Materials Database) is used in the study, which is an instance of quantitative computational research because the data are generated computationally. An internally consistent dataset of inorganic crystalline materials was developed, comprising 18,430 materials from various published databases, including both experimental and benchmark experiments, for a total of 64 experimental, structure- and simulation-derived properties. Five machine learning models, including linear regression, random forest, XGBoost, SVR, and feed-forward NN, were evaluated using root mean square error (RMSE), mean absolute error (MAE), mean square error (MSE), and coefficient of determination (R2). The results show that nonlinear ensemble models and neural networks significantly outperform linear regression, with XGBoost having the best performance across all models for predicting the band gap (RMSE = 0.31 eV; R2 = 0.91), and the random forest being a good model for the prediction of the elastic modulus (RMSE = 17.8 GPa; R2 = 0.88). The feature importance analysis confirms that electronegativity statistics, density, volume per atom, mean covalent radius and formation energy (DFT) statistics are the dominant predictors. The results lend significant support to the theory of finding in informatics of materials, for which the medium is physically meaningful descriptors that connect the atomic structure with the macro characteristics of their functionality. In practice, computational physics with ML could save screening dollars, inform experimental prioritisation, and augment design processes for semiconductors, batteries, alloys in the aerospace industry, nanomaterials and biomedical materials.