Explainable Machine Learning for Predicting Band Gaps and Formation Energies of Inorganic Materials: Evidence from the Materials Project

Abstract

Machine learning prediction of semiconductor properties provides an effective way to quickly screen materials and save on time-consuming simulations. In this work, explainable machine learning models have been developed and tested for predicting the band gaps of III–V inorganic semiconductors by using composition and crystallographic descriptors. The modelling framework consisted of Ridge Regression, Elastic Net, Random Forest, Extra Trees, Gradient Boosting and Support Vector Regression (SVR), with five-fold cross-validation for the formula grouping. The evaluation of model performance was done by calculating mean absolute error, root mean squared error and coefficient of determination. Gradient Boosting gave the best overall performance, which means that it can capture the nonlinear relationship between the chemical composition, lattice geometry, density, and unit-cell volume. The most influential predictors were identified by permutation-based interpretation as mean atomic number, periodic table period, lattice parameter a, mean atomic mass, volume and density. The results also indicated that, for certain GaN structures, errors in predictions were larger than others, which can indicate that the basic compositional and geometric descriptors are not sufficient to describe polymorphism, local bonding and unusual electronic configurations. In general, this work shows that explainable ensemble learning can give meaningful and interpretable band-gap predictions in specialized semiconductor systems.