Enhancing prediction accuracy of physical band gaps in semiconductor materials

Accurate band-gap prediction is essential for designing and discovering new materials with desired properties. However, current methods for calculating band gaps based on local and semilocal functionals lead to significant underestimation, hindering the effectiveness of in silico and high-throughput screening of materials. We present a machine learning model with domain adaptation to rapidly yield accurate band-gap prediction of semiconductors (oxides, chalcogenides, nitrides, phosphides, etc.). The approach circumvents the prerequisite for a large amount of physically measured band-gap data, which is notoriously scarce. It instead sources knowledge from a large dataset with underestimated band gaps and subsequently transfers knowledge to train a crystal graph convolution neural network (CGCNN) using a small dataset of accurate, physically measured band gaps. The prediction model shows a low mean absolute error (MAE) of 0.23 eV, outperforming those using Perdew-Burke-Ernzerhof (PBE) functionals (MAE = 0.87 eV). Visualization of the learned crystal graph using the t-distributed stochastic neighbor embedding (t-SNE) algorithm revealed that the crystal structure and composition have a strong influence on the material band gaps.