RealMat-BaG Benchmark for Bandgap Prediction in Semiconductors

A new benchmark, RealMat-BaG, has been introduced to evaluate machine learning models for bandgap prediction in semiconductors under experimentally relevant conditions. The benchmark addresses the poor generalization of models trained on computational data to experimental measurements. It includes an open-access dataset of experimental bandgaps with aligned crystal structures, comparing graph neural networks and classical machine learning baselines. Performance is assessed across statistical and domain-based splits, transfer from DFT-computed to experimental bandgaps, and interpretability at elemental-property and structural levels. Results highlight fundamental generalization limitations of current models.