Band Gap Prediction in h-BN Using MBJ Data and Physically Informed Descriptors

We present a reproducible pipeline for predicting the band gap of hexagonal boron nitride (h-BN) monolayers using high-fidelity Modified Becke–Johnson (MBJ) data from the JARVIS database. Departing from large statistical feature sets like Magpie, we construct a compact set of physically motivated descriptors—B–N electronegativity contrast, valence and radius metrics, bond length, areal density, and two tight-binding–inspired terms reflecting sublattice potential and π-hopping. Early trials with generic features yielded unstable performance (MAE ≈ 0.55 eV, R² ≈ 0.60). In contrast, our physical feature set achieves R² ≈ 0.9 and recovers the experimental monolayer gap within typical uncertainty, even using lightweight regression models. We also integrate a language model into the workflow to assist with parsing MBJ outputs, verifying units, formatting "material cards," and maintaining clean splits—without resorting to black-box prediction. The result is a compact, interpretable, and shareable ML workflow grounded in physical insight.