Explainable Machine Learning Framework for Bandgap Prediction and Virtual Screening of 2D Nanomaterials

Abstract

The discovery of two dimensional (2D) nanomaterials with tailored electronic bandgaps is critical for next generation optoelectronic devices, yet high throughput screening via density functional theory (DFT) remains computationally prohibitive. We present an end-to-end, cloud enabled machine learning framework that integrates Azure ML and MLflow experiment tracking for reproducible bandgap prediction and virtual screening of 2D materials. Using the JARVIS-DFT dataset comprising 1,103 two dimensional materials, we engineer 94 compositional and structural features free of DFT derived descriptors that risk feature leakage and benchmark three model architectures: Random Forest, XGBoost, and a graph neural network (GNN) surrogate. XGBoost achieves the best predictive performance with a mean absolute error (MAE) of 0.502 eV and R² = 0.582 on the held out test set. We apply SHapley Additive exPlanations (SHAP) to both tree based models, identifying the composition weighted mean Pauling electronegativity as the dominant predictor of the OptB88vdW bandgap, consistent with established chemical intuition. An ensemble based virtual screening pipeline identifies 269 candidate materials in the optoelectronic relevant bandgap window of 1.0–2.0 eV, with the top five candidates KTlO, CdSbS₂Cl, ZnPS₃, BiS₂, and MoS₂Cl₃ exhibiting ensemble predictions within 0.01 eV of the ideal 1.5 eV target. This work demonstrates that interpretable, leakage aware ML pipelines, deployed on scalable cloud infrastructure, can accelerate rational materials discovery while maintaining physical transparency.