MELRSNet for accelerating the exploration of novel ultrawide bandgap semiconductors

Ultrawide bandgap (UWBG) semiconductors, with bandgaps exceeding 3.4 eV of gallium nitride, offer the potential to overcome the limitations of conventional semiconductors and drive innovations in electronics and photovoltaics. However, discovering such materials remains a huge challenge due to the prohibitive cost of trial-and-error-based experiments and the complexity of cutting-edge quantum mechanical approaches. Here, we develop the Multistage Ensemble Learning Rapid Screening Network (MELRSNet), a data-driven hierarchical machine learning framework integrated with high-throughput first-principles calculations, designed for swift identification of UWBG semiconductors. Trained on the Materials Project dataset, MELRSNet utilizes elemental and structural features to classify, regress, and validate potential candidates. Its efficacy is underscored by the accurate prediction of bandgaps in UWBG oxides and the revelation of metric-bandgap relationships, aligning closely with first-principles calculations. Furthermore, MELRSNet's reliability is bolstered through the identification of eight novel ternary oxide compounds, derived from monoclinic hafnium oxide crystals, exhibiting high stability, desirable band gaps, and strong ultraviolet light absorption, marking them promising candidates for lab synthesis and subsequent applications. MELRSNet not only streamlines the discovery of UWBG semiconductors but also paves the way for high-throughput computational screening of other functional materials.