Physics-informed machine learning screening and validation of metal-organic framework/g-C₃N₄ heterojunction photocatalysts

To address the scarcity ( < 5%) of stable, visible-light-active (1.5–3.4 eV) semiconductors for sustainable hydrogen production via photocatalytic water splitting, we bypass the prohibitive DFT costs ( > 10⁷ CPU-hours) required to screen more than 20,000 MOF/g-C₃N₄ Type-II heterojunctions by employing physics-informed machine learning. Integrating Sanderson electronegativity theory and the Butler–Mulliken band alignment formalism, we engineered six descriptors from the Quantum MOF (20,372 structures with PBE band gaps) and CoRE-MOF databases, capturing metal node chemistry, linker electronics, and porosity while enforcing thermodynamically favorable band alignment without expensive quantum calculations. A consensus Random Forest/XGBoost ensemble achieved a ROC-AUC of 0.912 and a standard precision of 68.9% on 4,029 test structures, with a hybrid-cascade top-50 precision of 98%. External validation spanned 17 experimentally characterized MOF families (40 structures); on the closed-shell safe-label subset (N=11 families), the ensemble achieved a family-level recall of 100% and a precision of 85.7%, exceeding physics-rule baselines by 27.6 percentage points at the structure level (exact binomial McNemar p=0.004). Representative correct predictions include MOF-5, UiO-66, NU-1000, and NH₂-MIL-53(Al) (probability 0.75–0.99), with ZIF-8 and Cu-BTC correctly rejected. Explainable SHAP analysis identified metal ionic radius and linker π -conjugation as the dominant features, from which we extract synthesis rules favoring Ti/Zr/Fe nodes (r_ion=60–75 pm), high-stability aluminum-halide frameworks, electron-donating aromatic linkers (C/H =1.2–1.5), and pore diameters > 3.5 Å. A hybrid screening strategy combined ML probability filtering with thermodynamic and stability constraints. The pipeline narrows 20,152 candidates down to 983, cutting computational cost by roughly 20-fold against exhaustive HSE06 screening. What we end up with is a ranked shortlist of high-probability MOF/g-C₃N₄ heterojunctions for follow-up.