Skip to content
Farhat Ullah.
Research · Published July 2026

Physics-informed machine learning for photocatalyst discovery.

A physics-informed Random Forest and XGBoost ensemble that ranks MOF/g-C3N4 Type-II heterojunction candidates for photocatalytic hydrogen production without exhaustive hybrid-functional calculations.

Journal
Chinese Journal of Physics
Volume
103 · 1881-1895
DOI
10.1016/j.cjph.2026.07.025
Published research paper

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

The study addresses a costly materials-discovery problem: identifying metal-organic frameworks that can form useful Type-II heterojunctions with graphitic carbon nitride for photocatalytic hydrogen production. It combines physical band-alignment rules with interpretable machine learning to prioritize candidates before high-cost quantum-mechanical validation.

Authors
Abdullah Khan · Muhammad Saeed · Fakhrud Din · Farhat Ullah · Sami Ullah
Farhat Ullah's contribution
Methodology, formal analysis, validation, visualization, writing the original draft, and review and editing.
Method

What the paper contributes.

  1. 01

    Physics-grounded features

    Six interpretable descriptors encode metal-node chemistry, linker electronics and band alignment, reducing dependence on costly per-structure quantum calculations.

  2. 02

    Leakage-resistant validation

    Chemical fingerprints isolate 17 experimentally characterized MOF families as external holdouts instead of letting closely related structures leak into training.

  3. 03

    A hybrid decision pipeline

    Random Forest and XGBoost rankings are combined with thermodynamic, pore-accessibility and stability constraints to produce an experimentally useful shortlist.

Results

A shorter path to credible candidates.

The consensus ensemble was tested internally and against isolated experimental families. The final Physics-ML cascade traded a small amount of global recall for much stronger precision at the top of the ranking, where experimental resources would actually be spent.

MOFs analyzed
20,152Integrated QMOF and CoRE-MOF dataset
Ensemble ROC-AUC
0.912On 4,029 internal test structures
Top-50 precision
98%49 of 50 candidates were true positives
Final shortlist
983From 20,152 candidates
Compute reduction
20.5xVersus exhaustive HSE06 screening
External MOF families
1740 experimentally characterized structures
Applicability boundary

The validated framework is scoped to closed-shell and main-group metal nodes. Open-shell 3d transition-metal systems require higher-fidelity band-gap modeling, and every shortlisted material still needs prospective quantum and experimental validation.

Have a system that needs to work in production?

Tell me what's breaking — or what you're building.

Chat on WhatsApp