Generative Design and Graph Neural Network Screening of Novel Bimetallic Alloy Catalysts for the Hydrogen Evolution Reaction
Abstract
Identifying earth-abundant catalysts for the hydrogen evolution reaction (HER) is a central challenge in scalable green hydrogen production. We present a computational pipeline that couples diffusion-based generative modelling, machine-learning force-field relaxation, and graph neural network (GNN) property prediction to discover bimetallic HER catalysts beyond the structures and stoichiometries covered by existing adsorption databases. A generative model (MatterGen) is fine-tuned on bulk alloy structures labelled with their minimum hydrogen adsorption energy and steered towards adsorption energies near zero. From the generated pool we retain the stable structures that are novel bimetallic alloys: binary compositions built from the training-set metals but at stoichiometries or structures absent from that database. Each candidate's low-index facets are ranked by a machine-learning-force-field (MatterSim) surface-energy screen, and only the most stable terminations are converted into explicit slab models with hydrogen adsorbed at all high-symmetry surface sites and relaxed with the same force field. Two GNN surrogates (MEGNet and ALIGNN), trained on a high-throughput density functional theory (DFT) dataset, then screen the resulting slab candidates. Because the two models encode the local environment through different inductive biases, we retain only the alloys on which both agree, reducing the influence of any single model. This consensus screening, combined with a MatterSim energy-difference triangulation, resulted in a shortlist of near-thermoneutral bimetallic candidates. Eight are validated with plane-wave DFT, and four of them, Pd5Au, Pt5Rh, Cu5Pt3, and Pd13Ru3, bind hydrogen within the thermoneutral window. All eight alloys are dynamically stable by phonon calculations. These four alloys are highlighted as DFT-validated bimetallic HER candidates for experimental follow-up.
Related
- Sustainability-Aware Discovery of Dual-Atom Hydrogen Evolution Catalysts via a Unified Machine Learning Framework for Activity and Carbon Cost
- Improved screening of doped ZrO2 catalysts for CO2 hydrogenation from graph-based machine learning
- AI-guided high-throughput discovery of iridium- and ruthenium-free palladium-oxide catalysts for durable acidic oxygen evolution
- Dynamic Kinetic Evaluation Favors Compositionally Diverse Multicomponent Alloy Nanoparticles for Hydrogen Evolution
- Repairing PBE-Spurious Metallicity for HSE06-Level Screening of 2D Photocatalysts for Green Hydrogen Production
BibTeX
@misc{elavunkel2026generative,
title = {{Generative Design and Graph Neural Network Screening of Novel Bimetallic Alloy Catalysts for the Hydrogen Evolution Reaction}},
author = {Vipin Kurian Elavunkel and Sylvain Le Tonquesse and Hadrien Vroylandt},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.41932},
eprint = {2610.41932},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.41932},
}
RIS
TY - UNPB
TI - Generative Design and Graph Neural Network Screening of Novel Bimetallic Alloy Catalysts for the Hydrogen Evolution Reaction
AU - Vipin Kurian Elavunkel
AU - Sylvain Le Tonquesse
AU - Hadrien Vroylandt
PY - 2026
DA - 2026/10/30
PB - arXiv
DO - 10.48550/arxiv.2610.41932
UR - https://arxiv.org/abs/2610.41932
AB - Identifying earth-abundant catalysts for the hydrogen evolution reaction (HER) is a central challenge in scalable green hydrogen production. We present a computational pipeline that couples diffusion-based generative modelling, machine-learning force-field relaxation, and graph neural network (GNN) property prediction to discover bimetallic HER catalysts beyond the structures and stoichiometries covered by existing adsorption databases. A generative model (MatterGen) is fine-tuned on bulk alloy structures labelled with their minimum hydrogen adsorption energy and steered towards adsorption energies near zero. From the generated pool we retain the stable structures that are novel bimetallic alloys: binary compositions built from the training-set metals but at stoichiometries or structures absent from that database. Each candidate's low-index facets are ranked by a machine-learning-force-field (MatterSim) surface-energy screen, and only the most stable terminations are converted into explicit slab models with hydrogen adsorbed at all high-symmetry surface sites and relaxed with the same force field. Two GNN surrogates (MEGNet and ALIGNN), trained on a high-throughput density functional theory (DFT) dataset, then screen the resulting slab candidates. Because the two models encode the local environment through different inductive biases, we retain only the alloys on which both agree, reducing the influence of any single model. This consensus screening, combined with a MatterSim energy-difference triangulation, resulted in a shortlist of near-thermoneutral bimetallic candidates. Eight are validated with plane-wave DFT, and four of them, Pd5Au, Pt5Rh, Cu5Pt3, and Pd13Ru3, bind hydrogen within the thermoneutral window. All eight alloys are dynamically stable by phonon calculations. These four alloys are highlighted as DFT-validated bimetallic HER candidates for experimental follow-up.
ER -