Green hydrogen, made by splitting water with renewable electricity, depends on catalysts that can drive two half-reactions: hydrogen evolution at the cathode and oxygen evolution at the anode. Finding better electrocatalysts for these reactions has long been a slow, expensive process of trial and error, and even high-throughput computer screening with density functional theory (DFT) is often too costly to exhaustively explore the immense space of possible material combinations. A new review published in the journal Catal on 16 December 2025 argues that graph neural networks (GNNs), a branch of artificial intelligence, are now reshaping this field by learning material properties directly from atomic structure.
The review, led by Wenhao Dong, Qi Wang, Liping Ren, Jinjia Wei, Shaohua Shen and Jie Chen, systematically traces the evolution of machine learning in electrocatalyst design, from early hand-engineered descriptors to the emerging GNN-driven paradigm. The authors categorize GNN architectures into invariant, equivariant and Transformer-based models, and summarize recent applications to hydrogen evolution reaction (HER) and oxygen evolution reaction (OER) catalyst discovery. Their goal is to provide researchers with a methodological framework for using GNNs to accelerate the discovery of high-performance electrocatalytic materials.
Traditional machine learning approaches to electrocatalyst design rely on a multi-step workflow: data collection, feature engineering, model selection and training, and model evaluation. Feature engineering, in which the physical structure of a material is converted into numerical descriptors such as electronegativity, coordination number or d-band center, is described as the most critical and time-consuming step. Algorithms like support vector machines, Gaussian processes, random forests and gradient boosting decision trees have all been applied to predict catalytic properties, but their accuracy depends heavily on the quality of these manually designed features, which require deep domain expertise and can miss complex latent patterns in the data.
GNNs bypass this bottleneck by representing molecules or crystals as graphs, where atoms are vertices and chemical bonds or spatial proximity relationships are edges. Through a mechanism called message passing, atoms gather information from their neighbors, aggregate it, and update their own state representations over multiple iterations. This allows the model to automatically learn topological and geometric features, capturing atomic-scale connectivity, distance and angular relationships without manual feature engineering. According to the review, this preservation of topological integrity gives GNNs an advantage over traditional methods, which compress three-dimensional structures into fixed-length descriptor vectors and inevitably sacrifice spatial fidelity.
The review divides GNN architectures into three main paradigms. Invariant models, such as DTNN, SchNet, DimeNet and GemNet, convert three-dimensional coordinates into scalars that remain unchanged under rotation and translation, making them well suited for predicting scalar properties like formation energy or band gap. Equivariant models, including EGNN, PaiNN, NequIP and MACE, retain and transform geometric information throughout the network, handling both scalar and vector features so that outputs rotate in sync with input geometries. This is essential for tasks involving forces or molecular dynamics simulations. Transformer-based models, such as Graphormer, TorchMD-Net, SE(3)-Transformer and Equiformer, redesign self-attention mechanisms to possess geometric symmetry, combining global context modeling with physical accuracy.
For HER catalyst design, the review highlights several concrete applications. Li and colleagues combined active learning with the DimeNet-LSCG model to screen platinum alloys, identifying 12 materials with catalytic activity comparable to Pt(111), among which Cu3Pt and FeCuPt2 emerged as promising low-cost alternatives. Chen and coworkers built a high-throughput framework using SchNet to construct local potential energy surfaces as inputs to CGCNN, achieving a final mean absolute error of 0.231 eV and identifying 43 high-performance alloys from 2,973 candidates, some of which have been experimentally validated. Oguz and colleagues integrated Equiformer V2 with the AdsorbML framework to predict hydrogen adsorption energies on bimetallic alloy surfaces, screening approximately 5.6 × 10^5 initial adsorption sites and pinpointing four inexpensive candidates (Fe2Sb4, Cu6Sb2, Cu6Sn2, Ni2Sb2) with performance comparable to platinum.
For OER catalyst design, the review describes how GNNs have been applied to alloy, metal-oxide and high-entropy systems. Ghanekar and colleagues developed the ACE-GCN framework to model local environments through subgraphs, iteratively predicting complex surfaces of 11,500 Pt(221) step structures from only 600 data points. Chun and colleagues combined equivariant m-PaiNN models with active learning to explore over 30,000 bimetallic and trimetallic single-atom catalyst active sites, finding that Co and Ni sites generally outperform other metals in OER. Park and colleagues used CGCNN with active learning to screen 795 binary oxide supports and 21 transition metal atoms, identifying 14 novel single-atom catalysts with theoretical overpotentials lower than the benchmark commercial catalyst IrO2. Tran and colleagues, working with the OC22 dataset, used pre-trained GemNet-OC to predict adsorption energies for all facets of 4,119 oxide materials, identifying 122 bulk-stable and 68 nanoscale-stable potential OER catalysts with low overpotentials.
The review also emphasizes the role of large-scale public databases in enabling this progress. The Open Catalyst Project datasets OC20 and OC22 are particularly noteworthy: OC20 contains over 1.3 million molecular adsorption structures totaling approximately 8 million data points, while OC22 covers 60,000 DFT relaxation calculations and 9.88 million single-point calculations. These datasets encompass various reaction types involving small molecules, which the authors say is crucial for training machine learning models capable of generalizing across catalytic systems. However, they caution that data heterogeneity due to differences in computational accuracy or experimental conditions remains a challenge, and datasets must undergo systematic examination and preprocessing before being used for training.
Looking ahead, the authors identify several priorities. They argue that future databases must break down data silos, establish stricter quality control standards, and develop unified descriptors and metadata protocols to enhance model generalization. They also call for deeper interdisciplinary integration, suggesting that embedding established catalytic theories into the forward propagation or loss functions of GNNs could yield next-generation physics-informed models. Finally, they envision a shift from single performance metrics toward multi-objective optimization frameworks that simultaneously consider activity, stability, selectivity, cost and sustainability, with GNNs evolving from offline prediction tools into intelligent agents that autonomously drive scientific discovery through integration with robotic experimentation and high-throughput characterization.
Subject of Research: Application of graph neural networks to electrocatalyst design for green hydrogen production
Article Title: Graph neural networks reshaping the paradigm of electrocatalyst design for green hydrogen production
Article References: Dong, W., Wang, Q., Ren, L., Wei, J., Shen, S., & Chen, J. (2025). Graph neural networks reshaping the paradigm of electrocatalyst design for green hydrogen production. Catal, 1(1), Article 12. https://doi.org/10.1007/s44422-025-00013-7
Image Credits: AI Generated
DOI: 10.1007/s44422-025-00013-7
Keywords: graph neural networks, electrocatalysts, green hydrogen, water electrolysis, hydrogen evolution reaction, oxygen evolution reaction, machine learning, density functional theory, catalyst screening, artificial intelligence, materials discovery, open catalyst datasets
