Data-driven design of electrocatalysts: Machine learning workflows, paradigm shifts, and applications
- 1School of Information and Communication Engineering, Hainan University, Haikou 570228, China
- 2State Key Laboratory of Tropic Ocean Engineering Materials and Materials Evaluation, School of Materials Science and Engineering, Hainan University, Haikou 570228, China
Abstract
Machine learning (ML) is transforming the discovery of electrocatalysts for the carbon-neutral energy transition by overcoming the limitations of trial-and-error experimentation and costly quantum simulations. This review summarizes recent advances in data-driven electrocatalysis, focusing on two emerging paradigm shifts: autonomous literature mining assisted by large language models (LLMs) and materials representation via graph neural networks (GNNs). We first trace the evolution from manual data curation to LLM-assisted mining, which accelerates database construction by orders of magnitude; then, the shift from handcrafted descriptors to GNNs, which expands the computational screening space from thousands to millions of candidates. Subsequently, we discuss key algorithms—including active learning and symbolic regression—and their applications in the hydrogen evolution reaction (HER), oxygen evolution reaction, oxygen reduction reaction, and carbon dioxide reduction reaction, illustrating how they help identify non-intuitive descriptors that rationalize activity trends. We also highlight how these descriptor-based predictions guide the discovery of specific catalyst motifs or reaction pathways that can circumvent or break complex scaling relations inherent in multi-electron transfer processes. We then critically discuss challenges related to data heterogeneity, model interpretability, and the synthesis–performance gap, before offering a perspective on integrating physics-informed ML and autonomous self-driving laboratories to accelerate the rational design of next-generation electrocatalysts.
Related
- Cost-Effective Quantum-Mechanical Workflows for Molecular Machine Learning
- Machine-Learning-Assisted Carbon Dots: From Algorithms to Applications and Beyond
- NEBwalk: A transparent and calculator-agnostic framework for reproducible nudged elastic band workflows with machine-learned potentials and firstprinciples calculators
- Microbial Pyrolysis for the Sustainable Synthesis of Nanoparticle-Embedded Carbon Electrocatalysts
- Interfacial Effects on the Catalytic Performance of Water-Splitting Electrocatalysts
BibTeX
@article{liu2026datadriven,
title = {{Data-driven design of electrocatalysts: Machine learning workflows, paradigm shifts, and applications}},
author = {Ziwei Liu and Hanqing Gu and Liwei Yang and Wengcheng Wu and Zhong Huang and Haozhi Wang and Yang Wang and Yida Deng},
journal = {Advanced Powder Materials},
year = {2026},
month = may,
doi = {10.1016/j.apmate.2026.100433},
url = {https://doi.org/10.1016/j.apmate.2026.100433},
note = {CC BY},
}
RIS
TY - JOUR
TI - Data-driven design of electrocatalysts: Machine learning workflows, paradigm shifts, and applications
AU - Ziwei Liu
AU - Hanqing Gu
AU - Liwei Yang
AU - Wengcheng Wu
AU - Zhong Huang
AU - Haozhi Wang
AU - Yang Wang
AU - Yida Deng
PY - 2026
DA - 2026/05/21
JO - Advanced Powder Materials
DO - 10.1016/j.apmate.2026.100433
UR - https://doi.org/10.1016/j.apmate.2026.100433
AB - Machine learning (ML) is transforming the discovery of electrocatalysts for the carbon-neutral energy transition by overcoming the limitations of trial-and-error experimentation and costly quantum simulations. This review summarizes recent advances in data-driven electrocatalysis, focusing on two emerging paradigm shifts: autonomous literature mining assisted by large language models (LLMs) and materials representation via graph neural networks (GNNs). We first trace the evolution from manual data curation to LLM-assisted mining, which accelerates database construction by orders of magnitude; then, the shift from handcrafted descriptors to GNNs, which expands the computational screening space from thousands to millions of candidates. Subsequently, we discuss key algorithms—including active learning and symbolic regression—and their applications in the hydrogen evolution reaction (HER), oxygen evolution reaction, oxygen reduction reaction, and carbon dioxide reduction reaction, illustrating how they help identify non-intuitive descriptors that rationalize activity trends. We also highlight how these descriptor-based predictions guide the discovery of specific catalyst motifs or reaction pathways that can circumvent or break complex scaling relations inherent in multi-electron transfer processes. We then critically discuss challenges related to data heterogeneity, model interpretability, and the synthesis–performance gap, before offering a perspective on integrating physics-informed ML and autonomous self-driving laboratories to accelerate the rational design of next-generation electrocatalysts.
SN - 2772-834X
ER -