CatAgent: Multi-Agent Orchestration for Electrocatalyst Discovery

Seokhyun Choung, Hoyun Kim, Jongheun Kim, Jeong Woo Han

Keywords: Language model, Agentic workflow, Electrocatalysis, Machine learning interatomic potential, Benchmark

TL;DR: Large language models with critic feedback identify optimal adsorption energy electrocatalysts for hydrogen evolution 2x faster than random search.

Abstract: The discovery of efficient electrocatalysts is hindered by the combinatorial scale of candidate material space. Here we present CatAgent, an autonomous multi-agent workflow driven by large language models that achieves up to a 9.65-fold increase in discovery rates over adsorbate-specific random baselines. We benchmark 13 language models in single-shot and iterative modes across bimetallic alloy compositions. Critic-enabled iterations improve performance for most architectures, with top models concentrating proposals near zero theoretical overpotential. Our results suggest that catalyst screening can benefit from LLM-guided chemical reasoning.

Submission Track: Feedback-Based Learning for Materials Design - Tiny Paper

Submission Category: AI-Guided Design

Submission Number: 53