Autonomous multi-agent AI accelerates hydroxide exchange membrane discovery through physics-grounded inverse design

Hydroxide exchange membranes (HEMs) hold the key to platinum-free alkaline fuel cells and electrolyzers, yet their development remains constrained by a fundamental trilemma: ionic conductivity, alkaline stability and dimensional stability are governed by competing physical mechanisms, creating a complex Pareto frontier that has resisted decades of empirical optimization. Here we demonstrate that large language model (LLM) agents, when architected as collaborative multi-agent systems with access to physics-based simulation tools, can autonomously navigate this multi-objective landscape to discover materials inaccessible to conventional approaches. Our framework, OHMind, implements a cognitive hierarchy wherein a Supervisor agent decomposes high-level design goals into executable subtasks, a Design Engine explores a 56-dimensional chemical latent space via particle swarm optimization, and Physics Verifiers validate candidates through density functional theory and molecular dynamics---all coordinated through the Model Context Protocol. Critically, the agents do not merely execute predefined workflows; they reason about simulation failures, adapt search strategies and iteratively refine hypotheses, exhibiting problem-solving behaviours that mirror human researchers. From a chemical space encompassing millions of candidate architectures, OHMind identified a family of poly(fluorene)-based HEMs featuring novel piperidinium side chains that resolve the conductivity--stability--swelling trilemma. Experimental validation of the top candidate, PBF-Pip-2, confirmed the predictions: 113~mS~cm$^{-1}$ hydroxide conductivity at 80~$^{\circ}$C coupled with exceptional durability exceeding 2,000~h in 5~M KOH. These results establish that autonomous LLM agent systems can serve as genuine cognitive partners in materials discovery, capable of solving multi-objective optimization problems that have long challenged the field. Find more in: https://github.com/lunyang/OHMind and https://lunyang.github.io/