Catalyst-Agent: Autonomous heterogeneous catalyst screening and optimization with an LLM Agent

TL;DR

Catalyst-Agent is introduced, a Model Context Protocol server-based, LLM-powered AI agent that can explore vast material databases using the OPTIMADE API, make structural modifications, calculate adsorption energies using Meta FAIRchem's UMA (GNN) model via FAIRchem's AdsorbML workflow and slab construction, and make useful material suggestions to the researcher in a closed-loop manner.

Abstract

The discovery of novel catalysts tailored for particular applications is a major challenge for the twenty-first century. Traditional methods for this include time-consuming and expensive experimental trial-and-error approaches in labs based on chemical theory or heavily computational first-principles approaches based on density functional theory. Recent studies show that deep learning models like graph neural networks (GNNs) can significantly speed up the screening and discovery of catalyst materials by many orders of magnitude, with very high accuracy and fidelity. In this work, we introduce Catalyst-Agent, a Model Context Protocol (MCP) server-based, LLM-powered AI agent. It can explore vast material databases using the OPTIMADE API, make structural modifications, calculate adsorption energies using Meta FAIRchem's UMA (GNN) model via FAIRchem's AdsorbML workflow and slab construction, and make useful material suggestions to the researcher in a closed-loop manner, including surface-level modifications to refine near-miss candidates. It is tested on three pivotal reactions: the oxygen reduction reaction (ORR), the nitrogen reduction reaction (NRR), and the CO2 reduction reaction (CO2RR). Catalyst-Agent achieves a success rate of 23-34 percent among all the materials it chooses and evaluates, and manages to converge in 1-2 trials per successful material on average. This work demonstrates the potential of AI agents to exercise their planning capabilities and tool use to operationalize the catalyst screening workflow, provide useful, testable hypotheses, and accelerate future scientific discoveries for humanity with minimal human intervention.

Figures (8)

• Figure 1: Overview of Catalyst-Agent: (a) the LLM Agent supervisor acts as the chief orchestrator and the five specialized MCP servers act as the operating limbs (b) the end-to-end flowchart showing structure retrieval, modification, evaluation, and iteration.
• Figure 2: Optional surface-level modifications for catalysis after slab construction: (a) Top-layer substitution (doping) and (b) in-plane strain applied to the surface slab. All applied modifications are recorded in the results JSON.
• Figure 3: Screening campaign overview across the three target reactions. (a) Success rate $R_{\mathrm{succ}}$. (b) Distribution of trials to first success for successful materials, horizontal lines indicate mean and min--max range. (c) Fraction of materials that required at least one surface modification (trial count ${>}\,1$), compared between successful and failed material categories.
• Figure 4: Distribution of successful and unsuccessful surface modifications per material among materials that entered the modification loop (Total Trials $> 1$), grouped by outcome category: (a) modified successful materials and (b) modified failed materials. Plots show the full distribution with mean and min--max range; sample sizes $n$ are annotated below each group. The contrast between panels highlights the differential modification productivity across outcome categories.
• Figure 5: Plots of the number of trials per material, separated by reaction task and outcome category (success vs. failure). Horizontal lines indicate mean and min--max range. The distributions illustrate that convergence is rapid for most successful materials, while a subset of near-miss candidates and most CO$_2$RR failures require extended iteration.