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

Achuth Chandrasekhar, Janghoon Ock, Amir Barati Farimani · 2026

Read Catalyst-Agent: Autonomous heterogeneous catalyst screening and optimization with an LLM Agent by Achuth Chandrasekhar, Janghoon Ock, Amir Barati Farimani (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-01311.txt.

Introduction

Chandrasekhar, Ock, and Barati Farimani offer Catalyst-Agent, an LLM-powered agent that runs closed-loop catalyst screening alone for cell reactions like oxygen, nitrogen, and CO2 reduction. Built on an MCP server, it searches stores through OPTIMADE, builds slabs, finds sticking values with Meta FAIRchem's UMA model inside AdsorbML, and checks reaction markers. It refines near-miss hopefuls with shape tweaks. Across oxygen, nitrogen, and CO2 campaigns it performs strongly and settles in about one and a half to three and a half trials.

Catalog record

CREATORAchuth Chandrasekhar, Janghoon Ock, Amir Barati Farimani
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE23967
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.01311

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Catalyst-Agent: Autonomous heterogeneous catalyst screening and optimization with an LLM Agent Achuth Chandrasekhar, Janghoon Ock, Amir Barati Farimani 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 CO 2 reduction reaction (CO 2 RR). Catalyst-Agent achieves a success rate of 23–34% 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. keywords American Chemical Society, L a T e X † † affiliation: Mechanical Engineering, Carnegie Mellon University, Pittsburgh, PA 15213, USA † † affiliation: Department of Chemical and Biomolecular Engineering, University of Nebraska–Lincoln, Lincoln, NE 68588, USA † † email: [email protected] † † affiliation: Mechanical Engineering, Carnegie Mellon University, Pittsburgh, PA 15213, USA † † affiliation: Biomedical Engineering, Carnegie Mellon University, Pittsburgh, PA 15213, USA † † affiliation: Chemical Engineering, Carnegie Mellon University, Pittsburgh, PA 15213, USA † † affiliation: Materials Science and Engineering, Carnegie Mellon University, Pittsburgh, PA 15213, USA † † affiliation: Machine Learning Department, Carnegie Mellon University, Pittsburgh, PA 15213, USA † † abbreviations: IR,NMR,UV † † suppinfo: \mciteErrorOnUnknownfalse 1 Introduction Designing optimal catalyst materials is a central challenge in chemical technology. norskov2014fundamental ; dumesic1996principles ; zitnick2020introduction Catalyst discovery has traditionally proceeded through iterative experimental cycles of hypothesis, synthesis and performance measurement, with substantial time and monetary expenditure. Experimental researchers synthesize novel materials and characterize their activity, selectivity and stability, while computational researchers use first-principles electronic-structure calculations to predict catalytic trends before synthesis. Density functional theory (DFT) enables estimation of adsorption energies, activation barriers and other energetic descriptors on well-defined catalytic surfaces, supporting scaling relations and mechanistic insights. raccuglia2016machine ; wang2025density Despite this progress, conventional catalyst screening remains bottlenecked by the creation of physically consistent surface models and the execution of expensive calculations across diverse materials, facets and adsorption sites. The adsorption energy quantifies the thermodynamic driving force for adsorbate binding to a catalytic surface and corresponds to the global minimum energy over all configurations: ock2023beyond ; chen2025descriptor ; karimadom2021calculating \Delta E_{\mathrm{i}}=E_{\mathrm{slab}+\mathrm{i}}-E_{\mathrm{slab}}-E_{\mathrm{gas}}, (1) \Delta E_{\mathrm{ads}}=\min(\Delta E_{\mathrm{i}}) (2) where E_{\mathrm{slab}+\mathrm{i}} is the total energy of the relaxed adsorbate–slab system, E_{\mathrm{slab}} is the energy of the clean slab, and E_{\mathrm{gas}} is the gas-phase adsorbate reference energy. Negative \Delta E_{\mathrm{ads}} values indicate stronger binding, while positive values indicate weaker adsorption. Because adsorption energies of key intermediates correlate with rate-determining steps, they are widely adopted as activity and selectivity descriptors. Finding the global minimum energy across all configurations demands orchestrating multiple calculations over surface terminations, spin states and numerical settings. ock2026adsorb ; lan2023adsorbml ; chen2025multi High-throughput computational catalysis has focused on automating these evaluations across material families, but even with automation DFT relaxations remain slow (approximately 48 hours each), and the combinatorics of facets, sites and adsorbates multiplies this cost. This has motivated machine learning (ML) surrogate models that approximate DFT at a fraction of the computational expense. klawohn2023gaussian ; kyvala2023optimizing ; srinivasan2024electronic ; li2022deep ; thant2025kernel ; pun2019physically ; cao2024machine Among the various ML approaches, atomistic models that operate directly on atomic geometries are especially attractive. Graph Neural Networks (GNNs), which represent atoms as nodes and bonds as edges to capture local chemical environments, are particularly effective for predicting energies and forces of atomic systems. deng2023chgnet ; liao2023equiformerv2 ; batzner20223 ; batatia2022mace ; choudhary2021atomistic ; chen2019graph ; wood2025family Building on progress in GNNs, Universal machine-learned interatomic potentials (U-MLIPs) have been made possible by large databases consisting of DFT calculations and ab-initio molecular dynamics simulations (AIMD). kabyldamolecular ; maheshwari2025beyond ; han2025distmlip The AdsorbML framework by Lan et al. formalizes ML-accelerated adsorption evaluation by providing tools for slab construction, adsorbate placement, relaxation and energy computation. ML potentials relax catalyst-adsorbate systems from initial configurations. The FAIRchem initiative from Meta FAIRChem extends this ecosystem with the UMA family of universal ML interatomic potentials wood2025family , which provide transferable energy and force predictions across diverse chemistries. In this work, we employ the UMA small model (uma-s-1p1) as the primary surrogate within the AdsorbML workflow. First-principles validation is not eliminated but moved downstream to confirm ML-shortlisted candidates with higher-fidelity DFT calculations and experiments. Modern deep learning has also developed the transformer architecture, introduced in the seminal “Attention Is All you Need” work, which has revolutionized natural language processing. vaswani2017attention This architecture forms the backbone of large language models (LLMs) capable of advanced reasoning over goals, constraints and intermediate results in natural language. This creates an opportunity to connect scientific workflows with deep learning architectures that can predict outcomes in an iterative fashion. chaudhari2024alloybert ; jadhav2025llm ; badrinarayanan2025mofgpt ; chandrasekhar2025nanogpt ; chandrasekhar2024amgpt ; ock2023catalyst ; wang2022molecular For catalyst discovery, this opens the possibility of an autonomous agent that retrieves structures, evaluates energetics and refines candidates in response to quantitative feedback. Recent studies have endeavored to explore this intersection. CatGPT by Mok and Back mok2024generative , a fine-tuned GPT-2 radford2019language autoregressive model, generates heterogeneous catalyst surface structures as text sequences encoding lattice parameters, elemental symbols and fractional coordinates. Controllable generation is achieved through adsorbate-conditioned prompts, with validity filtering via a BERT-based devlin2019bert anomaly detector and validation using EquiformerV2 liao2023equiformerv2 and finally DFT, though the approach focuses on structure proposal rather than closed-loop energetic optimization. ChemReasoner by Sprueill et al. sprueill2024chemreasoner reframes screening as LLM-guided search with quantum-chemical feedback via beam search and context-aware planning achiam2023gpt ; the physics-informed reward steers exploration and reduces hallucination risk, though the system relies on simplifying assumptions for 3D structure conversion and a less accurate reward model than current universal potentials. Yao et al. yao2025llm presented an “LLM to Agent” framework for MgH 2 catalyst design combining literature mining ye2024prompt with supervised ML prediction and SHAP-based interpretability lundberg2017unified . These efforts establish that LLMs can contribute to catalyst discovery, but none provides a fully autonomous, tool-grounded framework spanning database retrieval through iterative surface refinement within a single closed-loop pipeline. The Model Context Protocol (MCP) framework mcp addresses this gap by providing a standardized interface for exposing external servers with tools and resources to an AI agent with tool-augmented reasoning. Agentic AI is poised to revolutionize scientific discovery. nigam2026polymeragentlargelanguagemodel ; pak2026agentic ; jadhav2026large ; zeng2025llm ; han2025tdflow ; ock2025large ; chandrasekhar2507automating ; chaudhari2025modular ;

Read the complete text