[Paper Review] CatMaster: An Agentic Autonomous System for Computational Heterogeneous Catalysis Research

Honghao Chen, Jiangjie Qiu+2 more|arXiv (Cornell University)|Jan 20, 2026

Machine Learning in Materials Science0 citations

TL;DR

CatMaster is an LLM-driven hierarchical agent system that converts natural-language requests into restartable, evidence-rich calculation workspaces for heterogeneous catalysis, integrating multi-fidelity workflows from surrogate relaxations to DFT validation.

ABSTRACT

Density functional theory (DFT) is widely used to connect atomic structure with catalytic behavior, but computational heterogeneous catalysis studies often require long workflows that are costly, iterative, and sensitive to setup choices. Besides the intrinsic cost and accuracy limits of first-principles calculations, practical workflow issues such as keeping references consistent, preparing many related inputs, recovering from failed runs on computing clusters, and maintaining a complete record of what was done, can slow down projects and make results difficult to reproduce or extend. Here we present CatMaster, a large-language-model (LLM)-driven agent system that turns natural language requests into complete calculation workspaces, including structures, inputs, outputs, logs, and a concise run record. CatMaster maintains a persistent project record of key facts, constraints, and file pointers to support inspection and restartability. It is paired with a multi-fidelity tool library that covers rapid surrogate relaxations and high-fidelity DFT calculations for validation when needed. We demonstrate CatMaster on four demonstrations of increasing complexity: an O2 spin-state check with remote execution, BCC Fe surface energies with a protocol-sensitivity study and CO adsorption site ranking, high-throughput Pt--Ni--Cu alloy screening for hydrogen evolution reaction (HER) descriptors with surrogate-to-DFT validation, and a demonstration beyond the predefined tool set, including equation-of-state fitting for BCC Fe and CO-FeN4-graphene single-atom catalyst geometry preparation. By reducing manual scripting and bookkeeping while keeping the full evidence trail, CatMaster aims to help catalysis researchers focus on modeling choices and chemical interpretation rather than workflow management.

Motivation & Objective

• Provide an autonomous, restartable workspace for atomistic catalysis studies that preserves an auditable record of all inputs, outputs, and decisions.
• Enable robust, interruptible execution across local and HPC resources with persistent project state.
• Bridge high-throughput surrogate screening with targeted high-fidelity DFT validation to accelerate discovery.
• Demonstrate multi-fidelity workflows and long-tail task handling through a series of catalysis demonstrations.

Proposed method

• Introduce a file-centric execution contract where milestone tasks emit tangible artifacts and a persistent project record.
• Employ a hierarchical planner-executor-summarizer agent loop with a persistent whiteboard memory for long-horizon workflows.
• Integrate a multi-fidelity tool library including geometry construction, retrieval, VASP input generation, surrogate relaxations (MACE), and DFT validation.
• Use strict schema validation for all tool inputs to ensure reproducibility and auditability.
• Enable deferred-resolution planning via whiteboard artifacts to handle unknown intermediate outcomes.
• Allow human-in-the-loop checkpoints to intervene at critical decisions while preserving the audit trail.

Experimental results

Research questions

• RQ1 Can an LLM-driven agent autonomously convert natural-language catalysis intents into restartable, auditable workspaces?
• RQ2 How does hierarchical agent orchestration perform long-horizon, multi-fidelity catalysis workflows across heterogeneous compute resources?
• RQ3 What is the impact of surrogate-based pre-screening followed by targeted DFT validation on identifying promising HER descriptors in alloy systems?
• RQ4 To what extent can composition of stable primitives enable long-tail tasks (EOS fitting, SAC geometry) without dedicated tools?
• RQ5 What are the limitations and failure modes of such an agent in real-world catalysis workflows?

Key findings

• CatMaster produces restartable workspaces that capture inputs, logs, and a machine-readable summary, enabling auditability and reproducibility.
• On O2 spin-state tasks, the system correctly identified triplet O2 as the ground state with a 0.38 eV energy difference and bond lengths around 1.233–1.234 Å.
• For BCC Fe surface energies, CatMaster's results matched literature benchmarks with relative errors within 3% for γ(110), γ(100), and γ(111).
• Under protocol-sensitivity planning with dispersion and center-fixed slabs, CatMaster preserved stability order while reflecting higher energies due to constraints.
• In multi-fidelity HER screening across Pt–Ni–Cu, surrogate screening with MACE narrowed 347 sites to top candidates with DFT validation showing close agreement (best ΔG_H* deviation +0.033 eV).
• The system demonstrates compositional autonomy by performing EOS fitting and SAC geometry construction using generic primitives without dedicated tools.