CatMaster: An Agentic Autonomous System for Computational Heterogeneous Catalysis Research

TL;DR

CatMaster addresses the reproducibility and cost challenges of DFT-based heterogeneous catalysis by delivering a file-centric, restartable workspace orchestrated by a hierarchical LLM agent. It combines a Planner/Executor/Summarizer loop with a persistent whiteboard and a multi-fidelity toolkit that couples surrogate relaxations (MACE) to targeted DFT validation, enabling complex workflows from O$_2$ spin-state checks to HER descriptor screening on Pt--Ni--Cu alloys. Key contributions include the file-centric execution contract, robust restartability and auditability, multi-fidelity screening, and tool-light compositional autonomy for long-tail tasks such as EOS fitting and SAC geometry construction. The approach reduces manual workflow overhead while preserving a full evidence trail, facilitating inspection, extension, and biologically interpretable modeling decisions in catalysis research.

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.

Figures (6)

• Figure 1: Overview of CatMaster’s architecture and file-centric data lineage. A high-level protocol request is decomposed into milestone tasks by the orchestration layer (dynamic execution, persistent whiteboard memory, and optional human-in-the-loop checkpoints). These tasks are executed through an extensible tool layer (geometry/input, data retrieval, and analysis utilities) and dispatched to heterogeneous compute backends (HPC clusters, dedicated GPU servers, and local resources). A unified evidence layer archives traces, artifacts, and final reports to support inspection, reproducibility, and restartability.
• Figure 2: Lifecycle of the O$_2$ spin-state workflow illustrating CatMaster’s sequential execution and evidence generation. (a) Milestone task sequence produced by the Planner and executed by the tool layer, including workspace setup, molecular construction, VASP input preparation, parallel execution of singlet/triplet branches, and automated parsing/reporting. (b) Scientific artifact from the run: the triplet state is recovered as the ground state, with $\Delta E_{\mathrm{singlet}-\mathrm{triplet}}=+0.38$ eV (PAW-PBE) and chemically sensible relaxed O--O bond lengths.
• Figure 3: BCC Fe surface workflow: protocol sensitivity and downstream CO adsorption. (a) Deferred-resolution planning via a whiteboard data contract: the most stable facet is unknown at plan time and is written to a persistent record after surface evaluation, then consumed by subsequent adsorption tasks. (b) Protocol sensitivity of Fe(110)/(100)/(111) surface energies: comparison between a standard symmetric-slab protocol (no D3) and a user-constrained protocol (center-fixed slabs + D3), showing a systematic increase in $\gamma$ while preserving the stability ordering. (c) CO adsorption relaxations on the Fe(110) $2\times 2$ slab: representative initial placements (bridge/hollow/on-top) are relaxed, and multiple initial placements converge to a common relaxed on-top motif; the final relaxed configuration shown has $E_{\mathrm{ads}}\approx-2.17$ eV.
• Figure 4: Outcomes of the multi-fidelity HER screening campaign in the Pt--Cu--Ni alloy space. (a) Funnel view of the pipeline from Materials Project retrieval and deduplication to slab generation, symmetry-unique adsorption-site enumeration, surrogate screening, and targeted DFT validation; numbers indicate the counts at each stage in this workflow run. (b) Distribution of $\Delta G_{\mathrm{H}^\ast}$ over all enumerated sites predicted by the surrogate, with the thermoneutral activity window highlighted (example criterion: $|\Delta G|<0.10$ eV). (c) Representative shortlisted candidate surfaces near the thermoneutral window (structures shown) with their validated $\Delta G_{\mathrm{H}^\ast}$ values.
• Figure 5: Equation-of-state (EOS) of BCC Fe generated via CatMaster using a volume scan. Markers denote DFT total energies as a function of volume per atom, and the solid line shows the Birch--Murnaghan fit used to extract the equilibrium volume and bulk modulus.