Catmaster Achieves Faster Heterogeneous Catalysis Research Using LLM-Driven Workflows

Researchers are addressing a major bottleneck in computational heterogeneous catalysis through a novel agentic autonomous system called CatMaster. Developed by Honghao Chen, Jiangjie Qiu, and Yi Shen Tew from the Beijing Key Laboratory of Artificial Intelligence for Advanced Chemical Engineering Materials at Tsinghua University, together with Xiaonan Wang and collaborators, CatMaster is a large-language-model–driven platform that automates complex computational workflows traditionally dependent on time-consuming, iterative, and often irreproducible manual scripting. The system translates natural-language requests into complete and executable calculation workspaces, including structures, input files, outputs, and detailed run records. By substantially reducing the overhead associated with workflow management, CatMaster enables researchers to focus on chemical interpretation and modelling decisions rather than the technical intricacies of setup and execution, thereby accelerating materials discovery and advancing the understanding of catalytic processes.

Although powerful libraries such as ASE and pymatgen, along with workflow managers for job orchestration, already exist, a significant operational gap remains in their day-to-day use. In practice, researchers frequently rely on manual intervention to chain tools together, convert data formats, and track intermediate structures. This manual oversight involves numerous micro-decisions, such as managing calculation files, tuning convergence parameters, diagnosing failures, and selecting candidates for higher-fidelity validation, all of which slow down iteration and reduce reproducibility. Even seemingly simple tasks, such as screening candidates for reactions like the hydrogen evolution reaction, can require extensive manual effort to execute calculations and extract chemically meaningful results, such as final energies or bond lengths.

To evaluate CatMaster’s capabilities, the authors selected the surface chemistry of body-centered cubic (BCC) iron as a representative test case, focusing on two key aspects: accuracy against established benchmarks and strict adherence to user-defined constraints in a multi-stage, end-to-end workflow. As a baseline test of physical correctness, CatMaster was instructed to compute surface energies for Fe(110), Fe(100), and Fe(111) using a standard protocol involving symmetric slabs, the PBE functional, and no dispersion corrections. The agent autonomously retrieved the appropriate bulk reference structure (mp-13), performed lattice relaxation, and constructed the required symmetric slab models. The resulting surface energies—γ₁₁₀ = 2.48, γ₁₀₀ = 2.56, and γ₁₁₁ = 2.76 J m⁻²—closely match literature values, with relative errors within 3%. This close agreement confirms that CatMaster can independently instantiate and execute standard thermodynamic workflows with high accuracy, validating its ability to handle geometry construction, parameter selection, and execution in a fully autonomous manner.

LLM-driven Workspaces for Catalysis Computation offer accelerated discovery

The research team engineered a hierarchical architecture comprising a Planner, an Executor, and a Summarizer, all anchored by a persistent whiteboard record, to manage complex catalytic computations. The Executor then carries out these tasks using schema-validated tools, while the Summarizer meticulously documents the process and results on the whiteboard, providing a clear audit trail. All demonstrations within the study were performed utilising GPT-5.2-thinking-medium or GPT-5.2-thinking-high, selected based on the complexity of each specific case. This tiered approach allows for efficient screening of a large chemical space, followed by rigorous verification of promising candidates with more computationally demanding methods. This allows for automatic human-in-the-loop intervention, enabling users to modify the task planning chain as needed. The system delivers a file-centric execution contract, persisting the authoritative project state in the file system rather than the LLM’s context window, ensuring resilience, inspectability, and reproducibility of results as Supporting Information.

CatMaster streamlines catalysis via LLM agents

This innovative approach addresses critical limitations in current workflows, which are often costly, iterative, and sensitive to setup choices. Experiments revealed that CatMaster effectively manages the complexities of catalysis research by treating the calculation workspace as the primary scientific record. The team demonstrated the system’s capabilities through four progressively complex demonstrations, beginning with an O2 spin-state check executed remotely. Measurements confirm that CatMaster’s multi-fidelity tool library enables rapid surrogate relaxations alongside accurate DFT calculations when needed.

Results demonstrate that the system’s hierarchical agent design, coupled with persistent whiteboard memory, supports long-horizon workflows and deferred resolution of intermediate outcomes. The team achieved a file-centric execution contract, transforming natural-language protocols into restartable and inspectable workspaces, ensuring a complete evidence trail. Scientists recorded that the workspace serves as an execution contract, with each task emitting a minimal evidence package including input decks, output logs, and key scalar summaries, facilitating downstream analysis and revision. This approach promises to accelerate discovery and enhance reproducibility in the field of computational catalysis.

CatMaster streamlines catalysis via LLM agents

Demonstrations across diverse workflows, including spin-state checks, surface thermodynamics, adsorption exploration, and high-throughput alloy screening, highlight its versatility and effectiveness. The authors acknowledge limitations in handling complex calculation failures, such as SCF non-convergence, and suggest developing specialised “debugger agents” as a future direction. Further research could focus on expanding the tool library and enhancing the agent’s ability to autonomously resolve intricate computational issues, ultimately enabling fully unattended operation.