An Agentic Framework for Autonomous Materials Computation

TL;DR

This work presents a domain-informed agent capable of autonomous first-principles materials computations by leveraging a library of validated workflows and modular components to generate executable VASP parameters. A new benchmark spanning SR, BS, AE, and TS tasks demonstrates that the agent substantially improves task completion and accuracy over baseline LLMs, with pronounced gains for open-source models when equipped with the agent. The study also analyzes failure modes, highlighting INCAR tag management and workflow context as critical bottlenecks, and introduces a rigorous, openly available benchmark for evaluating materials computation agents. Overall, the results establish a verifiable foundation for automated computational experimentation and mark a step toward fully automated scientific discovery in materials science.

Abstract

Large Language Models (LLMs) have emerged as powerful tools for accelerating scientific discovery, yet their static knowledge and hallucination issues hinder autonomous research applications. Recent advances integrate LLMs into agentic frameworks, enabling retrieval, reasoning, and tool use for complex scientific workflows. Here, we present a domain-specialized agent designed for reliable automation of first-principles materials computations. By embedding domain expertise, the agent ensures physically coherent multi-step workflows and consistently selects convergent, well-posed parameters, thereby enabling reliable end-to-end computational execution. A new benchmark of diverse computational tasks demonstrates that our system significantly outperforms standalone LLMs in both accuracy and robustness. This work establishes a verifiable foundation for autonomous computational experimentation and represents a key step toward fully automated scientific discovery.

Figures (6)

• Figure 1: Overview of the proposed agentic framework for autonomous materials computation. Given user’s simulation request and accompanied files, the agent selects the most appropriate workflow from a predefined workflow library based on established best practices. Each workflow is executed as a sequence of LLM-driven modular components that handle specific tasks such as parameter generation, file operations, and command execution. Upon completion of simulations, the results are parsed and compiled into user-readable outputs.
• Figure 2: Comparison of model performance with/without the agent.(a) shows the completion rates of six models without agent or with agent support. (b) shows the accuracy of the same models under the two settings. Together, the two plots demonstrate that introducing the agent leads to robust and consistent gains in both task completion capability and correctness across diverse model architectures.
• Figure 3: Task completion rate with/without the agent. Bar charts showing the completion rate of six LLMs across four tasks. For each task, results are reported for both the setting without agent support (orange) and the setting with agent support (blue). Each panel corresponds to one task type and displays the completion rate in percentage for all considered models.
• Figure 4: Task Result Accuracy With/Without Our Agent. For each category, two bars are provided per model, corresponding to the setting without agent support (orange) and the setting with agent support (blue). Each panel represents one task type and reports accuracy in percentage for all considered models.
• Figure 5: Bar plots comparing open-source and proprietary models under two evaluation metrics.(a) Completion rates for open-source and proprietary model groups, with separate bars indicating performance without agent support and with agent support. (b) Result accuracy for the same two model groups, again showing values for both settings. Each bar plot reports percentage scores for the two categories under the two evaluation conditions.