ChatMat: A Multi-Agent Chemist for Autonomous Material Prediction and Exploration

Abstract

Large language models (LLMs) have emerged as powerful tools for general-purpose tasks, but their performance in domain-specific applications, particularly in material properties prediction, is still constrained by limited access to specialized knowledge. To overcome these challenges, we introduce ChatMat, an artificial intelligence (AI) chemist powered by a multi-agent system, capable of performing complex material property predictions with minimal human intervention. Leveraging LLMs such as GPT-4o or local foundation models, ChatMat autonomously interprets unstructured textual prompts, plans scientific procedures, and executes complex materials workflows-from data retrieval to simulation and modeling-with minimal human input. The system is orchestrated by a Manager agent, which interfaces with human researchers and coordinates four role-specific agents: Property Depositor, Computing Designer, density functional theory (DFT) Operator, and Machine Learning-driven Potential Energy Surface (ML-PES) Performer. This modular, multi-agent architecture enables the seamless integration of data-driven and physics-based techniques, establishing a robust, autonomous pipeline for material prediction and novel material exploration. We demonstrate the versatility and efficacy of ChatMat through four experimental tasks of increasing complexity, including structure generation, charge density distribution acquisition, database operation, and ML-PES construction. Furthermore, a series of quantitative evaluation metrics have been designed to benchmark its performance, illustrating ChatMat's reliability and adaptability across diverse materials domains. Our work bridges the gap between autonomous experimental research and computational science, showcasing the potential of domain-specific autonomous research to accelerate material prediction and exploration.