Traditional process to discover new materials is complex, time-consuming, and costly, which often requires years of a sustained effort. Recent advances in large language models (LLMs) have demonstrated powerful capabilities in information processing, offering new opportunities for intelligent and autonomous materials research.

In a study published in Matter, a research team led by Prof.YU Xuefeng from the Shenzhen Institute of Advanced Technology (SIAT) of the Chinese Academy of Sciences developed a knowledge-driven multi-agent and robot system (MARS) for end-to-end autonomous materials discovery.

MARS is a knowledge-driven hierarchical architecture coordinating 19 LLM agents with 16 domain-specific tools organized into functional modules, achieving closed-loop autonomous materials discovery by integrating robotic experimentation. It features distinct functional groups, which enables specialized reasoning and mirrors the workflow of a human-led laboratory.

The groups include an Orchestrator for task coordination; a Scientist Group for knowledge retrieval and solution design; an Engineer Group that translates designs into executable protocols; an Executor Group controlling robotic platforms; and an Analyst Group for data interpretation and optimization strategy.

MARS offers professional guidance for materials development and relieves the hallucination inherent to current LLMs through hybrid retrieval-augmented generation. In experimental validation, the system optimized perovskite nanocrystal synthesis within 10 iterations. Besides, it designed a biomimetic "core-shell-corona" structure for water-stable perovskite composites in just 3.5 hours

This study develops an integrated artificial intelligence (AI)-driven framework MARS which helps to accelerate materials innovation.

Knowledge-driven multi-agent and robot system. (Image by SIAT)