SmartMat Explorer: Automated scientific research for materials discovery

Yeyun Chen, Yujie Liu, Zhuo Yang, Yi Fan, yiming ma, Yuqiang Li, Dongzhan Zhou, Junjun He

Keywords: AI4science, scientific agent, materials discovery

Abstract: Recent advances in large language models (LLMs) have transformed artificial intelligence by positioning language as a universal interface for reasoning and interaction. Building on this foundation, scientific agents have emerged that can autonomously conduct literature analysis, generate hypotheses, guide experiments, and perform reflective evaluation. Despite this progress, enabling such agents with powerful inspiration-mining capabilities, rigorous scientific reasoning, and reliable feasibility verification remains challenging. We present SmartMat Explorer, a scientific agentic framework for functional molecular materials discovery centered on host–guest supramolecular complexes. It integrates a knowledge graph with a causally-grounded RAG to capture deeper causal structures for inspiration mining and employs a beam search–driven multi-level scoring mechanism informed by computational tools to enhance rational design decisions. To rigorously evaluate autonomous research performance, we introduce SmartMat-Bench, an extensive benchmark built from cutting-edge materials science publications and designed to cover both guided innovation and unconstrained exploration tasks. Empirical results show that SmartMat Explorer powered by Claude-3.7-Sonnet improves reasoning accuracy by 35\% across five literature tasks and surpasses state-of-the-art deep research models by 6\%. Coupling it with domain-specific tools further yields 33\% gains in material discovery tasks. Overall, SmartMat Explorer offers a systematic and interpretable paradigm for robust and verifiable molecular materials discovery.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 1521