PhysAgent: A Multi-Agent Approach to the Automated Discovery of Physical Laws

The discovery of physical laws has traditionally relied on human intuition, analytical reasoning, and experimental observation. However, modern physics research is increasingly constrained by challenges such as high specialization, fragmented workflows, and limited computational resources, which impede scientific progress. To address these issues, we introduce PHYSAGENT, a novel multi-agent system powered by large language models (LLMs), designed to autonomously execute end-to-end scientific workflows—from hypothesis generation and computational modeling to data analysis and discovery. PhysAgent’s innovation lies in its specialized agent collaboration: a Mentor Agent guides scientific reasoning through Socratic questioning, a Student Agent handles technical execution (e.g., code implementation, DFT calculations), and a Leader Agent dynamically optimizes task scheduling and resource allocation. Integrating domain-specific tools like first-principles simulations (e.g., Quantum ESPRESSO, VASP), PhysAgent ensures reproducibility while maintaining human-in-the-loop refinement. We demonstrate its capability to autonomously derive physical laws—such as Kepler’s laws from orbital data and Newton’s second law from force-motion experiments—without prior knowledge. Furthermore, it extends to ab initio materials modeling, automating electronic structure calculations (e.g., GaAs band gaps). In addition, PhysAgent simulates complex real-world phenomena, such as raindrop flow on train windows, highlighting its adaptability beyond traditional physics problems. By harmonizing LLM-driven planning with domain-specific tools, PhysAgent establishes a trustworthy, scalable paradigm for AI-driven physics research, highlighting the transformative potential of multi-agent intelligence in accelerating discovery across classical and quantum systems.