PhysicsGPT: An Autonomous Scientific Discovery Agent Built on the Physics Intermediate Representation Framework

Key Points

• This research aims to develop an autonomous agent for scientific discovery in physics using advanced computational methods.
• Integrated PIR-based symbolic regression for data analysis.
• Employed a hybrid MSE+OT loss for optimizing model performance.
• Utilized conservation-law validation to ensure physical accuracy.
• Accumulated knowledge through a graph-based framework.
• Implemented active-inference for selecting experimental paths.
• Achieved 100% success on L1--L4 physics benchmark tasks.
• Demonstrated fully autonomous operation without human intervention.