Agentic Scientific Simulation: Execution-Grounded Model Construction and Reconstruction

Researchers introduce JutulGPT, an AI agent system for physics-based simulation that addresses the problem of underspecified natural language descriptions in scientific modeling. The system uses an execution-grounded approach where the simulator validates physical accuracy, but reveals limitations in tracking tacit assumptions made through simulator defaults.

• →JutulGPT combines code synthesis, static analysis, and systematic interpretation of solver diagnostics for scientific simulation.
• →The system explicitly detects underspecified modeling choices and resolves them autonomously or through user queries.
• →Agent-mediated model construction can be successfully grounded in simulator validation for improved accuracy.
• →Simulator defaults create invisible assumptions that aren't captured in logs, representing a structural limitation.
• →Autonomous reconstruction experiments reveal latent degrees of freedom and provide methodology for auditing reproducibility.