Agentic Scientific Simulation: Execution-Grounded Model Construction and Reconstruction

A team of researchers from SINTEF and the Norwegian University of Science and Technology (NTNU) has introduced a novel framework for 'agentic scientific simulation,' tackling a core challenge in using LLMs for physics-based modeling. Their paper, 'Agentic Scientific Simulation: Execution-Grounded Model Construction and Reconstruction,' argues that natural language descriptions of simulation models are inherently underspecified, leading to multiple valid but scientifically distinct configurations. To address this, they developed JutulGPT, a reference implementation that organizes model construction as an execution-grounded 'interpret-act-validate' loop. Here, the simulator itself (JutulDarcy) acts as the authoritative arbiter of physical validity, not just a runtime, grounding the AI agent's code generation in real-world physics.

The JutulGPT agent combines structured retrieval of documentation with code synthesis, static analysis, execution, and interpretation of solver diagnostics. Its key innovation is explicitly detecting underspecified modeling choices—like material properties or boundary conditions—and resolving them either autonomously (with logged assumptions) or through targeted user queries. A secondary experiment demonstrated the system's ability to autonomously reconstruct a reference model from abstract text, revealing that variability in reconstruction exposes latent degrees of freedom in simulation descriptions. This provides a practical, automated methodology for auditing the reproducibility of scientific models, a critical step for trust in AI-assisted research. The team has made all code, prompts, and agent logs publicly available, offering a foundational tool for reproducible computational science.