LLM Agents Generate Open-Source Multiphysics Models from Human Specifications
An arXiv preprint released October 7 2026 shows frontier large language model agents can build reproducible multiphysics models in open-source Julia for electrochemical devices…
- Researchers specify equations, parameters, methods and checkpoints; LLM agents then generate the models in Julia.
- Agent-built models agree with a COMSOL reference to within 0.7 percent of peak CO partial current density and with each other to 0.04 percent.
- Explicit specifications are shown to be critical for reproducibility when errors are deliberately introduced.
Framework Overview
The preprint presents a method in which a human-specified modeling harness encodes governing equations, parameters, numerical methods, build stages and checkpoints. Frontier LLM agents use this input to construct complete multiphysics continuum models.
A one-dimensional electrochemical CO2 reduction to CO in a porous gas diffusion electrode is used as the test case. The resulting models can resolve local pH, potential and concentration fields that are difficult to measure experimentally.
Validation Results
Independently built models, including fully autonomous agent-generated versions, match an equivalent COMSOL implementation to within 0.7 percent of the peak CO partial current density.
Agreement among different agent-built models reaches 0.04 percent. The underlying physics description remains under researcher control rather than depending on proprietary software.
Importance of Specifications
Experiments that systematically planted errors illustrate the value of explicit specifications for achieving reproducibility.
These tests also indicate the agent's capabilities and limitations when addressing issues in model physics.
Scope of the Evidence
This preprint, dated 2026-10-07, establishes a transparent approach in which physical descriptions become the primary input for model development.
The available evidence is the supplied abstract only; full paper details on implementation steps, further validation and exact agent behaviors remain unavailable.
