A new AI framework wraps large language models in a validation-and-repair loop that makes their simulation code actually run - most of the time.
Researchers built MOOSEnger, an agent system for the Multiphysics Object-Oriented Simulation Environment (MOOSE), an open-source ecosystem for multiphysics engineering simulations. Instead of asking a language model to write a simulation file and hoping for the best, MOOSEnger checks the model's output against MOOSE-specific syntax rules, pulls in relevant domain knowledge, and runs a generate-check-repair-run loop before accepting a result. Tested on 200 prompts across eight simulation families, the harness raised the rate of code that actually executes from 5 percent to 89.5 percent when paired with GPT 5.2, and from 0 percent to 76.5 percent with the smaller Gemma 4 31B model. A separate ten-case accuracy benchmark found all ten outputs matched the intended physics setup, and eight also hit the required numerical-accuracy threshold.
Executable code and correct code are not the same thing, and the paper is candid about that gap: passing a syntax check doesn't mean the simulation reflects real-world physics. The bigger takeaway is that the scaffolding around the model - the retrieval, the linting, the repair loop - did more work here than the model itself, since even a much smaller model got a similar lift once wrapped in the same harness.
It's a reminder that in scientific computing, as in most places LLMs are being pointed at real infrastructure, the model is the easy part - the validation layer is where the actual engineering happens.