LLM生成マルチフィジックスコードの意図検証をPDEで行う新手法 Your Simulation Runs but Solves the Wrong Physics: PDE-Grounded Intent Verification for LLM-Generated Multiphysics Simulation Code

As large language models increasingly write scientific computing code, automatic generation of simulation programs is moving from novelty to practical tool. This paper raises a pointed concern: code that compiles and runs is not the same as code that solves the intended physics, and the gap between the two can be dangerously invisible.

The authors focus on multiphysics simulations, where multiple coupled physical phenomena must be discretized and solved together. They argue that LLM-generated code in this domain frequently passes syntactic and runtime checks while silently solving the wrong partial differential equations. Typical failure modes include swapped boundary conditions, missing coupling terms between physics, sign errors in material coefficients, and inconsistent units. Because the resulting numerical outputs often look plausible, such semantic errors can slip past visual inspection and even past conventional unit tests.

The proposed approach reportedly grounds verification in the PDE itself rather than in the code surface. Conceptually, the system extracts the mathematical model that the generated code actually implements and compares it, at the equation level, against the user's stated intent. This shifts validation from a software-engineering question to a domain-knowledge question, which is arguably where the real risk lives. Frameworks such as FEniCS, Firedrake, and OpenFOAM, where PDEs are expressed through relatively explicit DSL constructs, may make this kind of symbolic extraction more tractable than for hand-rolled C++ or Fortran solvers.

The broader context is that scientific machine learning and AI-assisted coding are converging quickly. Tools like GitHub Copilot and ChatGPT are already used by computational scientists for everything from mesh setup to postprocessing, and vendors including NVIDIA with its Modulus framework and several DOE laboratories have begun to publish guidance on responsible LLM use in simulation workflows. Yet verification practices have not kept pace. Most existing benchmarks reward functional correctness on toy problems rather than fidelity to a specified mathematical model, which is exactly the dimension this paper targets.

If the method generalizes, it could complement emerging techniques such as physics-informed neural networks, formal verification of numerical kernels, and differentiable simulation, by providing a lightweight semantic gate before expensive runs are launched on HPC clusters. There are open questions, of course. Extracting a canonical PDE from arbitrary code is itself a hard symbolic problem, and matching against user intent presupposes that the user can articulate that intent formally, which is not always realistic in exploratory research. The approach may therefore work best as an assistant that flags suspicious discrepancies rather than as a definitive oracle.

Still, the framing is valuable. By insisting that we ask not only does the simulation run but does it solve the right equations, the work points toward a more rigorous standard for trusting LLM-generated scientific software, and it is likely to resonate with teams who have already been burned by silently wrong results.