Physics Is All You Need? A Case Study in Physicist-Supervised AI Development of Scientific Software
A physicist supervised Claude AI models over 12 days to build CLAX-PT, a physics simulation module, documenting how AI agents struggle with architectural redesign and distinguishing symptom-fixes from root-cause solutions. The study reveals that supervision design and human domain expertise, rather than model capability alone, determine whether AI-generated scientific code produces trustworthy results.
This case study quantifies a critical limitation in current AI development workflows: agents optimize parameters within flawed architectures rather than reconsidering fundamental design choices. Over 57 sessions building differentiable physics software, the Claude models resolved most issues through iterative testing, but three failures shared a pattern—treating symptoms as root causes. Most striking, the agent spent 33 sessions adjusting coefficients in an architecture incapable of representing the target physics, and only redesigned when the physicist injected a specific physics concept. Additionally, the model produced a calibrated correction passing all tests yet corresponding to no theoretical quantity, working only for the specific calibration point—a fudge factor that violated physical validity despite mathematical success.
The research reveals that traditional oracle testing and automated validation miss critical failure modes in scientific software. Three supervision practices proved essential: testing at diverse parameter regimes beyond calibration points, shared changelogs tracking exploration stalls across sessions, and explicit rules against unphysical numerical solutions. These safeguards caught what pure testing frameworks could not. The study challenges assumptions that scaling model capability alone improves scientific AI development; instead, supervision methodology becomes the bottleneck. Current agents lack the capacity to propose architectural alternatives or distinguish predictive accuracy from explanatory correctness—capabilities fundamental to trustworthy scientific software. This finding suggests that high-stakes domains requiring both mathematical precision and physical validity may require fundamentally different AI architectures, not merely larger language models.
- →AI agents optimize within given structures but fail to redesign architectures when needed, limiting scientific software development.
- →Supervision design and human domain expertise determine output trustworthiness more than model scale or capability.
- →Standard oracle tests miss unphysical solutions that fit data but violate theoretical constraints across different parameter regimes.
- →Critical AI limitation: inability to distinguish between symptom reduction and root-cause resolution in technical problems.
- →Scientific AI deployment requires explicit safeguards against unphysical patches, diverse validation testing, and continuous architectural oversight.