AI-Assisted Code Porting Speeds Up Nuclear Reactor Simulation Toolkit, But Human Expertise Remains Essential

A new software engineering study demonstrates how AI-assisted coding can accelerate the development of complex scientific software—but only when paired with sustained human expertise and rigorous validation.

Researchers working on Outram Park, an open-source simulation toolkit for advanced nuclear reactors, used Claude Code to port OpenFOAM libraries into Rust and build specialized thermal-hydraulics modules. The work reveals both gains and bottlenecks in applying agentic (AI-guided) development to mission-critical physics software.

The Porting Challenge#

The team undertook an ambitious conversion: migrating multiple established open-source libraries from their original languages into Rust, which offers safety guarantees and performance benefits for scientific computing. Rather than hand-writing all the code, they worked with an AI agent in a human-in-the-loop arrangement, where developers reviewed and validated each step.

This approach yielded measurable productivity improvements. However, verification and validation (V&V)—ensuring the ported code actually solves the physics correctly—quickly became the real bottleneck, rather than code generation itself.

Testing and Validation#

The team validated their new Outram-Foam libraries against standard test cases: the cavity problem and Sod shock tube, both classical benchmarks in computational fluid dynamics. Results aligned with published values within acceptable tolerances.

They also developed tampines-steam-tables, a library for two-phase flow calculations that showed good agreement with Moody's equilibrium-flow charts. A 1D choked-flow solver built using the same agentic approach was preliminarily tested against the Edwards blowdown case, a key scenario for reactor safety analysis.

Why Domain Knowledge Still Matters#

The findings underscore a critical insight: AI code generation can produce syntactically correct, even efficient implementations—but without physics expertise, the generated code may miss problem-specific requirements or subtle numerical issues. Expert reviewers had to ensure the algorithms matched the underlying physics and behaved correctly across the range of conditions the simulation needs to handle.

Further validation work remains ongoing, yet the methodology points to real potential for accelerating open-source scientific software development when human specialists remain in the loop to guide and verify AI-generated results.