Optimization of a Nuclear Reactor

by Paul Weimer (TUM)

Thursday 21 May 2026, 14:30 → 15:30 Europe/Berlin

E18/ENE Seminar Room 3268 (TUM PH)

This talk presents a framework that couples a custom GPU-accelerated Monte Carlo neutronics simulation tool (GAINS) with a large language model (LLM) -based agentic optimizer (NOVA) to automate and accelerate the reactor design process. The simulation side achieves a 5–23x speedup over conventional Serpent 2 calculations while maintaining geometric and physical consistency. On the optimization side, the LLM agent is equipped with a structured toolchain covering geometry generation, case management, simulation execution, and results parsing, while being deliberately constrained to mitigate hallucination and overconfidence.

As a first demonstration of the framework's capability, the system was able to independently construct a complex RBMK reactor model from scratch. Building on this, the talk then moves to the FRM II HEU-to-LEU fuel conversion challenge, a design space involving parameters such as number of fuel plates, involute radius, plate thickness, and fuel length, subject to criticality, cycle length, flux, and safety constraints. This talk will show the current state and capabilities but also highlight how LLMs like to “cheat the system” and pass of wrong or incomplete data.