Optimization through In-Context Learning and Iterative LLM Prompting for Nuclear Engineering Design Problems

TL;DR

The paper tackles multi-objective optimization of nuclear fuel lattice designs by introducing Optimization by Prompting (OPRO), which leverages in-context reasoning of LLMs guided by a meta-prompt to iteratively propose designs evaluated with high-fidelity CASMO-5 simulations, all without explicit hyperparameter tuning. In a GE-14 BWR lattice test, the approach demonstrates that reasoning-focused LLMs with detailed context can surpass traditional genetic algorithms in achieving $k_{\text{inf}}$ near 1.05 while maintaining a low PPF, illustrating faster and more explainable design exploration. A comparison across prompting strategies and model variants reveals that detailed context prompts with larger LLMs yield the best results, though smaller models can perform competitively under no-context prompts, highlighting trade-offs between reasoning depth, speed, and prompt design. The findings suggest LLM-driven optimization offers a practical, hyperparameter-free pathway for nuclear-design optimization, with implications for industrial adoption, while underscoring scalability and hallucination risks that motivate future domain-specific LLMs.

Abstract

The optimization of nuclear engineering designs, such as nuclear fuel assembly configurations, involves managing competing objectives like reactivity control and power distribution. This study explores the use of Optimization by Prompting, an iterative approach utilizing large language models (LLMs), to address these challenges. The method is straightforward to implement, requiring no hyperparameter tuning or complex mathematical formulations. Optimization problems can be described in plain English, with only an evaluator and a parsing script needed for execution. The in-context learning capabilities of LLMs enable them to understand problem nuances, therefore, they have the potential to surpass traditional metaheuristic optimization methods. This study demonstrates the application of LLMs as optimizers to Boiling Water Reactor (BWR) fuel lattice design, showing the capability of commercial LLMs to achieve superior optimization results compared to traditional methods.

Figures (6)

• Figure 1: Half symmetry of GE-14 Dominant (DOM) fuel lattice. The image is generated by the authors using Casmo-5 software.
• Figure 2: Optimization by LLM prompting flowchart with Casmo-5 evaluations. The image is generated by authors using Lucid App.
• Figure 3: Comparison of Average Best Scores Across Optimizers for Different Prompting Strategies. The yellow bar highlights the optimal approach.
• Figure 4: Average Steps Required for Convergence Across Optimizers and Prompting Strategies. The yellow bar highlights the optimal approach.
• Figure 5: Progression of Best Scores Over Optimization Steps for Different Models and Methods.