Reinforcement learning for multigroup energy grids

Multigroup transport codes collapse continuous-energy nuclear data onto a fixed set of energy groups, and where the group boundaries fall has a large effect on accuracy. A poor structure loses accuracy, and adding groups to make up for it costs run time and memory. Most structures in common use were picked by hand or by heuristics for one class of problem.

Approach

I treat building a group structure as a sequence of decisions. The agent starts from a high-fidelity energy grid and removes boundaries one at a time until it reaches the target number of groups. It is trained with a modified version of proximal policy optimization (PPO), and the reward favors accurate structures that use fewer groups. Because the agent learns which boundaries matter, it avoids the local minima that a greedy search can get stuck in, without restricting the starting grid.

Running a full transport solve for every candidate structure would make training far too slow. Instead, the reward comes from neural network surrogates that take energy, material, and spatial information as inputs. This removes the main computational limit on group structure optimization and also speeds up training.

Results

On the Godiva and BeRP ball criticality problems, the group structures built by the agent outperform commonly used group structures. They perform about as well as hierarchical agglomeration, with more flexibility in how the final structure is chosen.

Publications

Application of Reinforcement Learning for Multigroup Energy Grid Optimization for Neutron Transport Criticality Problems. Ben Whewell, Nathan Gibson, and Ajeeta Khatiwada, (2026). Under Review. arXiv: 2605.27895
Application of Reinforcement Learning to Multigroup Energy Grid Optimization for Criticality Problems. Ben Whewell, Nathan Gibson, and Ajeeta Khatiwada, (2026). International Conference on the Physics of Reactors (PHYSOR 2026), Turin, Italy.
MetaHeuristic Feature Selection for Energy Group Optimization and Analysis. Natalie Rouse, Benjamin Whewell, and Nathan Gibson, (2025). Technical Report LA-UR-25-29285, Los Alamos National Laboratory. DOI: 10.2172/2588817