All projects

FRIB 2023 - 2025

Static beam dump AI optimization

Genetic algorithms and reinforcement learning for a higher-power static beam dump.

My contribution
Developed a hybrid genetic algorithm and reinforcement learning approach to optimize the static beam-dump geometry.
Simulation result
72% higher average power-handling capability than the existing design in simulation.

Challenge

The FRIB static beam dump absorbs intense heat from a heavy-ion beam. As the beam size changes, a fixed geometry can develop localized hot spots. This work targets a 20 kW to 50 kW upgrade by finding a geometry that stays within thermal limits across different beam sizes.

Two-stage AI optimization

A genetic algorithm (GA) explores the overall geometry, then reinforcement learning (RL) refines the surface. A two-dimensional finite-difference thermal model evaluates candidate designs, followed by three-dimensional finite-element verification in COMSOL.

The GA varies the section dimensions, plate angle, and beam position, evaluating three representative beam sizes. The maximum temperature in the optimization model falls from 554.7 °C to 345.9 °C, providing the starting geometry for RL.

GA design evolution: maximum temperature by generation (left) and temperature fields for three beam sizes (right).

2. Reinforcement learning: local refinement

Starting from the GA design, the RL agent adjusts 34 surface design points using the temperature field as its state and lower peak temperature as its reward. Over 500 episodes, the maximum temperature decreases further to 324.1 °C in the optimization model.

RL refinement: maximum temperature by training episode (left) and the evolving temperature fields (right). The animations show optimization progress, not elapsed physical time.

Numerical verification and outcome

The optimized design is checked separately with a 3D finite-element model. Across the three representative beam sizes, the highest predicted temperature falls from 688.5 °C to 324.9 °C, below the 350 °C design limit. These verification results are distinct from the 2D optimization-model temperatures shown in the animations.

3D finite-element temperature comparison of the original and optimized static beam dump for three beam sizes
3D finite-element verification: original design (left) and optimized design (right) under the same three beam conditions.

The study reports 72% higher simulated average power-handling capability than the existing design, supporting the 50 kW upgrade under the evaluated conditions. These are numerical predictions, not measurements from an operating prototype.

Publications

SCIE

  • G. Lee, J. Song, R. Quispe-Abad, M. Patil, T. Kanemura, “Optimization of the FRIB Beam Dump: A Hybrid Genetic Algorithm and Reinforcement Learning Approach,” Nuclear Science and Techniques, 37, 184, 2026. https://doi.org/10.1007/s41365-026-02020-2

Conferences

  • G. Lee, J. Song, M. Patil, R. Quispe-Abad, N. Bultman, T. Kanemura, “Design Improvements of a Minichannel Beam Dump Wing through AI-Driven Genetic Algorithms,” 16th International Conference on Heavy Ion Accelerator Technology, 2025.
  • G. Lee, J. Song, M. Patil, R. Quispe-Abad, N. Bultman, T. Kanemura, “Design Optimization of the FRIB Beam Dump through Reinforcement Learning with Genetic Algorithm,” 2025 American Nuclear Society Annual Conference, 2025.

Research Project

(2023-2025) AIP Mini-channel and Water beam dump, FRIB.