Harnessing the Power of Gradient-Based Simulations for Multi-Objective Optimization in Particle Accelerators
K. Rajput, M. Schram, A. Edelen, J. Colen, A. Kasparian, R. Roussel, A. Carpenter, H. Zhang, and J. Benesch
In Review at Machine Learning: Science and Technology, 2024
Controlling particle accelerators requires simultaneous optimization of multiple potentially-competing objectives. We compared reinforcement learning (RL) methods to traditional multi-objective optimization techniques and found that RL algorithms that leverage differentiable environments achieve the highest performance. The key ingredient was a physics-based surrogate of the accelerator environment that enabled the RL agent to back-propagate gradients in a physics-informed way.