Applications of Fourier Neural Operators in the Ifmif-Dones Accelerator

Guillermo Rodríguez-Llorente, Galo Gallardo Romero, Roberto Gómez-Espinosa Martín

Keywords: Fourier Neural Operators, IFMIF-DONES accelerator, Deep Learning, Deep Reinforcement Learning, Surrogate Models, Parameter Optimization, Linear Accelerator, Beam Control, Inverse Problem, Accelerated Simulation, Nuclear Fusion, Neutron Irradiation, Partial Differential Equations, Sequential Decision Making, PyTorch, Stochastic Gradient Descent, Markov Decision Processes, Bellman Equations, Policy Function, Magnetic Control, Quadrupoles, OPAL Simulator, Gymnasium Framework, Proximal Policy Optimization, Magnetic Forces, Particle Distribution, Simulation Speed-up, Loss Function, AI for Science

TL;DR: Fourier Neural Operators (FNO) rapidly predict beam properties for the IFMIF-DONES accelerator, enabling faster DRL agent training and accelerator parameter optimization.

Abstract: In this work, Fourier Neural Operators are employed to improve control and optimization of an experimental module of the IFMIF-DONES linear accelerator, otherwise hindered by its simulations high complexity. The models are trained to predict beam envelopes along the lattice’s longitudinal axis, considering variations in quadrupole strengths and particle injections. They serve three purposes: enabling fast inference of beam envelopes, creating an environment for training a Deep Reinforcement Learning agent responsible for shaping the beam, and developing an optimizer for identifying optimal accelerator parameters. The resulting models offer significantly faster predictions (up to 3 orders of magnitude) com- pared to traditional simulators, with maximum percentage errors below 2 %. This accelerated simulation capability makes it feasible to train control agents, since the time per step taken is reduced from 3s to 4 × 10e−3s. Additionally, Stochastic Gradient Descent was applied to optimize one of the models itself, determining the best parameters for a given target and thus solving the inverse problem within seconds. These results demonstrate the synergy and potential of these Deep Learning models, offering promising pathways to advance control and optimization strategies in the IFMIF-DONES accelerator and in other complex scientific facilities.

Submission Number: 37