Optimizing the IFMIF-DONES Particle Accelerator with Differentiable Deep Learning Surrogate Models

Galo Gallardo Romero, Guillermo Rodríguez-Llorente, Lucas Magariños Rodríguez, Rodrigo Morant Navascués, Nikita Khvatkin Petrovsky, Roberto Gómez-Espinosa Martín

Keywords: Fourier Neural Operators, IFMIF-DONES accelerator, Deep Learning, Surrogate Models, Parameter Optimization, Linear Accelerator, Deep Learning Surrogate Models, Differentiable Surrogate Models, Fourier Neural Operator, Particle Accelerator Optimization, IFMIF-DONES, High Energy Beam Transport Line, Quadrupole Optimization, Neutron Irradiation, Gradient Descent, Beam Profile Prediction, Accelerator Physics, Machine Learning for Nuclear Fusion, Neutron Source Facility, Computational Efficiency, Inverse Problems in Physics, Beam Configuration Optimization, NVIDIA Modulus, OPAL Simulations, Physics-Informed Machine Learning, Scientific Experiment Optimization, Data-Driven Optimization, Beam Transport Modeling, Accelerator Design, Surrogate Modeling Techniques, Scientific Computing, High-Fidelity Simulations, Neural Network Architecture, Partial Differential Equations, Fast Inference Models, Deuteron Beam Dynamics, Computational Physics, Neural Operators, Simulation Speedup, Numerical Methods in Physics, Scientific Facility Enhancement, Inverse Problem, Accelerated Simulation, Nuclear Fusion, Partial Differential Equations, PyTorch, Gradient Descent

TL;DR: This paper introduces a method to optimize the IFMIF-DONES particle accelerator using Fourier Neural Operators, enabling fast and accurate beam profile predictions to efficiently adjust quadrupole values and improve performance

Abstract: In this work, Deep Learning Surrogate Models are employed to optimize the quadrupole values in the initial section of the High Energy Beam Transport Line of the IFMIF-DONES accelerator. Two Fourier Neural Operator models were trained: one for predicting two-dimensional beam profiles and another for forecasting one-dimensional beam statistics along the accelerator's longitudinal axis. These models offer up to 3 orders of magnitude speedup compared to traditional simulations, with a trade-off of maintaining accuracy within percentage errors below 6$\%$. Moreover, their differentiability allows seamless integration with optimization algorithms, enabling efficient tuning of quadrupole values to achieve specific beam objectives. This approach offers a robust solution for enhancing the performance of IFMIF-DONES accelerator and other scientific experiments.

Submission Number: 18