Skip to yearly menu bar Skip to main content


Poster
in
Workshop: D3S3: Data-driven and Differentiable Simulations, Surrogates, and Solvers

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

Galo Gallardo · Guillermo Rodriguez Llorente · Lucas Magariños · Rodrigo Morant Navascués · Nikita Kkhvatkin Petrovsky · Roberto Gómez-Espinosa Martín

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

[ ] [ Project Page ]
[ Poster]  [ OpenReview] 

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.

Chat is not available.