5th ICFA Beam Dynamics Mini-Workshop on Machine Learning for Particle Accelerators

Predicting Space-Charge Potentials and E-Fields Using CNN and KAN- 15'+5'

Not scheduled

20m

80/1-001 - Globe of Science and Innovation - 1st Floor (CERN)

Invited talks Surrogate Modelling and Digital Twins Surrogate Modelling and Digital Twins

Speaker

Isabella Vojskovic

Description

This study explores various neural network approaches to simulate beam dynamics, specifically addressing non-linear space charge effects. We introduce a convolutional encoder-decoder architecture with skip connections, achieving a relative error of 0.5% in predicting both transversal and coupled 3D electric self-fields. Additionally, to enhance interpretability and robustness, we investigate an auxiliary Kolmogorov-Arnold Network (KAN) designed to replicate the solution flow of a Finite Element Method (FEM). Our findings suggest that these advancements offer a potentially more efficient alternative to traditional numerical methods for non-linear space-charge calculations in beam dynamics simulations, delivering substantial speed-up.