IPAC'24 - 15th International Particle Accelerator Conference

Advancing non-linear Space Charge Simulations: Neural Networks and Analytical Approaches

SUPG004

19 May 2024, 14:00

4h

Bluegrass (MCC Exhibit Hall A)

Poster Presentation MC5.D02 Nonlinear Single Particle Dynamics Resonances, Tracking, Higher Order, Dynamic Aperture, Code Developments Student Poster Session

Speaker

Isabella Vojskovic (European Spallation Source ERIC)

Description

This study introduces a convolutional encoder-decoder architecture inspired by the skip connections used in ResNets, designed for predicting the transversal E-field. It has demonstrated impressive initial results, achieving a mean squared error (MSE) of $0.0054$, which further improves to $10^{-7}$ within just a few minutes of training. These results establish a strong foundation for advancing to 3D space charge simulations. Additionally, the potential of replacing traditional ellipsoidal methods with Gaussian envelope models for nonlinear space-charge calculations is explored, thereby potentially enhancing the accuracy of simulations. In parallel, polynomial neural networks are investigated alongside CNNs, aiming to accurately model both external and self-fields using simulation and measurement data, respectively.