Using Convolutional Neural Networks and Latent Conditional Diffusion to Speed Up 3D Coherent Synchrotron Radiation Simulations

An electron bunch undergoing circular motion emits synchrotron radiation. For the part of the spectrum where wavelengths exceed the bunch length, the radiation becomes coherent and ends up significantly distorting the phase space distribution of the electrons. Computing the effects of coherent synchrotron radiation (CSR) is one of the most computationally intensive tasks in accelerator physics. Here, we demonstrate how convolutional neural networks and latent conditional diffusion models trained on physics-based 3D simulations can be used to speed up the calculations of the electromagnetic wakefields in CSR. We get speed up factors of up to 1000 with errors of the order of a percent. The models generalize well for distributions never encountered before when they have spreads equal or greater than what they were trained on, but fail for distributions that are more compact.