Speaker
Description
Laser–plasma wakefield accelerators (LWFAs) are promising sources of high-energy electron beams, but optimizing their performance in terms of beam quality is a challenging task. We propose a novel reverse beam engineering approach using two modular neural network surrogates that can operate independently or in tandem. The first network is a surrogate model of a capillary discharge plasma source, trained on COMSOL Multiphysics hydrodynamic simulations, predicting the electron density profile given gas pressure, applied voltage, and capillary geometry. The second network is a surrogate for the LWFA process, trained on Particle-in-Cell (PIC) simulations, predicting electron beam parameters (energy spectrum, emittance, charge per bunch and bunch duration) given laser pulse properties and plasma density profiles. Together, these surrogates enable both forward predictions and inverse design, and the combined architecture can be linked to Monte Carlo (MC) simulations in Geant4 via ONNX to extend predictions toward dose and radiobiological endpoints.