Reconstructive AI Spectroscopy of Charged Particle Beams

Abstract

This work introduces a physics-informed neural network (PINN) framework for reconstructing electron energy spectra from sparsely sampled attenuation-curve data. Leveraging the NVIDIA PhysicsNeMo platform, the proposed mesh-free methodology operates directly on raw experimental datasets while explicitly incorporating all experimental uncertainties. Validation on subnanosecond electron beam measurements demonstrates that the approach accurately resolves complex, multi-peaked spectral features of energy distribution. The framework enforces physical consistency through embedded governing principles and exhibits substantial predictive capability for energy spectrum reconstruction from noisy, low-precision, and sparse experimental data.