ML-Driven Phase Space Tomography at LANSCE

Accurate knowledge of a beam’s phase space distribution is highly valuable for optimizing accelerator performance, diagnosing operational issues, and ensuring proper alignment and calibration. However, direct measurement of the full phase space is often challenging. In this work, we present a machine learning (ML)-based method for reconstructing the longitudinal phase space distribution from two-dimensional (2D) phase space scans at the Los Alamos Neutron Science Center (LANSCE).   Convolutional Neural Networks (CNNs) and U-Net architectures were trained on simulated datasets augmented with experimentally realistic noise. The models demonstrated robust learning and generalization from simulation to experiment. Using explainable ML techniques, we identified key regions of the 2D phase scans that contribute most to the reconstruction, enabling a reduction in required data acquisition to approximately one-third of the full scan. The models tested on experimental results give physically reasonable reconstructions that match those obtained from calibrating simulation. Additionally, we explored a super-resolution approach, achieving high-fidelity phase space reconstructions with significantly enhanced resolution relative to the input 2D scans.