Adaptively Guided Latent Diffusion for Time-Varying Inverse Problems in Particle Accelerators

Diffusion models have emerged as the state-of-the-art tools for generating accurate high-resolution representations of complex objects. One outstanding challenge faced by generative models, including diffusion models, is that of quickly time-varying systems with large distributions shifts. This work presents a general adaptive latent diffusion approach to solving extreme inverse problems for time-varying systems by building adaptive feedback control theory methods into generative latent diffusion models. The general approach is demonstrated for mapping vectors of non-invasive 1D beam loss or beam current monitor measurements to detailed views of projections of a charged particle beam's 6D phase space distribution. The results are demonstrated with multiparticle physics codes that simulate the evolution of intense time-varying proton beams in the kilometer-long Los Alamos Neutron Science Center (LANSCE) linear particle accelerator.