Some Recent Research Highlights

This is a summary of some recent research on the intersection of control theory and generative AI.

Latent Diffusion for Extreme Inverse Problems

A high level overview of the conditional diffusion-based beam diagnostic for the LANSCE accelerator at Los Alamos National Laboratory. In this case an adaptive conditionally guided diffusion model is solving an extreme inverse problem of mapping 1D beam loss and beam current measurements to high resolution detailed views of 2D projections of a charged particle beam's 6D phase space density. By adaptively tuning the conditional vector we are able to track time varying beam distributions based only on limited measurements.

Super-resolution diffusion for 6D phase space diagnostics

A high level overview of a conditionally guided adaptive diffusion process for tracking the time-varying 6D phase space density of a charged particle beam in the HiRES compact electron accelerator at Lawrence Berkeley National Laboratory. A VAE is first used to compress a combination of initial beam condition image-based measurements together with accelerator parameters into a low-dimensional latent embedding $\mathbf{z}\in\mathbb{R}^3$, from which the generative half of the VAE then creates a 6D tensor of size $32^6$ which represents the beam's 6D phase space density. Individual $32\times 32$ pixel 2D projections are then created in a physically consistent way by projecting from this single 6D tensor. The resolution of these images is then increased to $256\times 256$ so that fine beam phase space density perturbations can be seen by a super-resolution diffusion model. The entire process is made to be adaptive, to track time-varying beams based only on limited measurements, by adaptively tuning the latent space embedding using adaptive feedback control theory techniques.

A detailed view of the super resolution diffusion process.