Dr. Alexander Scheinker
Los Alamos National Laboratory
Lectures
Building Adaptive Feedback and Physics Constraints into Generative Diffusion Models
In generative AI, diffusion models are state-of-the-art for high resolution representations of highly diverse complex objects. This work presents an adaptive multimodal conditional diffusion approach which utilizes various modalities of data to generate accurate representations of complex systems. The approach is demonstrated for generating high resolution virtual views of the 6D phase space of intense charged particle beams in particle accelerators. The talk presents several modalities of using generative diffusion models as virtual diagnostics including directly in image space, latent diffusion, and super-resolution diffusion. This talk also discusses how this general diffusion approach can be combined with transformers and with LLMs for a wide range of science and engineering applications.
About
Dr. Alexander Scheinker is the adaptive machine learning team leader in the applied electrodynamics group (AOT-AE) at Los Alamos National Laboratory. He has master’s degrees in math and physics and a PhD in dynamic systems and control theory. At LANL, his adaptive ML team is developing novel adaptive and physics-constrained generative AI tools, such as generative diffusion models, that utilize adaptive feedback control theory and hard physics constraints to develop safe and robust generative AI tools for automatic tuning and control of complex time-varying systems such as high energy particle accelerators and for 3D dynamic imaging