May 1 seminar: Alexander Scheinker

May 1, 2026

Adaptive Generative Diffusion Models for Virtual Diagnostics of Complex Time-Varying Systemes

Alexander Scheinker, LANL

3:00 pm, UNM Centennial Engineering Center, Room 1026
Online Guests: Contact Prof. Santhanam <bsanthan@unm.edu> for a Zoom link

Abstract: Diffusion models are state-of-the-art for creating accurate representations of high dimensional complex objects, including everything from high resolution images to 3D protein structures. One of the main limitations of generative tools, including diffusion-based models, is an inability to adapt in real-time for time-varying systems. Typically, data-driven models must be brute-force re-trained for systems that undergo large distribution shifts. Another limitation is a lack of hard physics constraints. This talk gives a brief overview of the mathematics behind generative diffusion models and presents recent results on incorporating physics constraints and adaptive feedback within diffusion architectures to make them more robust for complex time-varying dynamic systems. Several experimental applications are shown including particle accelerator beams and electrochemistry.

Bio: Dr. Alexander Scheinker is the Adaptive Machine Learning Team Leader at Los Alamos National Laboratory with master’s degrees in math and physics and a PhD in adaptive control theory. He is associate editor for Physical Review Accelerators and Beams on AI/ML/control, a research affiliate at Lawrence Berkeley National Laboratory, an IEEE Control Systems Society Conference Associate Editor, and a Senior Member of IEEE. Alexander’s work is focused on developing robust adaptive control algorithms for analytically unknown time-varying nonlinear dynamic systems. He developed a bounded form of extremum seeking that is able to stabilize nonlinear systems with unknown and time-varying control direction. Recently he has been incorporating adaptive feedback control within generative AI tools for time-varying dynamic systems as well as hard physics constraints. His algorithms for autonomous accelerator control have been used at laboratories around the world including DESY, SLAC, LBNL, LANL, and CERN.