Calendar

Time interval: Events:

Monday, November 23, 2026

Posted November 5, 2026
Last modified November 13, 2026

Colloquium Questions or comments?

3:30 pm

Minglei Yang, Oak Ridge National Laboratory
Data-Driven Learning of Stochastic Dynamics with Applications to Plasma Physics

Stochastic dynamical systems arise in mathematical models of transport, diffusion, and particle motion. Their numerical simulation can be expensive, particularly when many trajectories are needed to resolve distributions, rare events, or boundary losses. This talk presents data-driven methods based on generative AI for efficient simulation, building on probabilistic numerical methods and motivated by problems in plasma physics.

I will first introduce probabilistic schemes for nonlocal transport and exit problems, together with convergence analysis and applications to heat transport and runaway electrons. Building on this foundation, I will discuss normalizing-flow and diffusion-based models that learn transition distributions from trajectory data. For fixed dynamics, these models can be reused across initial distributions and replace many small simulation steps with a learned sampling step. A hybrid construction accounts for both boundary exit and the motion of surviving particles. The discussion will emphasize the mathematical ideas behind these methods, convergence and error estimates for their key components, and numerical evidence of their accuracy and efficiency.

Finally, I will extend the generative AI approach from stochastic particle dynamics to high-dimensional, chaotic turbulence fields. A variational autoencoder coupled with learned latent dynamics provides a surrogate for plasma turbulence simulation, assessed through spatial structures, field correlations, and transport statistics.