Machine Learning of Turbulence Closures for the Simulation of the Dynamics of Transitions Between Metastable States

Dee Howard Lecture Series
Sponsored by the Wayne and Julie Fagan Endowed Fellowship
Presented on: 04 May 2026
Time and Venue: 1:00 pm BSE 2.102 -UT San Antonio

Machine Learning of Turbulence Closures for the Simulation of the Dynamics of Transitions Between Metastable States

Dr. Bisetti will present an overview of and select technical results from an ongoing DARPA project that seeks to develop and demonstrate a novel approach that discovers stochastic sub-grid scale (SGS) models for large-eddy simulation (LES) of meta-stable state transitions in a classical-fluid turbulent flow, using a newly proposed scientific multi-agent reinforcement learning framework. Our approach is unique in that it uses laboratory data from a compact tabletop turbulence simulator that is tunable over 4 orders of magnitude in the controlling parameters, e.g., the Reynolds number, and is characterized via non-intrusive optical measurements, including of multi-point/multi-time statistics of velocity.

Decades-long pursuits of models for multiscale systems such as turbulent flows have received a new boost from machine learning (ML). In this project, we aim to improve SGS models towards the simulation of turbulent flows that display switching between asymmetric meta-stable states and that are relevant to vehicles, ships, and propulsion systems of interest to the U.S. Department of Defense, e.g., wakes and separated & recirculating flows. Because those meta-stable states differ in topology, prediction of kinetic energy, mixing, and drag requires accurate simulation of transition frequency and state features. Transitions occur with relatively low frequency and are affected by small and large scales, which are distributed broadly in turbulent flows, necessitating closure approaches to model unresolved spatiotemporal scales and overcoming key challenges, i.e., a high Reynolds number and coherent broadband dynamics.