TEMPEST: closure models for turbulence across scales
TEMPEST (Transformative Explorations in Multi-Physics and Engineering of Scientific Turbulence) is an NSF Science and Technology Center, led by Michigan State University, that launched in September
- It brings experiments, theory, computation, and machine learning to bear on turbulence in the settings where it is hardest to predict: fusion energy, astrophysics, and advanced propulsion.
Within TEMPEST, my group works on the AI and machine learning that connects the center’s theory, simulations, and experiments. Passing from particles to kinetic equations to fluids to turbulence requires a closure at every step, and each closure discards information. We work out what is discarded, when the reduced model is valid, and what a learned model is actually approximating. To that end we are building an evaluation harness for closure models: a fixed suite of test problems, a common interface for analytic, learned, and hybrid methods, and accuracy, stability, and cost metrics computed the same way for every candidate. The goal is models built on the physics rather than around it.
The benchmark design and code are developing in the open in the TEMPEST model bake-off repository. See also the Yale News story on the collaboration.
Students interested in scientific machine learning with real mathematical content (kinetic theory, scale limits, closure) are welcome to get in touch.