CSE Community Seminar | March 20, 2026

Abstract

Numerical simulations of turbulent flows are relied upon in the aerospace, nuclear, automotive, chemical, wind, and architectural industries, as well as in weather prediction, climate modelling, and geophysical fluid dynamics. One particular source of error plagues all of these domains: turbulence modelling. We need turbulence models because it is prohibitively expensive to resolve all scales of turbulence for practical applications, and it will be for the foreseeable future. We have decades of experience developing heuristic-based, domain-specific turbulence models. However, the rise of machine learning has opened new possibilities for learning models directly from data. 

In this talk, I’ll provide an overview of the field and discuss some specific research directions I’m working on. I’ll cover: What is turbulence modelling? How can machine learning be used to improve turbulence models? How does it differ for RANS and LES? I’ll discuss the major issues facing data-driven turbulence models, such as generalizability, numerical stability, and over-optimism. I’ll focus on how we can incorporate physics-based biases, like equivariance, stochasticity, and physical realizability into data-driven turbulence models. Success in this area means more reliable predictions across all the domains that depend on turbulence modelling.