Data-driven A priori analysis of sub-grid scale stress closures

Better affordability of computational resources over the past decade has led to
increased use of scale-resolving simulation techniques such as Large-Eddy Simulations
(LES) by industries. These techniques involve the use of a model to account for the
effect of small-scale unresolved turbulence structures on large scale resolved
structures. The most commonly used models are often based on simplified assumptions
such as gradient viscosity hypothesis and turbulence quasi-equilibrium and often fail to
replicate higher-order statistics.

This talk will discuss the use of a data-driven approach i.e. neural nets for determining a
functional mapping that governs the physics behind the problem. We present an A priori
study conducted on the homogeneous isotropic turbulence to determine a model for
modeling sub-grid scale stress tensor and compare it with common gradient-viscosity
hypothesis based models. A comparison of higher-order statistics will shed light on the
universality of small-scale turbulent flow structures as hypothesized by Kolmogorov
1941 theory. The importance of selecting correct input features and the design of the
model form for obtaining a physically-consistent relation will also be discussed.