Turbulent flows contain a wide range of interacting spatial and temporal scales, making them difficult to model directly. Large Eddy Simulation (LES) resolves the large motions while modeling the influence of unresolved scales through a subgrid‑scale (SGS) model. Traditional physics‑based SGS models capture forward energy dissipation but fail to represent backscatter, the reverse transfer of energy from small to large scales, which is essential for accurate turbulence dynamics.
Recent work has shown that machine learning can overcome this limitation. A convolutional neural network (CNN) trained on high‑fidelity Direct Numerical Simulation data can learn both forward and backward energy transfer and remain stable when deployed inside an LES solver. However, despite its strong performance, the CNN provides little interpretability: its internal filters do not correspond to identifiable physical structures, leaving the learned turbulence mechanisms opaque.
This project focuses on improving interpretability through model distillation. We train a Gabor Neural Network (GaborNN), whose sinusoidal, frequency‑localized filters align naturally with turbulence scales, to reproduce the CNN's SGS predictions. To enable efficient training, we reformulated the learning problem from predicting π using vorticity and stream function to predicting the SGS stress τ from velocity components, allowing effective hyperparameter optimization with RayTune. This approach aims to combine the accuracy of the CNN with the transparency of the GaborNN, moving toward a physically interpretable SGS model for LES.