Invariant data-driven subgrid stress modeling in the strain-rate eigenframe for large eddy simulation

We present a new approach for constructing data-driven subgrid stress models large eddy simulation of turbulent flows. The key to our is representation model input and output tensors in the filtered strain rate eigenframe. Provided inputs outputs are selected non-dimensionalized suitable manner, this yields form that symmetric, Galilean invariant, rotationally reflectionally unit invariant. use train simple efficient neural network using only one time step direct numerical data from forced homogeneous isotropic turbulence simulation. demonstrate accuracy as well model's ability generalize previously unseen filter widths, Reynolds numbers, flow physics priori posteriori tests.