Using Machine Learning for Improving a Non-Linear k-eps Model: A First Attempt

Övrigt - rapport, 2023

Machine Learning (ML) is used for improving a non-linear $\ke$ model. A developing boundary-layer flow is used to train train the model. In boundary layer flow the model includes only two independent coefficients which are taken as output in a Neural Network (NN) model. Different inputs are evaluated including the velocity gradient and the square of the velocity gradient. They were scaled with either
the friction velocity and the viscosity or the turbulence kinetic energy and its dissipation.

The input variables are computed using DNS of developing boundary layer at Re_theta=8180. The NN model is created using Python's pytorch. It turns out that the NN model acts as a low-Re model, replacing all low-Re modifications which are present in non-ML non-linear eddy-viscosity turbulence models.

After having trained the NN model in the boundary-layer flow, it is validated in two channel flows at Re_tau = 550
and Re_tau = 5200$. Good agreement is obtained.