Inferring subsurface flow structures from surface manifestation in free-surface flows using neural network

For free-surface flows, subsurface turbulent structures can create distinguished and complex manifestations on the free surface. We have built a neural network based framework to infer the subsurface turbulent motions from the observations of the surface elevation and velocity fluctuations. Through a series of convolutional layers, the neural network extracts features from the surface measurements and then reconstructs the three-dimensional subsurface velocity field. As an example, we consider a turbulent open channel flow. We train the neural network using the data obtained from direct numerical simulations, which use a free-surface boundary-fitted grid to resolve the surface motions accurately. It is found that the neural network can infer certain turbulent coherent structures, such as the inclined vortices arising from the bottom boundary layer, and reveal their correlations with the surface manifestations.