Sub-grid-scale Modeling using Convolutional Neural Network

Predicting the behavior of a passive scalar in a turbulent flow is an important task in heat transfer and fluid dynamics field. To reduce the tremendous cost of simulating using the traditional method of direct numerical simulation, we propose a neural network algorithm that can significantly reduce the cost of the simulation. The network trains the partial derivative operators of the coarse-grained partial differential equation in order to capture the effect of the subgrid scales. Using such network, we are currently investigating a network that predicts the passive scalar transport of the fully turbulent channel flow using the low resolution grid.