Modeling of sub-grid scale eddies using machine learning

There are two main computer models for investigating turbulent flows, the most accurate of which is called Dynamic Numerical Simulation (DNS). This model resolves all scales of eddies directly, however, it requires the large number of grids and spends much time. The other model as an alternative to the former one is Large Eddy Simulation (LES) which resolve the larger-scale eddies directly and requires artificial modeling for the remaining smaller-scale eddies. Therefore, the results of the LES model can vary depending on the sub-grid model for modeling smaller-scale eddies. To make the performance of the LES model as similar as possible to that of the DNS, the correlation between large and small eddies is found using deep learning.

This study uses two representative deep learning models, Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN), which is based on the ANN algorithm. RNN has been used as a standard method for machine learning of sequence data (audio, natural language, etc.) (Yin et al., 2019). So, RNN needs a good memory which remembers all input in internal memory to forecast the future events by reminding previous data. While, CNN is a powerful tool to extract fluid dynamics features and predict flow fields. Strictly speaking, CNN recognizes geometrical topologies of jet flow. In this study, the flow structure characteristics are investigated using CNN, and the flow over time is examined using RNN.