Development of Deep Learning Methods for Inflow Turbulence Generation

Abstract: The present work proposes an inflow turbulence generation strategy using deep learning methods. This is achieved with the help of an autoencoder architecture with two different types of operational layers in the latent-space: a fully connected multi-layer perceptron and convolutional long short-term memory layers. A wall-resolved large eddy simulation of a turbulent channel flow at a high Reynolds number is performed to create a large database of instantaneous snapshots of turbulent flow-fields used to train neural networks. The training is performed with sequences of instantaneous snapshots so as to learn a snapshot at the next time instant. For the present work, velocity and pressure fluctuations are used for training to produce the same parameters at the next time instant. Within the autoencoders, different types of methods like average pooling and strided convolutions are tested for the reduction of spatial dimensions. For convolutional neural networks, the physical boundary conditions are implemented in the form of symmetric as well as periodic paddings. A priori simulations are performed with the trained deep learning models to check the accuracy of turbulence statistics produced. It is found that the use of convolutional long short-term memory layers in the latent space improved the quality of statistics, although issues related to stability for longer times are observed. Though instantaneous snapshots of the target flow are required for training, these a priori simulations suggest that the deep learning methods for generating inflow turbulence are one of the possible alternatives to existing methods.