Deep learning-based reduced-order model for turbulent flows

Reduced-order models (ROMs) are often developed to save computational cost of further analysis. These models perform well in laminar flow regime, however, their complexity increases for modeling turbulent flows. The dominating flow features are frequently extracted through Proper Orthogonal Decomposition (POD) in developing ROMs. High-dimensional snapshots are represented as low-dimensional subspaces in conventional ROMs employing optimal and orthogonal POD modes, followed by the Galerkin projection. In this article, we describe a data-driven model built on POD and recurrent neural network (RNN). We will evaluate the long-short-term memory (LSTM) network’s capacity to replicate the temporal POD basis coefficients. In particular, data are created for flow past a circular cylinder at Reynolds number 1000 using a parallel CFD code, and then LSTM in concert with POD is used to develop ROM. The LSTM network has a great ability to forecast turbulent flow’s statistical behaviour. The estimated coefficients are within the range of the actual DNS values, and it accurately reproduces the frequency content.