A data driven reduced order model of fluid flow by Auto-Encoder and self-attention deep learning methods

Abstract

This paper presents a new data-driven non-intrusive reduced-order model(NIROM) that outperforms the traditional Proper orthogonal decomposition (POD) based reducedorder model. This is achieved by using Auto-Encoder(AE) and attention-based deep learning methods. The novelty of the present work lies in that it uses Stacked AutoEncoder(SAE) network to project original high-dimensional dynamical systems onto a low dimensional nonlinear subspace and predict the fluid dynamics using an attentionbased deep learning method. A new model reduction neural network architecture for fluid flow problem is presented. The SAE network compresses high-dimensional physical information into several much smaller sized representations in a latent space. These representations are expressed by a number of codes in the middle layer of SAE neural network. Then, those codes at different time levels are trained to construct a set of hyper-surfaces with multi variable response functions using attention-based deep learning methods. The inputs of the attention-based network are previous time levels' codes and the outputs of the network are current time levels' codes. The codes at current time level are then projected back to the original full space by the decoder layers in the SAE network. The capability of this data-driven reduced order model (DDROM) is illustrated numerically by two test cases: flow past a cylinder, and a lock exchange, respectively. The results obtained show that the DDROM performs better in terms of accuracy than the popular model reduction method namely proper orthogonal decomposition. The improvement is assessed by comparison with a high fidelity POD model.