Predictions for Unsteady Flow Fields With Deep Learning Models of LSTM and GRU

Abstract

Deep learning methods based on time series prediction are widely applied to solving fluid dynamics problems. However, the application characteristics of Long Short-Term Memory (LSTM) and Gate Recurrent Unit (GRU) in fluid mechanics are lacking in the existing research. Therefore, this paper proposes two flow field prediction models combining the Proper Orthogonal Decomposition (POD) method, referred to as POD-LSTM and POD-GRU. Their performance in flow field prediction is compared with numerical simulation data and experimental data respectively. The numerical simulation was obtained by using OpenFOAM-v9, and the experimental data were obtained from Particle Image Velocimetry (PIV). The results show that LSTM performs excellently in predicting velocity field, while GRU excels in predicting pressure field. On the other hand, under the same computational resources and cost, the training time required by GRU is shorter than that of LSTM. Furthermore, as the number of POD modes increases, the overall deviation of LSTM network predictions from actual flow field values tends to increase. In contrast, GRU networks show the opposite trend. This study provides a detailed description of the performance of POD-LSTM and POD-GRU in fluid mechanics. The proposed prediction models demonstrate strong generalization capabilities, further exploring the potential of deep learning methods in fluid mechanics.

Keywords

• Deep learning (DL)
• gate recurrent unit (GRU)
• long short-term memory (LSTM)
• proper orthogonal decomposition (POD)
• unsteady flow