LKFlowNet: A deep neural network based on large kernel convolution for fast and accurate nonlinear fluid-changing prediction

Abstract

The rapid development of artificial intelligence has promoted the emergence of new flow field prediction methods. These methods address challenges posed by nonlinear problems and significantly reduce computational time and cost compared to traditional numerical simulations. However, they often struggle to capture the dynamic sparse characteristics of the flow field effectively. To bridge this gap, we introduce LKFlowNet, a new large kernel convolutional neural network specifically designed for complex flow fields in nonlinear fluid dynamics systems. LKFlowNet adopts a multi-branch large kernel convolution computing architecture, which can skillfully handle the complex nonlinear dynamic characteristics of flow changes. Drawing inspiration from the dilated convolution mechanism, we developed the RepDWConv block, a re-parameterized depthwise convolution that extends the convolutional kernel's coverage. This enhancement improves the model's ability to capture long-range dependencies and sparse structural features in fluid dynamics. Additionally, a customized physical loss function ensures accuracy and physical consistency in flow field reconstruction. Comparative studies reveal that LKFlowNet significantly outperforms existing neural network architectures, providing more accurate and physically consistent predictions in complex nonlinear variations such as velocity and pressure fields. The model demonstrates strong versatility and scalability, accurately predicting the flow field of various geometric configurations without modifying the architecture. This capability positions LKFlowNet as a promising new direction in fluid dynamics research, potentially revolutionizing flow field prediction by combining high efficiency and accuracy. Our results suggest that LKFlowNet could become an indispensable tool in intelligent flow field prediction, reshaping the analysis and processing of fluid dynamics.