AUTHOR=Yuan Yuchen , Song Ning , Nie Jie , Shi Xiaomeng , Chen Jingjian , Wen Qi , Wei Zhiqiang TITLE=PHI-SMFE: spatial multi-scale feature extract neural network based on physical heterogeneous interaction for solving passive scalar advection in a 2-D unsteady flow JOURNAL=Frontiers in Marine Science VOLUME=Volume 10 - 2023 YEAR=2023 URL=https://www.frontiersin.org/journals/marine-science/articles/10.3389/fmars.2023.1276869 DOI=10.3389/fmars.2023.1276869 ISSN=2296-7745 ABSTRACT=Fluid dynamic calculations have a profound impact on marine biochemical dynamic processes, influencing the behavior, interactions, and distribution of biochemical components within aquatic environments. The high precision numerical simulation of fluid dynamics poses a significant challenge due to the inherent complexity of fluid motion in real-world scenarios. Traditional numerical simulation methods typically achieve higher accuracy by increasing the resolution of the computational grid. However, this increase in grid resolution also introduces a higher computational demand. Recently, the advancement in machine learning has paved the way for overcoming aforementioned bottleneck issue by harnessing the power of deep learning techniques in numerical simulation. Numerical simulation methods that incorporate deep learning leverage discretized learned coefficients to attain high-precision solutions on low-resolution grids, which means that these methods can effectively alleviate the computational burden while maintaining simulation accuracy. Nevertheless, current fluid numerical simulation methods based on deep learning are constrained by their reliance on a single scale analysis of spatially correlated physical fields. This limitation prevents them from effectively capturing the diverse scale characteristics inherent in flow fields governed by complex motion laws in different physical spaces. Furthermore, different dynamic fields within the same system are subject to physical interactions, yet existing models lack an effective approach to enhance the correlation interactions among these dynamic fields. To tackle these challenges, we present spatial multi-scale featrue extract neural network based on physical heterogeneous interaction (PHI-SMFE). Specifically, the PHI module is designed to extract heterogeneity and interaction information from diverse dynamic fields. In addition, we propose the SMFE module that focuses on capturing multi-scale features in fluid dynamic fields. Notably, we utilize channel-biased convolution based on Partial Convolution to implement a separation strategy, effectively reducing the processing of redundant feature information. Compared to the current state-of-the-art model, our method exhibits a 41% increase in simulation accuracy and a 12.7% decrease in inference time during the iterative evolution of the unsteady flow field. These results demonstrate that our proposed model achieves SOTA-level performance in terms of both simulation accuracy and computational speedup.