CIESC Journal ›› 2022, Vol. 73 ›› Issue (12): 5483-5493.DOI: 10.11949/0438-1157.20221174

• Process system engineering • Previous Articles     Next Articles

Integrating physics-informed neural networks with partitioned coupling strategy for modeling conjugate heat transfer

Zhibin LU1,2(), Yimeng LI1,2, Chang HE1,2(), Bingjian ZHANG1,2, Qinglin CHEN1,2, Ming PAN3   

  1. 1.School of Materials Science and Engineering, Sun Yat-sen University, Guangzhou 510006, Guangdong, China
    2.Guangdong Engineering Center for Petrochemical Energy Conservation, Sun Yat-sen University, Guangzhou 510006, Guangdong, China
    3.Industrial Data Science and Technology (Guangzhou) Co. Ltd. , Guangzhou 510530, Guangdong, China
  • Received:2022-08-25 Revised:2022-10-07 Online:2023-01-17 Published:2022-12-05
  • Contact: Chang HE

集成分区耦合策略的物理信息神经网络模拟共轭传热过程研究

陆至彬1,2(), 李依梦1,2, 何畅1,2(), 张冰剑1,2, 陈清林1,2, 潘明3   

  1. 1.中山大学材料科学与工程学院,广东 广州 510006
    2.广东省石化过程节能工程技术研究中心,广东 广州 510006
    3.工数科技(广州)有限公司,广东 广州 510530
  • 通讯作者: 何畅
  • 作者简介:陆至彬(1998—),男,硕士,luzhb6@mail2.sysu.edu.cn
  • 基金资助:
    国家自然科学基金面上项目(22078372);广东省自然科学基金面上项目(2022A1515010479)

Abstract:

Physics-informed neural network (PINN) realizes deep learning with embedded physical knowledge by mathematically encoding partial differential equations, and has been successfully applied in the fields of fluid mechanics and heat transfer. However, due to the strong coupling of heat transfer in fluids and heat conduction in solids, regular PINN methods are difficult to effectively solve the conjugate heat transfer problems that commonly exist in the aforesaid fields. As a widely-used partitioned coupling strategy, the heat transfer coefficient forward temperature backward (hFTB) can feasibly deal with complex coupling relation in the interface by solving the fluid and solid domains separately. In this work, based on real physical property systems, a modeling strategy that combines partitioned coupling and PINN is proposed by using hFTB approach. Taking 2-D and 3-D conjugate heat transfer models as examples, the results of multi-physics fields obtained by the proposed strategy are compared with those by using conventional CFD simulation. The resulting maximum absolute errors of the solid temperature of the 2-D and 3-D models are only 0.19 K and 2.12 K, respectively, which reflects the effectiveness of the proposed strategy in modeling conjugate heat transfer under the real physical property systems.

Key words: physics-informed neural networks, partial differential equations, partitioned coupling, conjugate heat transfer, heat transfer coefficient forward temperature backward

摘要:

物理信息神经网络(PINN)通过对偏微分方程组进行数学编码,实现了内嵌物理知识的深度学习,已成功地应用于流体力学和传热领域。但是,由于固体导热和流体传热间强耦合关联,通用的PINN难以有效求解上述领域内普遍存在的共轭传热问题。作为应用较为广泛的分区耦合策略,传热系数正向温度反向法可通过分别独立求解流体域和固体域来灵活处理界面处的复杂耦合关系。本工作基于真实物性体系,利用传热系数正向温度反向法构建分区耦合PINN建模策略。以共轭传热二维和三维模型为例,将分区耦合PINN预测的多物理场结果与常规的CFD软件模拟结果进行对比,结果显示二维模型和三维模型的固体温度最大绝对误差分别为0.19 K和2.12 K,表明了分区耦合PINN策略处理真实物性下共轭传热建模问题的有效性。

关键词: 物理信息神经网络, 偏微分方程, 分区耦合, 共轭传热, 传热系数正向温度反向法

CLC Number: