MULTI-OUTPUT PHYSICS-INFORMED NEURAL NETWORKS MODEL BASED ON THE RUNGE-KUTTA METHOD

被引：0
|
作者：
Wei C. [1 ]
Fan Y. [1 ]
Zhou Y. [1 ]
Liu X. [1 ,2 ]
Zhang C. [2 ]
Wang H. [1 ]
机构：
[1] School of Mechanical Engineering, Tianjin University, Tianjin
[2] Yantai Longyuan Power Technology Co., Ltd., Shandong, Yantai
来源：
Lixue Xuebao/Chinese Journal of Theoretical and Applied Mechanics | 2023年 / 55卷 / 10期
关键词：
flow field prediction; multi-output neural network; parameter identification; Runge-Kutta method;
D O I：
10.6052/0459-1879-23-299
中图分类号：
学科分类号：
摘要：
Physics-informed neural networks (PINN) have attracted considerable attention in the field of intelligent scientific computing primarily due to their capacity to incorporate prior knowledge of physics. This outstanding integration allows PINNs to automatically satisfy physical constraints even with limited or zero labeled data. As a result, the applicability and effectiveness of PINN are vastly developed across numerous domains. However, it is worth noting that the discrete time models of PINN, also known as PINN-RK, face a significant limitation in their ability to approximate multiple physical quantities and solve coupled partial differential equation systems simultaneously. This shortcoming hinders its ability to handle complex multi-physics fields, which is a crucial drawback in various practical scenarios. To overcome this limitation, a multi-output physics-informed neural network based on Runge-Kutta method (MO-PINN-RK) is proposed in this paper.MO-PINN-RK, building upon the success of PINN-RK, incorporates a sophisticated parallel neural network architecture, boasting multiple output layers for enhanced performance and accuracy. By associating each output layer with a sub-network and assigning it with different physical quantities, MO-PINN-RK can accurately solve the coupled partial differential equation system and predict multiple physical quantities simultaneously. The MO-PINN-RK proposed in this paper overcomes the limitation of PINN-RK that is only applicable to one dimensional problems extending its applicability to more general multi-dimensional problems. To demonstrate the effectiveness of MO-PINN-RK, it is then applied to the flow field prediction and parameter identification of flow around a cylinder. The outcomes unequivocally reveal that MO-PINN-RK surpasses PINN in terms of flow field prediction precision, achieving an enhancement of no less than 2 times. At the same time, MO-PINN-RK reduces the relative error by an order of magnitude in the context of parameter identification. This highlights the exceptional capabilities of MO-PINN-RK in the field of fluid dynamics, offering a more accurate and efficient solution for solving complex problems. © 2023 Chinese Journal of Theoretical and Applied Mechanics Press. All rights reserved.
引用
收藏
页码：2405 / 2416
页数：11
相关论文
共 31 条
  • [11] Raissi M, Perdikaris P, Karniadakis GE., Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations, Journal of Computational Physics, 378, pp. 686-707, (2019)
  • [12] Hornik K, Stinchcombe M, White H., Multilayer feedforward networks are universal approximators, Neural Networks, 2, 5, pp. 359-366, (1989)
  • [13] Dissanayake M, Phan-Thien N., Neural-network-based approximations for solving partial differential equations, Communications in Numerical Methods in Engineering, 10, 3, pp. 195-201, (1994)
  • [14] Baydin AG, Pearlmutter BA, Radul AA, Et al., Automatic differentiation in machine learning: a survey, Journal of Marchine Learning Research, 18, pp. 1-43, (2018)
  • [15] Cuomo S, DI Cola VS, Giampaolo F, Et al., Scientific machine learning through physics –informed neural networks: where we are and what’s next, Journal of Scientific Computing, 92, 3, (2022)
  • [16] Yu J, Lu L, Meng X, Et al., Gradient-enhanced physics-informed neural networks for forward and inverse PDE problems, Computer Methods in Applied Mechanics and Engineering, 393, (2022)
  • [17] Wang S, Teng Y, Perdikaris P., Understanding and Mitigating Gradient Flow Pathologies in Physics-Informed Neural Networks, SIAM Journal on Scientific Computing, 43, 5, pp. A3055-A3081, (2021)
  • [18] Jagtap AD, Kawaguchi K, Karniadakis GE., Adaptive activation functions accelerate convergence in deep and physics-informed neural networks, Journal of Computational Physics, 404, (2020)
  • [19] Mao Z, Jagtap AD, Karniadakis GE., Physics-informed neural networks for high-speed flows, Computer Methods in Applied Mechanics and Engineering, 360, (2020)
  • [20] Meng X, Li Z, Zhang D, Et al., PPINN: Parareal physics-informed neural network for time-dependent PDEs, Computer Methods in Applied Mechanics and Engineering, 370, (2020)
1234