·

PINN综述


1 历史渊源

将神经网络用于求解微分方程的想法可追溯至上世纪 90 年代。[1] 采用浅层网络配合数值梯度近似 PDE,使用配点法和拟牛顿优化求解。[2] 提出“试探函数 + 神经网络参数化”的框架,研究了规则与不规则边界条件下的边值问题。然而,受限于当时的硬件条件和优化算法,这些早期工作并未引起广泛关注。

2 PINN 的核心技术与演进

物理信息神经网络(Physics-Informed Neural Networks, PINN)是一种将物理定律编码到神经网络中的科学机器学习方法。其核心思想是将求解微分方程转化为优化问题:训练神经网络逼近方程的解,损失函数由两部分构成——对已知数据(初始/边界条件)的拟合项,以及在时空域配点上强制满足物理方程的残差项。

现代 PINN 框架由 [3] 于 2017—2019 年正式确立。其成功的关键在于自动微分(Automatic Differentiation, AD)技术的引入,使得复杂 PDE 残差能够高效地集成到损失函数中,从而极大推动了该领域的发展。

同一时期,其他学者也提出了基于神经网络万能逼近性质求解 PDE 的方法。[4] 提出 Deep Ritz 方法,通过最小化能量泛函求解边值问题;[5] 提出 DGM 方法,采用 Galerkin 式残差最小化。这些方法——PINN(配点式)、Deep Ritz(能量式)、DGM(Galerkin 式)——共同构成了“科学机器学习”方法家族。

PINN 的发展并非单一模型的线性演进,而是一棵不断分枝的“家族树”。其演化主要沿两个方向展开:物理信息的融入方式神经网络架构的选择

2.1 基于物理信息融入方式的变体

标准 PINN 通过损失函数中的惩罚项“软约束”边界与初始条件,这需要精心调节各项权重。针对这一问题,研究者提出了不同的物理信息融入策略。

硬约束方法直接在网络结构中编码边界条件。PCNN [6] [7] [8] 通过特殊设计的网络结构,使输出在数学上自动满足边界/初始条件,损失函数仅包含物理残差项。这避免了权重调节的难题,但需要针对不同边界条件设计相应的网络结构。

变分形式方法从另一角度改进损失函数的构造。[9] 提出 hp-VPINN,结合域分解(h-refinement)与子域内近似阶数调整(p-refinement)。该方法不直接最小化 PDE 残差的点误差,而是最小化其变分形式或能量泛函。通过分部积分,可降低对解的导数阶数要求(如二阶 PDE 的损失函数仅需一阶导数),从而提升训练稳定性。

守恒型方法专注于保持物理守恒性质。[10] 提出 cPINN(Conservative PINN),将计算域分解为多个子域,每个子域由独立的 PINN 求解,并在子域交界面强制通量连续性,确保整体解在离散意义上守恒。这种分解策略也天然支持并行计算。

2.2 基于神经网络架构的变体

[3] 最初采用多层感知机(MLP)作为 PINN 的基础架构。随着研究深入,学者们探索了多种神经网络结构以适应不同类型的问题。

2.2.1 前馈网络变体

在保持前馈网络(Feed-forward Neural Network)结构的前提下,研究者尝试了不同的网络设计。[11] 将 PINN 与极限学习机(ELM)结合,提出 PI-ELM。ELM 的输入层权重随机初始化后固定不变,仅训练输出层权重,因此收敛速度极快。[12] 提出 SPINN,采用径向基函数(RBF)等核函数代替部分全连接层,在减少参数量的同时提供一定的可解释性。

算子学习代表了另一个重要方向——不再学习单个 PDE 的解,而是学习从输入函数到输出函数的映射。[13] 提出 DeepONet,由 Branch Net(编码输入函数)和 Trunk Net(编码输出坐标)两个子网络构成。[14] 将其扩展为 DeepM&Mnet,用于处理多物理场与多尺度耦合问题。

2.2.2 卷积与循环架构

针对具有特殊结构的问题,研究者引入了 CNN 和 RNN 架构。[15] 提出 PhyGeoNet,通过坐标变换将不规则物理域映射到规则计算域,从而发挥 CNN 在结构化网格上的优势。[16] 提出 TgAE,采用编码器-解码器结构对高维随机场进行非线性降维,同时融入物理约束。

对于动态系统,RNN 的序列处理能力尤为适用。[17] 设计了物理信息 RNN,其单元结构直接模仿数值积分格式(如欧拉法、龙格-库塔法),将物理模型硬编码到网络中。[18] 提出 PhyLSTM,使用多个 LSTM 分别建模状态变量与非线性恢复力,通过张量微分器实现物理耦合。

2.2.3 概率与集成架构

为实现不确定性量化,[19] 提出 B-PINN,采用贝叶斯神经网络(BNN)作为骨干。网络权重被建模为概率分布,通过后验采样(如 HMC 或变分推断)不仅给出预测值,还能量化预测的不确定性。

生成模型也被引入 PINN 框架。[20] 提出 PI-GAN,利用 GAN 框架求解随机微分方程:生成器产生符合物理定律的解场,判别器评估其真实性,物理约束通过自动微分嵌入生成器。[21] 提出 GatedPINN,采用专家混合(Mixture-of-Experts)架构,由门控网络动态选择或组合多个专家 PINN,实现对复杂问题域的自适应处理。

Bibliography

  • [1] M. G. Dissanayake and N. Phan-Thien, “Neural-network-based approximations for solving partial differential equations,” communications in Numerical Methods in Engineering, vol. 10, no. 3, pp. 195–201, 1994, doi: 10.1002/cnm.1640100303.
  • [2] I. Lagaris, A. Likas, and D. Fotiadis, “Artificial neural networks for solving ordinary and partial differential equations,” IEEE Transactions on Neural Networks, vol. 9, no. 5, pp. 987–1000, 1998, doi: 10.1109/72.712178.
  • [3] M. Raissi, P. Perdikaris, and G. Karniadakis, “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,” Journal of Computational Physics, vol. 378, pp. 686–707, 2019, doi: 10.1016/j.jcp.2018.10.045.
  • [4] W. E and B. Yu, “The Deep Ritz Method: A Deep Learning-Based Numerical Algorithm for Solving Variational Problems,” Communications in Mathematics and Statistics, vol. 6, no. 1, pp. 1–12, 2018, doi: 10.1007/s40304-018-0127-z.
  • [5] J. Sirignano and K. Spiliopoulos, “DGM: A deep learning algorithm for solving partial differential equations,” Journal of Computational Physics, vol. 375, pp. 1339–1364, 2018, doi: 10.1016/j.jcp.2018.08.029.
  • [6] Y. Zhu, N. Zabaras, P.-S. Koutsourelakis, and P. Perdikaris, “Physics-Constrained Deep Learning for High-dimensional Surrogate Modeling and Uncertainty Quantification without Labeled Data,” Journal of Computational Physics, vol. 394, pp. 56–81, 2019, doi: 10.1016/j.jcp.2019.05.024.
  • [7] L. Sun, H. Gao, S. Pan, and J.-X. Wang, “Surrogate modeling for fluid flows based on physics-constrained deep learning without simulation data,” Computer Methods in Applied Mechanics and Engineering, vol. 361, 2019, doi: 10.1016/j.cma.2019.112732.
  • [8] D. Liu and Y. Wang, “A Dual-Dimer method for training physics-constrained neural networks with minimax architecture,” Neural Networks, vol. 136, pp. 112–125, 2021, doi: 10.1016/j.neunet.2020.12.028.
  • [9] E. Kharazmi, Z. Zhang, and G. E. Karniadakis, “hp-VPINNs: Variational Physics-Informed Neural Networks With Domain Decomposition,” Computer Methods in Applied Mechanics and Engineering, 2020, doi: 10.1016/j.cma.2020.113547.
  • [10] A. D. Jagtap, E. Kharazmi, and G. E. Karniadakis, “Conservative physics-informed neural networks on discrete domains for conservation laws: Applications to forward and inverse problems,” Computer Methods in Applied Mechanics and Engineering, vol. 365, 2020, doi: 10.1016/j.cma.2020.113028.
  • [11] V. Dwivedi and B. Srinivasan, “Physics Informed Extreme Learning Machine (PIELM) - A rapid method for the numerical solution of partial differential equations,” Neurocomputing, pp. 96–118, 2019, doi: 10.1016/j.neucom.2019.12.099.
  • [12] A. A. Ramabathiran and P. Ramachandran, “SPINN: Sparse, Physics-based, and partially Interpretable Neural Networks for PDEs,” Journal of Computational Physics, vol. 445, 2021, doi: 10.1016/j.jcp.2021.110600.
  • [13] L. Lu et al., “Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators,” Nature Machine Intelligence, vol. 3, no. 3, pp. 218–229, 2021, doi: 10.1038/s42256-021-00302-5.
  • [14] S. Cai, Z. Wang, L. Lu, T. A. Zaki, and G. E. Karniadakis, “DeepM&Mnet: Inferring the electroconvection multiphysics fields based on operator approximation by neural networks,” Journal of Computational Physics, vol. 436, 2021, doi: 10.1016/j.jcp.2021.110296.
  • [15] H. Gao, L. Sun, and J.-X. Wang, “PhyGeoNet: Physics-informed geometry-adaptive convolutional neural networks for solving parameterized steady-state PDEs on irregular domain,” Journal of Computational Physics, vol. 428, 2020, doi: 10.1016/j.jcp.2020.110079.
  • [16] N. Wang, H. Chang, and D. Zhang, “Theory-guided Auto-Encoder for surrogate construction and inverse modeling,” Computer Methods in Applied Mechanics and Engineering, vol. 385, 2021, doi: 10.1016/j.cma.2021.114037.
  • [17] F. A. Viana, R. G. Nascimento, A. Dourado, and Y. A. Yucesan, “Estimating model inadequacy in ordinary differential equations with physics-informed neural networks,” Computers & Structures, vol. 245, 2021, doi: 10.1016/j.compstruc.2020.106458.
  • [18] R. Zhang, Y. Liu, and H. Sun, “Physics-informed multi-LSTM networks for metamodeling of nonlinear structures,” Computer Methods in Applied Mechanics and Engineering, vol. 369, 2020, doi: 10.1016/j.cma.2020.113226.
  • [19] L. Yang, X. Meng, and G. E. Karniadakis, “B-PINNs: Bayesian physics-informed neural networks for forward and inverse PDE problems with noisy data,” Journal of Computational Physics, vol. 425, 2021, doi: 10.1016/j.jcp.2020.109913.
  • [20] L. Yang, D. Zhang, and G. E. Karniadakis, “Physics-Informed Generative Adversarial Networks for Stochastic Differential Equations,” SIAM Journal on Scientific Computing, vol. 42, no. 1, pp. A292–A317, 2020, doi: 10.1137/18M1225409.
  • [21] P. Stiller et al., “Large-Scale Neural Solvers for Partial Differential Equations,” in Driving Scientific and Engineering Discoveries Through the Convergence of HPC, Big Data and AI, Cham: Springer International Publishing, 2020, pp. 20–34. doi: 10.1007/978-3-030-63393-6_2.