arrow
Volume 33, Issue 5
Investigating and Mitigating Failure Modes in Physics-Informed Neural Networks (PINNs)

Shamsulhaq Basir

Commun. Comput. Phys., 33 (2023), pp. 1240-1269.

Published online: 2023-06

Export citation
  • Abstract

This paper explores the difficulties in solving partial differential equations (PDEs) using physics-informed neural networks (PINNs). PINNs use physics as a regularization term in the objective function. However, a drawback of this approach is the requirement for manual hyperparameter tuning, making it impractical in the absence of validation data or prior knowledge of the solution. Our investigations of the loss landscapes and backpropagated gradients in the presence of physics reveal that existing methods produce non-convex loss landscapes that are hard to navigate. Our findings demonstrate that high-order PDEs contaminate backpropagated gradients and hinder convergence. To address these challenges, we introduce a novel method that bypasses the calculation of high-order derivative operators and mitigates the contamination of backpropagated gradients. Consequently, we reduce the dimension of the search space and make learning PDEs with non-smooth solutions feasible. Our method also provides a mechanism to focus on complex regions of the domain. Besides, we present a dual unconstrained formulation based on Lagrange multiplier method to enforce equality constraints on the model’s prediction, with adaptive and independent learning rates inspired by adaptive subgradient methods. We apply our approach to solve various linear and non-linear PDEs.

  • AMS Subject Headings

65D15, 35E05, 35E15, 68T20, 90C26, 90C29, 90C30, 90C31, 49N15, 90C46, 90C47, 90C90

  • Copyright

COPYRIGHT: © Global Science Press

  • Email address
  • BibTex
  • RIS
  • TXT
@Article{CiCP-33-1240, author = {Basir , Shamsulhaq}, title = {Investigating and Mitigating Failure Modes in Physics-Informed Neural Networks (PINNs)}, journal = {Communications in Computational Physics}, year = {2023}, volume = {33}, number = {5}, pages = {1240--1269}, abstract = {

This paper explores the difficulties in solving partial differential equations (PDEs) using physics-informed neural networks (PINNs). PINNs use physics as a regularization term in the objective function. However, a drawback of this approach is the requirement for manual hyperparameter tuning, making it impractical in the absence of validation data or prior knowledge of the solution. Our investigations of the loss landscapes and backpropagated gradients in the presence of physics reveal that existing methods produce non-convex loss landscapes that are hard to navigate. Our findings demonstrate that high-order PDEs contaminate backpropagated gradients and hinder convergence. To address these challenges, we introduce a novel method that bypasses the calculation of high-order derivative operators and mitigates the contamination of backpropagated gradients. Consequently, we reduce the dimension of the search space and make learning PDEs with non-smooth solutions feasible. Our method also provides a mechanism to focus on complex regions of the domain. Besides, we present a dual unconstrained formulation based on Lagrange multiplier method to enforce equality constraints on the model’s prediction, with adaptive and independent learning rates inspired by adaptive subgradient methods. We apply our approach to solve various linear and non-linear PDEs.

}, issn = {1991-7120}, doi = {https://doi.org/10.4208/cicp.OA-2022-0239}, url = {http://global-sci.org/intro/article_detail/cicp/21761.html} }
TY - JOUR T1 - Investigating and Mitigating Failure Modes in Physics-Informed Neural Networks (PINNs) AU - Basir , Shamsulhaq JO - Communications in Computational Physics VL - 5 SP - 1240 EP - 1269 PY - 2023 DA - 2023/06 SN - 33 DO - http://doi.org/10.4208/cicp.OA-2022-0239 UR - https://global-sci.org/intro/article_detail/cicp/21761.html KW - Constrained optimization, Lagrangian multiplier method, Stokes equation, convection equation, convection-dominated convection-diffusion equation, heat transfer in composite medium, Lid-driven cavity problem. AB -

This paper explores the difficulties in solving partial differential equations (PDEs) using physics-informed neural networks (PINNs). PINNs use physics as a regularization term in the objective function. However, a drawback of this approach is the requirement for manual hyperparameter tuning, making it impractical in the absence of validation data or prior knowledge of the solution. Our investigations of the loss landscapes and backpropagated gradients in the presence of physics reveal that existing methods produce non-convex loss landscapes that are hard to navigate. Our findings demonstrate that high-order PDEs contaminate backpropagated gradients and hinder convergence. To address these challenges, we introduce a novel method that bypasses the calculation of high-order derivative operators and mitigates the contamination of backpropagated gradients. Consequently, we reduce the dimension of the search space and make learning PDEs with non-smooth solutions feasible. Our method also provides a mechanism to focus on complex regions of the domain. Besides, we present a dual unconstrained formulation based on Lagrange multiplier method to enforce equality constraints on the model’s prediction, with adaptive and independent learning rates inspired by adaptive subgradient methods. We apply our approach to solve various linear and non-linear PDEs.

Basir , Shamsulhaq. (2023). Investigating and Mitigating Failure Modes in Physics-Informed Neural Networks (PINNs). Communications in Computational Physics. 33 (5). 1240-1269. doi:10.4208/cicp.OA-2022-0239
Copy to clipboard
The citation has been copied to your clipboard