Data-Guided Physics-Informed Neural Networks for Solving Inverse Problems in Partial Differential Equations

Wei Zhou,Y.F. Xu

2024-07-15

Abstract:Physics-informed neural networks (PINNs) represent a significant advancement in scientific machine learning by integrating fundamental physical laws into their architecture through loss functions. PINNs have been successfully applied to solve various forward and inverse problems in partial differential equations (PDEs). However, a notable challenge can emerge during the early training stages when solving inverse problems. Specifically, data losses remain high while PDE residual losses are minimized rapidly, thereby exacerbating the imbalance between loss terms and impeding the overall efficiency of PINNs. To address this challenge, this study proposes a novel framework termed data-guided physics-informed neural networks (DG-PINNs). The DG-PINNs framework is structured into two distinct phases: a pre-training phase and a fine-tuning phase. In the pre-training phase, a loss function with only the data loss is minimized in a neural network. In the fine-tuning phase, a composite loss function, which consists of the data loss, PDE residual loss, and, if available, initial and boundary condition losses, is minimized in the same neural network. Notably, the pre-training phase ensures that the data loss is already at a low value before the fine-tuning phase commences. This approach enables the fine-tuning phase to converge to a minimal composite loss function with fewer iterations compared to existing PINNs. To validate the effectiveness, noise-robustness, and efficiency of DG-PINNs, extensive numerical investigations are conducted on inverse problems related to several classical PDEs, including the heat equation, wave equation, Euler--Bernoulli beam equation, and Navier--Stokes equation. The numerical results demonstrate that DG-PINNs can accurately solve these inverse problems and exhibit robustness against noise in training data.

Machine Learning,Numerical Analysis

What problem does this paper attempt to address?

The paper primarily addresses the challenges faced by Physics-Informed Neural Networks (PINNs) in solving inverse problems of Partial Differential Equations (PDEs) and proposes a novel method. Specifically, the paper focuses on the following issues:

1. **Imbalanced Loss Terms**: At the early stages of training, the PDE residual loss in PINNs can quickly reduce to a low level, while the data loss remains high. This leads to an imbalance among the terms in the loss function, affecting the overall training efficiency.

2. **Efficiency and Accuracy in Solving Inverse Problems**: PINNs require additional iterations to reduce data loss and accurately estimate unknown parameters when solving inverse problems, which increases computational cost.

To address these issues, the paper proposes Data-Guided Physics-Informed Neural Networks (DG-PINNs). This method includes two stages:
- **Pre-training Stage**: Minimize only the data loss, i.e., train the neural network using a data-driven approach.
- **Fine-tuning Stage**: Initialize with the parameters obtained from the pre-training stage, then simultaneously minimize a composite loss function that includes data loss, PDE residual loss, and possibly initial and boundary condition losses.

In this way, DG-PINNs ensure that the data loss is already reduced to a low level before entering the fine-tuning stage, allowing the fine-tuning stage to converge more quickly to the global optimum, thereby improving the efficiency of solving inverse problems.

The paper further validates the effectiveness, noise resistance, and efficiency of DG-PINNs through numerical experiments and compares them with traditional PINNs, demonstrating their advantages in solving inverse problems of classic PDEs such as the heat equation and wave equation.