Adversarial Training for Physics-Informed Neural Networks

Yao Li,Shengzhu Shi,Zhichang Guo,Boying Wu

2023-10-18

Abstract:Physics-informed neural networks have shown great promise in solving partial differential equations. However, due to insufficient robustness, vanilla PINNs often face challenges when solving complex PDEs, especially those involving multi-scale behaviors or solutions with sharp or oscillatory characteristics. To address these issues, based on the projected gradient descent adversarial attack, we proposed an adversarial training strategy for PINNs termed by AT-PINNs. AT-PINNs enhance the robustness of PINNs by fine-tuning the model with adversarial samples, which can accurately identify model failure locations and drive the model to focus on those regions during training. AT-PINNs can also perform inference with temporal causality by selecting the initial collocation points around temporal initial values. We implement AT-PINNs to the elliptic equation with multi-scale coefficients, Poisson equation with multi-peak solutions, Burgers equation with sharp solutions and the Allen-Cahn equation. The results demonstrate that AT-PINNs can effectively locate and reduce failure regions. Moreover, AT-PINNs are suitable for solving complex PDEs, since locating failure regions through adversarial attacks is independent of the size of failure regions or the complexity of the distribution.

Machine Learning,Numerical Analysis

What problem does this paper attempt to address?

### Problems Addressed by the Paper

The paper primarily addresses the challenges faced by Physics-Informed Neural Networks (PINNs) when solving complex Partial Differential Equations (PDEs). Specifically:

1. **Robustness Issues**:
- The original PINNs lack sufficient robustness when dealing with problems that exhibit multi-scale behavior, sharp or oscillatory solutions, and temporal causality.

2. **Failure to Accurately Capture Failure Regions**:
- Existing PINNs methods fail to accurately identify regions where the model fails in certain complex problems, leading to a decline in model performance.

To tackle these issues, the authors propose an adversarial training-based method (AT-PINN), which improves the robustness and accuracy of PINNs by generating adversarial samples and fine-tuning the model with these samples. This method is particularly suitable for addressing problems with multi-scale behavior, sharp solutions, or oscillatory solutions, and can better capture temporal causality. Experimental results demonstrate that AT-PINN shows significant advantages in various complex PDE problems.