Adaptive Domain Decomposition for Finite Basis Physics-Informed Neural Networks
Organizations: University of North Carolina, Charlotte · Pacific Northwest National Laboratory · Delft University of Technology; Delft Institute of Applied Mathematics · University of Washington, Applied Mathematics ·
Abstract
In this work, we present adaptive domain decomposition finite basis physics-informed neural networks (ADD-FBPINNs). The proposed method integrates Partition of Unity Networks (POUNets), a data-driven approach for learning domain partitions, with finite Basis physics-informed neural networks (FBPINNs) to perform more efficient domain decomposition and achieve more accurate approximations. FBPINNs improve upon the accuracy of standard physics-informed neural networks (PINNs) through user-defined subdivisions of the computational domain. By adapting the domain decomposition based on the solution's features using a separable POUNet, the framework can automatically learn partitions that are stable and can be geometrically interpretable. We test this approach on the two-dimensional Helmholtz equation, two-dimensional Poisson equation with multiscale features, and viscous Burgers equation and observe improved accuracy compared with FBPINNs with fixed domain decomposition.
Figures & tables
| Overlap | Relative L2 error | Points active in 2 subdomains |
| 533 | ||
| 2206 | ||
| 1005 | ||
| 1509 |
| Overlap | Relative L2 error | Points active in 2 subdomains |
| 3353 | ||
| 42,346 | ||
| 4887 | ||
| 6225 | ||
| 7370 |
| Overlap | Relative L2 error | Points active in 2 subdomains |
| 7440 | ||
| 44,266 | ||
| 8100 | ||
| 11,245 | ||
| 14,876 |
| Case | Mode numbers | PINN | FBPINN | ADD-FBPINN | Number of subdomains |
| Case 1 | |||||
| Case 2 | |||||
| Case 3 | |||||
| Case 4 |
| Case | Frequencies | PINN | FBPINN | ADD-FBPINN | Number of subdomains |
| Case 1 | |||||
| Case 2 | |||||
| Case 3 | |||||
| Case 4 |
| Sharp gradient location | PINN | FBPINN | ADD-FBPINN | Number of subdomains |
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
| Architecture | Relative L2 Error | Number of subdomains |
| MLP | 9 | |
| Separable MLP | 9 | |
| RBF | 4 | |
| Separable RBF | 9 |
| FBPINN Parameter | 1D Sine ODE | 2D Helmholtz | 2D Poisson |
| Learning rate | |||
| Network size | |||
| Activation function | tanh | tanh | tanh |
| Collocation points (baseline) | 2,000 | 10,000 | 10,000 |
| Batch size | 1,000 | 10,000 | 10,000 |
| Iterations | 40,000 | 40,000 | 40,000 |
| PINN Parameters | 2D Helmholtz | 2D Poisson | Burgers |
| Learning rate | |||
| Network size | |||
| Activation function | tanh | tanh | tanh |
| Collocation points | 40,000 | 40,000 | 40,000 |
| Batch size | 10,000 | 10,000 | 10,000 |
| Iterations | 40,000 | 40,000 | 40,000 |
| FBPINN Parameters | 2D Helmholtz | 2D Poisson | Burgers |
| Learning rate | |||
| Network size | |||
| Activation function | tanh | tanh | tanh |
| Collocation points | 40,000 | 40,000 | 40,000 |
| Batch size | 10,000 | 10,000 | 10,000 |
| Iterations | 40,000 | 40,000 | 40,000 |
| ADD-FBPINN Parameters | 2D Helmholtz | 2D Poisson | Burgers |
| Learning rate | |||
| Network size | |||
| Activation function | tanh | tanh | tanh |
| Collocation points | 40,000 | 40,000 | 40,000 |
| Batch size | 10,000 | 10,000 | 10,000 |
| Iterations | 40,000 | 40,000 | 40,000 |
| ADD-FBPINN POUNet Parameters | 2D Helmholtz | 2D Poisson | Burgers |
| Polynomial degree | 2 | 2 | 2 |
| 40,000 | 40,000 | 40,000 | |
| Learning rate | |||
| Architecture | Separable RBF | Separable RBF | Separable RBF |
| Iterations | 2,000 | 2,000 | 2,000 |
| Collapse threshold |
Neural PDE Solvers
Neural PDE solvers use trainable models to approximate solutions or solution operators for partial differential equations, often with physical constraints.
See all 295 papers