Learning Only On Boundaries: a Physics-Informed Neural operator for Solving Parametric Partial Differential Equations in Complex Geometries

Zhiwei Fang, Sifan Wang, Paris Perdikaris

Abstract

Recently deep learning surrogates and neural operators have shown promise in solving partial differential equations (PDEs). However, they often require a large amount of training data and are limited to bounded domains. In this work, we present a novel physics-informed neural operator method to solve parametrized boundary value problems without labeled data. By reformulating the PDEs into boundary integral equations (BIEs), we can train the operator network solely on the boundary of the domain. This approach reduces the number of required sample points from $O(N^d)$ to $O(N^{d-1})$, where $d$ is the domain's dimension, leading to a significant acceleration of the training process. Additionally, our method can handle unbounded problems, which are unattainable for existing physics-informed neural networks (PINNs) and neural operators. Our numerical experiments show the effectiveness of parametrized complex geometries and unbounded problems.

Introduction

Solving partial differential equations (PDEs) is a cornerstone of physics and engineering. While traditional numerical methods like Finite Element Analysis are robust, they become computationally expensive for parametric problems where the equation must be solved repeatedly for different geometries or conditions. Physics-Informed Neural Networks (PINNs) offer a flexible alternative but suffer from the "curse of dimensionality," requiring dense sampling throughout the entire volume ($O(N^d)$). Furthermore, PINNs struggle with complex geometries and cannot easily handle unbounded domains, such as acoustic waves scattering into infinite space.

Methods

This paper introduces a physics-informed neural operator that bridges classical mathematics and modern AI by combining Boundary Integral Equations (BIEs) with the NOMAD (Nonlinear Manifold Decoders) architecture.

Instead of satisfying the PDE at every point within a volume, the authors use Green’s identities to transform the PDE into an integral equation defined exclusively on the boundary. This reduces the sampling requirement to $O(N^{d-1})$. The NOMAD framework uses an encoder to process boundary geometry and conditions, while a Fourier feature network decoder represents the unknown boundary potentials. Training is performed by minimizing the BIE residual, meaning the physics itself provides supervision without the need for labeled ground truth data.

Experiments and Results

The method was evaluated on three synthetic benchmarks: the 2D Laplace equation on complex "flower" geometries, the 2D Bi-harmonic equation (a 4th-order PDE used in elasticity), and the 3D Helmholtz equation for unbounded scattering.

Model prediction for the 2D Laplace equation on a complex flower geometry.
The difference between the prediction and ground truth, demonstrating high visual fidelity and low error.

Key findings include:

  • Accuracy: The model achieved high precision, with relative errors as low as ~0.7% for the Bi-harmonic equation and ~3.1% for 3D Helmholtz far-field problems.
  • Generalization: Once trained, the operator can instantly predict solutions for new geometries within its parameter range without re-training.
  • Unbounded Domains: By satisfying radiation conditions at infinity through the BIE formulation, the model successfully solves exterior scattering problems that are unattainable for standard PINNs.
  • Simplified Computation: High-order PDEs are reduced to 1st-order derivatives in the BIE formulation, simplifying the computational graph and stabilizing training.

Conclusion

By shifting the focus from the volume to the boundary, this physics-informed neural operator significantly reduces computational complexity and enables the solution of unbounded problems. While currently limited to linear PDEs with known Green's functions, future work aims to extend this framework to nonlinear problems and improve integration accuracy beyond Monte Carlo methods.

Summary updated

Abstract

Recently deep learning surrogates and neural operators have shown promise in solving partial differential equations (PDEs). However, they often require a large amount of training data and are limited to bounded domains. In this work, we present a novel physics-informed neural operator method to solve parametrized boundary value problems without labeled data. By reformulating the PDEs into boundary integral equations (BIEs), we can train the operator network solely on the boundary of the domain. This approach reduces the number of required sample points from $O(N^d)$ to $O(N^{d-1})$, where $d$ is the domain's dimension, leading to a significant acceleration of the training process. Additionally, our method can handle unbounded problems, which are unattainable for existing physics-informed neural networks (PINNs) and neural operators. Our numerical experiments show the effectiveness of parametrized complex geometries and unbounded problems.
BibTeX
@article{fang2024learning,
  author = {Zhiwei Fang and Sifan Wang and Paris Perdikaris},
  title = {{Learning Only On Boundaries: a Physics-Informed Neural operator for Solving Parametric Partial Differential Equations in Complex Geometries}},
  journal = {Neural Computation},
  year = {2024},
  doi = {10.1162/neco\_a\_01647},
}
Figure