Solving Partial Differential Equations in Different Domains by Operator Learning method Based on Boundary Integral Equations

Bin Meng, Yutong Lu, Ying JiangarXiv 2024 · OpenReview

TL;DR

This paper introduces two novel operator learning models, BI-DeepONet and BI-TDONet, that leverage Boundary Integral Equations (BIEs) to solve Partial Differential Equations (PDEs) across various domains without retraining. BI-TDONet, in particular, shows superior performance by using trigonometric coefficients to handle oscillatory functions and a structure inspired by the Singular Value Decomposition of operators.

What Problem Does This Paper Address?

Traditional numerical methods for solving Partial Differential Equations (PDEs) often struggle with computational complexity, especially for high-dimensional problems or when dealing with domains that change. While neural network-based methods like DeepONet have shown promise in learning operator mappings, they typically require retraining for new domains. This paper addresses the challenge of developing operator learning models that can efficiently solve PDEs on arbitrary domains without the need for retraining, specifically by integrating them with Boundary Integral Equations (BIEs). BIEs offer an advantage by reducing the problem's dimensionality to the boundary, and importantly, the solution domain remains fixed once the boundary is parameterized, making it ideal for operator learning.

Methods

The authors propose two operator learning frameworks:

  • Boundary Integral Type Deep Operator Network (BI-DeepONet): This model adapts the DeepONet architecture to learn the operator that maps boundary conditions and domain parameters to the solution of PDEs formulated via BIEs. It uses a trunk net to encode spatial points and a branch net to encode input functions.
  • Boundary Integral Trigonometric Deep Operator Neural Network (BI-TDONet): This is an advancement over BI-DeepONet. It addresses challenges with capturing high-frequency components and ensuring periodicity by:
  • Representing input functions (boundary parameter functions and right-hand side functions) as trigonometric series and using their coefficients as inputs.
  • Structuring the network based on the Singular Value Decomposition (SVD) of bounded linear operators, aligning network components with the SVD structure.
  • Using trigonometric coefficients as both inputs and outputs to maintain periodicity and better capture oscillatory/impulsive signals.

The core idea is to learn the inverse of the integral operator that defines the PDE solution, which is then expressed as a function of the boundary and input data.

Datasets

The paper generates custom datasets for training and testing. These datasets consist of:

  • Boundaries: A large number of 2D boundaries are generated, with specific filtering based on curvature to avoid overly narrow regions that demand excessive computational resources.
  • Input Functions: Various input functions (e.g., boundary conditions, forcing terms) are sampled and parameterized.
  • Solutions: Corresponding solutions to the PDEs are computed using BIEs and fast numerical methods.

The specific datasets are tailored for different problem types: Laplace Boundary Value Problems (LBVPs), potential flow, elastostatics, and acoustic wave scattering. The paper details the generation process for these datasets, including sampling distributions and parameter choices.

Key Insights and Ideas

  • BIEs as a Natural Fit for Operator Learning: The inherent property of BIEs where the solution domain remains fixed once the boundary is parameterized makes them exceptionally well-suited for operator learning models that aim to generalize across domains without retraining.
  • Trigonometric Coefficients for Oscillatory Signals: Representing functions as trigonometric series and using their coefficients as inputs/outputs in BI-TDONet is crucial for effectively handling oscillatory and impulse signals, leading to better accuracy and faster convergence.
  • SVD-Inspired Architecture for BI-TDONet: Aligning the BI-TDONet architecture with the Singular Value Decomposition of the underlying integral operator provides a principled way to model the operator's behavior, enhancing its performance.
  • BI-TDONet Outperforms BI-DeepONet: The BI-TDONet consistently achieves higher accuracy and faster convergence across various problems, especially those involving high-frequency components, demonstrating the benefits of its specialized architecture and input representation.
  • Operator Learning for BIEs is Fast: The proposed models achieve very fast inference times (on the order of 10⁻⁵ to 10⁻⁶ seconds), significantly outperforming traditional fast BIE algorithms.

Key Results

  • Superior Accuracy of BI-TDONet: BI-TDONet consistently achieves significantly lower Mean Absolute Error (MAE) and Mean Relative Error (MRE) compared to BI-DeepONet across all tested problems (LBVPs, potential flow, elastostatics, acoustic scattering). For instance, in LBVPs, BI-TDONet's MRE is approximately 10-2 to 10-1 of BI-DeepONet's. In the acoustic scattering problem, BI-TDONet achieves an error of 3.51 × 10⁻², while BI-DeepONet fails to converge.
  • Faster Convergence: BI-TDONet exhibits a faster convergence rate during training, reaching stable loss levels much earlier than BI-DeepONet.
  • Efficient Inference: Both models demonstrate very fast inference times, averaging less than 0.01 milliseconds per case, which is orders of magnitude faster than many established BIE algorithms.
  • Generalization Across Domains: The trained models can solve PDEs on new, unseen boundaries without requiring retraining, showcasing their generalization capabilities.
  • Handling High-Frequency Data: BI-TDONet effectively handles problems with high-frequency input functions, where BI-DeepONet struggles.

Limitations and Future Work

  • Scalability of BI-DeepONet for Vector-Valued Outputs: For vector-valued outputs, BI-DeepONet requires multiple instances, one for each component, which can increase complexity and resource demands for problems with many output components.
  • General PDE Solvability: The proposed BIE-based operator learning framework is primarily designed for problems that can be formulated using BIEs. Extending it to general linear or nonlinear PDEs where Green's functions are hard to obtain is a challenge.
  • Future Work: The authors suggest exploring the application of their BIE-based neural operators to nonlinear PDEs (like Navier-Stokes equations) and extending their use to three-dimensional problems (e.g., 3D LBVPs). They also mention the potential of learning Green's functions for general PDEs.

Summary updated

Abstract

This article explores operator learning models that can deduce solutions to partial differential equations (PDEs) on arbitrary domains without requiring retraining. We introduce two innovative models rooted in boundary integral equations (BIEs): the Boundary Integral Type Deep Operator Network (BI-DeepONet) and the Boundary Integral Trigonometric Deep Operator Neural Network (BI-TDONet), which are crafted to address PDEs across diverse domains. Once fully trained, these BIE-based models adeptly predict the solutions of PDEs in any domain without the need for additional training. BI-TDONet notably enhances its performance by employing the singular value decomposition (SVD) of bounded linear operators, allowing for the efficient distribution of input functions across its modules. Furthermore, to tackle the issue of function sampling values that do not effectively capture oscillatory and impulse signal characteristics, trigonometric coefficients are utilized as both inputs and outputs in BI-TDONet. Our numerical experiments robustly support and confirm the efficacy of this theoretical framework.
BibTeX
@misc{meng2024solving,
  author = {Bin Meng and Yutong Lu and Ying Jiang},
  title = {{Solving Partial Differential Equations in Different Domains by Operator Learning method Based on Boundary Integral Equations}},
  year = {2024},
  eprint = {2406.02298},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2406.02298},
}

Loading related content...

Figure