Abstract:Learning operators for parametric partial differential equations (PDEs) using neural networks has gained significant attention in recent years. However, standard approaches like Deep Operator Networks (DeepONets) require extensive labeled data, and physics-informed DeepONets encounter training challenges. In this paper, we introduce a novel physics-informed tailored finite point operator network (PI-TFPONet) method to solve parametric interface problems without the need for labeled data. Our method fully leverages the prior physical information of the problem, eliminating the need to include the PDE residual in the loss function, thereby avoiding training challenges. The PI-TFPONet is specifically designed to address certain properties of the problem, allowing us to naturally obtain an approximate solution that closely matches the exact solution. Our method is theoretically proven to converge if the local mesh size is sufficiently small and the training loss is minimized. Notably, our approach is uniformly convergent for singularly perturbed interface problems. Extensive numerical studies show that our unsupervised PI-TFPONet is comparable to or outperforms existing state-of-the-art supervised deep operator networks in terms of accuracy and versatility.

What problem does this paper attempt to address?

This paper attempts to address the problem of efficiently and accurately solving partial differential equations (PDEs) in parametric interface problems using neural networks without requiring a large amount of labeled data. Specifically, the paper focuses on elliptic interface problems, which have widespread applications in fields such as fluid mechanics, materials science, electromagnetics, and bionics. The characteristic of these interface problems is that the solution and its derivatives exhibit jump discontinuities at the interface, posing challenges for traditional numerical methods and standard neural networks. ### Main Issues 1. **High Data Demand**: Standard deep operator networks (such as DeepONet) require a large amount of labeled data for training. Although physics-informed deep operator networks (such as physics-informed DeepONet) reduce the need for labeled data, they encounter difficulties during training. 2. **Non-smooth Solutions**: Many existing methods are mainly suitable for PDEs with relatively smooth solutions. However, for interface problems with non-smooth or piecewise smooth solutions, the performance of existing methods is suboptimal. 3. **Multiscale Phenomena and Singular Perturbations**: For interface problems with multiscale phenomena or singular perturbations, existing methods struggle to effectively capture the details of the solution, leading to a decrease in accuracy. ### Solution The paper proposes a new physics-informed tailored finite point operator network (PI-TFPONet) method, whose main features include: - **No Labeled Data Required**: By fully utilizing the prior physical information of the problem, it avoids incorporating the PDE residual into the loss function, thereby addressing training difficulties. - **Local Basis Functions**: In each local region, the solution can be represented as a linear combination of a set of local basis functions, and the solution is reconstructed by learning the coefficients of these basis functions. - **Theoretical Guarantee**: Theoretical analysis shows that as long as the local mesh size is sufficiently small and the training loss is minimized, the solution of the PI-TFPONet method will converge to the true solution. - **High Accuracy and Robustness**: Extensive numerical experiments demonstrate that PI-TFPONet can achieve or exceed the accuracy of existing supervised deep operator networks in handling various interface problems, whether they are smooth problems or problems with singular perturbations. ### Application Scenarios - **Fluid Mechanics**: Simulating fluid flow in different media. - **Materials Science**: Studying the behavior of materials under different conditions. - **Electromagnetics**: Analyzing the propagation of electromagnetic fields in different media. - **Bionics**: Simulating physical processes in biological systems. In summary, by proposing the PI-TFPONet method, this paper aims to address the challenges of high data demand, non-smooth solutions, and multiscale phenomena in parametric interface problems, providing new tools and methods for scientific research and engineering applications in related fields.

Physics-Informed Tailored Finite Point Operator Network for Parametric Interface Problems

Tailored Finite Point Operator Networks for Interface problems

Physics-Informed Neural Operator for Learning Partial Differential Equations

Operator Learning Enhanced Physics-informed Neural Networks for Solving Partial Differential Equations Characterized by Sharp Solutions

Learning the solution operator of a nonlinear parabolic equation using physics informed deep operator network

Solving forward and inverse PDE problems on unknown manifolds via physics-informed neural operators

Solving parametric elliptic interface problems via interfaced operator network

Learning the solution operator of parametric partial differential equations with physics-informed DeepONets

Data-driven and physics-informed deep learning operators for solution of heat conduction equation with parametric heat source

Physics-Informed Deep Inverse Operator Networks for Solving PDE Inverse Problems

Learning Only on Boundaries: A Physics-Informed Neural Operator for Solving Parametric Partial Differential Equations in Complex Geometries

Physics-informed Discretization-independent Deep Compositional Operator Network

Physics-Informed Geometry-Aware Neural Operator

Fine-Tuning DeepONets to Enhance Physics-informed Neural Networks for solving Partial Differential Equations

Physics-Informed Deep Neural Operator Networks

Separable Operator Networks

A physics-informed transformer neural operator for learning generalized solutions of initial boundary value problems

Characterizing and Mitigating the Difficulty in Training Physics-informed Artificial Neural Networks under Pointwise Constraints

Deep fuzzy physics-informed neural networks for forward and inverse PDE problems

Neural Parameter Regression for Explicit Representations of PDE Solution Operators