Abstract
Interface problems pose significant challenges due to the discontinuity of their solutions, particularly when they involve singular perturbations or high-contrast coefficients, resulting in intricate singularities that complicate resolution. The increasing adoption of deep learning techniques for solving partial differential equations has spurred our exploration of these methods for addressing interface problems. In this study, we introduce Tailored Finite Point Operator Networks (TFPONets) as a novel approach for tackling parameterized interface problems. Leveraging DeepONets and integrating the Tailored Finite Point method (TFPM), TFPONets offer enhanced accuracy in reconstructing solutions without the need for intricate equation manipulation. Experimental analyses conducted in both one- and two-dimensional scenarios reveal that, in comparison to existing methods such as DeepONet and IONet, TFPONets demonstrate superior learning and generalization capabilities even with limited locations.
Introduction
In many physical systems, properties like heat conductivity or fluid density change abruptly at an "interface"—for instance, where oil meets water or metal meets ceramic. These interface problems are notoriously difficult to solve because the solutions are often discontinuous or have sharp "kinks" (singularities) at the boundary.
Traditional numerical methods, such as Finite Element Analysis, require very fine and carefully constructed meshes to capture these sharp changes. Standard deep learning approaches, like vanilla DeepONets, tend to "smooth out" these discontinuities, leading to poor accuracy near the interface. This challenge is exacerbated in singularly perturbed problems, where solutions change rapidly over tiny distances, and high-contrast problems, where material properties differ by orders of magnitude. TFPONets (Tailored Finite Point Operator Networks) address these issues by combining the flexible operator learning of DeepONets with the physics-rich local basis functions of the Tailored Finite Point Method (TFPM).
Methodology
The core innovation of TFPONets is a hybrid architecture that merges two distinct ideas: DeepONets for general operator learning and TFPM for capturing local singularities. Instead of relying solely on a standard neural network to learn the entire solution, TFPONets use a parallel structure.
As illustrated in Figure 1, the model consists of a standard DeepONet block working alongside several Basis Blocks. These Basis Blocks are shallow neural networks that take TFPM basis functions (such as Airy or Bessel functions) as inputs. These functions are mathematically "tailored" to represent the behavior of a PDE near singularities.
The DeepONet acts as an "error corrector," handling the general parts of the solution, while the Basis Blocks handle the sharp, difficult-to-learn features near the interface. The final output is the sum of these components. This approach "bakes" physical prior knowledge directly into the network's architecture rather than just relying on the loss function. The training process minimizes a composite loss function that includes both data loss ($L_{data}$) and jump condition loss ($L_{jump}$).
Experimental Evaluation
The researchers evaluated TFPONets using synthetic datasets generated for elliptic interface problems in 1D and 2D:
- 1D Scenarios: Problems featuring singular perturbations (with coefficients as small as $0.0001$) and high-contrast coefficients.
- 2D Scenarios: Complex geometries with jump conditions across an interface.
- Data Generation: Input functions ($f$) were sampled from Gaussian Random Fields. Ground truth solutions were generated using high-precision numerical TFPM on uniform meshes to ensure a "gold standard" for comparison.
Key Insights and Results
- Architecture over Loss: While many "physics-informed" networks (like PINNs) enforce physics through a penalty in the loss function, TFPONets incorporate physics into the network architecture itself via the basis blocks. This makes the learning process more stable and faster.
- Superior Accuracy: In 1D tests with singular perturbations, TFPONets captured sharp boundary and interior layers that vanilla DeepONets completely missed. In high-contrast tests (where coefficients jumped from 1 to $0.0001$), TFPONets maintained high precision, whereas standard models struggled to resolve the resulting singularities.
- Efficiency and Generalization: TFPONets achieved lower Mean Square Error (MSE) than existing methods like IONet, even when trained on fewer spatial locations. Because the model is based on DeepONet, it remains "meshless," allowing it to predict solutions at arbitrary points on a fine grid without retraining.
Conclusion
TFPONets demonstrate superior learning and generalization capabilities for interface problems compared to standard operator networks. By integrating tailored basis functions into the architecture, the model effectively resolves singularities and discontinuities. Future work will explore extending this method to 3D problems and automating the selection of basis functions to reduce the need for domain-specific expertise.
Summary updated