Tailored Finite Point Method (TFPM)

• TFPM is a meshless method that employs operator-adapted local trial spaces, constructed from exact solutions of frozen-coefficient PDEs, to solve interface problems.
• It assembles global solutions from local approximations via collocation or Galerkin techniques, effectively resolving discontinuities and steep gradients.
• Incorporated into deep neural operator architectures, TFPM enhances accuracy in high-contrast, singularly perturbed, and image segmentation applications.

The Tailored Finite Point Method (TFPM) is a meshless numerical technique for the solution of interface problems governed by partial differential equations (PDEs), characterized by discontinuities or steep gradients in the solution or its derivatives across interfaces. TFPM constructs local approximation spaces on small patches using exact (or nearly exact) solutions of the underlying homogeneous PDE with frozen coefficients, then assembles a global solution via collocation or Galerkin approaches. This methodology offers robust and uniformly convergent resolution of boundary and interface layers, singular perturbations, and high-contrast media without the need for intricate mesh refinement or equation manipulations. TFPM’s principles have recently been embedded within deep and physics-informed operator networks, leveraging its operator-adapted local trial spaces for enhanced accuracy in neural PDE solvers (Li et al., 2024, Du et al., 2024).

1. Mathematical Principles and Local Trial Spaces

Let $\Omega \subset \mathbb{R}^d$ ($d=1,2$) be partitioned by a smooth interface $\Gamma$ into subdomains $\Omega^{-}$ and $\Omega^{+}$. The archetypal elliptic interface problem seeks $u$ satisfying: $$-\nabla\cdot(a(x)\nabla u(x)) + b(x)u(x) = f(x),\quad x\in\Omega\setminus\Gamma,$$ with prescribed solution and flux jumps across $\Gamma$: $$[u]_{\Gamma}(x) = g_D(x), \qquad [ a\,\partial_n u ]_{\Gamma}(x) = g_N(x),$$ and Dirichlet or Neumann boundary conditions on $\partial\Omega$.

The core of TFPM is the construction, on each small local patch $\Omega^j$ (stencil), of a low-dimensional space spanned by functions $\{B_{j,m}\}$ that are exact solutions to the local homogeneous PDE with coefficients $(a,b)$ frozen to local averages. In 1D, these are typically Airy functions, exponentials, or linear functions, depending on the local character of $b(x)$; in 2D, Fourier–Bessel modes are used. Particular solutions for inhomogeneous terms are included as necessary. Coefficients for local expansions are fixed by enforcing solution continuity and interface jump conditions at patch interfaces (Li et al., 2024).

2. Discretization, Algorithmic Structure, and Meshless Implementation

The TFPM workflow consists of:

• Node Selection: Placement of global collocation nodes $\{x_i\}$; clustering may be used near interfaces or anticipated layers. In 2D, nodes often form a Cartesian grid or unstructured cloud.
• Patch Construction: Around each $x_i$, a patch $\Omega_i$ containing enough neighbors is selected (3–5 in 1D; 10–20 in 2D).
• Local Basis Evaluation: Coefficients $(a_i, b_i)$ are frozen. Tailored basis $\{B_{i,m}\}$ (e.g., Airy, exponential, Bessel) are constructed, possibly with particular integral terms.
• Coefficient Determination: One can employ (A) shape function methods, solving a local moment system for meshless MLS-type reconstruction, or (B) direct collocation, assembling a block-sparse linear system for all trial coefficients.
• Assembly and Solution: The global system is assembled enforcing the PDE, interface jumps, and boundary conditions at corresponding collocation points.
• Postprocessing: The solution $u_h(x)$ is evaluated via either shape functions or patched local expansions.

A high-level pseudocode and detailed assembly steps for both shape-function and direct-collocation variants are provided in (Li et al., 2024). In meshless form, the method can be cast as an MLS collocation so that

$u_h(x) = \sum_{i=1}^N \phi_i(x) u_i$

with shape functions $\phi_i(x)$ locally spanning the tailored basis and reproducing exact solutions to the frozen-coefficient homogeneous PDE.

3. Error Analysis, Convergence, and Robustness

If the physical coefficients and data are sufficiently smooth away from $\Gamma$, and the local trial space has polynomial order $k$, TFPM achieves global error of order $k+1$ in maximum norm. For second-order basis in 1D (Airy/exponential), one obtains

$\|u - u_h\|_{\infty,\Omega} \leq C h^2$

with $C$ independent of severe singular perturbations (e.g., small $\varepsilon$ in diffusion) or large coefficient jumps at the interface (Li et al., 2024, Du et al., 2024). In both regular and singularly perturbed regimes, this uniform convergence is observed in practice and is theoretically proven for the physics-informed operator variants (Du et al., 2024).

4. Integration with Operator Learning and Deep Architectures

The operator-adapted local bases of TFPM have been incorporated into neural operator architectures for parametric PDEs exhibiting interface-driven difficulties.

• TFPONet integrates TFPM with DeepONet, where the local tailored basis is used within the operator learning framework, enabling accurate function reconstruction from few collocation nodes. Learning and generalization performance on high-contrast and singularly perturbed 1D/2D elliptic problems is substantially enhanced versus pure DeepONet/IONet (Li et al., 2024).
• Physics-Informed TFPONet (PI-TFPONet) removes the need for labeled data by building the tailored basis into a neural network decoder; the only loss terms are interface jumps and boundary violations. Volumetric PDE residual is not used. This yields uniformly convergent, unsupervised operator learning for singular and high-contrast problems (Du et al., 2024).
• Variational Model Based Tailored UNet (VM_TUNet) employs TFPM to discretize the Laplacian term in a fourth-order Cahn–Hilliard segmentation PDE. At each grid point, the Laplacian is approximated locally by solving the exact linearized ODE, resulting in a point-adaptive formula involving neighbor values and tailored exponential coefficients. This slotting of TFPM into the PDE block of a UNet significantly sharpens boundary preservation in image segmentation tasks compared to standard five-point finite differences (Qi et al., 9 May 2025).

5. Numerical Experiments and Comparative Performance

TFPM and TFPM-powered neural operator variants demonstrate superior accuracy in learning, generalization, and singularity resolution:

Problem Type	Method	MSE / Error	Key Features
1D singular perturbation	TFPONet	$1.2 \times 10^{-5}$ (M=129)	Captures sharp layers; DeepONet error $5\times10^{-3}$
	TFPONet	$O(10^{-8})$ (M=641)	Error decays quadratically with resolution
1D high-contrast interface	TFPONet	Low MSE (O($10^{-4}$))	Resolves corner and boundary layers; DeepONet fails near singularities
2D interface (high-contrast)	TFPONet	O($10^{-3}$)	Outperforms IONet (O($10^{-2}$))
Image segmentation	VM_TUNet	Dice = $0.713\pm0.002$	Yields smoother, continuous vessel boundaries

Across tested domains, TFPM and its operator-embedded extensions enable fine boundary resolution, effective handling of jump conditions, and robust learning in the presence of strong singularities (Li et al., 2024, Qi et al., 9 May 2025, Du et al., 2024).

6. Comparison to Classical and Related Approaches

Classical meshless and finite point methods (e.g., MLSM, RBF-FDM) typically use general-purpose polynomial or RBF trial spaces that are not adapted to the local operator, requiring heavy grid refinement to resolve layers or singularities. TFPM departs fundamentally by employing operator-adapted (tailored) local spaces—exact solutions to the locally homogeneous PDE—that inherently capture the behavior of thin layers and interface singularities. This yields uniform convergence and numerically well-conditioned solvers. Neural architectures (IONet, PINNs) that learn interface conditions via penalty or loss augmentation using standard basis sets still struggle near sharp layers, while TFPM-augmented networks enforce interface mechanics directly at the discrete level (Li et al., 2024).

In variational deep learning–PDE hybrids, incorporating TFPM in place of classical finite differences demonstrably improves fine-structure and boundary preservation, especially in applications sensitive to sharp features such as image segmentation (Qi et al., 9 May 2025).

7. Extensions, Limitations, and Outlook

TFPM’s construction hinges on the ability to derive or efficiently approximate local homogeneous solutions to the (possibly linearized) PDE, limiting direct extension to strongly nonlinear or nonlocal problems without additional approximation. Most TFPM applications assume piecewise-smooth coefficients and interfaces. Neural operator variants alleviate some online computational burdens by shifting solution of the local collocation system to offline learning, but accurate basis construction remains essential.

A plausible implication is that further advances may arise by combining TFPM with physics-informed meta-learning and adaptivity, or by extending the tailored basis philosophy to handle more general classes of PDEs and interface geometries. The method’s robust convergence and meshless nature render it particularly suitable for problems with spatially complex, moving, or data-driven interfaces.

---

For a comprehensive exposition and mathematical detail on TFPM and its neural operator integrations, see (Li et al., 2024, Du et al., 2024), and (Qi et al., 9 May 2025).