A Neural-Enhanced Weak Galerkin Method for Second-Order Elliptic Problems with Low-Regularity Solutions

1. Problem Statement

The paper addresses the numerical approximation of second-order elliptic partial differential equations (PDEs) where the exact solution possesses low regularity.

• Context: Such problems arise in domains with geometric singularities (e.g., re-entrant corners) or discontinuous coefficients, where the solution $u$ fails to belong to the Sobolev space $H^{k+2}(\Omega)$.
• Limitation of Classical Methods: Standard Finite Element Methods (FEM) and Weak Galerkin (WG) methods, which rely on polynomial approximation spaces, suffer from degraded convergence rates when the solution is non-smooth. Achieving high accuracy typically requires excessive mesh refinement or complex enrichment strategies.
• Limitation of Pure Neural Methods: While Deep Learning approaches (e.g., Physics-Informed Neural Networks) show promise in approximating singularities, they often lack the rigorous variational structure, stability, and proven convergence guarantees of classical Galerkin methods.

2. Methodology: Neural-Enhanced Weak Galerkin (NE-WG)

The author proposes a hybrid method that integrates Neural Networks (NN) into the Weak Galerkin (WG) finite element framework in a structure-preserving manner.

A. The Weak Galerkin Foundation

The method builds upon the standard WG formulation for the problem $-\nabla \cdot (a\nabla u) = f$:

• Discretization: Uses a polytopal mesh $T_h$ and a discrete weak gradient $\nabla_w$.
• Space: The solution is approximated in a space $V_h^0$ consisting of pairs $\{v_0, v_b\}$ (interior and boundary values), allowing for discontinuities across element boundaries.
• Stabilization: A stabilizer term $s(u,v)$ is added to the bilinear form $a_w(u,v)$ to ensure coercivity and stability.

B. Neural Enrichment Strategy

The core innovation is the adaptive augmentation of the WG space with neural network functions:

1. Neural Space Definition: A class of feedforward neural networks $\mathcal{N}_M$ is defined. To satisfy homogeneous boundary conditions, these are multiplied by a function $\phi$ that vanishes on the boundary.
2. Lifting: Neural functions are "lifted" into the WG space via $L^2$ projections ($Q_0$ and $Q_b$), creating a lifted neural space $\hat{Z}_m \subset V_h^0$.
3. Augmented Space: The total approximation space is $W_h^m = V_h^0 + \text{span}\{\hat{n}_1, \dots, \hat{n}_m\}$.
4. Residual-Driven Adaptivity: The method employs an iterative enrichment procedure (Algorithm 1):
   • Compute the current WG solution $u_h^{(m)}$.
   • Calculate the residual functional $R_h(u_h^{(m)})$.
   • Select a new neural basis: Train a neural network $n_\theta$ to maximize the normalized residual $\eta_h$ (effectively finding the direction of maximum error reduction).
   • Update: Add the lifted neural function to the basis and solve the Galerkin problem in the enriched space $W_h^{m+1}$.

3. Key Contributions

1. Structure Preservation: The neural enrichment is formulated entirely within the WG variational framework. This preserves the symmetry, stability, and Galerkin orthogonality of the original WG method, unlike many "black-box" neural PDE solvers.
2. Adaptive Residual Maximization: The enrichment is not random; it is driven by a rigorous residual maximization principle that targets the dominant error components (typically singularities).
3. Compatibility: The method is naturally compatible with general polytopal meshes and non-convex domains, inheriting the geometric flexibility of WG methods.
4. Theoretical Error Analysis: The paper provides rigorous mathematical proofs for the convergence of the method, bridging the gap between neural approximation and finite element theory.

4. Theoretical Results and Error Estimates

The paper establishes quasi-optimal error estimates in the WG energy norm ($\|\cdot\|_{a_w}$).

• General Quasi-Optimality:
  For the exact solution $u$ and the neural-enhanced approximation $u_h^{(m)}$:
  $$\|Q_h u - u_h^{(m)}\|_{a_w} \leq C \left( \inf_{v \in W_h^m} \|Q_h u - v\|_{a_w} + h^{k+1}\|u\|_{H^{k+2}(\Omega)} \right)$$
  This shows the error is bounded by the best approximation in the enriched space plus the standard WG consistency error.

• Regular-Singular Decomposition:
  If the solution can be decomposed into a regular part $u_r \in H^{k+2}(\Omega)$ and a singular part $u_s$ (i.e., $u = u_r + u_s$), the error estimate becomes:
  $$\|Q_h u - u_h^{(m)}\|_{a_w} \leq C \left( h^{k+1}\|u_r\|_{H^{k+2}(\Omega)} + \varepsilon_m \right)$$
  Where $\varepsilon_m$ is the approximation error of the singular component $u_s$ by the neural basis functions in the enriched space.

Implication: If the neural network can effectively approximate the singular component ($\varepsilon_m \to 0$), the method recovers the optimal convergence rate determined by the regular part $u_r$, effectively overcoming the bottleneck caused by low regularity.

5. Significance

• Bridging Two Worlds: This work successfully combines the robustness and theoretical guarantees of classical finite element methods with the flexible approximation power of deep learning.
• Solving Low-Regularity Problems: It offers a viable path to solving elliptic problems with singularities without the need for extreme mesh refinement (which is computationally expensive) or manual identification of singularity functions.
• Rigorous Framework: By embedding neural networks into a variational framework, the method avoids the stability issues often associated with pure neural network solvers, providing a mathematically sound approach for scientific computing.

In summary, the paper presents a novel, adaptive, and theoretically grounded numerical method that enhances the Weak Galerkin framework with neural networks to efficiently and accurately solve elliptic problems with low-regularity solutions.