Boundary neuron method for solving partial differential equations

1. Problem Statement

The paper addresses the challenge of solving Partial Differential Equations (PDEs), specifically boundary value problems (BVPs) for elliptic equations (Laplace and Helmholtz) on complex geometries and unbounded domains.

• Context: Traditional Boundary Element Methods (BEM) reduce the problem dimension by moving unknowns from the domain to the boundary. However, BEM requires fine discretization for high accuracy, leading to dense linear systems with high computational and storage costs.
• Limitations of Existing ML Methods: Recent neural network approaches (e.g., BINN, BI-GreenNet) that combine neural networks with boundary integral equations (BIE) rely heavily on gradient-based optimization. This introduces issues such as sensitivity to optimization efficiency, slow convergence, and susceptibility to local minima.
• Gap: There is a lack of methods that combine the dimensionality reduction of BIE with the linear solution structure of Random Feature Methods (RFM), which avoids non-convex optimization.

2. Methodology: Boundary Neuron Method with Random Features (BNM-RF)

The authors propose BNM-RF, a hybrid framework that approximates the unknown boundary functions using a shallow neural network with fixed random features.

Core Concept

Instead of training weights via backpropagation, the hidden layer parameters (weights $\omega_j$ and biases $b_j$) are randomly sampled and fixed. The output coefficients ($\beta_j$) are determined by solving a linear least squares problem.

Algorithmic Steps

1. Boundary Integral Formulation: The PDE is transformed into a Boundary Integral Equation (BIE) using layer potentials (single and double layer). The unknowns are the boundary values ($u$) and normal derivatives ($\partial u/\partial n$) on Dirichlet and Neumann boundaries.
2. Neural Approximation: The unknown boundary functions are approximated by a shallow network:
   $$u_M(x) = \sum_{j=1}^M \beta_j \sigma(\langle \omega_j, x \rangle + b_j)$$
   where $\sigma$ is an activation function (e.g., cosine or tanh), and $\omega_j, b_j$ are fixed random parameters.
3. Linear System Construction:
   • Substitute the neural approximation into the BIE.
   • Enforce the equation at a set of collocation points $\{x_i\}$ on the boundary.
   • This results in a linear system $A\beta = b$, where matrix $A$ contains integrals of the basis functions against the fundamental solution, and vector $b$ contains known boundary condition terms.
4. Solution: Solve for the optimal coefficients $\beta^*$ using a linear least squares solver (e.g., normal equations or QR decomposition).
5. Reconstruction: The final solution in the domain is reconstructed using the boundary integral representation formula with the solved coefficients.

3. Key Contributions

1. Novel Framework: The paper introduces BNM-RF, the first method to systematically combine Random Feature Methods with Boundary Integral Equations. It bridges the gap between classical BEM and modern scientific machine learning.
2. Avoidance of Non-Convex Optimization: By fixing hidden parameters, the method transforms the PDE solving process into a linear least squares problem, eliminating the need for iterative gradient-based training and its associated instability.
3. Theoretical Convergence Analysis:
   • The authors extend error analysis from strong-form PDEs to the boundary integral setting.
   • They establish convergence rates and error bounds for both the boundary function and the interior solution.
   • The error bound is shown to depend on the fill-in distance of collocation points ($h^\gamma$) and the number of neurons ($1/\sqrt{M}$).
4. Dimensionality Reduction: The method retains the advantage of BIE by reducing the problem dimension (solving only on the boundary) while handling complex and unbounded domains naturally.

4. Numerical Results

The method was tested on Laplace and Helmholtz problems in 2D and 3D, comparing performance against BINN (gradient-based) and classical BEM.

• Convergence Rate (2D Helmholtz): BNM-RF demonstrated convergence rates significantly faster than theoretical predictions. It achieved high accuracy with fewer degrees of freedom compared to traditional BEM.
• Accuracy vs. BINN (2D Laplace & Helmholtz):
  • In 2D tests, BNM-RF achieved superior accuracy compared to BINN (e.g., relative $L_2$ error of $2.02 \times 10^{-5}$ for BNM-RF vs. $5.94 \times 10^{-4}$ for BINN in a flower-shaped domain).
  • BNM-RF showed high stability and low sensitivity to random parameter initialization ranges, whereas BINN suffers from training instability.
• Exterior Problems: BNM-RF performed comparably to BEM in exterior Helmholtz problems (acoustic scattering) and significantly outperformed BINN.
• 3D Acoustic Scattering: In a 3D sphere scattering benchmark, BNM-RF achieved accuracy comparable to BEM using only a small number of neurons (32–64).
  • Limitation: Slightly higher errors than BEM were observed, attributed to the insufficient treatment of singular integrals and geometric approximation errors in the current implementation.

5. Significance and Future Work

• Practicality: BNM-RF offers a practical, theoretically grounded framework for solving PDEs on complex geometries without the computational overhead of training deep networks.
• Efficiency: It provides a "one-shot" solution via linear algebra, making it highly efficient for problems where gradient-based training is too slow or unstable.
• Future Directions:
  • Improving the numerical treatment of singular and nearly singular integrals to enhance 3D accuracy.
  • Integrating Fast Multipole Methods (FMM) to accelerate the solution of the dense linear systems arising from BIE.
  • Extending the framework to time-dependent, variable-coefficient, and nonlinear PDEs.

In summary, the paper presents a robust alternative to gradient-based neural PDE solvers by leveraging the linear solvability of random features within the boundary integral formulation, achieving competitive accuracy with significantly reduced training complexity.