Benchmarking Physics-Informed Neural Networks and Boundary Elements Methods for Wave Scattering

1. Problem Statement

The study addresses the computational challenge of solving 2D acoustic wave scattering problems governed by the Helmholtz equation in an unbounded domain. Specifically, it models the scattering of a time-harmonic plane wave by a circular, sound-hard obstacle.

• The Core Challenge: Traditional numerical methods (like Finite Element Method) struggle with unbounded domains because they require artificial truncation and fine spatial discretization, which increases computational cost, especially at high frequencies.
• The Alternative: Physics-Informed Neural Networks (PINNs) offer a mesh-free alternative that incorporates physical laws directly into the loss function. However, their performance is often claimed without rigorous comparison to established solvers, leading to questions about their true accuracy and efficiency trade-offs.
• Research Question: How do PINNs compare to the Boundary Element Method (BEM)—a state-of-the-art solver for unbounded wave problems—in terms of accuracy and computational efficiency?

2. Methodology

The authors conducted a controlled benchmark comparing two distinct approaches applied to the same problem under identical conditions.

A. Boundary Element Method (BEM)

• Formulation: The BEM solves the problem by discretizing only the boundary of the scatterer ($\Gamma_I$) rather than the entire domain. It uses the fundamental solution (Green's function) for the 2D Helmholtz equation.
• Radiation Condition: The Sommerfeld radiation condition (ensuring waves decay correctly at infinity) is satisfied exactly by construction through the use of the fundamental solution, eliminating the need for artificial boundary truncation.
• Implementation: The boundary is discretized into elements using Gaussian quadrature ($n_q=8$). The system of integral equations is assembled and solved to find boundary values, which are then used to evaluate the field at interior points.

B. Physics-Informed Neural Networks (PINNs)

• Formulation: A neural network approximates the scattered field $u_{sct}(x, y)$ (outputting real and imaginary parts).
• Domain Truncation: Since PINNs cannot inherently handle infinite domains, the computational domain is truncated at an artificial boundary ($\Gamma_E$). The Sommerfeld radiation condition is approximated using an Absorbing Boundary Condition (ABC): $\nabla u \cdot n + ik u = 0$.
• Loss Function: The network is trained to minimize the sum of squared residuals of:
  1. The Helmholtz equation in the interior ($\Omega_P$).
  2. The Neumann boundary condition on the obstacle ($\Gamma_I$).
  3. The ABC on the artificial boundary ($\Gamma_E$).
• Optimization: The authors employed Optuna for hyperparameter optimization (learning rate, layers, neurons, activation functions). The optimal configuration found was: 3 hidden layers, 25 neurons, Sine activation, and a learning rate of $10^{-2}$.

C. Comparison Strategy

• Fair Baseline: To avoid bias, the authors did not compare a "best-case" PINN against a "worst-case" BEM. Instead, they tuned the PINN hyperparameters to find the best accuracy, then adjusted the BEM discretization (number of integration points) to match that specific accuracy level.
• Metrics:
  • Accuracy: Relative error against an analytical solution.
  • Efficiency: Wall-clock time for assembly/solution (BEM) vs. training (PINN), and evaluation time for both.
  • Generalization: Performance tested outside the training domain (far-field behavior).

3. Key Results

A. Accuracy vs. Computational Cost

• Training/Assembly Time: BEM is significantly faster for the initial solution.
  • BEM: Assembly and solution take $\approx 10^{-2}$ seconds.
  • PINN: Training takes $\approx 10^2$ seconds (roughly 4 orders of magnitude slower than BEM assembly).
• Evaluation Time: Once trained, PINNs are faster for inference.
  • PINN: Evaluation at new points takes $\approx 10^{-2}$ seconds.
  • BEM: Evaluating the solution at interior points takes $\approx 10^0$ seconds (slower than PINN inference).
• Total Cost: For a single forward problem, BEM is vastly superior (Total BEM: ~2.06s vs. Total PINN: ~289.92s).

B. Spatial Error Distribution

• BEM: Errors are slightly higher near the obstacle boundary but remain consistent throughout the domain.
• PINN: Errors are concentrated near the artificial boundary ($\Gamma_E$) where the domain was truncated and the ABC applied.

C. Generalization and Far-Field Behavior

• BEM: Maintains high accuracy even when evaluated far outside the original computational region. The error remains low ($\approx 7.27 \times 10^{-2}$) because the radiation condition is mathematically exact.
• PINN: Suffers a drastic loss of accuracy outside the training domain. The relative error jumps to $\approx 7.09$ (nearly 100% error) in the far field.
  • Reason: The PINN relies on the ABC at the truncated boundary. Without explicit enforcement of the Sommerfeld condition, the network fails to capture the correct asymptotic decay of the wave at large distances.

D. Frequency Sensitivity

• BEM maintains consistent low error across varying wavenumbers ($k r_i$).
• PINN accuracy degrades more significantly as the wavenumber increases, requiring more complex network architectures to maintain performance.

4. Key Contributions

1. Reproducible Benchmark: The study establishes a rigorous, reproducible framework for comparing ML-based solvers with classical numerical methods, emphasizing the need for "strong baselines" (using an optimal BEM rather than a suboptimal one).
2. Quantification of Trade-offs: It provides concrete data showing that while PINNs offer fast inference, their training cost is prohibitive for single forward problems compared to BEM.
3. Identification of Generalization Limits: The paper highlights a critical limitation of standard PINNs in scattering problems: the inability to generalize to the far field due to the lack of exact radiation condition enforcement.
4. Open Source: The code and trained models are made publicly available to ensure transparency and reproducibility.

5. Significance and Conclusion

• Current State: For standard forward wave scattering problems in unbounded domains, BEM remains the superior choice due to its speed, accuracy, and exact handling of radiation conditions. PINNs are currently too computationally expensive for single-shot simulations.
• Potential Utility: PINNs may still be valuable in scenarios requiring repeated evaluations (e.g., inverse problems, optimization, or real-time control) where the high initial training cost is amortized over thousands of fast inference steps.
• Future Directions: The authors propose hybrid approaches as the most promising path forward. Specifically, replacing the differential operator in the PINN loss function with a boundary integral formulation. This would allow PINNs to inherit the exact radiation condition handling of BEM while retaining the mesh-free flexibility of neural networks.

In summary, the paper argues that while PINNs are a flexible tool, they are not yet a direct replacement for classical methods like BEM in wave scattering. Their practical application requires careful consideration of the specific problem type (forward vs. inverse) and the necessity of hybridizing them with classical physics-based formulations to overcome generalization limits.