Virtual boundary integral neural network for three-dimensional exterior acoustic problems

Ilia Marchevskiy; Jiahao Li; Qiang Xi; Zhuojia Fu

arxiv: 2604.18636 · v1 · submitted 2026-04-19 · 💻 cs.SD · cs.LG

Virtual boundary integral neural network for three-dimensional exterior acoustic problems

Jiahao Li , Qiang Xi , Ilia Marchevskiy , Zhuojia Fu This is my paper

Pith reviewed 2026-05-10 06:07 UTC · model grok-4.3

classification 💻 cs.SD cs.LG

keywords virtual boundary integralneural networkexterior acousticsacoustic scatteringSommerfeld radiation conditionBurton-Miller formulationthree-dimensional acoustics

0 comments

The pith

A neural network parametrizes source density on an internal virtual boundary to solve three-dimensional exterior acoustic problems without singular integrals.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces a virtual boundary integral neural network for three-dimensional exterior acoustic problems. It places a virtual surface inside the scatterer or vibrating body and represents the associated source density with a neural network. When coupled to the acoustic fundamental solution, this construction satisfies the Sommerfeld radiation condition by design and allows direct evaluation of pressure and its normal derivative at arbitrary field points. Because the integration surface lies apart from the physical boundary, the method avoids singular and near-singular kernel evaluations that occur in conventional boundary integral approaches. Numerical tests on scattering, multiple-body interaction, and underwater propagation match analytical solutions and commercial software results, with a Burton-Miller extension further stabilizing solutions near characteristic frequencies.

Core claim

By introducing a virtual boundary inside the scatterer or vibrating body and representing the associated source density with a neural network coupled to the acoustic fundamental solution, the formulation satisfies the Sommerfeld radiation condition by construction and enables direct evaluation of the acoustic pressure and its normal derivative at arbitrary field points. Because the integration surface is separated from the physical boundary, the method avoids singular and near-singular kernel evaluations. The geometric parameters of the virtual boundary are optimized jointly with the source density during training to reduce sensitivity to boundary placement.

What carries the argument

Virtual boundary integral representation, in which a neural network parametrizes the source density on an internally placed and jointly optimized virtual surface that is coupled to the acoustic fundamental solution.

Load-bearing premise

The geometric parameters of the virtual boundary can be optimized jointly with the source density during training without introducing new instabilities or biasing the solution.

What would settle it

A test case in which joint optimization of virtual-boundary geometry with source density produces divergence or markedly worse agreement with analytical solutions than a fixed virtual boundary would falsify the robustness claim.

Figures

Figures reproduced from arXiv: 2604.18636 by Ilia Marchevskiy, Jiahao Li, Qiang Xi, Zhuojia Fu.

**Figure 2.** Figure 2: FIG. 2. Schematic of plane wave incidence on a rigid sphere [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗

**Figure 3.** Figure 3: presents contour maps of the total pressure magnitude around the rigid sphere for an incident wavenumber of 1.0 rad/m . The VBINN result in [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗

**Figure 4.** Figure 4: FIG. 4. Convergence of the VBINN under [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗

**Figure 5.** Figure 5: FIG. 5. Comparison of pressure responses before [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗

**Figure 6.** Figure 6: FIG. 6. Geometry of the pea shaped scatterer. [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗

**Figure 7.** Figure 7: FIG. 7. Reconstructed acoustic field and relative [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗

**Figure 8.** Figure 8: FIG. 8. Convergence of the virtual boundary [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗

**Figure 9.** Figure 9: FIG. 9. Geometry of the four spheres scattering system. [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗

**Figure 10.** Figure 10: FIG. 10. Far-field directivity of the four-sphere array for normal incidence and oblique incidence at [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗

**Figure 11.** Figure 11: FIG. 11. Geometric model of the capsule shell and shallow ocean. [PITH_FULL_IMAGE:figures/full_fig_p020_11.png] view at source ↗

**Figure 12.** Figure 12: FIG. 12. Comparison of SPL in the shallow [PITH_FULL_IMAGE:figures/full_fig_p020_12.png] view at source ↗

**Figure 13.** Figure 13: FIG. 13. Spatial distribution of SPL for the capsule shell case. [PITH_FULL_IMAGE:figures/full_fig_p021_13.png] view at source ↗

read the original abstract

This paper presents a virtual boundary integral neural network (VBINN) for exterior acoustic problems in three dimensions. The method introduces a virtual boundary inside the scatterer or vibrating body and represents the associated source density with a neural network. Coupled with the acoustic fundamental solution, this representation satisfies the Sommerfeld radiation condition by construction and enables direct evaluation of the acoustic pressure and its normal derivative at arbitrary field points. Because the integration surface is separated from the physical boundary, the formulation avoids the singular and near singular kernel evaluations associated with coincident source and collocation points in conventional boundary integral learning methods. To reduce sensitivity to boundary placement, the geometric parameters of the virtual boundary are optimized jointly with the source density during training. Numerical examples for acoustic scattering, multiple body interaction, and underwater acoustic propagation show close agreement with analytical solutions and COMSOL results, and the Burton Miller extension further improves stability near characteristic frequencies. These results demonstrate the potential of VBINN for exterior acoustic analysis in three dimensions.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

VBINN puts a trainable virtual surface inside the scatterer and lets an NN stand in for the source density, which dodges singular kernels and satisfies radiation conditions by construction, though the joint geometry optimization still needs checks for stability.

read the letter

The paper's main contribution is a boundary-integral style method for 3D exterior acoustics that moves the integration surface inside the body and approximates the source density with a neural network. Because the virtual surface never coincides with the physical boundary, the kernels stay regular, and the fundamental solution automatically enforces the Sommerfeld condition at infinity. The authors also let the geometric parameters of that virtual surface become trainable variables optimized alongside the network weights, which they say reduces the usual trial-and-error in placing the surface. Numerical tests on scattering, multiple scatterers, and underwater propagation line up with analytical solutions and COMSOL runs, and the Burton-Miller formulation is added to handle fictitious frequencies. Those results are the strongest part of the work so far. The soft spot is the joint optimization itself. Treating boundary location and shape as free parameters is clever on paper, but the abstract gives no ablation on different initial placements, no monitoring of whether the surface stays strictly interior, and no evidence that the optimizer cannot drift into a configuration that simply compensates for network error. Without those checks it is hard to know how robust the method is outside the reported examples. This is aimed at computational engineers who already work with boundary integrals or physics-informed networks and want a mesh-free option for exterior wave problems. Readers who need practical 3D acoustic predictions will get the most out of it. The idea is coherent and the examples are relevant, so it deserves a serious referee who can ask for the missing optimization diagnostics.

Referee Report

2 major / 1 minor

Summary. The paper presents a virtual boundary integral neural network (VBINN) for three-dimensional exterior acoustic problems. A virtual boundary is placed inside the scatterer, with the associated source density represented by a neural network. The acoustic fundamental solution is used to satisfy the Sommerfeld radiation condition by construction, and the separation of virtual and physical boundaries avoids singular kernel evaluations. Geometric parameters of the virtual boundary are optimized jointly with the neural network weights and biases to reduce placement sensitivity. Numerical examples for acoustic scattering, multiple-body interactions, and underwater propagation demonstrate close agreement with analytical solutions and COMSOL results; a Burton-Miller extension is shown to improve stability near characteristic frequencies.

Significance. If the central claims hold, the work provides a mesh-free, singularity-avoiding framework for exterior acoustics that automatically handles radiation conditions and adapts boundary placement. The combination of boundary integrals with neural networks and the Burton-Miller stabilization represents a practical advance for complex 3D problems. Credit is given for the multi-scenario numerical validations and the explicit incorporation of the Burton-Miller formulation to address fictitious frequencies.

major comments (2)

[Abstract and method formulation] Abstract and method description: The central claim that joint optimization of virtual boundary geometric parameters with neural network weights reduces placement sensitivity without introducing instabilities, bias, or degeneracy is load-bearing for the method's reliability, yet no analysis of the optimization landscape, well-posedness of the augmented loss, or ablation studies across random initializations and scatterer shapes is provided to support it.
[Numerical examples] Numerical examples: Reported agreement with analytical solutions and COMSOL is promising, but the absence of detailed error tables, convergence studies, network architecture specifications, and training hyperparameters prevents rigorous assessment of whether the results support the accuracy and robustness claims.

minor comments (1)

[Abstract] The abstract states 'close agreement' without quantitative error metrics; including representative L2 or pointwise errors would strengthen the presentation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments and positive assessment of the potential of VBINN. We address each major comment point by point below and will revise the manuscript to incorporate additional supporting material.

read point-by-point responses

Referee: [Abstract and method formulation] Abstract and method description: The central claim that joint optimization of virtual boundary geometric parameters with neural network weights reduces placement sensitivity without introducing instabilities, bias, or degeneracy is load-bearing for the method's reliability, yet no analysis of the optimization landscape, well-posedness of the augmented loss, or ablation studies across random initializations and scatterer shapes is provided to support it.

Authors: We agree that the joint optimization is central to the method and that further supporting analysis would strengthen the manuscript. The numerical results already show reduced sensitivity through improved agreement when geometric parameters are optimized jointly, but we acknowledge the lack of explicit ablation or landscape analysis. In the revision we will add ablation studies across multiple random initializations and scatterer shapes, together with a discussion of the augmented loss and why the separation of virtual and physical boundaries helps avoid degeneracy or instability. These additions will provide the requested empirical and explanatory support. revision: yes
Referee: [Numerical examples] Numerical examples: Reported agreement with analytical solutions and COMSOL is promising, but the absence of detailed error tables, convergence studies, network architecture specifications, and training hyperparameters prevents rigorous assessment of whether the results support the accuracy and robustness claims.

Authors: We concur that more quantitative detail is needed for rigorous assessment. The current manuscript reports visual and qualitative agreement, but lacks the requested tables and specifications. In the revised version we will insert detailed error tables (including L2 and pointwise errors versus analytical and COMSOL references), convergence studies with respect to network size and training iterations, complete network architecture descriptions (layers, neurons, activations), and a full list of training hyperparameters (optimizer, learning rate, epochs, etc.). This will allow direct evaluation of accuracy and reproducibility. revision: yes

Circularity Check

0 steps flagged

No circularity: derivation uses standard fundamental solution and NN approximation of source density

full rationale

The VBINN construction places a virtual interior surface whose source density is represented by a neural network, then evaluates the exterior field via the known acoustic fundamental solution (which satisfies the radiation condition by construction). Boundary conditions on the physical scatterer are enforced through a loss that is minimized during training, including optional joint optimization of virtual-boundary geometry parameters. This is a standard collocation-style approximation whose output is not algebraically identical to its inputs; the numerical examples compare against independent analytical solutions and COMSOL reference data. No self-definitional equations, fitted-input predictions, or load-bearing self-citations appear in the derivation chain.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The approach rests on standard acoustic theory plus neural-network universal approximation; no new physical entities are introduced.

free parameters (2)

virtual boundary geometric parameters
Optimized jointly with neural-network weights during training to reduce placement sensitivity.
neural network weights and biases
Learned from data to represent source density on the virtual surface.

axioms (2)

standard math The acoustic fundamental solution satisfies the Sommerfeld radiation condition at infinity.
Invoked to ensure the representation satisfies radiation by construction.
domain assumption A neural network can accurately approximate the source density distribution on the virtual boundary.
Central to the method's ability to represent arbitrary source fields.

pith-pipeline@v0.9.0 · 5471 in / 1434 out tokens · 39244 ms · 2026-05-10T06:07:45.237441+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages

[1]

To further show that this converged value is reasonable, we also performed a global sweep over the virtual sphere radius using the conventional virtual boundary el ement method

Convergence of the virtual boundary radius under different initial conditions. To further show that this converged value is reasonable, we also performed a global sweep over the virtual sphere radius using the conventional virtual boundary el ement method. As shown in Fig. 8(b), the optimal radius is found to be 0.88R  by comparing the mean squared error...

work page 2000
[2]

Methods for reconstructing acoustic quantities based on acoustic pressure measurements,

Spatial distribution of SPL for the capsule shell case. Furthermore, Fig. 13 shows the spatial distributi on of the sound pressure level at an excitation frequency of 300 Hz. The spatial distributions predicted by VBINN also agree well with the COMSOL results. These comparisons validat e the effectiveness of VBINN for predicting underwater acoustic propag...

work page 2008

[1] [1]

To further show that this converged value is reasonable, we also performed a global sweep over the virtual sphere radius using the conventional virtual boundary el ement method

Convergence of the virtual boundary radius under different initial conditions. To further show that this converged value is reasonable, we also performed a global sweep over the virtual sphere radius using the conventional virtual boundary el ement method. As shown in Fig. 8(b), the optimal radius is found to be 0.88R  by comparing the mean squared error...

work page 2000

[2] [2]

Methods for reconstructing acoustic quantities based on acoustic pressure measurements,

Spatial distribution of SPL for the capsule shell case. Furthermore, Fig. 13 shows the spatial distributi on of the sound pressure level at an excitation frequency of 300 Hz. The spatial distributions predicted by VBINN also agree well with the COMSOL results. These comparisons validat e the effectiveness of VBINN for predicting underwater acoustic propag...

work page 2008