Physics-Informed Acoustic Model Overview

• PIAM is a modeling paradigm that integrates governing acoustic laws such as the wave and Helmholtz equations into learning-based inference.
• It enhances sound field reconstruction by embedding physical constraints through methods like boundary integrals, covariance kernels, or neural operators.
• Various architectures including PINNs, boundary-integral networks, and constrained Gaussian processes enable robust simulation and inverse problem solving.

A Physics-Informed Acoustic Model (PIAM) is a class of acoustic modeling methods in which the estimator is constrained by acoustical structure rather than trained as a purely data-driven interpolator. Across the literature, this structure may be the wave equation, the Helmholtz equation, the Webster or horn equation, boundary integral representations, geometry-aware covariance constructions, or explicit propagation mechanisms such as reflection and scattering. In that sense, PIAM denotes a family of models rather than a single architecture: coordinate-based PINNs for wavefields, boundary-integral networks for source-free sound field reconstruction, boundary-constrained Gaussian processes for acoustic emission mapping, neural operators for scattering, and mechanistic simulators for bioacoustic propagation all fall within the same design principle of combining learning or inference with governing acoustics (Moseley et al., 2020, Damiano et al., 4 Jun 2025, Jones et al., 2022, Koyama et al., 2024, Shen et al., 7 Jul 2026).

1. Definition and conceptual scope

The unifying premise of PIAM is that acoustic estimation should not be treated as generic regression when the target field is known to obey wave physics. In the sound-field-estimation literature, this is stated explicitly: purely data-driven interpolation may violate acoustic wave physics, produce physically impossible spatial patterns, and generalize poorly when observations are sparse, whereas acoustic physics priors are valuable a priori information (Koyama et al., 2024). The same logic appears in seismic wavefield prediction, room sound-field reconstruction, duct acoustics, and scattering, where the governing PDE or boundary representation is elevated from background knowledge to an explicit part of the model (Moseley et al., 2020, Karakonstantis et al., 2024, Nair et al., 2024).

PIAMs differ in what counts as the “physics” being imposed. In standard PINN formulations, the governing PDE appears as a residual penalty evaluated by automatic differentiation. In boundary-integral formulations, the field is not learned directly throughout the domain; instead, the model learns a boundary quantity and reconstructs the interior by the Kirchhoff–Helmholtz integral, so the acoustic equation is satisfied by construction in the source-free region (Damiano et al., 4 Jun 2025). In acoustic emission mapping, the “physics-informed” content is weaker but still explicit: a Gaussian-process prior is restricted to the physical domain and satisfies a homogeneous Neumann boundary condition through Laplacian eigenfunctions defined on the geometry (Jones et al., 2022).

This breadth matters because the label “physics-informed” is often associated only with PINNs. The cited work shows a broader taxonomy: basis-expansion methods, kernel methods, neural implicit representations, operator-learning systems, and geometric/path-based propagation simulators can all instantiate the same principle when acoustic laws, boundary conditions, or propagation constraints define the admissible solution space (Koyama et al., 2024, Nair et al., 2024, Shen et al., 7 Jul 2026). A plausible implication is that PIAM is best understood as a modeling doctrine—learning under acoustic admissibility—rather than as the name of one canonical network.

2. Governing acoustic priors

The most common PIAM prior is the acoustic wave equation. In 2D seismic PINN work, the enforced PDE is the constant-density acoustic wave equation

$\nabla^{2}u - \frac{1}{v^{2}}\partial_{tt}u = 0,$

with the network representing the pressure response $u(t,x)$ as a continuous function of space-time coordinates and, in the Marmousi extension, source position (Moseley et al., 2020). Room impulse response reconstruction in the time domain uses the same physical principle, with the inhomogeneous wave equation written as

$$\nabla^2 p(t,r) - \frac{1}{c^2}\frac{\partial^2 p(t,r)}{\partial t^2} = \chi(t,r_s),$$

and the PINN residual enforced in homogeneous form inside the modeled region (Karakonstantis et al., 2024).

In frequency-domain acoustics, the dominant prior is the Helmholtz equation. For source-free sound field reconstruction, the interior pressure satisfies

$$(\nabla^2 + k^2)p(\mathbf r,\omega)=0, \qquad k=\frac{\omega}{c},$$

and this equation underlies both Helmholtz PINNs and the boundary-integral formulation used in PIBI-Net (Damiano et al., 4 Jun 2025). In scattering, the same equation governs the scattered pressure field with sound-hard internal boundaries and absorbing or impedance outer boundaries (Nair et al., 2024, Nair et al., 2024). For low-frequency room acoustics with absorption, the field is complex-valued, and one study splits the inhomogeneous Helmholtz problem into coupled real and imaginary residual equations under homogeneous Neumann boundary conditions on all cube faces (Schoder et al., 18 May 2025).

Tube and vocal-tract acoustics introduce 1D reduced-order PDEs. ResoNet and related tube-reconstruction models enforce a damped horn or Webster-type equation with viscous and thermal losses,

$$\phi_{xx} + \frac{1}{A}A_x\phi_x = GR\phi + \left(\frac{G\rho}{A} + \frac{RA}{K}\right)\phi_t + \frac{\rho}{K}\phi_{tt},$$

along with input-flow, radiation, and periodicity constraints (Yokota et al., 2023, Luan et al., 18 May 2025). In singing-voice synthesis, the time-domain Webster equation is written in velocity-potential form as

$$\frac{1}{c^2}\partial_{tt}\psi - \frac{1}{A(x)}\partial_x\!\big(A(x)\partial_x\psi\big)=0,$$

with a Robin radiation condition at the lip end and a glottal volume-flow boundary at the input (Lu et al., 14 Feb 2026).

Other PIAMs use specialized priors. Acoustic VTI wave propagation is formulated as a coupled anisotropic frequency-domain system for pressure and an auxiliary field, with the network learning the scattered field to avoid the point-source singularity (Song et al., 2020). Acoustic emission mapping constructs the covariance kernel from Laplacian eigenfunctions satisfying

$$-\nabla^2\phi_j(\mathbf x)=\lambda_j^2\phi_j(\mathbf x), \qquad d\phi_j(\mathbf x)=0 \ \text{on}\ \delta\Omega,$$

thereby embedding geometry and homogeneous Neumann behavior into every GP sample (Jones et al., 2022). In a different use of the acronym, a multimodal speech model regularizes a latent sound representation by the lossless Westervelt equation,

$$\nabla^2 p - \frac{1}{c_0^2}\frac{\partial^2 p}{\partial t^2} = - \frac{\beta}{\rho_0 c_0^4}\frac{\partial^2 p^2}{\partial t^2},$$

to capture nonlinear acoustic distortions in teleconference audio (Chen et al., 26 Aug 2025).

3. Model families and representations

The literature contains several distinct PIAM architectures.

Family	Core physical mechanism	Representative papers
Coordinate-based PINN	PDE residual on continuous field $u(x,t)$ or $p(x)$	(Moseley et al., 2020, Yokota et al., 2023, Luan et al., 18 May 2025)
Boundary-integral network	Kirchhoff–Helmholtz reconstruction from learned boundary density	(Damiano et al., 4 Jun 2025)
Boundary-constrained GP	Laplacian eigenbasis satisfying domain and boundary constraints	(Jones et al., 2022)
Operator and point-cloud models	Geometry-parameterized solution operator or multi-domain PDE residuals	(Nair et al., 2024, Wang et al., 2024)
Mechanistic propagation simulator	Explicit direct, reflected, and scattered path assembly	(Shen et al., 7 Jul 2026)

The canonical PIAM form is the coordinate-based PINN. In seismic wavefield prediction, the network is a fully connected feedforward model with 10 layers, 1024 hidden channels, softplus activations in all hidden layers, and a linear output layer; in the Marmousi setting its input becomes $u(t,x)$0, so one trained model can represent a family of wavefields indexed by source position (Moseley et al., 2020). ResoNet uses two neural-network blocks—a wave-equation network $u(t,x)$1 and a radiation network $u(t,x)$2—with residual blocks and the Snake activation $u(t,x)$3, chosen because it outperformed tanh and sin in preliminary tests (Yokota et al., 2023). The tube-reconstruction PINN adopts a deep residual network with Random Fourier Feature Embedding, Snake activation, and 643,401 trainable parameters (Luan et al., 18 May 2025).

Boundary-integral PIAMs move the learned quantity from the interior to the boundary. PIBI-Net uses a shallow multilayer perceptron with 3 fully connected layers, 64 neurons per hidden layer, and a single-neuron output layer to estimate the boundary pressure density $u(t,x)$4. The interior field is then reconstructed by the Kirchhoff–Helmholtz integral rather than by direct network output (Damiano et al., 4 Jun 2025).

Geometry-aware operator models generalize across domains. MPIPN builds on a PointNet-style architecture in which explicit physical quantities are concatenated with point-cloud coordinates, local and global features are extracted, implicit quantities such as density and Young’s modulus are separately encoded, and the fused representation is trained by adaptive multi-domain physics-informed losses (Wang et al., 2024). PGI-DeepONet represents scatterer shape by 16 NURBS control-point parameters and evaluates the solution operator through a branch–trunk architecture composed of two ResNets with 5 residual blocks, 3 linear layers per block, 100 neurons, and Sine activations (Nair et al., 2024).

Some PIAMs are not neural PDE solvers at all. ForestIR is a path-based simulator in which the source–microphone impulse response is assembled as a sum of direct propagation, ground reflection, single trunk scattering, and optional branch/leaf scattering terms. The environment influences the recording only through these impulse responses, which are computed from geometry, atmospheric conditions, and array layout (Shen et al., 7 Jul 2026). The acoustic emission GP is similarly non-neural: its key representational object is a covariance expansion on 256 Laplacian eigenbases defined on the actual plate geometry (Jones et al., 2022).

4. Objectives, optimization, and inference

Most PINN-style PIAMs use hybrid objectives combining sparse observations with physics residuals. In the 2D wave-equation model, the loss adds an FD-data term to a PDE-residual term,

$u(t,x)$5

with $u(t,x)$6 in the constant-density experiments (Moseley et al., 2020). The room-impulse-response PINN uses an adaptive-weighted loss with learned $u(t,x)$7 and $u(t,x)$8 parameters balancing data and PDE residuals (Karakonstantis et al., 2024). In scattering, the loss is a weighted sum of interior Helmholtz residual, rigid-boundary residual, and outer-boundary residual, optimized first with Adam and then with L-BFGS (Nair et al., 2024).

A distinctive subgroup augments the PDE term with periodicity and coupling constraints. ResoNet minimizes

$u(t,x)$9

where $$\nabla^2 p(t,r) - \frac{1}{c^2}\frac{\partial^2 p(t,r)}{\partial t^2} = \chi(t,r_s),$$0 enforces equality of pressure, velocity, and second time derivatives between $$\nabla^2 p(t,r) - \frac{1}{c^2}\frac{\partial^2 p(t,r)}{\partial t^2} = \chi(t,r_s),$$1 and one period $$\nabla^2 p(t,r) - \frac{1}{c^2}\frac{\partial^2 p(t,r)}{\partial t^2} = \chi(t,r_s),$$2, allowing the network to learn the steady resonant cycle directly instead of simulating a long transient (Yokota et al., 2023). The tube reconstruction model uses the same pattern—PDE, input boundary, periodicity, and observation losses—and extends it to inverse identification of unknown radiation coefficients $$\nabla^2 p(t,r) - \frac{1}{c^2}\frac{\partial^2 p(t,r)}{\partial t^2} = \chi(t,r_s),$$3 by PINN Fine-Tuning Method (PINN-FTM) (Luan et al., 18 May 2025).

Training schedules are often problem-specific. The seismic PINN first trains only on FD boundary data, then turns on the physics residual halfway through training, while linearly increasing the temporal horizon of residual collocation points toward $$\nabla^2 p(t,r) - \frac{1}{c^2}\frac{\partial^2 p(t,r)}{\partial t^2} = \chi(t,r_s),$$4 s. This curriculum is intended to mimic time stepping and improve convergence on an oscillatory PDE (Moseley et al., 2020). In low-frequency room acoustics, discrepancy learning first pretrains the network on a Green’s-function-generated baseline and then fine-tunes under PDE and boundary-condition residuals; only selected layers may be unfrozen in some variants (Schoder et al., 18 May 2025).

Boundary-integral PIAMs alter the optimization problem fundamentally. PIBI-Net minimizes only the mismatch between measured microphone pressure and the pressure reconstructed from the discretized boundary integral,

$$\nabla^2 p(t,r) - \frac{1}{c^2}\frac{\partial^2 p(t,r)}{\partial t^2} = \chi(t,r_s),$$5

with no PDE-residual hyperparameter $$\nabla^2 p(t,r) - \frac{1}{c^2}\frac{\partial^2 p(t,r)}{\partial t^2} = \chi(t,r_s),$$6, because the Helmholtz equation is enforced through the representation itself rather than as a penalty (Damiano et al., 4 Jun 2025). Bayesian PIAMs replace gradient training by posterior inference: the neural-physical and Bayesian spectral propagation framework uses forward and backward propagation relations in the likelihood and employs NUTS/HMC to infer the attenuation and wavenumber coefficients $$\nabla^2 p(t,r) - \frac{1}{c^2}\frac{\partial^2 p(t,r)}{\partial t^2} = \chi(t,r_s),$$7, yielding uncertainty over room impulse response and relocalization quantities (Huang et al., 2023).

5. Application domains and empirical behavior

A major early demonstration of PIAM in acoustics is wavefield simulation for heterogeneous media. The 2D acoustic PINN was tested on homogeneous, layered, and Earth-realistic Marmousi velocity models. It accurately simulated the wavefield across these cases, generalized far outside the first $$\nabla^2 p(t,r) - \frac{1}{c^2}\frac{\partial^2 p(t,r)}{\partial t^2} = \chi(t,r_s),$$8 or $$\nabla^2 p(t,r) - \frac{1}{c^2}\frac{\partial^2 p(t,r)}{\partial t^2} = \chi(t,r_s),$$9 FD time steps used for training, and in the homogeneous case predicted wavefields more than $$(\nabla^2 + k^2)p(\mathbf r,\omega)=0, \qquad k=\frac{\omega}{c},$$0 later in time than the boundary data. Once trained, it queried arbitrary space-time locations in about $$(\nabla^2 + k^2)p(\mathbf r,\omega)=0, \qquad k=\frac{\omega}{c},$$1 s, whereas a 2D FD simulation of the full medium took about $$(\nabla^2 + k^2)p(\mathbf r,\omega)=0, \qquad k=\frac{\omega}{c},$$2 s per source, so inference was roughly $$(\nabla^2 + k^2)p(\mathbf r,\omega)=0, \qquad k=\frac{\omega}{c},$$3 times faster per query (Moseley et al., 2020).

In source-free sound field reconstruction, PIBI-Net reported stronger high-frequency behavior than standard Helmholtz PINNs. In a 2D rectangular room simulated with Pyroomacoustics, it outperformed PIDL across all tested frequencies, significantly outperformed PINN above about $$(\nabla^2 + k^2)p(\mathbf r,\omega)=0, \qquad k=\frac{\omega}{c},$$4 Hz, kept NCC close to $$(\nabla^2 + k^2)p(\mathbf r,\omega)=0, \qquad k=\frac{\omega}{c},$$5 over the full frequency range, remained more robust when the number of microphones decreased, and achieved the same NMSE as PINN with fewer microphones (Damiano et al., 4 Jun 2025). Time-domain room reconstruction shows a complementary pattern: with 900 measured receiver positions available and only 100 used for training, the mMLP-PINN reconstructed the early part of the RIR particularly well and achieved about $$(\nabla^2 + k^2)p(\mathbf r,\omega)=0, \qquad k=\frac{\omega}{c},$$6 dB lower RMSE than the other tested neural models in the model-selection comparison (Karakonstantis et al., 2024).

Tube and resonance applications emphasize inverse identification under limited observations. ResoNet matched CTCS finite-difference solutions closely in the time domain, with most regions below $$(\nabla^2 + k^2)p(\mathbf r,\omega)=0, \qquad k=\frac{\omega}{c},$$7 error, and identified the cutoff-frequency-based loss parameter to within about $$(\nabla^2 + k^2)p(\mathbf r,\omega)=0, \qquad k=\frac{\omega}{c},$$8 error after $$(\nabla^2 + k^2)p(\mathbf r,\omega)=0, \qquad k=\frac{\omega}{c},$$9 epochs, remaining similarly accurate under $$\phi_{xx} + \frac{1}{A}A_x\phi_x = GR\phi + \left(\frac{G\rho}{A} + \frac{RA}{K}\right)\phi_t + \frac{\rho}{K}\phi_{tt},$$0 Gaussian noise (Yokota et al., 2023). In the unknown-radiation tube problem, PINN-FTM reconstructed the full field from noisy pressure measurements only at the open end and estimated $$\phi_{xx} + \frac{1}{A}A_x\phi_x = GR\phi + \left(\frac{G\rho}{A} + \frac{RA}{K}\right)\phi_t + \frac{\rho}{K}\phi_{tt},$$1, $$\phi_{xx} + \frac{1}{A}A_x\phi_x = GR\phi + \left(\frac{G\rho}{A} + \frac{RA}{K}\right)\phi_t + \frac{\rho}{K}\phi_{tt},$$2 against ground truth $$\phi_{xx} + \frac{1}{A}A_x\phi_x = GR\phi + \left(\frac{G\rho}{A} + \frac{RA}{K}\right)\phi_t + \frac{\rho}{K}\phi_{tt},$$3, $$\phi_{xx} + \frac{1}{A}A_x\phi_x = GR\phi + \left(\frac{G\rho}{A} + \frac{RA}{K}\right)\phi_t + \frac{\rho}{K}\phi_{tt},$$4, whereas the post hoc traditional optimization method failed badly on $$\phi_{xx} + \frac{1}{A}A_x\phi_x = GR\phi + \left(\frac{G\rho}{A} + \frac{RA}{K}\right)\phi_t + \frac{\rho}{K}\phi_{tt},$$5 under the same noisy conditions (Luan et al., 18 May 2025).

Boundary-constrained statistical PIAMs have been effective when training coverage is sparse or incomplete. For acoustic emission $$\phi_{xx} + \frac{1}{A}A_x\phi_x = GR\phi + \left(\frac{G\rho}{A} + \frac{RA}{K}\right)\phi_t + \frac{\rho}{K}\phi_{tt},$$6 mapping on a complex plate with holes, the constrained GP used geometry-aware Laplacian eigenfunctions and homogeneous Neumann boundary behavior to improve predictions near inaccessible edges. In the hardest reported case—30 mm spacing with no boundary measurements—the average nMSE across all 28 sensor pairs was 7.30 for the constrained GP and 16.91 for the standard GP; computational complexity was also reduced from $$\phi_{xx} + \frac{1}{A}A_x\phi_x = GR\phi + \left(\frac{G\rho}{A} + \frac{RA}{K}\right)\phi_t + \frac{\rho}{K}\phi_{tt},$$7 for a standard GP to $$\phi_{xx} + \frac{1}{A}A_x\phi_x = GR\phi + \left(\frac{G\rho}{A} + \frac{RA}{K}\right)\phi_t + \frac{\rho}{K}\phi_{tt},$$8 for the constrained model (Jones et al., 2022).

Scattering and operator-learning PIAMs target forward simulation across geometries. For multiple rigid-scatterer Helmholtz problems, the superposition PINN improved substantially over a single baseline PINN: in a $$\phi_{xx} + \frac{1}{A}A_x\phi_x = GR\phi + \left(\frac{G\rho}{A} + \frac{RA}{K}\right)\phi_t + \frac{\rho}{K}\phi_{tt},$$9 lattice at 500 Hz, the $$\frac{1}{c^2}\partial_{tt}\psi - \frac{1}{A(x)}\partial_x\!\big(A(x)\partial_x\psi\big)=0,$$0-PINN reached $$\frac{1}{c^2}\partial_{tt}\psi - \frac{1}{A(x)}\partial_x\!\big(A(x)\partial_x\psi\big)=0,$$1 and $$\frac{1}{c^2}\partial_{tt}\psi - \frac{1}{A(x)}\partial_x\!\big(A(x)\partial_x\psi\big)=0,$$2, which was described as a 75% improvement in prediction accuracy over the corresponding $$\frac{1}{c^2}\partial_{tt}\psi - \frac{1}{A(x)}\partial_x\!\big(A(x)\partial_x\psi\big)=0,$$3-PINN case (Nair et al., 2024). PGI-DeepONet generalized across unseen NURBS-defined rigid shapes and reduced average simulation time from $$\frac{1}{c^2}\partial_{tt}\psi - \frac{1}{A(x)}\partial_x\!\big(A(x)\partial_x\psi\big)=0,$$4 s to $$\frac{1}{c^2}\partial_{tt}\psi - \frac{1}{A(x)}\partial_x\!\big(A(x)\partial_x\psi\big)=0,$$5 s, roughly a $$\frac{1}{c^2}\partial_{tt}\psi - \frac{1}{A(x)}\partial_x\!\big(A(x)\partial_x\psi\big)=0,$$6 speedup, while reporting mean $$\frac{1}{c^2}\partial_{tt}\psi - \frac{1}{A(x)}\partial_x\!\big(A(x)\partial_x\psi\big)=0,$$7 for circular shapes and $$\frac{1}{c^2}\partial_{tt}\psi - \frac{1}{A(x)}\partial_x\!\big(A(x)\partial_x\psi\big)=0,$$8 for arbitrary shapes (Nair et al., 2024).

Other PIAMs export physics-constrained parameters rather than full fields. The Webster-based singing backend learned a positive area function $$\frac{1}{c^2}\partial_{tt}\psi - \frac{1}{A(x)}\partial_x\!\big(A(x)\partial_x\psi\big)=0,$$9 and a radiation coefficient $$-\nabla^2\phi_j(\mathbf x)=\lambda_j^2\phi_j(\mathbf x), \qquad d\phi_j(\mathbf x)=0 \ \text{on}\ \delta\Omega,$$0, then validated them out of graph with an independent FDTD Webster solver. For /a/ and /u/, post-rendering improved LSD by about 6–9 dB relative to the DDSP-only baseline; the learned $$-\nabla^2\phi_j(\mathbf x)=\lambda_j^2\phi_j(\mathbf x), \qquad d\phi_j(\mathbf x)=0 \ \text{on}\ \delta\Omega,$$1 clustered around $$-\nabla^2\phi_j(\mathbf x)=\lambda_j^2\phi_j(\mathbf x), \qquad d\phi_j(\mathbf x)=0 \ \text{on}\ \delta\Omega,$$2, and the exported controls remained stable under discretization changes, moderate source variations, and about ten percent pitch shifts (Lu et al., 14 Feb 2026). ForestIR plays a similar role for bioacoustics: it linked forest layout, atmospheric state, ground type, and array geometry to synthetic recordings, and in field comparison obtained mean EDC Pearson correlation 0.837 versus 0.251 for a legacy model (Shen et al., 7 Jul 2026).

6. Limitations and open technical issues

The dominant technical difficulty across PIAMs is the oscillatory, multi-scale nature of acoustic fields. Wavefields contain sharp fronts, interference, wide amplitude ranges, and high frequencies; PINNs can struggle to represent this efficiently, and acoustics is repeatedly described as harder than smoother PDE settings (Moseley et al., 2020). In the layered seismic example, errors were larger near velocity interfaces, and without smoothing the layered-model accuracy dropped significantly (Moseley et al., 2020). In room-acoustic convergence studies, at least six training points per wavelength were necessary for accurate training and subsequent predictions, while highly localized sources with $$-\nabla^2\phi_j(\mathbf x)=\lambda_j^2\phi_j(\mathbf x), \qquad d\phi_j(\mathbf x)=0 \ \text{on}\ \delta\Omega,$$3 caused complete failure of convergence under the tested architectures and point densities (Schoder et al., 18 May 2025).

Many PIAMs rely on restrictive validity conditions. PIBI-Net assumes a source-free reconstruction region and known boundary $$-\nabla^2\phi_j(\mathbf x)=\lambda_j^2\phi_j(\mathbf x), \qquad d\phi_j(\mathbf x)=0 \ \text{on}\ \delta\Omega,$$4, was evaluated only in 2D simulation, and does not yet handle regions containing sources or explicit boundary conditions (Damiano et al., 4 Jun 2025). The acoustic-emission GP is explicit that it is not a complete physics model of AE propagation, because it incorporates boundary-condition knowledge rather than full wave physics (Jones et al., 2022). ForestIR is likewise not a full wave simulator; it assumes linear time-invariant source–microphone channels, single scattering for trunks, image-source ground reflection, and simplified ground reflectivity (Shen et al., 7 Jul 2026).

Inverse problems add identifiability issues. The vocal-tract Webster model states that the inverse problem is not uniquely identifiable from single-channel sustained vowels, so the recovered $$-\nabla^2\phi_j(\mathbf x)=\lambda_j^2\phi_j(\mathbf x), \qquad d\phi_j(\mathbf x)=0 \ \text{on}\ \delta\Omega,$$5 is to be interpreted as a spectrally equivalent control parameterization rather than a guaranteed anatomical reconstruction (Lu et al., 14 Feb 2026). In tube-loss identification, the dominant thermal-loss constant could be recovered accurately, while the weaker viscous-loss constant remained poorly identified because the boundary pressure waveform was much less sensitive to it (Yokota et al., 2024). MPIPN explicitly notes that PDE residuals alone admit degenerate solutions such as $$-\nabla^2\phi_j(\mathbf x)=\lambda_j^2\phi_j(\mathbf x), \qquad d\phi_j(\mathbf x)=0 \ \text{on}\ \delta\Omega,$$6 or $$-\nabla^2\phi_j(\mathbf x)=\lambda_j^2\phi_j(\mathbf x), \qquad d\phi_j(\mathbf x)=0 \ \text{on}\ \delta\Omega,$$7, so some labeled observations remain necessary even in a physics-informed regime (Wang et al., 2024).

There are also architecture-specific failure modes. High-frequency scattering requires more neurons, more residual blocks, and more epochs; very high frequency degrades accuracy if sampling is not sufficiently dense (Nair et al., 2024). Uniform boundary discretization produces larger errors at sharp edges in PGI-DeepONet (Nair et al., 2024). In room impulse response reconstruction, performance decreases when extrapolating far from the measured region, and late reverberant fields are harder to model accurately than the early RIR (Karakonstantis et al., 2024). In the singing renderer, the in-graph waveform remains breathier than the reference, a “periodicity gap” that motivated periodicity-aware objectives and explicit glottal priors (Lu et al., 14 Feb 2026).

7. Terminological scope and acronym reuse

The cited literature does not use PIAM with a single narrow meaning. In acoustics proper, the term covers models that enforce physics exactly through admissible basis or integral representations, softly through PDE residual penalties, structurally through covariance or architecture design, or mechanistically through explicit propagation formulas (Koyama et al., 2024, Damiano et al., 4 Jun 2025, Jones et al., 2022, Shen et al., 7 Jul 2026). This suggests that “physics-informed acoustic model” is best treated as an umbrella designation for acoustically constrained inference.

The acronym is also reused outside conventional sound-field or wave-propagation tasks. A multimodal financial-risk system names its speech front end PIAM: a 300M-parameter, multi-task, bilingual acoustic model that predicts both emotion labels and transcriptions from raw teleconference audio, while regularizing a latent sound-pressure quantity by a Westervelt-equation residual. In that setting, PIAM is not used for field simulation or reconstruction, but for extracting paralinguistic signatures robust to clipping and codec distortion; the full multimodal framework explained up to 43.8% of the out-of-sample variance in 30-day realized volatility while not predicting directional stock returns (Chen et al., 26 Aug 2025). The reuse is terminological rather than methodological continuity.

A common misconception is therefore that PIAM is synonymous with “PINN for acoustics.” The literature does not support that restriction. Boundary-integral networks eliminate interior PDE penalties entirely by reconstructing through the Kirchhoff–Helmholtz equation; constrained Gaussian processes encode boundary admissibility in the prior; DeepONet and PointNet variants learn geometry-conditioned solution operators; mechanistic simulators such as ForestIR are physics-informed without being learned surrogates in the usual sense (Damiano et al., 4 Jun 2025, Jones et al., 2022, Nair et al., 2024, Wang et al., 2024, Shen et al., 7 Jul 2026). The stable core meaning is not the choice of learner but the decision to let acoustics define the feasible hypothesis class.