Helmholtz-Informed Neural Networks

• HINN is a class of neural models that embed the Helmholtz equation to enforce physically valid wave or diffusion solutions.
• They leverage tailored strategies such as PDE residual enforcement, Helmholtz-adapted parameterization, and operator surrogates for robust field reconstruction.
• Applications span acoustics, seismic inversion, optical scattering, and thermal imaging, demonstrating physics-informed improvements over pure data fitting.

A Helmholtz-Informed Neural Network (HINN) is a neural model whose inductive bias, objective, or representation is explicitly structured by the Helmholtz equation or by closely related Helmholtz-derived physics. In the literature represented here, the label itself is not universal: several papers use “physics-informed neural network,” “acoustics-informed neural network,” or more specialized names, while nonetheless embedding the Helmholtz equation, Helmholtz decomposition, or a frequency-domain Helmholtz-like reformulation into training or architecture (Ma et al., 2024, Song et al., 2021, Chen et al., 2023, Saba et al., 2022, Zhu et al., 4 Sep 2025). The common idea is to restrict learned fields toward physically valid wave or diffusion solutions, typically through PDE residuals, structured decompositions, or analytic field splitting, rather than relying on data fitting alone.

1. Conceptual scope and terminology

The most direct form of a HINN treats a field as a neural function constrained by the homogeneous or inhomogeneous Helmholtz equation. In acoustics, this appears as reconstruction of complex pressure fields or acoustic transfer functions under constraints such as

$\nabla^2 P + k^2 P = 0$

in a source-free region (Ma et al., 2024), or as sound-field estimation around a rigid sphere with both a Helmholtz residual and the zero radial velocity condition on the sphere (Chen et al., 2023). In seismic inversion, a PINN-based wavefield reconstruction method embeds the frequency-domain acoustic Helmholtz equation, specifically in scattered-wave form, directly into the loss (Song et al., 2021). In optical diffraction tomography, a U-Net-like forward model is trained with the Helmholtz equation as a physical loss for the scattered field (Saba et al., 2022).

A broader usage includes models that are not strict residual-minimizing Helmholtz solvers but are still organized around Helmholtz structure. “HDNet” is informed by Helmholtz decomposition rather than the scalar Helmholtz PDE: it decomposes a flow field into irrotational and solenoidal parts and is described as a Physics-Inspired Neural Network using Helmholtz synthesis and a learned scalar potential (Qi et al., 2024). The MR-EPT method reconstructs complex electrical properties through a Helmholtz-constrained neural representation of the RF field $B_1^+$ (Yu et al., 2022). The thermal-diffusion work converts the heat equation to the frequency domain and trains what it explicitly calls a HINN on a pseudo-Helmholtz equation (Zhu et al., 4 Sep 2025).

This suggests that “HINN” functions best as an umbrella term for neural methods in which Helmholtz physics is not merely background knowledge but an explicit design ingredient. A plausible implication is that the term covers three recurring mechanisms: residual-based PDE enforcement, Helmholtz-adapted field parameterization, and Helmholtz-derived intermediate structure.

2. Mathematical formulations

The canonical frequency-domain form used across several domains is the Helmholtz equation

$\nabla^2 u + k^2 u = 0$

or its inhomogeneous/scattered variant. In the seismic wavefield reconstruction inversion method, the starting point is

$$\omega^2 m\,u(\mathbf{x},\omega) + \nabla^2 u(\mathbf{x},\omega) = s(\mathbf{x}_s),$$

with the scattered-wave reformulation

$$\omega^2 m\,\delta u + \nabla^2 \delta u = -\omega^2 \delta m\,u_0, \qquad \delta m = m-m_0,$$

which removes the point-source singularity from the learned quantity (Song et al., 2021). The central reconstruction loss combines receiver-data mismatch with the scattered Helmholtz residual, so the network is intentionally allowed to violate the PDE somewhat in order to better fit data, exactly as in WRI (Song et al., 2021).

In acoustics-informed sound field reconstruction, the field inside a source-free region is constrained by

$P + \frac{1}{(\omega/c)^2}\nabla^2 P = 0,$

equivalently

$\nabla^2 P + k^2 P = 0,\qquad k=\frac{\omega}{c},$

with the real and imaginary parts enforced separately because the network is real-valued (Ma et al., 2024). In HRTF upsampling, the governing PDE is likewise written as

$\nabla^2 P + (\omega/c)^2 P = 0,$

and then rearranged as

$\frac{\nabla^2P}{(\omega/c)^2} + P = 0,$

so that the PDE residual has the same physical unit as the pressure itself (Ma et al., 2023).

The rigid-sphere acoustics formulation supplements Helmholtz physics with a Neumann-type boundary condition: $$\left.\frac{\partial P(\mathbf r,\omega)}{\partial r}\right|_{r=a}=0,$$ implemented in Cartesian coordinates through the chain rule and a boundary loss $B_1^+$0 (Chen et al., 2023). This is a direct example of a HINN combining a Helmholtz operator with geometry-specific boundary physics.

Some works move one level below or adjacent to Helmholtz. The FOA room-impulse-response model is time-domain rather than frequency-domain; it enforces linearized momentum and continuity,

$B_1^+$1

and derives FOA-specific losses linking the $B_1^+$2 channel to $B_1^+$3 through spatial and temporal derivatives (Masuyama et al., 9 Jul 2025). The authors explicitly note that this is not a direct Helmholtz PINN, but it is conceptually close because the wave equation follows from those first-order PDEs (Masuyama et al., 9 Jul 2025). The heat-transfer HINN transforms

$B_1^+$4

into the pseudo-Helmholtz equation

$B_1^+$5

or, with $B_1^+$6,

$B_1^+$7

and then learns the real and imaginary parts of the transformed thermal field (Zhu et al., 4 Sep 2025).

This suggests a useful technical distinction. Direct HINNs enforce Helmholtz itself; adjacent formulations enforce PDEs that imply a Helmholtz structure after time-harmonic or Fourier transformation.

3. Neural parameterization strategies

One recurrent strategy is the coordinate-based neural field. The compact acoustics-informed neural network takes Cartesian coordinates $B_1^+$8 as input and returns pressure values, with one network or two decoupled networks for the real and imaginary parts (Ma et al., 2024). The rigid-sphere method similarly uses a fully connected neural network whose input is the 3D Cartesian position $B_1^+$9 and whose output is the estimated sound pressure $\nabla^2 u + k^2 u = 0$0 at a fixed frequency (Chen et al., 2023). The seismic WRI method uses spatial coordinates and source location as input to represent multi-source scattered wavefields with one network (Song et al., 2021).

A second strategy is operator-style image or volume mapping. In diffraction tomography, “MaxwellNet” takes a refractive-index distribution on a Cartesian grid and predicts the corresponding scattered field on the grid, while a discrete Helmholtz residual is computed by finite-difference convolutions rather than coordinate autodiff (Saba et al., 2022). This is not a canonical coordinate PINN; it is better described as a Helmholtz-informed operator surrogate (Saba et al., 2022).

A third strategy is to build Helmholtz structure into the representation itself. The plane-wave activation based neural network replaces conventional activations with

$\nabla^2 u + k^2 u = 0$1

so that a one-hidden-layer network becomes

$\nabla^2 u + k^2 u = 0$2

a trainable superposition of plane waves (Wang et al., 2020). The paper explicitly interprets this as a generalization of the plane wave partition of unity method, with learned amplitudes and directions (Wang et al., 2020). This is a particularly strong form of Helmholtz-informed parameterization because each neuron acts like a Helmholtz-adapted basis function.

A fourth strategy uses decomposition rather than direct field regression. HDNet predicts a scalar potential $\nabla^2 u + k^2 u = 0$3, recovers the irrotational field as $\nabla^2 u + k^2 u = 0$4, and then defines the solenoidal field as the residual, thereby building Helmholtz decomposition into the architecture (Qi et al., 2024). The taper-based scattering formulation rewrites the total field as

$\nabla^2 u + k^2 u = 0$5

where $\nabla^2 u + k^2 u = 0$6 is a $\nabla^2 u + k^2 u = 0$7 taper and $\nabla^2 u + k^2 u = 0$8 is the learned remainder. This converts the homogeneous Helmholtz equation into an inhomogeneous equation for the scattered field, which the authors report enhances and accelerates training (Dörfler et al., 2024).

Capacity selection can also be Helmholtz-informed. In HRTF upsampling, the hidden width is linked to the spherical-harmonic order

$\nabla^2 u + k^2 u = 0$9

and the final width prescription is

$$\omega^2 m\,u(\mathbf{x},\omega) + \nabla^2 u(\mathbf{x},\omega) = s(\mathbf{x}_s),$$0

after splitting the field into four subnetworks (Ma et al., 2023). In sound field reconstruction, the width is tied to the cylinder-harmonic order parameter

$$\omega^2 m\,u(\mathbf{x},\omega) + \nabla^2 u(\mathbf{x},\omega) = s(\mathbf{x}_s),$$1

and only one or two hidden layers with fewer than ten neurons per hidden layer were sufficient in the reported experiments (Ma et al., 2024). This suggests that some HINNs are not only loss-informed by Helmholtz structure but capacity-informed as well.

4. Training mechanisms and physics enforcement

The most common enforcement mechanism is a composite loss with data fidelity and a Helmholtz residual. In compact acoustics-informed reconstruction, for the real part,

$$\omega^2 m\,u(\mathbf{x},\omega) + \nabla^2 u(\mathbf{x},\omega) = s(\mathbf{x}_s),$$2

and

$$\omega^2 m\,u(\mathbf{x},\omega) + \nabla^2 u(\mathbf{x},\omega) = s(\mathbf{x}_s),$$3

with analogous imaginary-part terms (Ma et al., 2024). In the rigid-sphere problem, the loss is

$$\omega^2 m\,u(\mathbf{x},\omega) + \nabla^2 u(\mathbf{x},\omega) = s(\mathbf{x}_s),$$4

combining measurement fit, Helmholtz residual, and rigid-boundary consistency (Chen et al., 2023).

Some models encode additional physical symmetries structurally. Region-to-region sound field reconstruction uses a deep-set architecture

$$\omega^2 m\,u(\mathbf{x},\omega) + \nabla^2 u(\mathbf{x},\omega) = s(\mathbf{x}_s),$$5

which preserves acoustic reciprocity because the representation is permutation invariant in source and receiver positions (Chen et al., 27 Jan 2026). The same model adds a Helmholtz residual term

$$\omega^2 m\,u(\mathbf{x},\omega) + \nabla^2 u(\mathbf{x},\omega) = s(\mathbf{x}_s),$$6

and combines it with supervised data loss as

$$\omega^2 m\,u(\mathbf{x},\omega) + \nabla^2 u(\mathbf{x},\omega) = s(\mathbf{x}_s),$$7

with $$\omega^2 m\,u(\mathbf{x},\omega) + \nabla^2 u(\mathbf{x},\omega) = s(\mathbf{x}_s),$$8 in the reported experiments (Chen et al., 27 Jan 2026).

Automatic differentiation is central in coordinate-based HINNs. It is used to compute second derivatives in sound field reconstruction (Ma et al., 2024), the Laplacian and radial derivatives around the rigid sphere (Chen et al., 2023), the RF-field Laplacian in MR-EPT (Yu et al., 2022), and the first- and second-order derivatives needed in FOA physics-informed priors (Masuyama et al., 9 Jul 2025). Discrete derivative operators appear when the network outputs full grids rather than pointwise values, as in diffraction tomography (Saba et al., 2022).

Several papers show that Helmholtz-informed training benefits from nonstandard optimization or architectural choices. The activation-function study for scattered-field Helmholtz PINNs reports that swish yields superior performance compared to $$\omega^2 m\,u(\mathbf{x},\omega) + \nabla^2 u(\mathbf{x},\omega) = s(\mathbf{x}_s),$$9, $$\omega^2 m\,\delta u + \nabla^2 \delta u = -\omega^2 \delta m\,u_0, \qquad \delta m = m-m_0,$$0, and ELU within a fixed Helmholtz setup (Al-Safwan et al., 2021). Hyper-parameter optimization for Helmholtz PINNs finds that shallow networks, sine activation, and moderate-to-large width are preferred on bounded forward problems (Escapil-Inchauspé et al., 2022). Domain decomposition via FBPINNs is proposed as a way to mitigate the limitations of standard PINNs in the homogeneous Helmholtz equation by using overlapping subdomains and smooth windowing (Dolean et al., 19 Nov 2025).

5. Application domains

The acoustics literature is the densest concentration of explicit HINN-like work in this set of papers. A compact acoustics-informed neural network reconstructs 2D sound pressure and pressure gradients inside a region of interest from measured boundary pressures, outperforming cylinder harmonics and SVD baselines in the reported experiments (Ma et al., 2024). A rigid-sphere PINN estimates the sound field outside a sphere from only $$\omega^2 m\,\delta u + \nabla^2 \delta u = -\omega^2 \delta m\,u_0, \qquad \delta m = m-m_0,$$1 microphones on the sphere, enforcing Helmholtz physics and the zero radial velocity condition (Chen et al., 2023). HRTF upsampling treats measured HRTFs as samples of an acoustic wave field on a sphere and uses a Helmholtz-regularized MLP sized according to spherical-harmonic bandwidth (Ma et al., 2023). Region-to-region sound field reconstruction extends the idea from point-to-region interpolation to continuously varying source and receiver positions under reciprocity-preserving architecture and Helmholtz regularization (Chen et al., 27 Jan 2026).

Seismic inversion provides a different but closely related formulation. PINN-based wavefield reconstruction inversion uses a first network for scattered wavefields and a second network for squared slowness, both trained through the scattered Helmholtz equation (Song et al., 2021). The method is explicitly positioned as avoiding repeated matrix inversions in classical frequency-domain WRI while yielding smooth low-wavenumber velocity models useful for subsequent FWI (Song et al., 2021).

Optics and electromagnetics show that HINN principles transfer beyond acoustics. In diffraction tomography, a U-Net is trained to map refractive-index distributions to scattered fields under a Helmholtz residual, and the learned forward model is then used inside an iterative reconstruction loop (Saba et al., 2022). In MR-based electrical properties tomography, two coordinate MLPs jointly represent the complex RF field $$\omega^2 m\,\delta u + \nabla^2 \delta u = -\omega^2 \delta m\,u_0, \qquad \delta m = m-m_0,$$2 and the complex electrical-property field $$\omega^2 m\,\delta u + \nabla^2 \delta u = -\omega^2 \delta m\,u_0, \qquad \delta m = m-m_0,$$3, coupled by

$$\omega^2 m\,\delta u + \nabla^2 \delta u = -\omega^2 \delta m\,u_0, \qquad \delta m = m-m_0,$$4

so that Laplacians are taken on a smooth neural field rather than by finite differences on noisy measurements (Yu et al., 2022).

The flow and radio-map works broaden the conceptual boundary of HINN. HDNet uses Helmholtz decomposition as a physics prior for divergence-free or curl-free flow estimation (Qi et al., 2024). RadioDiff-$$\omega^2 m\,\delta u + \nabla^2 \delta u = -\omega^2 \delta m\,u_0, \qquad \delta m = m-m_0,$$5 is not a classical residual-based Helmholtz PINN; instead, it uses Helmholtz analysis to derive singularity maps corresponding to regions with $$\omega^2 m\,\delta u + \nabla^2 \delta u = -\omega^2 \delta m\,u_0, \qquad \delta m = m-m_0,$$6, then conditions a dual diffusion model on those physics-derived structures for multipath-aware radio map construction (Wang et al., 22 Apr 2025). This suggests that Helmholtz information can enter not only through PDE residuals but also through intermediate targets and generative conditioning.

The thermal HINN is notable because it explicitly uses the term HINN. It predicts internal temperature distributions without internal measurements by training in the Fourier domain on the pseudo-Helmholtz equation, then reconstructing the transient 3D field by inverse Fourier transform (Zhu et al., 4 Sep 2025). A plausible implication is that Helmholtz-informed design can apply to diffusion problems once time dependence is converted into a frequency-domain complex spatial operator.

6. Empirical behavior, limitations, and recurring misconceptions

The empirical record across these papers is favorable but highly qualified. In 2D sound field reconstruction, the decoupled acoustics-informed network achieves average pressure-reconstruction errors of $$\omega^2 m\,\delta u + \nabla^2 \delta u = -\omega^2 \delta m\,u_0, \qquad \delta m = m-m_0,$$7 dB in the anechoic chamber, $$\omega^2 m\,\delta u + \nabla^2 \delta u = -\omega^2 \delta m\,u_0, \qquad \delta m = m-m_0,$$8 dB in the medium room, and $$\omega^2 m\,\delta u + \nabla^2 \delta u = -\omega^2 \delta m\,u_0, \qquad \delta m = m-m_0,$$9 dB in the small room at 3 kHz, and also improves radial-gradient reconstruction relative to CH and SVD (Ma et al., 2024). Around a rigid sphere, the PINN yields $P + \frac{1}{(\omega/c)^2}\nabla^2 P = 0,$0–$P + \frac{1}{(\omega/c)^2}\nabla^2 P = 0,$1 dB lower NMSE than SH and plane-wave decomposition for $P + \frac{1}{(\omega/c)^2}\nabla^2 P = 0,$2 (Chen et al., 2023). For FOA room impulse responses, the proposed PI-DANF performs best consistently on both $P + \frac{1}{(\omega/c)^2}\nabla^2 P = 0,$3 and $P + \frac{1}{(\omega/c)^2}\nabla^2 P = 0,$4, while the naive $P + \frac{1}{(\omega/c)^2}\nabla^2 P = 0,$5-channel wave prior improves over vanilla DANF (Masuyama et al., 9 Jul 2025). In diffraction tomography, MaxwellNet inference is reported as $P + \frac{1}{(\omega/c)^2}\nabla^2 P = 0,$6 ms in 2D and $P + \frac{1}{(\omega/c)^2}\nabla^2 P = 0,$7 ms in 3D, versus $P + \frac{1}{(\omega/c)^2}\nabla^2 P = 0,$8 s and $P + \frac{1}{(\omega/c)^2}\nabla^2 P = 0,$9 s for COMSOL in the reported settings (Saba et al., 2022).

At the same time, several misconceptions are directly contradicted by the literature. HINNs are not uniformly superior to classical numerical solvers. Hyper-parameter tuning of Helmholtz PINNs on bounded domains shows that FEM remains dramatically faster and often more accurate, especially as frequency increases (Escapil-Inchauspé et al., 2022). HINNs are also not all the same method family. Some are residual-based coordinate networks (Ma et al., 2024), some are two-network inverse formulations (Song et al., 2021), some are operator surrogates (Saba et al., 2022), and some are decomposition-based architectures (Qi et al., 2024). Not every “physics-informed” neural acoustic field is a direct Helmholtz PINN; the FOA model is explicitly time-domain and first-order, though conceptually close (Masuyama et al., 9 Jul 2025). Conversely, not every method with “Helmholtz” in the name concerns the Helmholtz PDE at all; the paper on the Helmholtz machine and the Free Energy Principle is about variational generative modeling rather than wave physics (Liu, 2023).

A recurring limitation is high-frequency difficulty. Hyper-parameter tuning for bounded Helmholtz problems finds that training deteriorates as $\nabla^2 P + k^2 P = 0,\qquad k=\frac{\omega}{c},$0 increases, consistent with spectral bias (Escapil-Inchauspé et al., 2022). The taper-based scattering formulation improves the usable wave-number range from roughly $\nabla^2 P + k^2 P = 0,\qquad k=\frac{\omega}{c},$1 to $\nabla^2 P + k^2 P = 0,\qquad k=\frac{\omega}{c},$2 in the reported waveguide experiments, but does not eliminate spectral bias entirely (Dörfler et al., 2024). Domain-decomposed FBPINNs improve convergence in the homogeneous Helmholtz problem, yet high-frequency cases remain challenging and are sensitive to PML thickness and optimizer choice (Dolean et al., 19 Nov 2025). The activation-function study likewise presents swish as a partial optimization improvement rather than a complete solution to Helmholtz PINN training pathologies (Al-Safwan et al., 2021).

Another recurrent limitation is that many methods are trained per scene, per room, or per frequency rather than amortized across broad distributions. The compact acoustics-informed model is fitted to one sound field instance per frequency/source condition (Ma et al., 2024). The rigid-sphere method is single-frequency and simulation-based (Chen et al., 2023). The thermal HINN reconstructs frequency-domain components and then returns to time domain, but the provided text does not establish a universal cross-material model (Zhu et al., 4 Sep 2025). This suggests that many current HINNs function as physics-regularized inverse solvers rather than as broadly pretrained foundation models.

7. Research directions

Several directions emerge repeatedly. One is stronger Helmholtz-adapted parameterization: plane-wave activations (Wang et al., 2020), scattering-field decompositions (Dörfler et al., 2024), and decomposition architectures (Qi et al., 2024) all indicate that explicit structural bias can outperform generic MLP regression. Another is better optimization: swish activations (Al-Safwan et al., 2021), Gaussian-process-based hyper-parameter optimization (Escapil-Inchauspé et al., 2022), adaptive task weighting (Masuyama et al., 9 Jul 2025), and domain decomposition (Dolean et al., 19 Nov 2025) each target different aspects of training instability. A third is richer physical conditioning: reciprocity-preserving deep sets (Chen et al., 27 Jan 2026), FOA channel priors (Masuyama et al., 9 Jul 2025), and Helmholtz-derived singularity maps for diffusion models (Wang et al., 22 Apr 2025) show that PDE information can enter the model beyond a plain residual norm.

A broader synthesis is that the most effective HINNs do not merely append a Helmholtz penalty to a generic neural network. They reshape the learning problem so that the network sees a reduced, better-conditioned, or symmetry-aware target. The literature here supports that conclusion across acoustics, seismic inversion, optical scattering, RF property reconstruction, flow decomposition, radio-map generation, and thermal imaging (Ma et al., 2024, Song et al., 2021, Saba et al., 2022, Yu et al., 2022, Qi et al., 2024, Wang et al., 22 Apr 2025, Zhu et al., 4 Sep 2025).