Wave-PDE Nets: Neural Wave Dynamics

• Wave-PDE Nets are a family of neural architectures that embed wave propagation dynamics into their design, enabling efficient simulation, operator learning, and solver correction.
• They incorporate techniques such as differentiable PDE solvers, recursive operator learning, and wavelet-based filters to capture both physical trajectories and internal propagation.
• Performance benefits include significant error reduction in solver corrections, faster computations through learned propagators, and enhanced interpretability via physics-informed constructs.

Wave-PDE Nets denotes a family of neural constructions that embed wave-propagation structure into prediction, operator learning, solver correction, or the network layer itself. In the cited literature, the term is used or suggested for data-driven predictors of PDE-generated wave dynamics from observed fields, hybrid schemes of the form $$\Psi_{\Delta t}^\theta=\mathcal I^\theta \mathcal G_{\Delta t}\mathcal R$$ that lift a fast coarse solver toward fine-solver fidelity, architectures whose elementary operation is a differentiable simulation of a wave equation, and wavelet- or filter-structured models that encode multiresolution or differential-operator priors directly into training and representation (Sorteberg et al., 2018, Kaiser et al., 2024, Vejendla, 5 Oct 2025, Long et al., 2017, Ernst et al., 2022).

1. Conceptual scope

Across this literature, the central mathematical object is an evolution or solution operator rather than a static predictor. In the solver-integrated setting, the target map is written as

$$\mathfrak u_{n+1}=\mathcal F_{\Delta t}\mathfrak u_n \approx \Psi_{\Delta t}^\theta \mathfrak u_n=\mathcal I^\theta \mathcal G_{\Delta t}\mathcal R\mathfrak u_n,$$

where restriction, coarse propagation, and learned prolongation are composed end-to-end. In recursive operator learning, the same short-time neural propagator is reused block-by-block over a long interval, so that long-time prediction is obtained by repeated application of a single learned map (Kaiser et al., 2024, Liu et al., 2022).

The cited works do not impose a single training doctrine. Some are fully supervised on trajectories generated by numerical solvers or analytic solutions. Some are trained from PDE residuals or integral conservation forms without large labeled datasets. Others treat a PDE solver as the neural primitive and optimize it on downstream language, vision, or graph-learning tasks. This suggests that “Wave-PDE Nets” is best understood as a structural category defined by how wave dynamics enters the architecture, rather than by a single loss, dataset type, or application domain (Chang et al., 2021, Yue et al., 26 May 2025, Vejendla, 5 Oct 2025).

A recurring distinction within the category is between models that learn wave evolution externally and models that use wave propagation internally. External models approximate physical trajectories, numerical propagators, or Green operators. Internal models propagate hidden states through a trainable medium, with learned wave speed, damping, or Laplacian structure, and use that propagation as the mixing mechanism of the network itself (Aldirany et al., 2023, Vejendla, 5 Oct 2025).

2. Data-driven wave forecasting and operator learning

An early precursor is the 2D shallow-water surrogate in “Approximating the solution to wave propagation using deep neural networks,” which learns future rendered wave fields governed by the Saint-Venant equations in non-conservative form without Coriolis or viscous forcing. The system is bounded by solid-wall boundary conditions, the network observes only five input frames of rendered wave patterns, and the architecture is an encoder–propagator–decoder with four convolutional layers, a 1000-dimensional latent vector, and three fully connected LSTMs that share hidden and cell states across initialization, self-propagation, and re-insertion of rolled-out predictions. It is trained with mean squared error and ADAM in PyTorch, predicts 20 frames ahead during training, and is rolled out to 80 frames at test time by feeding back the last five predicted frames. The reported behavior is a gradual SSIM decline rather than immediate failure, recognizable wave structure up to 80 time steps, qualitative generalization to an unseen propagating linear wave front, and a direct speed comparison of under half a second for 80 predicted frames versus about 35 seconds for the numerical solver (Sorteberg et al., 2018).

A different operator-learning viewpoint appears in the method for time-dependent wave equations over unbounded domains. There the network learns the finite-time solution map

$$u(\cdot,t)\big|_{\Omega}=\mathcal SU_0,\qquad U_T\approx \mathcal N_\theta(U_0),$$

using FCNNs or ResNets trained on Gaussian wave packets parameterized by wave numbers and band width. The method is presented as an alternative to artificial boundary conditions such as absorbing boundary conditions, perfectly matched layers, and Dirichlet-to-Neumann approximations. For the wave equation, interpolation errors are typically around $10^{-3}$, the experiments extend up to 8D, and extrapolation is reported as meaningful but weaker than interpolation, with degradation as the test condition moves farther from the training distribution (Xie et al., 2021).

GreenONets specialize operator learning further by building the Green’s-function representation into the architecture. For the homogeneous wave equation with zero initial velocity, the solution is written as

$u(x,t)=\int_\Omega g(x,t,\xi)\,u_0(\xi)\,d\xi,$

and the network approximates this through the discrete form

$$\hat Q(u_0)(x,t)=\frac1m\sum_{i=1}^m G(x,t,x_i)\,u_0(x_i).$$

Relative to physics-informed DeepONets with the same residual-based loss, GreenONets are reported to decrease loss faster, converge in fewer epochs, and generalize better to higher-frequency test initial conditions; in 1D homogeneous media with GRF length scale $l=0.5$, the maximum pointwise error for test function $(1-x^2)^{10}$ is about $0.04$ for GreenONets and about $0.14$ for DeepONets (Aldirany et al., 2023).

DeepPropNet addresses long-time evolution by learning only a short-time wave propagator and reusing it recursively. For the forced 1D wave equation, it combines a POD-style spatial basis with a causality-DeepONet for temporal forcing history, then advances across time blocks using the same learned operator. On the variable-wave-speed example, the reported maximum relative error is about $$\mathfrak u_{n+1}=\mathcal F_{\Delta t}\mathfrak u_n \approx \Psi_{\Delta t}^\theta \mathfrak u_n=\mathcal I^\theta \mathcal G_{\Delta t}\mathcal R\mathfrak u_n,$$0, and in a five-block recursive setting it is about $$\mathfrak u_{n+1}=\mathcal F_{\Delta t}\mathfrak u_n \approx \Psi_{\Delta t}^\theta \mathfrak u_n=\mathcal I^\theta \mathcal G_{\Delta t}\mathcal R\mathfrak u_n,$$1 with true blockwise initial data and about $$\mathfrak u_{n+1}=\mathcal F_{\Delta t}\mathfrak u_n \approx \Psi_{\Delta t}^\theta \mathfrak u_n=\mathcal I^\theta \mathcal G_{\Delta t}\mathcal R\mathfrak u_n,$$2 when the model’s own prediction supplies the next block’s initial condition (Liu et al., 2022).

3. Solver-integrated correction and time-parallel refinement

A major strand of Wave-PDE Nets does not replace numerical solvers outright; it augments them. In “Numerical wave propagation aided by deep learning,” the PDE is the 2D scalar wave equation

$$\mathfrak u_{n+1}=\mathcal F_{\Delta t}\mathfrak u_n \approx \Psi_{\Delta t}^\theta \mathfrak u_n=\mathcal I^\theta \mathcal G_{\Delta t}\mathcal R\mathfrak u_n,$$3

with periodic boundary conditions. A coarse solver $$\mathfrak u_{n+1}=\mathcal F_{\Delta t}\mathfrak u_n \approx \Psi_{\Delta t}^\theta \mathfrak u_n=\mathcal I^\theta \mathcal G_{\Delta t}\mathcal R\mathfrak u_n,$$4 on $$\mathfrak u_{n+1}=\mathcal F_{\Delta t}\mathfrak u_n \approx \Psi_{\Delta t}^\theta \mathfrak u_n=\mathcal I^\theta \mathcal G_{\Delta t}\mathcal R\mathfrak u_n,$$5 is combined with a learned corrector in energy variables

$$\mathfrak u_{n+1}=\mathcal F_{\Delta t}\mathfrak u_n \approx \Psi_{\Delta t}^\theta \mathfrak u_n=\mathcal I^\theta \mathcal G_{\Delta t}\mathcal R\mathfrak u_n,$$6

and the network target is

$$\mathfrak u_{n+1}=\mathcal F_{\Delta t}\mathfrak u_n \approx \Psi_{\Delta t}^\theta \mathfrak u_n=\mathcal I^\theta \mathcal G_{\Delta t}\mathcal R\mathfrak u_n,$$7

The correction is implemented by the U-Net-like JNet architecture, trained on about $$\mathfrak u_{n+1}=\mathcal F_{\Delta t}\mathfrak u_n \approx \Psi_{\Delta t}^\theta \mathfrak u_n=\mathcal I^\theta \mathcal G_{\Delta t}\mathcal R\mathfrak u_n,$$8 examples with Adam, batch size 128, 2000 epochs, and a learning rate schedule of $$\mathfrak u_{n+1}=\mathcal F_{\Delta t}\mathfrak u_n \approx \Psi_{\Delta t}^\theta \mathfrak u_n=\mathcal I^\theta \mathcal G_{\Delta t}\mathcal R\mathfrak u_n,$$9 followed by $$u(\cdot,t)\big|_{\Omega}=\mathcal SU_0,\qquad U_T\approx \mathcal N_\theta(U_0),$$0. The paper emphasizes that causality must be appropriately sampled in the training data, that phase accuracy rather than amplitude accuracy primarily controls Parareal convergence, and that the learned network can reduce coarse-solver errors often by around 90% in difficult media. In the unscaled Marmousi experiment, the fine solver step is about 4.53 s, JNet feedforward about 2.74 s, and a hypothetical 20-core Parareal run could yield about $$u(\cdot,t)\big|_{\Omega}=\mathcal SU_0,\qquad U_T\approx \mathcal N_\theta(U_0),$$1 speed-up, discounting parallel overhead (Nguyen et al., 2021).

The 2024 refinement of this framework formulates the learned propagator explicitly as

$$u(\cdot,t)\big|_{\Omega}=\mathcal SU_0,\qquad U_T\approx \mathcal N_\theta(U_0),$$2

for high-frequency wave propagation on heterogeneous media. It replaces raw $$u(\cdot,t)\big|_{\Omega}=\mathcal SU_0,\qquad U_T\approx \mathcal N_\theta(U_0),$$3 variables by the transformed representation $$u(\cdot,t)\big|_{\Omega}=\mathcal SU_0,\qquad U_T\approx \mathcal N_\theta(U_0),$$4 via $$u(\cdot,t)\big|_{\Omega}=\mathcal SU_0,\qquad U_T\approx \mathcal N_\theta(U_0),$$5 and $$u(\cdot,t)\big|_{\Omega}=\mathcal SU_0,\qquad U_T\approx \mathcal N_\theta(U_0),$$6, studies end-to-end propagators E2E-JNet3, E2E-JNet5, E2E-Tira, and E2E-Trans, and trains with the discretized energy semi-norm MSE under multi-step supervision. A key claim is that temporal wave dynamics must be learned, not just one-step mappings. Quantitatively, the best overall test result is E2E-Tira with energy MSE $$u(\cdot,t)\big|_{\Omega}=\mathcal SU_0,\qquad U_T\approx \mathcal N_\theta(U_0),$$7, compared with $$u(\cdot,t)\big|_{\Omega}=\mathcal SU_0,\qquad U_T\approx \mathcal N_\theta(U_0),$$8 for the prior modular baseline NE2E-JNet3; E2E-JNet3 trained on $$u(\cdot,t)\big|_{\Omega}=\mathcal SU_0,\qquad U_T\approx \mathcal N_\theta(U_0),$$9 reaches $10^{-3}$0; and the end-to-end setup improves average validation accuracy by more than 46% and test accuracy by about 53% relative to NE2E-JNet3. The paper also reports that E2E-JNet3 is about 41% faster to train than average, whereas E2E-Tira is about $10^{-3}$1 slower than average (Kaiser et al., 2024).

Both papers extend the learned propagator into Parareal. The 2024 study reports that direct training with $10^{-3}$2 is unstable and diverges, while fine-tuning with $10^{-3}$3 improves accuracy and stabilizes the Parareal process. This places Wave-PDE Nets, in this strand, within numerical-analysis workflows where the learned component is a coarse-to-fine corrector and the fine solve remains the reference dynamics (Kaiser et al., 2024).

4. Trainable wave-equation layers and graph wave dynamics

A more recent strand turns the wave equation into the layer itself. “Wave-PDE Nets: Trainable Wave-Equation Layers as an Alternative to Attention” treats the hidden representation as a field $10^{-3}$4 and advances it by the 1D damped wave equation

$10^{-3}$5

The trainable parameters $10^{-3}$6 and $10^{-3}$7 are produced by $10^{-3}$8 convolutions followed by softplus, the Laplacian is computed spectrally by FFT and inverse FFT in $10^{-3}$9 time, and the time update uses a symplectic velocity-Verlet integrator. The paper states a universality theorem for a single Wave-PDE layer followed by a linear readout, evaluates on WikiText-103, Long-Range Arena, and CIFAR-10/100, and reports on WikiText-103 that Wave-PDE Net reaches perplexity 18.49 with 9.5 GFLOPs, 9.1 ms/batch, and 9.1 GB peak memory, compared with Transformer perplexity 18.52, 15.6 GFLOPs, 12.1 ms/batch, and 9.5 GB. The accompanying claims are wall-clock reduction by up to 30%, peak-memory reduction by 25%, and ablation evidence that symplectic integration and a spectral Laplacian are critical for stability and performance (Vejendla, 5 Oct 2025).

The graph analogue is developed in Graph Wave Networks, where message passing is modeled by the graph wave equation

$u(x,t)=\int_\Omega g(x,t,\xi)\,u_0(\xi)\,d\xi,$0

with graph gradient and divergence defined from neighbor feature differences. The explicit update

$u(x,t)=\int_\Omega g(x,t,\xi)\,u_0(\xi)\,d\xi,$1

is proved constantly stable for any $u(x,t)=\int_\Omega g(x,t,\xi)\,u_0(\xi)\,d\xi,$2 under the stated Laplacian choices, both for GWN-sym and for the frequency-adaptive GWN-fa variant. The paper positions the second-order-in-time wave equation against heat-equation GNNs such as GRAND, GRAND++, and HiD-Net, and reports strong benchmark performance on homophilic, heterophilic, and large-scale datasets, together with depth robustness and improved handling of over-smoothing and heterophily (Yue et al., 26 May 2025).

These architectures differ from surrogate simulators in a precise way: they do not approximate an external physical solver for later reuse, but rather use oscillatory, global, and in part energy-preserving propagation as the internal communication mechanism of the model itself. That distinction is explicit in the attention-alternative framing of Wave-PDE Nets and the message-passing framing of Graph Wave Networks (Vejendla, 5 Oct 2025, Yue et al., 26 May 2025).

5. Wavelet, filter, and multiresolution formulations

Wave-PDE Nets also includes methods in which the wave prior appears through filters, moments, or wavelets rather than through an explicit wave equation. PDE-Net learns evolution PDEs with stacked $u(x,t)=\int_\Omega g(x,t,\xi)\,u_0(\xi)\,d\xi,$3-blocks of the form

$u(x,t)=\int_\Omega g(x,t,\xi)\,u_0(\xi)\,d\xi,$4

where the convolution operators $u(x,t)=\int_\Omega g(x,t,\xi)\,u_0(\xi)\,d\xi,$5 approximate derivatives and are constrained by moment matrices and sum rules so that each filter retains a differential interpretation. The point-wise network approximates the nonlinear response $u(x,t)=\int_\Omega g(x,t,\xi)\,u_0(\xi)\,d\xi,$6, and the architecture is designed to both predict dynamics and uncover hidden PDE structure. Although not wave-propagation-specific, this wavelet-inspired filter theory is a direct antecedent for later interpretable PDE-learning networks (Long et al., 2017).

A certified wavelet-based PINN extends this logic to parameterized PDEs by replacing heuristic pointwise residual penalties with a wavelet approximation of the residual dual norm. In the ultra-weak setting, the paper uses the exact identity

$u(x,t)=\int_\Omega g(x,t,\xi)\,u_0(\xi)\,d\xi,$7

and approximates the dual norm through weighted wavelet coefficients. The resulting loss

$u(x,t)=\int_\Omega g(x,t,\xi)\,u_0(\xi)\,d\xi,$8

serves simultaneously as training objective and computable upper error bound. The reported estimators are strict upper bounds once the wavelet level is sufficiently large, with very good quantitative effectivity in the elliptic examples considered (Ernst et al., 2022).

PIMWNN moves further toward an explicit multiresolution solver. It expands the unknown in a Shannon-wavelet basis,

$u(x,t)=\int_\Omega g(x,t,\xi)\,u_0(\xi)\,d\xi,$9

substitutes this ansatz into the PDE, and solves the resulting overdetermined linear system by least squares. The paper emphasizes that only the outer-layer coefficients are trained, that the basis derivatives are applied analytically rather than through standard deep-network autodiff, and that the method is mesh-free. On the reported problems, it is claimed to achieve higher accuracy and faster speed than PINNs; for the 1D Helmholtz test, PIMWNN achieves $$\hat Q(u_0)(x,t)=\frac1m\sum_{i=1}^m G(x,t,x_i)\,u_0(x_i).$$0 error $$\hat Q(u_0)(x,t)=\frac1m\sum_{i=1}^m G(x,t,x_i)\,u_0(x_i).$$1 with runtime about 80 s, while the compared PINN takes about 282 s and is less accurate (Han et al., 11 Aug 2025).

WaveLiT transfers the multiresolution prior to large-scale neural PDE solving through a discrete wavelet transform for lossless tokenization, augmented linear attention, a shared-weight multiscale feature pyramid, and an $$\hat Q(u_0)(x,t)=\frac1m\sum_{i=1}^m G(x,t,x_i)\,u_0(x_i).$$2 wavelet-domain auxiliary loss. On eight TheWell benchmarks, bespoke 1–10M-parameter models compete with foundation models 100–1000$$\hat Q(u_0)(x,t)=\frac1m\sum_{i=1}^m G(x,t,x_i)\,u_0(x_i).$$3 larger, with the largest gains on acoustic_scattering_maze and helmholtz_staircase, where the paper states that the wavelet-multiscale prior fits the dominant dynamical structure. The strongest transfer of the 10M foundation variant is likewise reported where that prior matches the dynamics, and the weakest transfer on chaotic advection-dominated flows (Sankaran et al., 25 May 2026).

6. Accuracy, interpretability, and limitations

The empirical appeal of Wave-PDE Nets is usually a combination of speed, acceptable fidelity, and structural inductive bias. The 2018 shallow-water surrogate produces 80 predicted frames in under half a second, versus about 35 seconds for the numerical solver, but it is explicitly described as a fast approximate simulator rather than a replacement for a high-fidelity PDE solver. The 2025 wave-equation layer reports practical efficiency gains over Transformers together with better stability under symplectic integration, while WaveLiT shows that small models with a wavelet prior can train on a single GPU and still match much larger PDE foundation models on selected benchmarks (Sorteberg et al., 2018, Vejendla, 5 Oct 2025, Sankaran et al., 25 May 2026).

Interpretability appears in several technically distinct forms. In the attention-alternative Wave-PDE layer, visualizations show lower learned wave speed around nouns, verbs, and punctuation and higher damping at clause or sentence boundaries, which the authors interpret as adaptive routing of information through the medium. In the certified wavelet PINN, interpretability is tied instead to computable a posteriori error control, since the loss is aligned with a residual norm that upper-bounds the solution error. In WaveLiT, the transfer pattern itself is treated as physically interpretable: the model is strongest where the wavelet-multiscale prior matches the dynamics and weakest where chaotic advection dominates (Vejendla, 5 Oct 2025, Ernst et al., 2022, Sankaran et al., 25 May 2026).

The limitations are equally consistent across the literature. Long-horizon accuracy degrades, reflections and boundary effects are difficult, and training-distribution coverage is decisive. The 2018 surrogate reports that boundary reflections are the hardest part to match and that a linear-wave-front test still retains a residual prior toward circular patterns from training. The solver-correction papers note sensitivity to out-of-distribution media, instability when the Parareal training distribution is mismatched, and the need to sample states that actually occur along causal propagation trajectories. The unbounded-domain solution-map approach reports no rigorous error theory, and its extrapolation deteriorates with distance from the training range (Sorteberg et al., 2018, Nguyen et al., 2021, Kaiser et al., 2024, Xie et al., 2021).

A second limitation concerns numerical structure. The trainable wave-equation layer is sensitive to the integration step size $$\hat Q(u_0)(x,t)=\frac1m\sum_{i=1}^m G(x,t,x_i)\,u_0(x_i).$$4, with stability deteriorating around $$\hat Q(u_0)(x,t)=\frac1m\sum_{i=1}^m G(x,t,x_i)\,u_0(x_i).$$5 in the reported sweep, and its implementation is optimized for 1D and 2D grid-like data rather than irregular graphs. Graph Wave Networks prove constant stability only under the stated Laplacian constructions and assume an undirected simple graph. WaveLiT highlights a different failure mode: even when one-step prediction is strong, autoregressive rollout on chaotic systems can still deteriorate because small per-step errors grow geometrically when the effective Lipschitz constant exceeds one (Vejendla, 5 Oct 2025, Yue et al., 26 May 2025, Sankaran et al., 25 May 2026).

Taken together, these works define Wave-PDE Nets less as a single architecture than as a technical program: approximate or internalize wave-governed propagation by imposing physically structured dynamics, operator factorizations, or multiresolution priors. The main divergence inside the program is methodological. Some models learn from raw wave movies, some correct numerical solvers, some propagate hidden states by a trainable wave equation, and some replace generic neural representations with constrained filters or wavelet bases. The common claim is narrower than general “physics-informed AI”: wave structure is not merely a source of data or regularization, but the organizing principle of the network itself.