LatentPDE: Latent-Space PDE Modeling

• LatentPDE is a family of methods that represent PDE-governed fields in a compact latent space, separating state compression from evolution and inference.
• They use encoder–latent-operator–decoder pipelines to reduce computational cost, enable long-horizon rollouts, and effectively manage sparse observations.
• These approaches extend traditional reduced-order models and neural operators, offering innovative solutions for both forward simulations and inverse problems.

Searching arXiv for recent and related papers on latent-space PDE methods to ground the article. LatentPDE denotes a family of methods that represent partial-differential-equation-governed fields in a learned or physically structured latent space and perform the principal inference, evolution, or operator approximation there rather than directly on the original discretized domain. In the broad literature, the term encompasses reduced-order latent simulators, latent neural operators, latent continuous-time PDE models, latent generative inverse models, and latent diffusion frameworks. A canonical 2024 instantiation is the Latent Neural PDE Solver, which learns time-dependent PDE dynamics on a mesh-reduced latent grid via a non-linear autoencoder and a latent propagator (Li et al., 2024). Across the wider landscape, latent-space formulations are used to reduce computational cost, improve long-horizon rollout feasibility, accommodate partial or irregular observations, enable bidirectional forward/inverse inference, and in some cases impose physics through interpretable latent variables or embedded coarse PDE solvers (Iakovlev et al., 2023, Wang et al., 2024, Chatzopoulos et al., 16 Feb 2026).

1. Concept and scope

LatentPDE methods begin from the observation that many neural PDE solvers still carry out temporal evolution or operator application on full-order discretized fields, even when they use multiresolution encoders internally. The resulting cost scales with the spatial resolution, so long rollouts and high-resolution domains become expensive in memory and runtime (Li et al., 2024). The central LatentPDE idea is therefore to decouple representation from evolution: compress the physical state to a smaller latent object, operate on that latent object, and decode back only when needed (Li et al., 2024, Wu et al., 2022).

In one common formulation, the physical state $u_t$ is mapped by an encoder $\phi$ to a latent state $z_t$, evolved by a latent propagator $\gamma$, and reconstructed by a decoder $\psi$, so that

$$u_{t+1} = \psi\big(\gamma(\phi(u_t))\big), \qquad u_{t+k} \approx \psi\big(\gamma^{(k)}(\phi(u_t))\big).$$

This pattern appears explicitly in the Latent Neural PDE Solver (Li et al., 2024), and conceptually reappears in latent global evolution models (Wu et al., 2022), latent neural operators (Wang et al., 2024), and parameter-conditioned latent dynamics models (Liang et al., 13 Mar 2026).

The term also covers formulations in which the latent variable does not merely compress a state but carries a more structured role. In "Learning Space-Time Continuous Neural PDEs from Partially Observed States," the latent state is a space-time continuous field $z(t,x)$ defined by interpolation over latent values at anchor points, with dynamics governed by a local neural PDE on fixed continuous neighborhoods (Iakovlev et al., 2023). In GenPANIS, the latent variable jointly encodes discrete microstructures and PDE solutions, while a physics-aware decoder embeds a coarse-grained PDE solver (Chatzopoulos et al., 16 Feb 2026). In the 2026 paper titled simply "LatentPDE," the latent variables are parameterized directly as PDE coefficients and source terms, so the latent space is explicitly interpretable (Tsao et al., 26 Apr 2026). This suggests that "LatentPDE" is best understood not as a single architecture but as a design principle: move PDE inference to a compressed latent representation, while choosing how much physical structure that latent representation should retain.

2. Core architectural patterns

A recurrent architectural pattern is the encoder–latent-operator–decoder pipeline. In the Latent Neural PDE Solver, a non-linear autoencoder maps a full grid $D$ to a coarse latent grid $D_l$, and a residual CNN propagates latent states on that coarse grid (Li et al., 2024). The latent remains a coarse field rather than a flat vector, preserving spatial structure for convolutions. By contrast, LE-PDE uses a global latent vector $z^k \in \mathbb{R}^{d_z}$ and evolves it autoregressively via an MLP residual dynamics model, so rollout updates no longer scale with the number of physical cells (Wu et al., 2022).

Another pattern is latent operator learning. The Latent Neural Operator maps geometric-space samples to a shorter latent sequence through Physics-Cross-Attention, applies transformer blocks in latent space, and decodes back to arbitrary query positions through an inverse Physics-Cross-Attention map (Wang et al., 2024). Latent Spectral Models similarly factor the operator into coordinate-to-latent projection, latent spectral solving, and latent-to-coordinate reconstruction, with the latent solve implemented by a trigonometric neural spectral block (Wu et al., 2023). LaMO replaces Fourier or attention-based latent operators by selective state-space models, giving a latent neural operator in which the principal operator acts on a reduced set of latent tokens (Tiwari et al., 25 May 2025).

A third pattern is latent continuous-time dynamics. The space-time continuous latent neural PDE model defines

$\phi$0

with interpolation providing spatial continuity and an ODE solver providing temporal continuity (Iakovlev et al., 2023). DLDMF uses a parameter-conditioned Neural ODE in latent space, with a parameter encoder producing an embedding that both initializes and conditions the latent flow (Liang et al., 13 Mar 2026). This contrasts with static coordinate-based surrogates that treat time as just another input (Liang et al., 13 Mar 2026).

Generative variants add probabilistic inference. Di-BiLPS uses a VAE to compress joint coefficient–state data, a latent diffusion model to represent uncertainty, and contrastive learning to align sparse-conditioning embeddings with full-field embeddings (Li et al., 13 May 2026). L-DPS uses a VAE, an unconditional latent diffusion prior, diffusion posterior sampling, and a differentiable surrogate to perform approximate Bayesian inversion in latent space (Wang et al., 25 Jun 2026). GenPANIS combines a latent-variable generative model with a physics-aware decoder and a RealNVP prior, supporting both forward prediction and inverse recovery in one architecture (Chatzopoulos et al., 16 Feb 2026).

3. Relationship to reduced-order modeling and operator learning

LatentPDE methods are closely related to classical reduced-order modeling, but they replace linear modal assumptions by learned non-linear representations. The Latent Neural PDE Solver explicitly contrasts itself with POD–Galerkin ROMs: classical ROM collects snapshots, computes leading modes via SVD, represents the state as $\phi$1, and projects the PDE onto those modes, whereas LNS uses a non-linear convolutional autoencoder for $\phi$2 and a learned latent dynamics network $\phi$3 instead of Galerkin projection (Li et al., 2024). The latent field remains structured as a coarse grid rather than a vector of global modes, preserving locality (Li et al., 2024).

Compared with Koopman-style models, some latent PDE solvers forgo latent linearity in favor of expressivity. LNS makes no linearity assumption in latent space and instead uses a residual CNN with dilated convolutions (Li et al., 2024). LE-PDE learns a global latent dynamical system without assuming a pre-specified reduced basis, while emphasizing long-horizon latent rollouts and latent consistency losses (Wu et al., 2022). This suggests that the modern LatentPDE literature is less concerned with closed-form reduced dynamics than with learnable low-dimensional manifolds that remain stable under repeated rollout.

In the operator-learning literature, LatentPDE methods can be viewed as compressed neural operators. LNO explicitly learns the operator in latent space rather than the geometric space, which reduces the cost of attention from dependence on $\phi$4 physical samples to dependence on a latent length $\phi$5 (Wang et al., 2024). LSM pushes this further by giving the latent operator a spectral interpretation, factorizing the operator into

$\phi$6

and providing a latent trigonometric basis block with a stated approximation rate for Lipschitz functions in one-dimensional latent space (Wu et al., 2023). LaMO similarly positions its selective SSM in latent space as a learnable integral kernel operator (Tiwari et al., 25 May 2025). A plausible implication is that latent-space operator learning serves two distinct purposes at once: computational reduction and an inductive bias toward low-dimensional operator structure.

4. Observation regimes, conditioning, and inverse problems

A major branch of the LatentPDE literature addresses settings in which the full PDE state is not observed. The space-time continuous latent neural PDE model is designed for noisy, partially observed data on irregular spatiotemporal grids, with observations $\phi$7 interpreted as a noisy partial mapping of an unobserved latent field $\phi$8 (Iakovlev et al., 2023). The model uses amortized variational inference, multiple shooting, and a local spatiotemporal encoder to infer latent initial conditions from partial observations (Iakovlev et al., 2023). The paper reports that many baselines degrade heavily on partially observed data, whereas the latent neural PDE maintains strong performance (Iakovlev et al., 2023).

Di-BiLPS targets even more extreme sparsity, down to 3% observations and tested to 1%, by combining latent diffusion, contrastive alignment between sparse and full representations, and geometry-aware decoding for zero-shot super-resolution (Li et al., 13 May 2026). Its conditioning space is learned from sparse coefficient or state observations and boundary conditions, and observation-guided denoising imposes consistency at the observed points (Li et al., 13 May 2026). The paper emphasizes that removing condition guidance causes catastrophic degradation, while PDE guidance in latent space has comparatively limited effect (Li et al., 13 May 2026). This suggests that in ultra-sparse regimes, representation alignment may be more decisive than direct latent PDE residual guidance.

Inverse problems also motivate latent generative models. GenPANIS formulates a joint distribution over latent variables $\phi$9, discrete microstructures $z_t$0, and PDE solutions $z_t$1, then performs both forward prediction $z_t$2 and inverse inference $z_t$3 through latent-space conditioning (Chatzopoulos et al., 16 Feb 2026). The decoder for $z_t$4 includes a differentiable coarse-grained PDE solver, allowing gradients with respect to $z_t$5 while preserving discrete-valued microstructures in the decoder for $z_t$6 (Chatzopoulos et al., 16 Feb 2026). L-DPS performs posterior-guided inversion by sampling in the latent space of parameter fields and computing likelihood gradients through the decoder–surrogate composition, thus avoiding repeated numerical PDE solves during posterior sampling (Wang et al., 25 Jun 2026). In both cases, the latent variable mediates between complex priors and PDE-constrained likelihoods.

The 2021 latent space solver for PDE generalization addresses a different inverse-like use case: it solves steady PDEs in latent space via iterative fixed-point refinement, starting from a coarse-grid PDE initialization rather than random latent initialization (Ranade et al., 2021). Because geometry, boundary conditions, source terms, and solutions all have latent encodings, the solver uses coarse numerical solutions to initialize a conditioned solution manifold and refines that estimate in latent space (Ranade et al., 2021). This is not Bayesian inversion, but it illustrates another LatentPDE theme: latent representations can mediate between sparse or coarse physical evidence and high-resolution predictions.

5. PDE classes, benchmarks, and empirical behavior

LatentPDE methods have been evaluated on a wide spectrum of PDEs and domains. The Latent Neural PDE Solver studies 2D Navier–Stokes vorticity on a $z_t$7 grid, shallow water equations over a global domain, and tank sloshing multiphase flow with Reynolds-averaged continuity and momentum equations plus VOF advection (Li et al., 2024). It uses latent grids such as $z_t$8 for Navier–Stokes, $z_t$9 for shallow water, and $\gamma$0 for tank sloshing (Li et al., 2024). The paper reports that LNS is competitive in accuracy while enabling training rollouts of 10–20 steps for shallow water that are infeasible for full-grid FNO/FNO-Mixer/UNet baselines (Li et al., 2024).

LE-PDE evaluates on a 1D nonlinear PDE family, 2D Navier–Stokes into turbulent regimes, and a 2D Navier–Stokes inverse boundary optimization problem (Wu et al., 2022). It reports up to 128x reduction in the dimensions to update and up to 15x improvement in speed, while remaining competitive with state-of-the-art surrogates (Wu et al., 2022). The latent representation is global rather than field-like, and the paper highlights latent backpropagation through time for inverse optimization (Wu et al., 2022).

Latent operator models have been tested across solid and fluid mechanics. LNO reports state-of-the-art accuracy on four out of six forward benchmarks and one inverse benchmark, while reducing GPU memory by 50% and speeding up training 1.8 times (Wang et al., 2024). LSM reports a relative gain of 11.5% averaged on seven benchmarks covering both solid and fluid physics (Wu et al., 2023). LaMO reports a 32.3% improvement over existing baselines in solution operator approximation across regular grids, structured meshes, and point clouds (Tiwari et al., 25 May 2025). These results indicate that latent compression can be beneficial even when the task is not explicit time stepping but function-to-function operator approximation.

For parameterized PDEs, DLDMF targets simultaneous parameter generalization and temporal extrapolation, reporting on convection–diffusion–reaction equations and 2D Navier–Stokes (Liang et al., 13 Mar 2026). The paper argues that standard parameterized models that treat time as just another input struggle with both tasks together, whereas a parameter-conditioned latent Neural ODE with disentangled manifold fusion performs well on unseen parameter settings and long-term extrapolation (Liang et al., 13 Mar 2026). This suggests that latent dynamics are not only a computational device but also an inductive bias for temporal extrapolation.

Generative and reconstruction-oriented LatentPDE models use different benchmarks. GenPANIS studies Darcy flow and Helmholtz equations in multiphase media, with sparse and noisy observations, unseen boundary conditions, and varying volume fractions (Chatzopoulos et al., 16 Feb 2026). The 2026 LatentPDE framework studies advection–diffusion, Klein–Gordon, and Helmholtz equations on $\gamma$1, focusing on sparse-observation reconstruction and super-resolution from structured missingness (Tsao et al., 26 Apr 2026). Di-BiLPS evaluates Darcy, Helmholtz, Poisson, and both bounded and unbounded Navier–Stokes under 3% observations (Li et al., 13 May 2026). Together these works show that "LatentPDE" now spans both surrogate simulation and generative reconstruction.

6. Interpretability, limitations, and open directions

A persistent issue in latent-space PDE modeling is interpretability. Many methods compress fields into latent vectors or tensors with no direct physical meaning. LNS notes that latent variables have no direct physical meaning and that latent design choices such as grid size and channel count are tuned empirically (Li et al., 2024). The 2021 latent-space solver similarly learns compact encodings of geometry, boundary conditions, source terms, and solutions, but its latent variables are not tied to identifiable PDE parameters (Ranade et al., 2021). LNO and LSM learn latent operator spaces whose coordinates are useful computationally but not inherently interpretable (Wang et al., 2024, Wu et al., 2023).

Some recent work explicitly addresses this limitation. GenPANIS gives the latent variable a probabilistic semantics as a joint continuous embedding for discrete microstructures and PDE solutions, and its decoder includes a differentiable coarse-grained PDE solver (Chatzopoulos et al., 16 Feb 2026). The 2026 LatentPDE paper goes further by parameterizing the latent variables directly as the coefficients and source terms of an assumed governing PDE, making the latent space inherently interpretable (Tsao et al., 26 Apr 2026). This suggests an emerging split within the field between opaque latent manifolds optimized for expressivity and structured latent spaces optimized for physical identifiability.

Several limitations recur across the literature. Many models remain restricted to uniform grids or regular discretizations, especially when they rely heavily on CNNs (Li et al., 2024). Some lack formal stability or error guarantees, even when they perform well empirically over long rollouts (Li et al., 2024, Ranade et al., 2021). Data-driven latent dynamics often do not enforce hard conservation laws (Li et al., 2024, Tiwari et al., 25 May 2025). Generative methods that use latent diffusion or variational inference can still be limited by surrogate or decoder accuracy (Li et al., 13 May 2026, Wang et al., 25 Jun 2026). Di-BiLPS notes that performance deteriorates drastically at 0.1% sparsity and that PDE guidance in latent space is only weakly effective in its current form (Li et al., 13 May 2026). L-DPS notes that its posterior is approximate because of the VAE range restriction, Tweedie plug-in likelihood, surrogate substitution, and finite-step guided reverse diffusion (Wang et al., 25 Jun 2026).

Open directions are explicit in multiple papers. LNS suggests extending to irregular meshes and geometries and integrating physics-informed elements or operator-learning ideas into the latent framework (Li et al., 2024). LSEM proposes scalable assembled latent dynamical systems built from reusable local elements and points toward "foundation-model surrogate solvers" built from reusable local models (Chung et al., 5 Jan 2026). LaMO proposes extending SSM-based latent operators toward large multi-physics and multi-geometry foundation models (Tiwari et al., 25 May 2025). DLDMF suggests that parameter-conditioned latent dynamics can be further extended to more complex parameter spaces and longer time extrapolation (Liang et al., 13 Mar 2026). A plausible implication is that the next stage of LatentPDE research will focus less on whether to use latent spaces at all, and more on what structure those latent spaces should encode: geometry, uncertainty, interpretability, compositionality, or exact physical constraints.

7. Position within the broader PDE-ML landscape

LatentPDE occupies a position between reduced-order modeling, neural operators, latent dynamics models, and generative inverse methods. Relative to classical ROM, it replaces linear bases and explicit projection with learned non-linear encoders and latent propagators (Li et al., 2024). Relative to standard neural operators, it shifts the main operator application from geometric space to a compressed latent space (Wang et al., 2024, Wu et al., 2023). Relative to PINNs, it often enforces physics through latent-space structure, decoder design, or operator-guided inference rather than direct collocation residual minimization on every new instance (Iakovlev et al., 2023, Ranade et al., 2021). Relative to diffusion-based PDE models, it can embed the generative process in a lower-dimensional latent manifold for efficiency and sometimes for interpretability (Li et al., 13 May 2026, Wang et al., 25 Jun 2026, Tsao et al., 26 Apr 2026).

The diversity of architectures under the label is therefore substantial. Some methods are deterministic latent simulators for time stepping (Li et al., 2024, Wu et al., 2022). Some are latent neural operators for forward maps (Wang et al., 2024, Wu et al., 2023, Tiwari et al., 25 May 2025). Some are probabilistic latent PDE models for irregular partial observations (Iakovlev et al., 2023). Some are joint latent generative models for forward and inverse tasks (Chatzopoulos et al., 16 Feb 2026, Li et al., 13 May 2026, Wang et al., 25 Jun 2026). Some emphasize parameterized dynamics and disentanglement (Liang et al., 13 Mar 2026), while others emphasize physically interpretable latent parameterizations (Tsao et al., 26 Apr 2026).

Taken together, the literature suggests a unifying definition: LatentPDE is a methodological paradigm in which the essential computational object for PDE learning is not the high-dimensional field itself, but a latent representation that is evolved, manipulated, inferred, or sampled under architectural or probabilistic constraints tied to PDE structure. The specific choice of latent representation—coarse latent grid, global vector, continuous latent field, latent token set, diffusion latent, or interpretable PDE-parameter latent—determines the balance among efficiency, physical structure, uncertainty quantification, and extrapolative behavior.