Posterior-First Neural PDE Simulation

• Posterior-first neural PDE simulation is a design paradigm that first infers the posterior over hidden states or parameters prior to performing PDE rollouts.
• It integrates Bayesian inference, neural posterior estimation, and latent diffusion techniques to address uncertainty in complex, high-dimensional PDE systems.
• Empirical studies demonstrate improved accuracy and efficiency in simulations, with benchmarks showing significant reductions in error and faster per-sample computation times.

Posterior-first neural PDE simulation denotes a family of methods in which posterior inference is the primary object of the simulation pipeline rather than a by-product of a deterministic field-to-future map. Across recent formulations, the common pattern is to infer a posterior over latent problem state, physical parameters, or constrained solution manifolds from partial observations and then condition rollout, posterior prediction, or inverse reconstruction on that posterior. In one line of work, the central object is the posterior over the minimal task-sufficient hidden state from a single observed field (Wang et al., 5 May 2026). In another, it is a single $\beta$-conditioned neural posterior estimator $q_\phi(\theta \mid x,\beta)$ for generalized Bayesian inference over simulator parameters (Sun et al., 29 Jan 2026). Related variants target posterior sampling in latent spaces for PDE inverse problems (Wang et al., 25 Jun 2026), enforce posterior consistency under hard PDE constraints via a co-area correction (Xu et al., 3 Jun 2026), or construct posterior ensembles with differentiable simulators and PINN-based solvers (Zeghal et al., 2022, Agata et al., 8 May 2026, Yang et al., 2022). The unifying principle is that uncertainty over hidden PDE-relevant structure is represented explicitly and propagated through the simulator, solver, or surrogate.

1. Conceptual definition and problem setting

Posterior-first formulations arise when the deployed interface is informationally impoverished relative to the latent state controlling PDE evolution. A representative example is the single-observation setting, where the model receives only one observed field $x_t$ but the downstream target $y$ depends on hidden field components, coefficients, operators, boundary conditions, or memory. In that setting, the relevant latent object is the minimal task-sufficient state $Z_t^\star$, defined through

$Y_t^\tau \perp W_t \mid Z_t^\star,$

with posterior

$$\pi_t(\cdot\mid x_t) := \mathbb P\!\left(Z_t^\star\in\cdot\mid X_t=x_t\right).$$

A direct deterministic map from $x_t$ to future fields cannot represent non-Dirac ambiguity over $Z_t^\star$; this is the core failure mode identified in the single-field setting (Wang et al., 5 May 2026).

A complementary formulation appears in generalized Bayesian inference for simulator parameters. There the posterior-first object is the tempered posterior family

$$p_\beta(\theta \mid x) \propto \pi(\theta)\,p(x\mid \theta)^\beta,$$

with $q_\phi(\theta \mid x,\beta)$0. Learning a single $q_\phi(\theta \mid x,\beta)$1-conditioned estimator

$q_\phi(\theta \mid x,\beta)$2

makes posterior sampling the first step of the PDE workflow: parameter samples are drawn in one forward pass and then pushed through a PDE solver or neural emulator to generate posterior predictive fields (Sun et al., 29 Jan 2026).

For PDE inverse problems with high-dimensional spatial parameters, posterior-first can also mean that each reverse-diffusion step explicitly follows an approximate posterior score,

$q_\phi(\theta \mid x,\beta)$3

in a latent space learned by a VAE and a diffusion prior. This is the organizing idea of latent diffusion posterior sampling for Darcy flow inversion (Wang et al., 25 Jun 2026).

These formulations share an operational claim: the simulator, surrogate, or predictor should be conditioned on an inferred posterior over hidden PDE-relevant variables before rollout. This suggests that “posterior-first” is not a single algorithm but a design principle spanning amortized SBI, score-based inverse sampling, physics-constrained generation, and Bayesian PINN pipelines.

2. Statistical objects and theoretical structure

The strongest abstract statement appears in the task-sufficient-state formulation. For task $q_\phi(\theta \mid x,\beta)$4, action $q_\phi(\theta \mid x,\beta)$5, and latent-state conditional risk $q_\phi(\theta \mid x,\beta)$6, the observation-conditional Bayes value factors through the posterior: $q_\phi(\theta \mid x,\beta)$7 Accordingly, Bayes-optimal downstream decisions depend on $q_\phi(\theta \mid x,\beta)$8 only through $q_\phi(\theta \mid x,\beta)$9. In the same framework, deterministic point-latent interfaces incur an ambiguity barrier

$x_t$0

which vanishes only for Dirac posteriors (Wang et al., 5 May 2026).

For generalized Bayes, the central theoretical statement is a forward-KL characterization of the SNIS-weighted objective. With weights

$x_t$1

the objective

$x_t$2

satisfies

$x_t$3

This gives a mass-covering fit to the tempered posterior. For NRE weights and $x_t$4, the paper further states

$x_t$5

hence $x_t$6 in that regime (Sun et al., 29 Jan 2026).

A different theoretical issue emerges when the posterior is defined by a hard PDE constraint $x_t$7. Conditioning a generative prior $x_t$8 on the measure-zero manifold $x_t$9 is not uniquely defined unless a limiting procedure is specified. The small-residual-noise limit yields the co-area-correct posterior

$y$0

where $y$1. Projection- and guidance-based methods that enforce $y$2 but omit the Fixman factor instead target the Hausdorff restriction $y$3 on $y$4, which is generally biased (Xu et al., 3 Jun 2026).

For differentiable-simulator NPE, the relevant theoretical object is the posterior score. If $y$5 is accessible, one can regularize $y$6 via

$y$7

combined with the standard negative log-likelihood objective. The point is that PDE adjoints expose local posterior geometry and can shape the learned posterior more sample-efficiently (Zeghal et al., 2022).

3. Amortized posterior estimation for PDE simulators

The most explicit posterior-first amortization scheme is the $y$8-amortized neural posterior estimator for generalized Bayes. The workflow assumes a bounded temperature domain $y$9 with $Z_t^\star$0, and trains a single conditional density estimator $Z_t^\star$1 across both observations and temperatures. At inference, for any new observation $Z_t^\star$2 and any $Z_t^\star$3, one draws $Z_t^\star$4 in a single forward pass, with no simulator calls or inference-time MCMC (Sun et al., 29 Jan 2026).

Two training routes are specified.

Route A synthesizes tempered pairs by first learning a joint score $Z_t^\star$5 with denoising score matching on base-joint samples $Z_t^\star$6, then running short-run Langevin dynamics on the tempered joint

$Z_t^\star$7

using

$Z_t^\star$8

The resulting tempered dataset $Z_t^\star$9 is then used for conditional maximum likelihood training of $Y_t^\tau \perp W_t \mid Z_t^\star,$0.

Route B reuses a fixed base dataset $Y_t^\tau \perp W_t \mid Z_t^\star,$1 and applies self-normalized importance sampling. In the NRE variant, one trains a classifier $Y_t^\tau \perp W_t \mid Z_t^\star,$2 on joint versus product-of-marginals samples and forms a ratio estimator

$Y_t^\tau \perp W_t \mid Z_t^\star,$3

For each $Y_t^\tau \perp W_t \mid Z_t^\star,$4 on a grid $Y_t^\tau \perp W_t \mid Z_t^\star,$5, one computes globally normalized SNIS weights and optimizes

$Y_t^\tau \perp W_t \mid Z_t^\star,$6

The paper states that global per-$Y_t^\tau \perp W_t \mid Z_t^\star,$7 normalization keeps the gradient unbiased for the full-data weighted objective (Sun et al., 29 Jan 2026).

Within PDE pipelines, this estimator is explicitly “posterior-first”: first sample $Y_t^\tau \perp W_t \mid Z_t^\star,$8, then push $Y_t^\tau \perp W_t \mid Z_t^\star,$9 through the PDE solver or emulator to obtain posterior predictive fields. The same source recommends field encoders such as CNNs, U-Nets, Vision Transformers, or spatio-temporal transformers for high-dimensional $$\pi_t(\cdot\mid x_t) := \mathbb P\!\left(Z_t^\star\in\cdot\mid X_t=x_t\right).$$0, and MDNs or conditional normalizing flows such as MAF/NSF for $$\pi_t(\cdot\mid x_t) := \mathbb P\!\left(Z_t^\star\in\cdot\mid X_t=x_t\right).$$1 depending on dimension (Sun et al., 29 Jan 2026).

Empirically, the method is reported on Gaussian Mixture, Two Moons, SLCP, and Lorenz–96. Across these benchmarks, “a single $$\pi_t(\cdot\mid x_t) := \mathbb P\!\left(Z_t^\star\in\cdot\mid X_t=x_t\right).$$2-amortized $$\pi_t(\cdot\mid x_t) := \mathbb P\!\left(Z_t^\star\in\cdot\mid X_t=x_t\right).$$3 matches non-amortized power posterior samplers (MCMC/tempering) over a wide $$\pi_t(\cdot\mid x_t) := \mathbb P\!\left(Z_t^\star\in\cdot\mid X_t=x_t\right).$$4-range in two-sample discrepancies (MMD and C2ST),” including Lorenz–96 with chaotic dynamics and without inference-time MCMC (Sun et al., 29 Jan 2026).

4. Posterior sampling, latent diffusion, and manifold-correctness

For high-dimensional PDE inverse problems, latent diffusion posterior sampling constructs the posterior in a compressed latent space. A VAE maps the parameter field $$\pi_t(\cdot\mid x_t) := \mathbb P\!\left(Z_t^\star\in\cdot\mid X_t=x_t\right).$$5 to a latent $$\pi_t(\cdot\mid x_t) := \mathbb P\!\left(Z_t^\star\in\cdot\mid X_t=x_t\right).$$6 for a $$\pi_t(\cdot\mid x_t) := \mathbb P\!\left(Z_t^\star\in\cdot\mid X_t=x_t\right).$$7 Darcy-flow problem, reducing dimension from $$\pi_t(\cdot\mid x_t) := \mathbb P\!\left(Z_t^\star\in\cdot\mid X_t=x_t\right).$$8 to $$\pi_t(\cdot\mid x_t) := \mathbb P\!\left(Z_t^\star\in\cdot\mid X_t=x_t\right).$$9. An unconditional latent DDPM learns the prior score, and a differentiable surrogate $x_t$0 supplies likelihood gradients through the decoder–surrogate chain. The reverse update uses deterministic DDIM plus normalized likelihood guidance

$x_t$1

with

$x_t$2

and calibrated constants $x_t$3, $x_t$4, $x_t$5 (Wang et al., 25 Jun 2026).

The reported cost comparison is explicit. With an A100-class GPU, L-DPS+FNO costs “$x_t$6 minute” per sample, versus “$x_t$7 minutes” for full-space DPS+FNO, described as “$x_t$8 faster per sample.” Accuracy is similar in noiseless regimes and superior under sparse or noisy observations. For example, at $x_t$9, $Z_t^\star$0, the reported relative $Z_t^\star$1 error is $Z_t^\star$2 for L-DPS+FNO, $Z_t^\star$3 for full-space DPS, and $Z_t^\star$4 for InvFNO; at $Z_t^\star$5, $Z_t^\star$6, L-DPS and full-space DPS are $Z_t^\star$7 and $Z_t^\star$8, respectively (Wang et al., 25 Jun 2026).

A different correctness issue appears when the posterior is restricted to a hard-constraint manifold. The co-area analysis states that satisfying the physics is not sufficient for posterior consistency. The correct target on $Z_t^\star$9 is

$$p_\beta(\theta \mid x) \propto \pi(\theta)\,p(x\mid \theta)^\beta,$$0

with log-density

$$p_\beta(\theta \mid x) \propto \pi(\theta)\,p(x\mid \theta)^\beta,$$1

The CoCoS sampler enforces feasibility, uses tangent proposals, projects along normal directions, and accepts proposals with the co-area potential

$$p_\beta(\theta \mid x) \propto \pi(\theta)\,p(x\mid \theta)^\beta,$$2

On controlled benchmarks, omitting the co-area factor inflates posterior error to “about $$p_\beta(\theta \mid x) \propto \pi(\theta)\,p(x\mid \theta)^\beta,$$3 the sampling-noise floor,” while minimal-displacement projection is biased at “about $$p_\beta(\theta \mid x) \propto \pi(\theta)\,p(x\mid \theta)^\beta,$$4 the floor”; CoCoS is reported to match the gold-standard posterior “to within sampling noise” (Xu et al., 3 Jun 2026).

These two strands address different aspects of posterior-first PDE simulation. L-DPS emphasizes scalable approximate posterior sampling under expensive forward models, whereas co-area-correct sampling emphasizes that the target measure itself must be correct when PDE constraints are treated as hard manifold conditions. This suggests a division between computational acceleration and measure-theoretic fidelity that is often conflated in inverse-problem practice.

5. Differentiable simulators, PINNs, and posterior ensembles

When the simulator is differentiable, posterior-first inference can exploit PDE sensitivities directly. In neural posterior estimation with differentiable simulators, the training loss combines conditional log-likelihood and score matching: $$p_\beta(\theta \mid x) \propto \pi(\theta)\,p(x\mid \theta)^\beta,$$5 For PDE-constrained models defined by a residual $$p_\beta(\theta \mid x) \propto \pi(\theta)\,p(x\mid \theta)^\beta,$$6 and observations $$p_\beta(\theta \mid x) \propto \pi(\theta)\,p(x\mid \theta)^\beta,$$7, implicit differentiation and the adjoint system provide

$$p_\beta(\theta \mid x) \propto \pi(\theta)\,p(x\mid \theta)^\beta,$$8

and, under Gaussian observation noise,

$$p_\beta(\theta \mid x) \propto \pi(\theta)\,p(x\mid \theta)^\beta,$$9

The paper states that gradient information “helps constrain the shape of the posterior and improves sample-efficiency,” with “often $q_\phi(\theta \mid x,\beta)$00–$q_\phi(\theta \mid x,\beta)$01 fewer simulations” needed once proposals are narrowed (Zeghal et al., 2022).

Bayesian PINN approaches shift the prior from weight space toward function space. In the functional-prior-based framework, the unknown field $q_\phi(\theta \mid x,\beta)$02 is endowed with a Gaussian process prior

$q_\phi(\theta \mid x,\beta)$03

typically with RBF kernel

$q_\phi(\theta \mid x,\beta)$04

Random Fourier features are used through

$q_\phi(\theta \mid x,\beta)$05

to align the network representation with the functional prior (Agata et al., 8 May 2026).

Two variants are reported. FPI-BPINN first learns a weight prior consistent with the functional prior by minimizing an MMD objective between neural function samples and target stochastic-process samples, then performs posterior inference in weight space with pSGLD+R and inner PINN solves. fParVI-PINN performs particle-based variational inference directly in function space and updates weights via the Jacobian transpose. On 1D seismic traveltime tomography, the paper reports MMD to a semi-analytical posterior of $q_\phi(\theta \mid x,\beta)$06 for FPI-BPINN and $q_\phi(\theta \mid x,\beta)$07 for fParVI-PINN at $q_\phi(\theta \mid x,\beta)$08, with fParVI-PINN generally described as more accurate due to analytic functional priors. On 2D Darcy inversion, reported unoptimized compute times are “$q_\phi(\theta \mid x,\beta)$09 hours on $q_\phi(\theta \mid x,\beta)$10 NVIDIA A100” for FPI-BPINN and “$q_\phi(\theta \mid x,\beta)$11 hours” for fParVI-PINN (Agata et al., 8 May 2026).

MO-PINN takes a different route: it constructs a multi-output neural solver whose outputs are interpreted as an empirical posterior. For output index $q_\phi(\theta \mid x,\beta)$12, bootstrap-perturbed targets are generated from Gaussian noise, and each output $q_\phi(\theta \mid x,\beta)$13 is trained to satisfy both data and physics. The resulting empirical posterior is

$q_\phi(\theta \mid x,\beta)$14

with posterior moments extracted directly from the ensemble (Yang et al., 2022). In the reported inverse 1D diffusion-reaction example, the predicted parameter distribution for small noise has mean $q_\phi(\theta \mid x,\beta)$15 and standard deviation $q_\phi(\theta \mid x,\beta)$16 around true $q_\phi(\theta \mid x,\beta)$17; for large noise, mean $q_\phi(\theta \mid x,\beta)$18 and standard deviation $q_\phi(\theta \mid x,\beta)$19. In the 2D inverse diffusion-reaction example with true $q_\phi(\theta \mid x,\beta)$20, the reported estimates are $q_\phi(\theta \mid x,\beta)$21 and $q_\phi(\theta \mid x,\beta)$22 for small and large noise, respectively (Yang et al., 2022).

PDE-NetGen is not, by itself, a posterior inference method, but it provides a symbolic-to-neural route for differentiable PDE solvers that can serve as the computational backbone of posterior-first workflows. It translates symbolic PDEs into convolutional stencils and time-stepping networks in Keras, enabling calibration, data assimilation, and uncertainty quantification through automatic differentiation. The paper makes the Bayesian interpretation explicit under Gaussian assumptions,

$q_\phi(\theta \mid x,\beta)$23

where $q_\phi(\theta \mid x,\beta)$24 is a variational data-assimilation objective (Pannekoucke et al., 2020).

6. Empirical behavior, implementation patterns, and limitations

The single-field posterior-first benchmark on metadata-hidden PDEBench tasks provides the most direct evidence for the design principle in neural PDE simulation. With hidden metadata withheld at inference, posterior recovery reduces pooled rollout nRMSE from $q_\phi(\theta \mid x,\beta)$25 for Direct-Point to $q_\phi(\theta \mid x,\beta)$26 for Posterior-$q_\phi(\theta \mid x,\beta)$27, with oracle reference $q_\phi(\theta \mid x,\beta)$28, thereby closing $q_\phi(\theta \mid x,\beta)$29 of the direct-to-oracle gap; on the high-ambiguity subset, the reported closure is $q_\phi(\theta \mid x,\beta)$30 (Wang et al., 5 May 2026). Family-level results are also reported: Diffusion–Reaction $q_\phi(\theta \mid x,\beta)$31 with oracle $q_\phi(\theta \mid x,\beta)$32, Diffusion–Sorption $q_\phi(\theta \mid x,\beta)$33 with oracle $q_\phi(\theta \mid x,\beta)$34, Shallow Water $q_\phi(\theta \mid x,\beta)$35 with oracle $q_\phi(\theta \mid x,\beta)$36, and Incompressible Navier–Stokes $q_\phi(\theta \mid x,\beta)$37 with oracle $q_\phi(\theta \mid x,\beta)$38.

The same source reports that ambiguity-stratified gains increase monotonically with held-out ambiguity score: $q_\phi(\theta \mid x,\beta)$39 is $q_\phi(\theta \mid x,\beta)$40 in the low regime, $q_\phi(\theta \mid x,\beta)$41 in mid, $q_\phi(\theta \mid x,\beta)$42 in high, and $q_\phi(\theta \mid x,\beta)$43 in very high ambiguity. Synthetic exact-ambiguity experiments further show point-versus-posterior gaps tracking the theoretical barrier: near-Dirac $q_\phi(\theta \mid x,\beta)$44 with point gap $q_\phi(\theta \mid x,\beta)$45; mid $q_\phi(\theta \mid x,\beta)$46 vs $q_\phi(\theta \mid x,\beta)$47; high $q_\phi(\theta \mid x,\beta)$48 vs $q_\phi(\theta \mid x,\beta)$49; very high $q_\phi(\theta \mid x,\beta)$50 vs $q_\phi(\theta \mid x,\beta)$51 (Wang et al., 5 May 2026).

Implementation patterns recur across the literature. High-dimensional fields are typically encoded by CNNs, U-Nets, transformers, or VAE encoders; latent or parameter posteriors are represented by MDNs, normalizing flows, diffusion samplers, particle ensembles, or multi-output neural heads. Inference-time acceleration is obtained either by full amortization, as in $q_\phi(\theta \mid x,\beta)$52 (Sun et al., 29 Jan 2026), or by substituting differentiable surrogates for repeated PDE solves, as in L-DPS (Wang et al., 25 Jun 2026). Physics enters either through the simulator likelihood, through differentiable residuals and adjoints, through hard manifold constraints, or through PINN training loops (Zeghal et al., 2022, Xu et al., 3 Jun 2026, Agata et al., 8 May 2026).

Several limitations are explicit. In $q_\phi(\theta \mid x,\beta)$53-amortized NPE, SNIS weights can become heavy-tailed for $q_\phi(\theta \mid x,\beta)$54 far from $q_\phi(\theta \mid x,\beta)$55, especially as $q_\phi(\theta \mid x,\beta)$56, so ESS monitoring is recommended and Route A may be preferred when weights collapse (Sun et al., 29 Jan 2026). In L-DPS, the posterior is approximate because of restriction to the decoder range, the Tweedie plug-in likelihood, surrogate substitution for the true PDE solver, and finite-step reverse diffusion (Wang et al., 25 Jun 2026). In hard-constraint generation, naive projection or guidance can be severely biased if constraint sensitivity is heterogeneous; the co-area factor is not optional if one claims posterior correctness (Xu et al., 3 Jun 2026). In differentiable-simulator NPE, inaccurate gradients or poorly conditioned adjoints can bias score penalties (Zeghal et al., 2022). In posterior-first single-field prediction, benefits diminish when the observation is effectively identifying, because the posterior becomes near-Dirac (Wang et al., 5 May 2026).

A concise taxonomy of the main posterior-first variants is as follows.

Paradigm	Posterior object	Distinctive mechanism
Single-field posterior-first simulation	$q_\phi(\theta \mid x,\beta)$57 over minimal task-sufficient state	Proper-scoring posterior head before rollout
$q_\phi(\theta \mid x,\beta)$58-amortized SBI for PDE simulation	$q_\phi(\theta \mid x,\beta)$59	One-pass conditional posterior sampling across temperatures
Latent diffusion posterior sampling	$q_\phi(\theta \mid x,\beta)$60 in latent space	DDIM prior steps plus surrogate likelihood guidance
Co-area-correct constrained sampling	$q_\phi(\theta \mid x,\beta)$61	Tangent proposals and Fixman-correct Metropolis
Functional-prior Bayesian PINNs	$q_\phi(\theta \mid x,\beta)$62 in function or weight space	GP/RFF priors with ParVI or pSGLD+R
MO-PINN posterior ensembles	Empirical posterior over $q_\phi(\theta \mid x,\beta)$63 and $q_\phi(\theta \mid x,\beta)$64	Multi-output bootstrap ensemble with physics constraints

Taken together, these works define posterior-first neural PDE simulation as a shift in target: from predicting a single rollout directly from incomplete observations to explicitly representing and propagating posterior uncertainty over the hidden variables that make PDE evolution non-identifiable, misspecified, or computationally intractable when treated monolithically.