Error-Conditioned Neural Solvers

• Error-conditioned Neural Solvers are iterative neural PDE solvers that use the residual field as a conditioning signal rather than as a target for optimization.
• The method employs a predictor–corrector architecture where a predictor generates an initial guess and a recurrent corrector refines the solution based on spatial residuals.
• Empirical studies demonstrate that ENS can significantly outperform traditional residual minimization methods in terms of speed and reconstruction accuracy under challenging conditions.

Error-conditioned Neural Solvers (ENS) are iterative neural PDE solvers in which the current PDE residual field is used as a direct conditioning signal for correction, rather than as an objective for an inner numerical optimizer. In the formulation introduced in "Error-Conditioned Neural Solvers" (Jiang et al., 25 Jun 2026), a predictor produces an initial field, a recurrent corrector reads the spatial structure of the residual at each iteration, and the solver learns an update policy that improves reconstruction accuracy under known governing equations. ENS therefore occupies a distinct position between feed-forward neural surrogates that are physics-blind at inference and hybrid methods that perform test-time residual minimization through gradient descent or Gauss–Newton steps (Jiang et al., 25 Jun 2026).

1. Definition and scope

ENS is formulated for operator-learning settings in which one seeks a fast map from PDE-defining inputs to solution fields. The governing problem is written either as

$\mathcal L_\theta(u)=f,$

or through a residual operator

$r(u)=F(u;f,g)\approx 0,$

where the inputs may include forcing $f$, boundary conditions $g$, coefficients or parameters $\theta$, and the output is a discretized field $u$ or a spatiotemporal trajectory (Jiang et al., 25 Jun 2026).

The defining ENS update is iterative. A predictor $P_\theta$ produces an initialization $\hat u^{(0)}$; the residual of the current iterate is computed as

$r^{(k)} = F(\hat u^{(k)};f),$

and a corrector $C_\phi$ outputs a learned correction

$r(u)=F(u;f,g)\approx 0,$0

followed by

$r(u)=F(u;f,g)\approx 0,$1

The crucial distinction is that $r(u)=F(u;f,g)\approx 0,$2 is passed into the network as an input feature map, so the residual is a conditioning signal rather than an optimization target (Jiang et al., 25 Jun 2026).

This definition excludes several nearby but non-identical families. Post hoc neural error estimators, auxiliary error-equation solvers, residual-minimization hybrids, indirect neural correctors, and formally certified neural solvers all share aspects of error awareness, but they do not in themselves instantiate the residual-as-input refinement policy that characterizes ENS in the narrow sense (Dogra et al., 2020).

2. The residual–reconstruction gap

The central theoretical motivation for ENS is the claim that low PDE residual does not, in general, imply accurate reconstruction. Proposition 1 of (Jiang et al., 25 Jun 2026) states that if $r(u)=F(u;f,g)\approx 0,$3 is the true solution with $r(u)=F(u;f,g)\approx 0,$4, then for any $r(u)=F(u;f,g)\approx 0,$5 there exists $r(u)=F(u;f,g)\approx 0,$6 such that

$r(u)=F(u;f,g)\approx 0,$7

When the smallest singular value of the residual Jacobian is small, residual minimization becomes a poor proxy for reconstruction. The appendix gives explicit linear and nonlinear constructions, and the paper identifies Helmholtz near resonances and advection-dominated Navier–Stokes/Kolmogorov regimes as concrete settings where this gap is amplified (Jiang et al., 25 Jun 2026).

This argument is directed against a common misconception: that any method driving $r(u)=F(u;f,g)\approx 0,$8 downward is necessarily approaching the desired solution. ENS rejects that identification. PINO test-time optimization, diffusion-guided methods, and PCFM all use residuals at inference, but they still treat residual as something to optimize externally, and therefore inherit the failure mode in which the solver moves along nearly flat low-residual directions without improving reconstruction (Jiang et al., 25 Jun 2026).

A second misconception is that residuals are never useful as error surrogates. The certification framework of (Mukherjee et al., 19 Mar 2026) is more qualified: it shows that residual, boundary, and initial-condition violations can be converted into explicit solution-error guarantees when two extra ingredients are available, namely compactness of the candidate solution family and an equation-specific stability estimate. In that setting one obtains bounds of the form

$r(u)=F(u;f,g)\approx 0,$9

for sampled residuals, or

$f$0

when a global residual certificate is available (Mukherjee et al., 19 Mar 2026). ENS is therefore best understood not as a denial of residual information, but as a re-interpretation of residuals: informative, but not intrinsically aligned with reconstruction unless additional structure is present.

3. Architecture, training, and residual conditioning

The ENS architecture comprises two learned components: a predictor $f$1 and a recurrent corrector $f$2. The corrector is reused across all correction steps, so the method is an unrolled recurrent refinement procedure rather than a sequence of independently trained stages. Predictor and corrector are trained jointly, but the corrector inputs are detached across iterations, so gradients are not propagated through the full correction chain; this lets the corrector learn from the actual distribution of intermediate residuals encountered during rollout (Jiang et al., 25 Jun 2026).

The training loss supervises all iterates: $f$3 In the reported implementation, ENS is trained with $f$4 unrolled correction steps and then run for many more iterations at inference, with stable improvement reported past 100 iterations until a residual floor is reached (Jiang et al., 25 Jun 2026).

How the residual is computed depends on the PDE family. For static PDEs—Helmholtz, Poisson, and Darcy—the paper uses second-order central finite differences on a uniform $f$5 grid with spacing $f$6. For nonlinear Helmholtz,

$f$7

the residual is the finite-difference Laplacian plus $f$8 at each grid point. For Navier–Stokes and Kolmogorov flow, the state is a vorticity trajectory, temporal derivatives are computed by finite differences, and spatial derivatives spectrally, giving

$f$9

as the residual field used for correction (Jiang et al., 25 Jun 2026).

Backbone choice is not incidental. For static equations, both predictor and corrector are modified FNOs with four Fourier layers, 64 hidden channels, and 20 Fourier modes per spatial dimension; the usual lifting and projection layers are replaced with 3-layer CNNs, or 5-layer CNNs for Darcy, to preserve local and high-frequency residual structure. For Navier–Stokes and Kolmogorov flow, both components use a transformer-based VideoPDE backbone. Inputs to the corrector are concatenated along channels as $g$0 (Jiang et al., 25 Jun 2026).

The training setup reported in (Jiang et al., 25 Jun 2026) uses AdamW with learning rate $g$1, weight decay $g$2, and batch size $g$3. The correction step size is fixed during training, with $g$4 for static PDEs and Kolmogorov flow, and $g$5 for Navier–Stokes. The main computational claim is that one ENS refinement step requires one residual computation plus one forward pass of the corrector, which is substantially cheaper than inference-time optimization requiring repeated backward passes or Jacobian-based projection (Jiang et al., 25 Jun 2026).

4. Empirical behavior and operational regimes

ENS was evaluated on four PDE families: linear and nonlinear Helmholtz, Poisson, Darcy flow, and Navier–Stokes/Kolmogorov flow. Static datasets were generated on $g$6 grids; training sets contained 1,000 samples per static PDE and Navier–Stokes, and 5,000 for Kolmogorov flow, with 100 test samples. The reported metrics are relative $g$7 reconstruction error and PDE residual MSE, precisely because the two can disagree severely in ill-conditioned regimes (Jiang et al., 25 Jun 2026).

The reported results illustrate the residual–reconstruction gap directly. On Kolmogorov flow in-distribution, ENS attains relative error $g$8 in $g$9 s/sample, whereas PINO(TTOP) attains a much lower residual $\theta$0 than ENS’s $\theta$1, yet has reconstruction error $\theta$2, roughly $\theta$3 worse. Under forcing shift, ENS reaches $\theta$4, while PINO(TTOP) remains at $\theta$5, about $\theta$6 worse, matching the headline claim that ENS gains can reach $\theta$7 on turbulent Kolmogorov flow (Jiang et al., 25 Jun 2026).

On Helmholtz, ENS is reported as best on both reconstruction and residual metrics in all regimes listed in Table 2 of (Jiang et al., 25 Jun 2026). In-distribution it reaches $\theta$8 and $\theta$9. Under wavenumber shift, ENS obtains $u$0 relative error, whereas PINO(TTOP) remains at $u$1 despite a low residual, again exemplifying the mismatch between residual reduction and reconstruction quality (Jiang et al., 25 Jun 2026).

On Navier–Stokes, ENS is best in-distribution and under viscosity and forcing shift, with the reported super-resolution exception where PINO(TTOP) and POSEIDON can outperform it. In-distribution ENS reaches $u$2 relative error in $u$3 s/sample, compared with PINO(TTOP)’s $u$4 at $u$5 s/sample. Under viscosity shift, ENS attains $u$6, while PINO(TTOP) attains $u$7 despite lower residual (Jiang et al., 25 Jun 2026).

The empirical interpretation offered in (Jiang et al., 25 Jun 2026) is not merely that residual conditioning adds extra iterations. Ablations replacing the residual input with a zero field while preserving the same architecture and correction budget cause both reconstruction and residual improvements to stall. Conditioning on the gradient of the physics loss rather than the residual field accelerates residual decrease but not reconstruction improvement. This indicates that the main ENS effect is not generic iterative computation, but access to the residual field as structured spatial context (Jiang et al., 25 Jun 2026).

5. Precursors and adjacent error-aware solver families

ENS was formalized in 2026, but several earlier and adjacent lines of work developed components that later became ENS-relevant. The following grouping reflects the terminology used in those papers.

Work	Relation to ENS	Key mechanism
(Dogra et al., 2020)	Early and fairly explicit precursor	Internal local error reconstruction from residual-induced error dynamics; second-stage learned correction
(Dogra, 2020)	Conceptual precursor, not modern ENS	Derives an auxiliary DE for $u$8; trains a second network $u$9 and returns $P_\theta$0
(Wright, 2022)	Recursive error-correction network	Trains successive correction networks on derived error PDEs for Poisson-type problems
(Mukherjee et al., 19 Mar 2026)	Certification framework	Converts residual, boundary, and initial errors into explicit solution-error guarantees under compactness and stability
(Lu et al., 1 Jul 2026)	Error-aware operator-conditioned neural residual solver	For second-kind BIEs, residual controls error through uniformly bounded $P_\theta$1; multilevel schedule uses residual tolerances
(Wei et al., 16 Nov 2025)	Error-mitigating hybrid solver	Inserts neural correction into the PDE RHS rather than directly into the state, reducing perturbation amplification on the order of $P_\theta$2
(Rout, 8 Aug 2025)	Conditioning-aware feature design	Identifies ill-conditioning of the activation/design matrix as the dominant bottleneck in PIELM and improves rank via Shifted Gaussian Encoding
(Kim et al., 8 May 2026)	Explicit error decomposition without conditioning	Separates residual, collar, policy, and learned-model error in stabilized neural HJB policy iteration
(Gorard et al., 16 Feb 2026)	Certified bounded-error neural solver	Uses characteristic-based regularity analysis, compositional error bounds, and theorem proving for worst-case $P_\theta$3 control

The ODE-oriented works (Dogra et al., 2020) and (Dogra, 2020) are especially close in spirit to ENS because they elevate the solution error to an explicit latent object. In (Dogra et al., 2020), the residual is interpreted as the forcing term of an approximate error ODE, yielding a discrete recursive estimator for pointwise state error and a second neural network trained on the estimated error trajectory. In (Dogra, 2020), the original differential equation is rewritten into an auxiliary error equation

$P_\theta$4

and a second network $P_\theta$5 is trained to approximate the error field. Both are post hoc internal error estimation and correction frameworks rather than jointly conditioned residual-to-update policies, but they supply a direct intellectual lineage for ENS (Dogra et al., 2020).

The remaining works broaden the ENS landscape in complementary directions. (Mukherjee et al., 19 Mar 2026) explains when residuals can be trusted as error signals at all; (Lu et al., 1 Jul 2026) shows that this trust can become structurally robust when the underlying operator is uniformly well-conditioned, as in second-kind boundary integral equations; (Wei et al., 16 Nov 2025) shows that long-horizon stability depends not only on what error signal is available, but also on where neural correction enters the solver; (Rout, 8 Aug 2025) identifies low rank and ill-conditioning of the feature matrix as a primary source of solver error amplification; (Kim et al., 8 May 2026) gives an explicit decomposition of residual, boundary-collar, policy, and model-identification errors in neural HJB policy iteration; and (Gorard et al., 16 Feb 2026) demonstrates that error-bounded compositional design and formal verification can be built into neural PDE solvers even without explicit runtime conditioning.

6. Limitations, misconceptions, and research directions

ENS inherits several limitations from the operator-learning setting in which it was introduced. The published experiments are restricted to relatively simple 2D PDE families on regular grids, and the method assumes that the governing equations are known at inference so that a residual can be computed. Performance also depends on the backbone’s ability to represent high-frequency residual structure; in the reported static experiments, pure FNO backbones could diverge for Helmholtz, whereas CNN-augmented FNOs converged (Jiang et al., 25 Jun 2026).

A broader limitation is conceptual. ENS uses residuals as context rather than as an optimization target, but this does not make residuals universally reliable. The certification results of (Mukherjee et al., 19 Mar 2026) imply that residual-derived quantities become trustworthy error indicators only when supported by compactness and PDE-specific stability estimates. The multilevel BIE analysis of (Lu et al., 1 Jul 2026) sharpens the same point in a more operator-specific way: for second-kind Nyström systems, residual controls discrete solution error because $P_\theta$6 stays uniformly bounded, so residual minimization has mesh-independent a posteriori meaning. This suggests that ENS-style conditioning should be strongest in regimes where the residual admits a structurally justified error interpretation (Mukherjee et al., 19 Mar 2026).

Another misconception is that any error-aware neural solver is an ENS. The literature surveyed here separates several non-equivalent ideas: post hoc error-PDE solvers (Dogra, 2020), internal local error transport and learned correction (Dogra et al., 2020), indirect low-gain correction pathways for autoregressive PDE solvers (Wei et al., 16 Nov 2025), and bounded-error formally verified architectures (Gorard et al., 16 Feb 2026). All are relevant to ENS, but none is identical to the 2026 formulation in which the residual field itself is the conditioning channel for a recurrent corrector.

The most plausible future directions are already implicit in the precursor literature. Joint solution–error networks, adaptive collocation driven by estimated error magnitude, error-conditioned correction blocks, and residual-to-error supervision without ground truth are all suggested explicitly in the discussions surrounding post hoc error-equation methods (Dogra, 2020). A second plausible direction is to combine explicit error conditioning with low-amplification insertion strategies of the kind used by indirect correctors, so that the corrector is informed by residual or uncertainty features but its own mistakes enter the numerical dynamics through a filtered RHS pathway rather than a direct state overwrite (Wei et al., 16 Nov 2025). A third is to integrate certified residual-to-solution bounds or operator-specific conditioning diagnostics into the correction policy itself, moving ENS from empirical residual awareness toward calibrated error awareness (Kim et al., 8 May 2026).

In this sense, ENS marks a reclassification of PDE residuals within neural scientific computing. Earlier methods either minimized residuals, bounded their consequences, or used them to derive auxiliary error equations. ENS turns the residual field into a first-class inference-time signal from which a neural solver learns how to update its own prediction. The concept is therefore both architectural and epistemic: architectural because it defines a predictor–corrector refinement loop, and epistemic because it treats the solver’s current constraint violation as readable structure rather than as a scalar loss to be reduced blindly (Jiang et al., 25 Jun 2026).