acceptodds
Under review as a conference paper at ICLR 2027

A Greedy PDE Router for Blending Neural Operators and Classical Methods

Abstract

When solving PDEs, classical numerical solvers are often computationally expensive, while machine learning methods can suffer from spectral bias, failing to capture high-frequency components. Designing an optimal hybrid iterative solver–where, at each iteration, a solver is selected from an ensemble of solvers to leverage their complementary strengths–poses a challenging combinatorial problem. While greedy selection is desirable for its constant-factor approximation guarantee to the optimal solution under Lipschitz assumptions, it requires knowledge of the true error at each step, which is unavailable in practice. We address this by proposing an approximate greedy router that efficiently mimics a greedy approach to solver selection. Empirical results on the Poisson, convection–diffusion, and anisotropic diffusion equations show that our method consistently reduces time-to-solution relative to single-solver baselines and existing hybrid approaches such as HINTS.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
hybrid solver, pdes, model routing

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_qf5COpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a greedy iterative PDE solver that dynamically switches between classical solvers and NOs at each iteration. The central motivation is that classical iterative methods are effective at reducing certain error components, particularly high-frequency errors, while neural operators can provide rapid corrections but may suffer from spectral bias. The authors propose a greedy solver selection routine with theoretical guarantees and learn an approximate greedy router. Experiments on three 2D elliptic PDEs show some performance improvement.

Critical strengths
  • The proposed greedy solver selection routine with theoretical guarantees is insightful and interesting.
  • The evaluations on 2D elliptic PDEs show some performance gains.
Critical weaknesses
  • The theoretical guarantee relies on strict conditions such as zero-preserving contraction. These conditions may not hold for practical PDE-solving scenarios.
  • Limited scope. Only three elliptic PDEs on 128x128 uniform grid, and periodic Dirichlet BCs. This setting is insufficient to demonstrate the results.
  • Limited baselines. Only legacy baselines are compared with. How does the method compare with a pure NO, such as Transolver [1] or LaMO [2], or other neural multigrid solver methods such as [3]?
  • Above-"oracle" performance. In Table 1, for Conv equations, Learned Greedy-GS (35.60) outperforms True-Greedy-GS (35.77, smaller is better), while the authors claim "True-Greedy is an oracle". How is an oracle ("ideal" strategy) surpassed? I would recommend the authors proofread the numbers in theTables.
  • The Figures are too small and hard to read.
  • References:
    • [1] Wu et al., Transolver: A Fast Transformer Solver for PDEs on General Geometries, ICML 2024.
    • [2] Tiwari et al., Latent Mamba Operator for Partial Differential Equations, ICML 2025.
    • [3] Greenfeld et al., Learning to Optimize Multigrid PDE Solvers, ICML 2019.
Questions for authors
  • Please see "weaknesses" above.
3 comments

We thank the reviewer for the careful assessment. We address each point below.

Critical Weakness 1: The theoretical guarantee relies..

Zero preservation is natural for residual-based classical iterations: a zero residual produces no update, and the same property can be enforced in a learned corrector. Contraction in the stated norm also holds for suitably damped iterations on many discretized linear elliptic problems; weighted Jacobi in our setting provides an example. We agree, however, that it is not automatic for every solver, neural correction, or PDE. In particular, the proposition does not cover an action whose error map fails contraction in that norm, and contraction alone does not establish postfix monotonicity for a mixed ensemble. We have made these conditions explicit, extended the greedy bound to weak μ\mu-postfix monotonicity, and included an example showing that weak supermodularity can hold even when global contraction fails.

Critical Weakness 2: Limited scope.

We have expanded the evaluation to Darcy flow and reaction–diffusion with spatially varying coefficients and forcing functions, using homogeneous Dirichlet boundary conditions on a 128 grid. These experiments also use LaMO neural corrections rather than DeepONet. The results appear in Section 5 (“Varying Coefficients and Forcings”) and Table 2 of the revised manuscript. Please see our response to Reviewer 84BD’s Critical Weakness 3 for details.

Critical Weakness 3: Limited baselines.

We have added comparisons with dedicated one-shot LaMO and Transolver models, alongside our standalone DeepONet, on the same 64 instances for each of the three PDEs. The table reports relative solution error and inference time as mean ± standard deviation. Although one-shot inference is inexpensive, none of the models achieves the 10−1210^{-12} relative solution-error target used in our time-to-solution experiments. We additionally tested whether repeated neural residual corrections, u(k+1)=u(k)+Nθ(f−Au(k))u^{(k+1)}=u^{(k)}+\mathcal N_\theta(f-Au^{(k)}), could reach this accuracy without numerical steps. DeepONet’s error plateaus above the target, with no successful cases within the 60-second evaluation budget. Undamped LaMO and Transolver residual iterations diverge on all evaluated cases. These findings illustrate why low-cost neural predictions and high-accuracy hybrid solves serve different purposes, and support combining neural corrections with numerical refinement.

PDEModelOne-shot relative L2 errorOne-shot time (ms)Iterative terminal relative L2 errorIterative outcome
PoissonDeepONet0.000598 ± 0.001375.56 ± 0.020.000598 ± 0.00137Budget limit: 64/64
PoissonLaMO0.06 ± 0.02055.94 ± 0.08Diverged (>10⁶ guard)Diverged: 64/64
PoissonTransolver0.513 ± 0.1938.40 ± 0.10Diverged (>10⁶ guard)Diverged: 64/64
ConvDiffDeepONet0.00133 ± 0.003035.55 ± 0.020.00133 ± 0.00303Budget limit: 64/64
ConvDiffLaMO0.0962 ± 0.0416.17 ± 0.20Diverged (>10⁶ guard)Diverged: 64/64
ConvDiffTransolver0.641 ± 0.1667.76 ± 0.02Diverged (>10⁶ guard)Diverged: 64/64
AnisoDiffDeepONet4.33e-05 ± 7.67e-056.06 ± 0.594.31e-05 ± 7.68e-05Budget limit: 64/64
AnisoDiffLaMO0.0742 ± 0.03056.75 ± 0.05Diverged (>10⁶ guard)Diverged: 64/64
AnisoDiffTransolver0.0436 ± 0.03638.03 ± 0.04Diverged (>10⁶ guard)Diverged: 64/64

Critical Weakness 4: Above-"oracle" performance.

``True-Greedy’’ means that each decision uses the true error to select the action with the largest immediate error reduction. It is a one-step oracle, not the globally optimal sequence or an oracle for minimum wall-clock time-to-accuracy. The learned router may therefore follow a different trajectory that reaches the tolerance slightly sooner. In the Conv–GS row, the reported difference is 0.17 ms, much smaller than the standard errors. We, therefore, do not interpret it as evidence that learned routing outperforms the oracle greedy policy.

Thank you for your detailed explanation. The additional numerical results also look promising. I am raising my score from weak reject to weak accept.

Thank you for reviewing our response and updating your assessment. We appreciate your feedback.

Reviewer_qTrXOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies hybrid iterative linear solvers in which, at every iteration, a router picks one "preconditioning function" from an ensemble (classical Jacobi/GS/SymGS/SOR/MG and learned neural corrections), generalizing HINTS.

The three main contributions are:

  1. Theory for an oracle greedy rule that knows the true error.
  2. A learned router that imitates the oracle greedy rule.
  3. Experiments on periodic 2D Poisson, convection-diffusion and anisotropic diffusion on a 128*128 grid show some promising results.
Critical strengths
  1. A clean and general formulation, and a convincing empirical case that adaptive timing matters. Writing the hybrid solver as u + C_{S_t}(f − Lu) with a per-step choice over an arbitrary ensemble subsumes HINTS and parameterized solver families. The paired experiments, the convergence histories and the action-usage plots show a consistent pattern. A badly timed neural correction can undo progress. The learned router avoids this and lands within a few percent of the oracle greedy policy across five solver families. The ensemble result also shows that a small set of complementary actions gives most of the benefit.
Critical weaknesses
  1. The theory does not cover the deployed method. Specifically, the bound holds for an oracle greedy rule that knows e^(0). The learned router inherits it only through per-step Bayes consistency under teacher forcing (Thm 4.1). Nothing controls how routing errors compound over T steps, which is the exposure-bias problem that DAgger is then used to patch.

  2. The benchmarks and baselines are weak to support the time-to-solution claim. All main experiments use constant coefficients on periodic 128 grids, with one fixed operator per PDE and only f varying. An FFT solves such systems in O(N log N), and the "neural operator" is a least-squares map onto a Fourier basis, i.e., a truncated spectral solve. Neither an FFT solve nor strong iterative baselines (e.g., MG-preconditioned CG) are reported.

Questions for authors

Q1. Can the authors report wall-clock times for an FFT direct solve and for MG-preconditioned CG or FMG on the same 128 and 512 problems?

Q2. Can you show results where the coefficient field a varies across samples (e.g., heterogeneous Darcy diffusion), so that the neural correction must generalize across operators, not just forcings?

Q3. How does performance change if the router also sees (a, f, u^(t)), as the theory assumes? Conversely, how well does a simple threshold rule do, e.g., "apply the neural correction when the last error ratio exceeds some threshold"?

3 comments

We thank the reviewer for their thoughtful feedback. We address each question below.

Critical Weakness 1: The theory does not cover the…

We agree that our theory does not provide an end-to-end guarantee, but it offers a principled, three-step path from optimal solver sequencing to practical learned routing. Our approach consists of three steps to turn an intractable optimization problem into a practical algorithm:

  1. Optimal sequence -> Oracle greedy (Theorem 3.1): We prove a constant-factor approximation guarantee under structural assumptions with access to true errors.
  2. Oracle greedy -> Learned greedy (Theorem 4.1): Without true errors at inference time, we train a router using a surrogate loss. We establish Bayes consistency for one-step greedy decisions on teacher-forced states, but this does not control errors accumulated along learned trajectories.
  3. Teacher forcing -> Deployment (DAgger): Teacher forcing exposes the router only to oracle-generated states, while deployment follows states induced by its own decisions. We therefore roll out the learned router, obtain greedy-oracle labels for the states it visits, and add them to training. This follows the dataset-aggregation principle of Ross et al. (2011), whose analysis motivates training on learner-induced states to address compounding errors.

Together, these steps provide a principled progression from a hard combinatorial optimization problem to practical learned routing: two steps have formal guarantees in our paper, the third appeals to established imitation-learning theory, and the complete approach is empirically shown to closely track oracle greedy performance across PDEs and solver families.

Critical Weakness 2/Question 1: Can the authors report…

We have evaluated an FFT and multigrid-preconditioned Krylov baselines on the same 64 instances and accuracy target. FFT is decisively faster on these constant-coefficient periodic problems, while comparisons with Krylov methods depend on the solver family. Our objective is to demonstrate that, given a fixed numerical–neural solver pair, adaptive routing can improve on both HINTS scheduling and using the numerical solver alone. For example, a hybrid restricted to Jacobi and a neural correction is not expected to outperform SOR with ω=1.5\omega=1.5, an update unavailable to its router. Thus, the relevant within-family comparisons are Jacobi alone and HINTS using that same pair. FFT and Krylov methods do not fit directly into the numerical component of our current framework: FFT computes a direct solution rather than an iterative update, while Krylov methods maintain search directions or a basis across iterations. Inserting a neural correction therefore requires explicitly handling this state, for example through restarting. Although these are valuable standalone baselines, assessing the benefit of routing within those solver families would require incorporating them into the hybrid action space. We leave this additional algorithmic design to future work.

MethodPoissonConvDiffAniso
FFT0.72 ± 0.140.58 ± 0.020.55 ± 0.02
MG-preconditioned Krylov26.10 ± 1.7541.04 ± 0.30152.75 ± 4.68
Learned Greedy-Jacobi1,280.80 ± 519.792,900.52 ± 1,453.841,117.90 ± 607.95 [62/64]
Learned Greedy-Weighted Jacobi19.95 ± 3.0021.23 ± 1.5427.47 ± 1.11
Learned Greedy-GS35.63 ± 5.7235.60 ± 6.0355.53 ± 18.22
Learned Greedy-SymGS35.04 ± 5.4336.69 ± 5.8469.44 ± 34.26
Learned Greedy-SOR44.60 ± 7.8045.97 ± 8.4747.98 ± 6.39
Learned Greedy-MG29.93 ± 3.7033.71 ± 4.4682.76 ± 40.43

Critical Weakness 2/Question 2: Can you show results where…

We have added experiments on Darcy flow and reaction–diffusion with spatially varying coefficients and forcing functions, evaluated on a 128 grid with homogeneous Dirichlet boundary conditions. These results appear in Section 5 (“Varying Coefficients and Forcings”) and Table 2 of the revised manuscript. Please see our response to Reviewer 84BD’s Critical Weakness 3 for details.

Question 3: How does performance change…

We have added comparisons with a residual-threshold rule and a compact router using pooled (a, f, u) inputs. Please see our response to Reviewer 84BD’s Critical Weakness 2 for the results and experimental details.

Thank you for your detailed response. The numbers look good, so I'm raising my score from 2 to 3 (weak acceptance).

The results against mg-preconditioned krylov look encouraging for some of the hybrid solvers. I also appreciate the added experiments on varying coefficients and the comparisons with simpler routing rules.

My concern about the theory gap still remains, but the clarification is helpful. Since fft is still much faster on these benchmarks, I think the efficiency claim should focus more clearly on the benefit of adaptive routing within the selected solver families. Please make these limitations clear in the revised paper.

Thank you for reconsidering the paper and for raising your score. We agree that the efficiency claim should focus on adaptive routing within the evaluated solver families. If accepted, we will make this scope explicit in the camera-ready main text.

Reviewer_84BDOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes an adaptive hybrid PDE solver that dynamically selects between neural operators and classical numerical methods using a greedy routing strategy. The authors establish theoretical approximation guarantees based on weak sequence supermodularity and introduce a surrogate loss to train the router without access to true errors at inference time. Experiments on three PDE benchmarks demonstrate improved computational efficiency over fixed scheduling strategies such as HINTS.

Critical strengths
  1. The proposed adaptive routing strategy addresses the limitations of fixed scheduling in existing hybrid PDE solvers, offering a promising approach to improving computational efficiency.
  2. The paper provides theoretical analysis of the greedy routing strategy and introduces a surrogate loss to approximate optimal solver selection, strengthening the methodological foundation.
  3. Experiments across three PDE benchmarks and multiple classical numerical solvers demonstrate promising improvements in convergence speed and computational efficiency.
Critical weaknesses
  1. The theoretical guarantees rely on strict contraction and postfix monotonicity assumptions. However, Appendix E shows that several evaluated solvers violate these assumptions, leaving a gap between the theoretical analysis and practical performance.
  2. The experimental evaluation lacks sufficient ablation studies and comparisons with simpler adaptive routing strategies. In particular, the individual contributions of DAgger and the proposed surrogate loss remain unclear, making it difficult to assess the necessity of these components.
  3. The experiments primarily focus on two-dimensional linear PDEs with constant coefficients. The lack of evaluations on nonlinear PDEs, variable-coefficient problems, and out-of-distribution scenarios limits the evidence for the method's generalization capability.
Questions for authors
  1. Given that several evaluated solvers violate the contraction and postfix monotonicity assumptions, can the authors clarify whether the theoretical approximation guarantees remain applicable in practice or establish guarantees under weaker assumptions?
  2. Could the authors provide additional ablation studies to isolate the contributions of DAgger and the surrogate loss, and compare the proposed method with simpler residual-based or threshold-based adaptive routing strategies?
  3. How well does the proposed routing strategy generalize to nonlinear PDEs, variable-coefficient problems, and different neural operator architectures?
2 comments

We thank the reviewer for recognizing our three key contributions. Below, we address each concern

Critical Weakness 1/Question 1: The theoretical guarantees rely on…

We agree that the postfix monotonicity and contraction assumptions are not necessarily established for all solvers in our experiments. In the revised submission, we prove a greedy bound under weak μ\mu-postfix monotonicity, a relaxed notion of postfix monotonicity. Appendix E reports sampled one-action postfix ratios as empirical estimates for μ\mu. The contraction-based result also has practical scope. Many numerical solvers for elliptic problems can satisfy contraction in the stated norm, as our weighted-Jacobi diagnostics illustrate. Moreover, contraction is sufficient rather than necessary for weak supermodularity. A new example in Appendix A.3 has a noncontractive error map but satisfies the defining inequality with α(O)=1\alpha(O)=1. Thus the revised theory covers a meaningful class of solver ensembles, while its applicability to any particular hybrid ensemble depends on verifying the Theorem 4.1’s stated conditions.

Critical Weakness 2/Question 2: The experimental evaluation lacks sufficient…

The table compares a residual-threshold rule, our learned router, and a compact router using pooled (a,f,u) inputs across three PDEs and six numerical solver families. The threshold rule applies a neural correction when residual reduction is insufficient: ∣rt∣2∣rt−w∣2>θ\frac{|r_t|_2}{|r_{t-w}|_2}>\theta, provided at least c routing decisions have elapsed since the previous neural correction; otherwise, it applies the numerical solver block. Here, w is the lookback window, controls the required progress, and c prevents overly frequent costly neural corrections. We selected these parameters on separate development instances and froze them before evaluation. Entries in the table report mean ± standard deviation over successful solves, with incomplete success counts indicated. Our learned router has a lower successful-case mean than the threshold rule in 17 of 18 settings; ConvDiff with Jacobi is the exception. These results support the benefit of learned routing over this simple residual-progress heuristic.

FamilyMethodPoissonConvDiffAniso
JacobiThreshold1,913.37 ± 719.192,858.28 ± 1,075.543,816.06 ± 866.76
Compact (a,f,u)5,862.54 ± 7,434.71 [31/64]3,280.87 ± 1,446.49 [51/64]1,675.12 ± 862.97 [59/64]
Learned Greedy1,280.80 ± 519.792,900.52 ± 1,453.841,117.90 ± 607.95 [62/64]
Weighted JacobiThreshold37.98 ± 2.1340.24 ± 1.9243.38 ± 5.28
Compact (a,f,u)20.21 ± 1.5025.15 ± 4.00 [63/64]32.76 ± 2.16
Learned Greedy19.95 ± 3.0021.23 ± 1.5427.47 ± 1.11
GSThreshold75.24 ± 9.7885.86 ± 2.10186.33 ± 8.25
Compact (a,f,u)56.78 ± 118.8042.18 ± 8.73 [62/64]78.83 ± 59.08
Learned Greedy35.63 ± 5.7235.60 ± 6.0355.53 ± 18.22
SymGSThreshold71.08 ± 5.2568.42 ± 8.25128.08 ± 7.66
Compact (a,f,u)83.38 ± 213.0939.70 ± 11.98 [61/64]72.54 ± 39.42
Learned Greedy35.04 ± 5.4336.69 ± 5.8469.44 ± 34.26
SORThreshold150.77 ± 17.01173.12 ± 7.95130.00 ± 3.27
Compact (a,f,u)67.88 ± 104.3450.47 ± 10.2862.16 ± 22.82
Learned Greedy44.60 ± 7.8045.97 ± 8.4747.98 ± 6.39
MGThreshold35.03 ± 2.8138.29 ± 3.01420.77 ± 40.05
Compact (a,f,u)32.93 ± 6.1334.93 ± 4.95152.30 ± 97.33
Learned Greedy29.93 ± 3.7033.71 ± 4.4682.76 ± 40.43

Critical Weakness 3/Question 3: The experiments primarily focus…

We have expanded the experiments to address this concern. Section 5 (Varying Coefficients and Forcings) and Table 2 now report results for Darcy flow and reaction–diffusion with spatially varying coefficients and forcing, using homogeneous Dirichlet boundaries on a 128 grid. These experiments use LaMO (Tiwari et al., ICML 2025) rather than the DeepONet used in the original experiments, extending the evaluation to another neural architecture. Architecture and router-training details are provided in the appendix. Learned routing reduces mean solve time relative to tuned HINTS by 19.4–38.2% for Darcy paired with GS, SymGS, and SOR, with all instances converging. For reaction–diffusion, SOR improves by 23.4%, while learned Jacobi converges on all 64 instances and HINTS-Jacobi converges on none within the evaluation budget. Reaction–diffusion SymGS is slower than HINTS on average, and multigrid-only remains faster than its hybrid on both PDEs. The new results therefore demonstrate applicability beyond constant-coefficient periodic problems and a single neural architecture.

Our framework assumes that the PDE is linear in the unknown solution. Extending the framework to nonlinear PDEs is an important direction for future work.

Thank you for your detailed response. The additional theoretical clarifications and experimental results are helpful. Although some concerns remain, I appreciate the authors' efforts to address them and will maintain my original score of 3.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.