acceptodds
Under review as a conference paper at ICLR 2027

PARA: Patch-based Adaptive multi-Resolution, time-Advancing and scale-equivariant neural PDE solver

Abstract

We present PARA, a simple yet effective unified framework for advancing multi-resolution states of hyperbolic PDEs. By representing the spatial solution as a tree of independent local patches at varying resolutions (TreeGrid), PARA shifts structural complexity out of the neural network. The future state of each patch is predicted concurrently using a lightweight network from its localized neighborhood (stencil). Restricting the network to the stencil of a patch gives it an inductive bias towards locality while errors fall and flatten once the stencil captures the physical domain of dependence, in contrast to whole-frame models which capture the entire domain yet yield higher errors. Crucially, we introduce scale normalization, which uses the scaling symmetry of the Euler equations to make the solver scale-equivariant by construction, so a single network learns the solution map in dimensionless units. By expanding the TreeGrid into a hierarchy of dense grids, one per resolution level (MultiGrid), PARA cuts every patch with its halo into a stencil of one shape, so all levels advance in a single batched tensor operation. On the Well's challenging compressible Euler benchmark, our symmetry-driven, patch-based approach achieves exceptional accuracy and generalization across spatial resolutions, outperforming whole-frame single-resolution SOTA neural solvers in one-step accuracy and across long-horizon autoregressive rollouts under identical capacity restrictions.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural PDE solver, multi-resolution, scientific machine learning, adaptive mesh refinement, compressible Euler equations

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

This manuscript proposes the PARA framework, which is a neural solver for hyperbolic PDEs. PARA adapts to multi-resolution cases while preserving locality and physical scale equivariance. The authors propose a TreeGrid structure where the solution is stored as non-overlapping patches at different resolutions, with each patch advanced independently and processed from different resolution levels in a single batched operation. A MultiGrid representation is introduced to convert the irregular TreeGrid into dense grids at each resolution for efficient extraction of these stencils. Instead of using dataset-level mean/std statistics, PARA derives local reference quantities such as mean density, characteristic energy velocity, and stencil length, and normalizes them into dimensionless quantities. The authors argue that this makes the solver scale-equivariant by construction. Experiments on the Well compressible Euler dataset show that PARA substantially improves one-step error for all four baselines and generally improves long-horizon rollout error. However, the experimental settings are, in general, less fair, and the results lack many critical components/metrics. It is thus difficult to recommend acceptance as-is. The authors are encouraged to provide the missing experimental data/metrics, mostly on those they already conducted/reported, to persuade me.

Critical strengths
  • The decomposition of locality, adaptivity, and symmetry, with TreeGrid, multigrid, and normalization, etc., is well-motivated and insightful.
  • The scale-normalization formulation embeds intuitive physical meaning.
  • The experiments show that PARA works with and enhances multiple standard baselines.
Critical weaknesses
  • The computational efficiency claim is not well demonstrated. The results report only the number of stencils processed rather than the actual end-to-end inference time/latency. Since this is an irregular multigrid, it consists of many irregular and small grids/stencils. On a GPU, such small/irregular batches can be much more inefficient than large ones. This means that merely "few stencils", such as reducing the number of stencils by a ratio of x%, does NOT guarantee the end-to-end latency to be decreased by x%, or even decreased. Other efficiency-related metrics, such as FLOPS, memory traffic, or GPU utilization rate, are also not reported. Moreover, the cost for multigrid hierarchy construction, restriction/prolongation, stencil extraction, etc., is also missing. None of these efficiency metrics are reported or discussed.
  • The training cost is not computationally matched. For example, the UNet uses 31h by default, yet 85h with PARA.
  • The experimental scope does not match the claims. Only γ=1.4\gamma = 1.4 subset is evaluated.
Questions for authors
  • What is the actual end-to-end inference speedup? In the overall pipeline, which components offer true speedup, and which are wall-clock bottlenecks?
  • How do legacy numerical frameworks such as MATLAB or AMGCL perform on the γ=1.4\gamma = 1.4 subset? If the proposed method fails to outperform legacy non-neural frameworks, this overall paper will be significantly weakened.
2 comments
AuthorsOpenReview ↗

Response to Reviewer TAFp (1/2)

We thank the reviewer for the careful review, and for the positive remarks on the decomposition into locality, adaptivity and symmetry. You asked for the missing experimental data and metrics, mostly on the experiments we had already run. We have added them to the manuscript, marked in blue, and answer each point below.

  1. End-to-end inference cost (see also our replies to Reviewer wFcJ, point 3, and to the fourth review, relayed by the AC, point 2): You raise an excellent point that fewer stencils need not mean a faster step. We have added the wall clock, FLOPs, GPU utilization, peak memory and memory traffic of a step to the manuscript (Section 4.4, Appendix A.5.7). On one RTX 4090 the step time falls in proportion to the stencil count, and at all three thresholds a PARA step takes less wall clock than the whole-frame step (Table 18, reproduced below).
ConfigurationStencilsStep (ms)GFLOPsGPU utilizationMemory (GB)
Whole frame--907640.990.83
PARA uniform25615230540.983.03
PARA threshold 11307715460.882.72
PARA threshold 2864910310.822.09
PARA threshold 359337050.741.52

The MultiGrid exists to avoid the irregular small batches you describe. It expands the TreeGrid into one dense grid per level, from which every stencil is cut at one shape, so a standard backbone advances all levels in a single batched tensor operation. The network's forward pass takes most of a step, at 0.5 to 0.6 ms per stencil, so reducing the stencil count is what gives the speedup (Figure 17). PARA's own operations, including the MultiGrid construction and the refinement, are a fixed cost of about 5 ms per step at every threshold and are not a bottleneck.

  1. Training cost (see also our reply to the fourth review, relayed by the AC, point 1): In Table 3, the PARA U-Net trained for 10 epochs, about 21 hours ​​(Appendix A.3.2), reaches a one-step VRMSE of 0.0167, below the whole-frame U-Net at 0.0366 after 40 epochs and 32 hours. Therefore, with less training time, the PARA U-Net outperforms the whole-frame U-Net, both with statistical normalization.
ModelEpochsTraining time (h)One-step VRMSE
Whole-frame U-Net40320.0366
PARA U-Net10210.0167
PARA U-Net40850.0126

In our main results we match the training data and the network, the same U-Net at 17.5 M parameters in both forms, not the training time. A surrogate is trained once and used many times, so what matters more is how much it learns from the same data, and the training time matters less.

AuthorsOpenReview ↗

Response to Reviewer TAFp (2/2)

  1. Experimental scope and claims (see also our replies to Reviewer wFcJ, point 2, and to the fourth review, relayed by the AC, point 4): We agree. Our results are on the gamma = 1.4 subset of the benchmark, and the abstract and conclusion now say so. We chose this subset, the value for air at standard conditions, to separate learning the solution map from identifying gamma from the frame (Appendix A.3.1). All our networks, whole-frame and PARA, are trained on this subset. For the Euler equations at gamma = 1.4, we then demonstrate scale symmetry (Section 4.3, extended in this revision) and generalization to an unseen flow (Section 4.5, added in this revision). The published results on all ten values of gamma are listed for reference (Tables 1 and 2). Appendix A.6 shows how scale normalization, and with it PARA, extends to three other hyperbolic PDEs, the Burgers equation, the shallow water equations and ideal MHD.

  2. Comparison against a classical solver: We agree that a neural solver should be placed against a classical solver in wall clock time, and we have added this comparison. We compare PARA with a modern classical DGSEM solver, Trixi.jl, on an unseen flow, the Kelvin-Helmholtz instability (Section 4.5, Appendix A.7). The table below is an excerpt of Table 21.

SolverHardwareStencilsTime per frame (ms)VRMSE at step 100
DGSEM, Trixi.jl16-core CPU, 24 threads--932reference
Whole-frame U-NetRTX 4090--903.19
PARA U-Net, threshold 3RTX 4090134780.55

The DGSEM solver, at a CFL number of 0.95, requires on average 34 time steps to cover the Well's frame interval of 0.015 s, while the PARA and whole-frame U-Nets cover it in one step. As shown in the table, PARA at threshold 3 advances one frame faster than the whole-frame U-Net and the classical solver. The solver runs on a CPU and the networks on a GPU, so the costs are not like-for-like. After 100 steps the VRMSE of PARA against the solver's solution is 0.55, below the whole-frame U-Net at 3.19 and below persistence, the initial frame held fixed, at 0.97.

We are glad to discuss any of these points further.

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

This paper presents a method (PARA) for neural hyperbolic PDE solving. The method combines three main components: (1) spatial stencils on the domain, which exploit the local nature of hyperbolic PDEs; (2) adaptive processing of data at multiple levels, to account for multiscale physics; and (3) scale normalization based on symmetries of the Euler equations. With this design, the authors address three limitations of existing neural PDE solvers: the need for the network to learn which parts of a global input are relevant for the prediction task, the computational cost of uniform spatial resolution, and the lack of scale equivariance.

The authors test their method on the Well’s 2D compressible Euler dataset (periodic boundary conditions) and report improvements over four whole-frame baselines in the one-step prediction setting. In contrast to PARA, which operates on patches, the baselines used for comparison are whole-frame neural PDE solvers that process the entire domain.

Critical strengths

I find the idea of combining locality, adaptive multiresolution processing, and scale normalization interesting, although, as the authors themselves state, the individual components are not new. Still, the result is a well-engineered framework. It is a nice feature that the proposed method can be applied to different neural network architectures without requiring major modifications. The authors also show performance improvements (in VRMSE) over the considered whole-frame baselines.

Critical weaknesses
  1. The authors motivate their method through the locality property of hyperbolic PDEs, but I am not fully convinced that the experiments show locality to be the component responsible for the performance improvements. As presented, many things change between PARA and the baselines (for example, multiresolution training and residual prediction, the number of training samples, and the number of optimizer steps). I appreciate the ablation study on halo width and the comparison of training procedures (Table 15). However, these experiments do not fully disentangle the effect of locality from the other changes. I would be happy to see a more detailed ablation study, as it would help to judge where the improvement comes from.

  2. PARA is evaluated on only one dataset (the Euler dataset from the Well) with periodic boundary conditions and a fixed value of gamma. It is unclear how well the approach works on other hyperbolic PDEs and with different boundary conditions. In this regard, I find the title of the paper quite misleading: the paper does not present a general neural PDE solver, but rather a neural hyperbolic PDE solver for periodic boundary conditions. Moreover, I am not sure that the method generalizes to other hyperbolic PDEs, as the scale normalization and the refinement criterion are specifically designed for the Euler equations.

Questions for authors
  1. Could the authors provide further ablation studies to show the effect of locality more clearly?

  2. In Table 5, the authors claim large reductions in the number of processed stencils through adaptive refinement. How do these reductions translate into actual inference time, which is what practitioners in the field care about? In particular, how do the inference times of whole-frame prediction, uniform PARA, and adaptive PARA compare when multiresolution processing is included?

4 comments
AuthorsOpenReview ↗

Response to Reviewer wFcJ (1/2)

We thank the reviewer for the careful review, and for recognizing that the framework applies across standard deep learning architectures. You asked for a clearer attribution of the gains, for evidence beyond one dataset, and for inference times. We have added these to the manuscript, marked in blue, and answer each point below.

  1. Attribution of PARA's gains (see also our reply to the fourth review, relayed by the AC, point 1): We agree that the comparison between PARA and the whole frame changes several things at once. We have revised Section 4.1 (marked in blue) to attribute the gain to the PARA framework as a whole, in which the stencil reduces the filtering the network must learn and the MultiGrid creates training samples at every level from each frame. Locality is a necessary part of it, not the sole cause.

In PARA, locality means predicting each patch from a bounded neighborhood, and the halo width H sets the size of that neighborhood, so varying H is varying locality directly. The halo width study (Table 3, reproduced below) is a controlled ablation within PARA. Its rows H = 0 to 32 share the network, the training on the stencils of every level, the patch residual as target, and the training procedure, including the number of samples and optimizer steps, and differ only in the halo width (the rows with H > 16 omit level 1, as the caption notes). The H = 0 row is therefore PARA with every component you listed but without the local neighborhood, and it scores 0.1080, three times worse than the whole-frame U-Net at 0.0366. Adding the halo alone lowers the error to 0.0167 at H = 16 and it saturates beyond that, so PARA outperforms the whole frame only once the neighborhood is added.

Halo width HStencilc (%)m (%)One-step VRMSE
032^201370.1080
848^25.216.10.0289
1664^281.51.330.0167
2480^295.00.0900.0151
3296^299.50.00370.0150
Whole frame512^210000.0366

Three further measurements show that the error saturates as the stencil covers the domain of dependence, the locality property of hyperbolic PDEs that motivates PARA.

(a) Coverage (Table 3, columns c and m). c is the share of training patches whose domain of dependence lies inside the stencil, and m is the cells outside the stencil that reach the patch, as a percentage of the stencil's cells. As H grows, c rises toward 100 percent and m falls toward zero, both in step with the fall and saturation of the error.

(b) Signal reach (Table 7, Figure 7). The farthest a signal travels from a cell in one step has a median of 12.1 cells, a 90th percentile of 17.9 and a 99th percentile of 28.8. A halo of 16 cells therefore holds the reach of most cells, and one of 32 nearly all. Figure 7 visualizes the directional signal travel toward a patch, showing how much of its domain of dependence lies inside the stencil at H = 8 to 32.

(c) Sensitivity (Figures 4 and 14). The response of the trained network to an input cell falls by an order of magnitude over the first 16 cells from the patch, so the network has learned the extent of the domain of dependence.

The training procedure is not the cause. We trained each baseline under its authors' default procedure. Under PARA's procedure the whole-frame U-Net changes by only 7 percent, from 0.0366 to 0.0342, and PARA's gain holds under both (Table 16).

AuthorsOpenReview ↗

Response to Reviewer wFcJ (2/2)

  1. Generalization and title (see also our replies to Reviewer TAFp, point 3, and to the fourth review, relayed by the AC, point 4): We agree with the scoping. We have changed the title to "PARA: Patch-based Adaptive multi-Resolution, time-Advancing and scale-equivariant neural hyperbolic PDE solver", and the conclusion now names the setting of our results, the gamma = 1.4 subset with periodic boundary conditions.

To test PARA beyond the one dataset, we have added a flow outside the training set, the Kelvin-Helmholtz instability (Section 4.5, added in this revision). We generate the reference with Trixi.jl, a DGSEM code whose numerical method differs from that of Clawpack, the code behind the Well's data. Without retraining, the VRMSE of PARA after 100 steps is 0.55, below persistence, the initial frame held fixed, at 0.97, while the whole-frame U-Net reaches 3.19. The scale symmetry test (Section 4.3) now also covers time and length scaling, and the error of PARA with scale normalization is unchanged under all three.

Our scale normalization relies on localized dimensional analysis rather than dataset statistics, so it extends to other hyperbolic PDEs with a scaling symmetry. We have added Appendix A.6, which derives the local reference quantities for the Burgers equation, the shallow water equations and ideal MHD. The refinement indicator takes no part in training, so for another PDE the user can choose any scalar suited to it (Appendix A.6). We fixed gamma = 1.4 to separate learning the solution map from identifying gamma from the frame, as Appendix A.3.1 explains, and all our networks, whole-frame and PARA, are trained on this subset. Non-periodic boundary conditions can be supported by populating halos with ghost cells, a standard extension we leave to future work.

  1. Inference time (see also our replies to Reviewer TAFp, point 1, and to the fourth review, relayed by the AC, point 2): The reduction in stencils carries over to inference time. We have added the wall clock, FLOPs, GPU utilization, peak memory and memory traffic of a step to the manuscript (Section 4.4, Appendix A.5.7). At thresholds 1, 2 and 3 of Table 5, an adaptive PARA step takes less wall clock than the whole-frame step, and a uniform PARA step takes more (Table 18, reproduced below, measured on one RTX 4090).
ConfigurationStencilsStep (ms)GFLOPsGPU utilizationMemory (GB)
Whole frame--907640.990.83
PARA uniform25615230540.983.03
PARA threshold 11307715460.882.72
PARA threshold 2864910310.822.09
PARA threshold 359337050.741.52

The forward pass dominates at every setting, at 0.5 to 0.6 ms per stencil (Figure 17). PARA's own operations, the MultiGrid construction, the refinement, the stencil cut, and the TreeGrid assembly, are a fixed cost of about 5 ms per step. Uniform PARA is 1.7 times slower than the whole frame because the halos give it four times the cells, so the speedup comes from the adaptive rollout.

We are glad to discuss any of these points further.

Thank you. I raised for score to 3.

Thank you. We appreciate your time and feedback.

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

The paper introduced PARA, a neural solver for hyperbolic PDEs. By integrating patch based prediction with hirarchical multigrid representation and a scale normalization method based on the Euler equations, the authors address three fundamental mismatches in existing neural solvers, namely locality, AMR compatibility and scale equivariance. The framework is validated on the Well benchmark across multiple backbone models and shows substantial improvements in both single-step accuracy and long-horizon rollouts.

Critical strengths
  1. Justification of the receptive field. Each patch in the domain was extended by a halo of width H as an inductive bias for locality. This decision was justified both theoretically, for hyperbolic equations, and empirically, by exaxmining various sizes of H. Through experiments, the authors showed that information from the global context is redundant for this task.
  2. Physics oriented scale normalization. Previous methods typically rely on statistical normalization (e.g. in the Well), which, as the authors note, loses the physical ratios between fields. Instead, the authors use the scaling symmetry of the Euler equations to define the normalization. As a result, they show that in long-horizon prediction, models with statistical normalization diverge, whereas their scale normalization produces more reliable predictions.
Critical weaknesses

None.

Questions for authors
  1. Table 1 shows that multi-frame history substantially improves model performance under the same capacity restriction. Did you consider integrating it with PARA?
  2. In the halo width sensitivity section, it is not clear whether H=16H = 16 applies only to the most refined AMR level or to all levels. Is the same halo width used at every level? Intuitively coarser levels would need a smaller HH but this is not explicitly mentioned in the text.
1 comment

We thank the reviewer for recommending acceptance, and for the positive remarks on scale normalization and on the justification of the local receptive field. We agree with the intuition behind both of your questions. We answer them below, and we are glad to discuss anything further.

  1. Multi-frame history: Yes, we considered it, and chose a single input frame for this work to keep memory and inference time low. With one frame, PARA already has a lower error than the whole-frame networks trained on the same data at a similar parameter count (Table 1). Classical multistep methods use more past steps to raise the order of the method, and in the neural context we share your hypothesis that a multi-frame history provides more information that can be used to improve PARA's performance. Integrating it into PARA's stencils is feasible, and we leave it to future work.

  2. Halo width across levels: Yes, the same halo width H, in cells, is used at every level. Your intuition is correct that a coarser level needs fewer halo cells for the same domain of dependence, since its cells are larger and the same H covers a larger physical distance. We keep H constant so that every stencil has one shape, and all levels advance in a single batched tensor operation. The text did not state this, and we have made it explicit in the coverage paragraph of Section 4.2, which now reads "As the halo width increases from 0 to 32 cells at every level, the one-step VRMSE at level 4 drops and then flattens, ..." (marked in blue).

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers

We thank all reviewers for their constructive and detailed feedback. We are encouraged that the reviewers found the decomposition of the PARA framework into locality, adaptivity and symmetry well-motivated and insightful (TAFp), the framework well-engineered and applicable to different neural network architectures (wFcJ), the local receptive field justified both theoretically and empirically (YxzT), and the accuracy gains substantial, with useful ablations and transparent reporting of limitations (the fourth review).

The main changes in the revised manuscript, marked in blue, are:

  1. A generalization test on an unseen flow, the Kelvin-Helmholtz instability (Section 4.5). (TAFp, wFcJ, the fourth review)
  2. Time and length scaling in the scale symmetry test (Section 4.3). (the fourth review)
  3. End-to-end inference cost measurements of the adaptive rollouts (Section 4.4, Appendix A.5.7). (TAFp, wFcJ, the fourth review)
  4. Title revised to "neural hyperbolic PDE solver". (wFcJ)
  5. Abstract and conclusion now name the gamma = 1.4 subset and add the generalization to an unseen flow. (TAFp, wFcJ, the fourth review)

The full list of changes, in the order of appearance in the revised paper:

  • Title. Revised to "neural hyperbolic PDE solver". (wFcJ)
  • Abstract. Names the gamma = 1.4 subset of the benchmark, and adds the generalization to an unseen flow. (TAFp, wFcJ, the fourth review)
  • Section 4.1. The gain is attributed to the PARA framework as a whole, not to locality alone. (wFcJ, the fourth review)
  • Section 4.2. States that the halo width applies at every level. (YxzT)
  • Section 4.3, Table 4. The scale symmetry test now covers time and length scaling in addition to density scaling. (the fourth review)
  • Section 4.4. Wall clock of a step at thresholds 1, 2 and 3, against the uniform grid and the whole-frame U-Net, and the effect of half the time step on the diverging rollout. (TAFp, wFcJ, the fourth review)
  • Section 4.5. New. Generalization to the Kelvin-Helmholtz instability over a continuous rollout of 100 steps, scored against a classical DGSEM solver, with the solver's cost. (TAFp, wFcJ, the fourth review)
  • Conclusion. Names the gamma = 1.4 subset and the periodic boundary conditions, and reports the divergence of one uniform rollout with its remedy. (TAFp, wFcJ, the fourth review)
  • Appendix A.3.2. Training time of the ten-epoch PARA U-Net. (TAFp, the fourth review)
  • Appendix A.5.5. New. A rollout at half the time step, which removes the divergence of the scale-normalized U-Net. (the fourth review)
  • Appendix A.5.7. New. Inference cost: wall clock by operation, FLOPs, GPU utilization, peak memory and memory traffic. (TAFp, wFcJ, the fourth review)
  • Appendix A.5.8. New. Details of the scale symmetry test. (the fourth review)
  • Appendix A.6. New. Reference quantities of scale normalization for the Burgers equation, the shallow water equations and ideal MHD. (wFcJ)
  • Appendix A.7. New. Setup, reference reliability, full results and cost of the Kelvin-Helmholtz test. (TAFp, wFcJ, the fourth review)

We address the specific remarks and questions of each reviewer in the individual replies.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.