acceptodds
Under review as a conference paper at ICLR 2027

Scale-Split Neural Operator for Memory- and Data-Efficient 3D Turbulence Prediction

Abstract

Neural surrogates have emerged as fast alternatives to the numerical simulation of three-dimensional turbulence. However, training them at high resolution remains challenging, since the memory of full-field models grows with the resolution. In addition, full-resolution training data are expensive to simulate and store, and therefore scarce. We introduce ScaleSplit-NO (Scale-Split Neural Operator), which exploits the scale structure of turbulence with two neural operators: a Parent predicts the global coarse field at the next time step, and a Child predicts full-resolution local patches conditioned on this prediction. Neither model operates on the full-resolution field. The Child is pretrained alone and then attached to the Parent's coarse prediction through zero-initialized connections. On two complex high-resolution turbulence benchmarks, ScaleSplit-NO surpasses all competing baselines in both prediction accuracy and data efficiency. On the higher-resolution dataset JHTDB256 (), its normalized mean squared error (NMSE) is 53% lower than that of the strongest baseline, and its training memory is 79% lower than that of the most memory-efficient baseline. We further demonstrate its effectiveness for urban wind prediction in a real district of Montreal on a grid, reducing one-step NMSE by 65.8% relative to the baseline. Moreover, swapping in a Parent trained on additional coarse fields improves prediction without retraining the Child, providing further accuracy gains at a small storage cost. Our code is included in the supplementary material.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Neural operators, AI for science, fluid dynamics, partial differential equations

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_TN6POpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper provides a new method for modeling turbulent flows. The proposed approach uses models for predicting coarse and fine regions which is intuitively similar to large-eddy simulations in traditional computational fluid dynamics. Results in the main paper and appendix show better quality of predictions than a few popular baselines. However there are many important results that are hidden away in the appendix, and the main paper only shows mean squared error metrics. Thus, the approach and results and promising, but the main manuscript could be revised for better reporting of a valuable contribution.

Edit 1: Increasing score to accept since all concerns are addressed. Good paper with method that beats strong baselines in stats and physics metrics.

Critical strengths
  1. Proposed modeling approach is intuitive and is shown to be stronger at predicting turbulent flow fields than a variety of baselines.
  2. Paper is well documented with extensive set of physics-related analyses and training details and experiments have been designed to be reproducible.
Critical weaknesses
  1. Many of the important results are hidden away in the appendices.
Questions for authors
  1. Baseline configurations: Several baselines have been downsized to satisfy GPU memory constraints. To what extent do the authors anticipate that the conclusions would change if larger baseline configurations were used?

  2. Physics-based evaluation: The paper primarily uses NMSE of the predicted flow fields to evaluate model performance. The appendix includes additional physics-related analyses, such as turbulence energy spectra. It would be useful to include quantitative metrics based on turbulence spectral statistics, divergence errors, and conservation-law residuals in the main paper. Could the authors expand the evaluation to include these metrics?

  3. Connection to classical turbulence modeling: The proposed approach appears closely related to classical large-eddy simulation (LES), subgrid-scale closures, and multiscale turbulence modeling. Could the authors include a brief discussion connecting these concepts to their approach, particularly highlighting the physical principles motivating the architecture and how it differs from conventional filtered-equation methods?

4 comments
AuthorsOpenReview ↗

Thank you for your thoughtful and constructive feedback

Thank you for your careful reading and the constructive review. We appreciate your recognition that the proposed approach is intuitive and stronger at predicting turbulent flow fields than a variety of baselines, and that the paper is well documented and designed to be reproducible. We take this opportunity to address each of your concerns below.

Q1. Larger baseline configurations

Response. Thank you for this important question. To test larger baseline configurations, we used GPUs with 96 GB of memory to retrain all six baselines with larger configurations on JHTDB256 using 1/8 of the training data. The detailed results are given in the common response C1 (Tables R1 and R2 in C1). As shown in Table R2, all enlarged baselines remain less accurate than ScaleSplit-NO. The strongest of them, P3D-L, reaches an NMSE of 2.27 × 10⁻³, compared with 1.25 × 10⁻³ for ScaleSplit-NO. With the larger configurations, U-Net, EddyFormer, and P3D improve, while FNO, MG-TFNO, and ReViT change little. The enlarged baselines use 52 to 88 GiB of training memory, 17 to 29 times the 3.0 GiB required by ScaleSplit-NO. The conclusions of the paper therefore hold with the larger baseline configurations. We have included these results in the appendix of the revised manuscript (Appendix B.2, pp. 16–17).

W1 & Q2. Physics-based evaluation in the main paper

Response. Thank you for your suggestion. We have added these metrics to the main paper (Table 2 in Section 4.1, p. 7, Table R8). For all methods trained with 1/8 of the data, Table 2 reports the spectral error in a low, a middle, and a high wavenumber band, the divergence error, and the residual error of the mass equation on MHD64 and of the momentum equation on JHTDB256. The divergence and residual errors compare the prediction with the ground truth, because the ground truth itself does not satisfy the discretized equations on the evaluation grid. ScaleSplit-NO has the lowest error in every metric on both datasets. At low wavenumbers, the most accurate methods are close to each other, and the differences arise at high wavenumbers. There, several baselines have errors close to or above 100%, so they predict the small scales hardly better than leaving them out, whereas ScaleSplit-NO keeps the error of these scales near 10%. The divergence and the residual involve spatial derivatives, which give greater weight to high-wavenumber components, and are therefore most sensitive to errors at small scales. ScaleSplit-NO also achieves the lowest divergence and residual errors on both datasets. The definitions of the metrics are given in Appendix B.6, pp. 27–28.

Table R8. Physics-based metrics of one-step predictions with 1/8 of the training data, averaged over all test pairs (Table 2 in the revised manuscript, p. 7). The spectral error is given in three wavenumber bands, k ≤ 8, 8 < k ≤ 16, and k > 16 on MHD64, and k ≤ 32, 32 < k ≤ 64, and k > 64 on JHTDB256. The divergence error is that of the magnetic field on MHD64 and of the velocity on JHTDB256, and the residual error is that of the mass equation on MHD64 and of the momentum equation on JHTDB256.

DatasetMethodSpectral, low (%)Spectral, mid (%)Spectral, high (%)Divergence (%)Residual (%)
MHD64U-Net1.624.6424.95.7924.6
FNO7.3845.392.413.868.8
MG-TFNO0.8152.5428.710.122.0
EddyFormer0.8323.1920.37.9421.2
P3D0.8875.4942.65.8628.8
ReViT5.4547.21084.9673.2
ScaleSplit-NO0.7281.9311.04.8415.7
JHTDB256U-Net0.3462.9435.422.435.9
FNO7.4441.510627.988.2
MG-TFNO0.71338.889.526.459.4
EddyFormer0.80777.010122.062.8
P3D0.004410.81017.518.323.5
ReViT0.34945.713129.967.1
ScaleSplit-NO0.004280.5148.1312.716.6

Q3. Connection to LES and classical turbulence modeling

Response. Thank you for your thoughtful question. We agree that the connection to large-eddy simulation (LES) and to multiscale turbulence modeling helps to explain the physical principles behind our design.

The design of ScaleSplit-NO rests on the scale structure of turbulence. Large structures are correlated across broad regions of the domain and can be represented on a coarse grid. In contrast, small structures are correlated only over short distances. LES also rests on this scale structure. LES resolves the large scales on a coarse grid and represents the effect of the unresolved small scales with a subgrid-scale model. In ScaleSplit-NO, the Parent predicts the large scales on a coarse grid. The coarse fields on which the Parent is trained are obtained by truncating the full-resolution fields to the low frequencies and sampling the result on the coarse grid. These coarse fields correspond to the filtered fields of an LES with a sharp spectral filter.

Beyond the large scales, ScaleSplit-NO also predicts the full-resolution field, with the Child predicting each local patch from the coarse prediction of the Parent. The Child receives this coarse prediction as a condition, interpolated to the fine grid and cropped at the patch location. The condition supplies the large-scale motion that the patch does not contain. The physical basis of this design is that the small scales at one location are shaped mainly by the large-scale motion around this location and only weakly by the small scales elsewhere in the domain [R3.3].

In numerical simulation, a method more closely related to this design is grid nesting, which rests on the same physical principle. In grid nesting, a coarse grid covers the whole domain and drives the computation on fine grids placed on local regions. Grid nesting has long been used in numerical weather prediction [R3.1] and is also applied to LES [R3.2]. In grid nesting, a fine-grid computation needs to cover only a local region, because the rest of the domain influences the local region mainly through the large-scale motion provided by the coarse grid. The cost of the fine-grid computation then depends on the size of the local region and not on the size of the whole domain. For ScaleSplit-NO, this principle allows the training memory to be set by the patch size and the coarse grid, so that the training memory does not grow with the resolution of the field.

ScaleSplit-NO differs from filtered-equation methods such as LES in two ways. First, LES advances the large scales by solving the filtered equations, in which a subgrid-scale model represents the effect of the removed small scales. In ScaleSplit-NO, the Parent instead learns the evolution of the large scales directly from the filtered fields of the data, whose evolution already includes this effect, so no subgrid-scale model is needed. Second, LES does not compute the small scales, whereas in ScaleSplit-NO the Child predicts them in each patch, which yields the full-resolution field.

Following your helpful suggestion, we have added the connection to LES and grid nesting to Section 3 (p. 3) of the revised manuscript.

Thank you again for your careful reading and helpful comments, which have helped us strengthen the physical evaluation and clarify the connections to classical turbulence modeling.

[R3.1] N. A. Phillips and J. Shukla, "On the strategy of combining coarse and fine grid meshes in numerical weather prediction," Journal of Applied Meteorology and Climatology, 12(5), 763–770, 1973.

[R3.2] C.-H. Moeng, J. Dudhia, J. Klemp, and P. Sullivan, "Examining two-way grid nesting for large eddy simulation of the PBL using the WRF model," Monthly Weather Review, 135(6), 2295–2311, 2007.

[R3.3] H. Tennekes, "Eulerian and Lagrangian time microscales in isotropic turbulence," Journal of Fluid Mechanics, 67(3), 561–567, 1975. (cited in the paper)

Thank you on extending the analysis and the discussion. I learnt a lot from this exchange. All concerns addressed. Raising score from 3->4. Good luck!

AuthorsOpenReview ↗

Thank you for your feedback and for raising your score!

We sincerely appreciate your thorough review and positive feedback. It is especially encouraging to hear “All concerns addressed.” Your thoughtful questions and constructive suggestions have helped us strengthen the manuscript. Thank you again for the time and effort you have devoted to this helpful exchange.

Reviewer_5U56OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents scalesplit-NO (Scale-Split Neural Operator) an approach for training on large scale turbulence data that exploits the spatial structure of the problem by integrating a Parent network to learn on coarse data which then can be used to condition a child network on full resolution data. One key part of the scheme is the patching scheme applied to the full resolution children network which allows training on a small spatial context and the mapping done with the parent network to associate it with the correct patch. The main contribution is the scheme to exploit the multiscale structure of the data. Experiments in the paper present accuracy comparison against baselines trained on full resolution data with limited or no patching and analysis of data training efficiency provided by the coarse resolution parent network.

Overall, the paper's strongest contribution is the scheme for cheaply being able to add coarse resolution information to improve performance. However, there are major issues with the baselines. The baselines use full resolution inputs with either no or very limited patching. The analysis of accuracy improvement is done such a way that the full model fits on 40GB of GPU memory, therefore the baseline network architectures needed to be shrunk in order to train with the full resolution. Due to this it's difficult to show how much of the increased accuracy is due to patched implementation or the coarse resolution conditioning. In order to get a proper comparison these baselines should be run in a patched mode with a deeper network to understand this dynamic. The baselines also do not use the 40GB of training memory available to them. If the 40GB constraint is the factor that needs to be worked around the baseline should utilize all of the available memory, unless it was shown that a larger network didn't provide anymore accuracy. Proper baselines would need to be performed to show a true comparison.

Minor Comments:

  1. On Figure 2, the bold labels under each box are difficult to read
Critical strengths

The method makes training on high-resolution 3D fields feasible under the constraint of modest GPU memory sizes and the coarse resolution correction provides a reasonable path for creating more signal to improve learning on the turbulence problem.

Critical weaknesses

The baselines do not provide a useful comparison to show how much value the coarse resolution correction provides.

Questions for authors

Is there a reason you didn't use a patched version of the baseline?

4 comments
AuthorsOpenReview ↗

Thanks for your review and insightful comments

Thank you for your careful review and for the close examination of the baselines. We appreciate your recognition that the scheme for cheaply adding coarse-resolution information is the strongest contribution of the paper, and that the method makes training on high-resolution 3D fields feasible with modest GPU memory. We take this opportunity to address each of your concerns below.

W1 & Q1. Patched baselines

Response. Thank you for this insightful question. A patched baseline brings one element of the design of ScaleSplit-NO to the original model. In ScaleSplit-NO, the Child network learns on small full-resolution patches, a choice that rests on the observation that the small scales at one location depend mainly on the nearby flow. In the submission, we trained each baseline with the scheme of its original work. U-Net, FNO, ReViT, and EddyFormer take the whole field as input. MG-TFNO splits the field into patches but computes one loss on the assembled field, so every update needs the whole field. P3D is trained on large crops of 128³, following the setting of its authors on the same JHTDB simulation. Small patches are a new setting for these models, and training a model in this setting involves choices that its original work leaves open. Following your constructive suggestion, we built patched versions of all six baselines on JHTDB256 with 1/8 of the training data, using the random patch sampling and overlapping, smoothly weighted assembly of ScaleSplit-NO. ScaleSplit-NO surpasses all six patched baselines in prediction accuracy, achieving 46% lower NMSE with 34% less training memory than the strongest patched baseline, P3D-L (Tables R5–R6). We have included these experiments and their analysis in Appendix B.3, pp. 16–18; Tables 9–11, p. 18 of the revised manuscript.

Experimental setup. The patched baselines are trained on random 32³ patches of the 256³ fields. We chose 32³ so that none of the baselines has to be reduced, while the patches stay close to the 16³ patches of ScaleSplit-NO (Table R5). On 16³ patches, U-Net and P3D would have to use fewer layers, since they downsample the input layer by layer. We also chose the learning rate, the batch size, the training schedule, and the boundary treatment of a non-periodic patch cut from a periodic field to suit training on patches. For U-Net and P3D, we also apply the trained network to the whole 256³ field at once (Table R6).

Table R5. Configurations and training memory (GiB) of the baselines trained on 32³ patches.

MethodConfigurationTraining memory
U-Netwidth 64, depth 42.7
FNOwidth 48, modes 12, 8 layers4.6
MG-TFNOwidth 64, modes 32, padding 88.6
EddyFormerwidth 32, 4 layers, 4 subdomains per axis, 13 modes8.3
P3DP3D-L4.6
ReViTtoken patch size 13.0
ScaleSplit-NO3.0

Table R6. One-step test NMSE (×10⁻³) on JHTDB256 with 1/8 of the training data. The first column gives the baselines in their original settings, and the other two columns the baselines trained on 32³ patches with two ways of inference. Whole-field inference applies the trained network to the entire 256³ field at once. Only U-Net and P3D support this inference. Patch inference assembles the patches with an overlap of 27% and blends them with the same weights as in ScaleSplit-NO. MG-TFNO keeps its own input context when stitched. For ScaleSplit-NO, patch inference refers to its Child patches.

MethodOriginal settingPatch training, whole-field inferencePatch training, patch inference
U-Net10.16.7617.5
FNO93.3N/A3.43
MG-TFNO28.3N/A3.38
EddyFormer40.3N/A2.77
P3D2.653.102.31
ReViT32.1N/A7.09
ScaleSplit-NO1.25N/A1.25

As shown in Table R6, every baseline reaches a lower error after training on patches than with its original training, with whole-field inference for U-Net and with stitching for the other baselines. These results are consistent with the physical motivation behind ScaleSplit-NO: local small-scale dynamics are expected to depend more strongly on nearby flow than on distant small-scale fluctuations. All patched baselines nevertheless remain less accurate than ScaleSplit-NO. The most accurate patched baseline, P3D, reaches an NMSE of 2.31 × 10⁻³, compared with 1.25 × 10⁻³ for ScaleSplit-NO. Thanks for your constructive comments. We have added the above results to the appendix of the revised manuscript (Appendix B.3, pp. 16–18; Tables 9 and 10, p. 18).

To show how much of the accuracy comes from the coarse-resolution conditioning, we also compare ScaleSplit-NO with the Child without Parent, which is pretrained on the same 16³ patches without the Parent's coarse prediction (Table R7). With 1/8 of the training data on JHTDB256, the Child alone reaches an NMSE of 2.12 × 10⁻³. Fine-tuning with the Parent's coarse prediction lowers the NMSE to 1.25 × 10⁻³. For every amount of training data, this fine-tuning lowers the NMSE by 27% to 41%. The relative reduction is largest when the training data are most limited.

Table R7. One-step test NMSE (×10⁻³) on JHTDB256 of the Child without Parent and of ScaleSplit-NO for different amounts of training data. Parentheses give the reduction of the NMSE by ScaleSplit-NO relative to the Child alone.

Training dataChild without ParentScaleSplit-NO
1/82.121.25 (↓40.8%)
1/41.811.17 (↓35.1%)
1/21.591.08 (↓32.3%)
All1.411.03 (↓27.1%)

Thank you again for your helpful suggestion. We have added Tables R5, R6, and R7 and this analysis to the appendix of the revised manuscript (Appendix B.3, pp. 16–18; Tables 9–11, p. 18).

Concern Raised in the Paper Summary: Baseline Memory Budget

Response. Thank you for this suggestion. The patched-baseline experiments above directly address your request to evaluate all six baselines with patch training (Tables R5–R6). We also evaluated larger baseline configurations under the training schemes used in their original works, to examine whether increasing model capacity and the memory budget closes the accuracy gap. Please see our Common response C1 for the configurations and results of all six enlarged baselines (Tables R1–R2 in C1). In short, larger baselines do not close the gap to ScaleSplit-NO, even with far more memory than the 40 GB budget. Specifically, we used GPUs with 96 GB of memory to retrain all six baselines with larger configurations on JHTDB256 using 1/8 of the training data, with 52 to 88 GiB of training memory, 17 to 29 times the 3.0 GiB required by ScaleSplit-NO. For FNO, MG-TFNO, and ReViT, the larger configurations produce little change in accuracy. U-Net, EddyFormer, and P3D improve, but all of them remain less accurate than ScaleSplit-NO. The strongest enlarged baseline, P3D-L, reaches an NMSE of 2.27 × 10⁻³, compared with 1.25 × 10⁻³ for ScaleSplit-NO. We have included these results in the appendix of the revised manuscript (Appendix B.2, pp. 16–17).

Minor 1. Readability of Figure 2

Response. Thank you for this helpful suggestion. We have removed the bold formatting from the formulas in Figure 2 (p. 4) to improve readability.

We thank you again for your careful reading and constructive comments, which have helped us strengthen the experimental evaluation and improve the clarity of the manuscript.

I'm satisfied with how the authors have addressed my concerns. Since these concerns have been addressed I am raising my score to 3.

AuthorsOpenReview ↗

Thank you for your feedback and for raising your score!

We sincerely appreciate your positive feedback. We are pleased that our responses have addressed your concerns, and we thank you for raising your score. Your suggestions have helped us strengthen the baseline comparisons and experimental evaluation. Thank you again for your time and thoughtful input.

Reviewer_rP4wOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Scale-Split Neural Operator, a multi scale neural operator that can be trained without processing the full-resolution grid. Two neural operators (Parent and Child) are used for predicting global and local fields, respectively, with the Child conditioned on Parent. The authors demonstrated accuracy gains while drastically reducing the memory requirement on large-scale 3D turbulence prediction tasks.

Critical strengths
  1. The method is well motivated. The authors identified the memory scaling issue as one of the main bottlenecks to scaling ML models to large-scale turbulence. The memory footprint is an important factor that limits real-world ML for CFD applications.
  2. The proposed method is principled. The scale separation in turbulence dynamics is used to construct the Parent-Child architecture. It achieves promising performance despite its simplicity.
Critical weaknesses
  1. The authors are missing an important baseline, Pixel-Resolved Long-Context Learning for Turbulence at Exascale. This prior work also focused on JHTDB and proposed a three-way scale separated ViT that also avoids processing the full-resolution field.

It would be unrealistic to perform baseline comparison as this related work is based on cluster-scale training, and didn't open source. However, the authors should discuss the relevance of this work, and ablate the key architectural difference. Otherwise, the paper's contribution remains unclear.

  1. The inference speed of the proposed model is on average 2-5x slower than most baselines, because the model needs to at least perform two forward loops. In 3D turbulence prediction tasks, numerical solvers are not significantly slower than a ML model. Therefore, the speed disadvantage remains something to be addressed.
Questions for authors
  1. Can you clarify the difference between the baseline mentioned above?
  2. Did you compare efficiency with solvers that achieve similar accuracy?
  3. I'm a bit confused by the inconsistent baseline performance. The authors shrunk baseline size on JHTDB256 in order to fit in memory, but this leads to imbalanced baseline capacity as reported in their papers. For example, FNO as baseline has hidden dimension = 12, while as backbone has 32; EddyFormer is the best baseline on MHD64 but the worst baseline on JHTDB256. Could you justify this?
4 comments
AuthorsOpenReview ↗

Thank you for your thorough review and encouraging feedback

Thank you for your detailed and constructive review. We appreciate your recognition that the method is well motivated and principled, and that the memory footprint is an important factor that limits real-world ML for CFD applications. We take this opportunity to address each of your concerns below.

W1 & Q1. Relation to the exascale scale-separated ViT

Response. Thank you for pointing us to this relevant work. Yin et al. [R1.1] train a ViT, the Turbulence Transformer, on the 1024³ JHTDB data. To shorten the input from billions to a few million tokens, the ViT receives each state at three scales, which are a coarse version of the whole field for the large scales, a coarse version of a subdomain for the intermediate scales, and a full-resolution local window for the small scales.

The two works differ mainly in how the models are trained, in the scale of the computation, and in the amount of training data. (i) The Turbulence Transformer is trained end to end as one model, which takes all three scales of the current state as input. In contrast, ScaleSplit-NO is trained in separate parts. The Parent neural operator first predicts the coarse field of the next time step, and the Child neural operator then predicts each full-resolution patch conditioned on this predicted coarse field. (ii) Yin et al. scale the computation up, and their main focus is the efficient parallelization of attention over millions of tokens on up to 32,768 GPUs to train a ViT at full resolution. Our work instead keeps the computation as small as possible. The training memory of ScaleSplit-NO does not grow with the resolution and is at most 3.0 GiB on JHTDB256, which fits on a single GPU. (iii) Our work also studies the setting where training data are scarce. The paper evaluates ScaleSplit-NO with as little as 1/8 of the training data. The small patches of ScaleSplit-NO yield many training samples from the same data, so the model can be trained well in this setting.

These differences show that ScaleSplit-NO addresses a different goal, which is accurate prediction of full-resolution turbulence with the memory of a single GPU and with limited training data. Following your helpful suggestion, we have added this discussion to the Related Work section of the revised manuscript (p. 3).

[R1.1] J. Yin et al., "Pixel-Resolved Long-Context Learning for Turbulence at Exascale: Resolving Small-scale Eddies Toward the Viscous Limit," 2026 IEEE International Parallel and Distributed Processing Symposium (IPDPS), 347–362, 2026.

W2 & Q2. Inference speed and comparison with numerical solvers

Response. Thank you for raising these points on (i) inference speed and (ii) numerical solvers, which we answer in turn.

(i) We first address inference speed. We agree that inference speed is important for a surrogate model. In the submission, the whole inference of ScaleSplit-NO used a relatively naive implementation, which motivated us to optimize it. We increase the parallelism by predicting many Child patches in one batch, reduce the arithmetic cost with FP16 matrix multiplications and a tanh approximation of GELU, reduce the memory traffic by fusing consecutive operations, and remove the overhead of launching many small operations by recording the whole inference as one CUDA Graph. Without retraining the model, the inference becomes about 6 times faster than in Table 3 of the submission, while the relative increase of the NMSE is negligible (a few parts in ten thousand).

Table R3. One-step test NMSE (×10⁻³) with 1/8 of the training data and inference time for one complete field. All times are measured under the same conditions on A100 GPUs. ScaleSplit-NO is given for several overlaps of the Child patches. The default overlap is in bold.

MethodMHD64 OverlapMHD64 NMSEMHD64 Time (ms)JHTDB256 OverlapJHTDB256 NMSEJHTDB256 Time (s)
U-Net15.52810.10.31
FNO97.61593.30.41
MG-TFNO10.35528.30.71
EddyFormer8.8334140.30.38
P3D14.5352.650.41
ReViT93.82332.10.44
ScaleSplit-NO0%8.875.40%2.170.22
20%6.447.716%1.330.36
33%6.2012.327%1.250.54
43%6.1019.338%1.230.88
50%6.0528.650%1.211.61

As shown in Table R3, with the optimized inference, ScaleSplit-NO with the default overlap is faster than all baselines on MHD64 and comparable to them on JHTDB256. ScaleSplit-NO therefore no longer requires significantly more inference time than the baselines.

(ii) We next compare the speed of ScaleSplit-NO with that of numerical solvers. Based on the available records and our estimates, compared with the solvers that generated the data, ScaleSplit-NO is about three orders of magnitude faster on JHTDB256 and about five orders of magnitude faster on MHD64. On a single A100 GPU, ScaleSplit-NO with the default overlap predicts one step in 0.54 s on JHTDB256 and in 12 ms on MHD64 (Table R3). For both datasets, the solver codes are not publicly available, and the records of the data generation do not report enough detail to measure or reproduce the cost of the solvers.

JHTDB256 is downsampled from a pseudo-spectral DNS of forced isotropic turbulence on a 1024³ grid, with a time step of 2 × 10⁻⁴ and a snapshot stored every 10 time steps [R1.2, R1.3]. We use every tenth stored snapshot, so one step of our model spans 100 time steps of the DNS. The DNS was run on a Linux cluster with over 40 nodes [R1.2], but its run time and core hours are not reported. The database paper gives a rough estimate. Tracking particles over 50 stored snapshots takes less than 20 minutes, and generating the underlying data takes an order of magnitude more time [R1.2], about 40 minutes per 100 time steps of the DNS. As a rough cross-check, NVIDIA reports 23.9 s per time step for another pseudo-spectral DNS code at 1024³ on a single 64-core CPU node [R1.4], which corresponds to about 40 minutes for 100 time steps, the same order of magnitude.

MHD64 is downsampled from simulations on a 256³ grid, computed with a Fortran and MPI code [R1.5, R1.6]. The Well reports 48 hours on 64 CPU cores for one simulation of 100 frames [R1.6], which gives about 29 minutes per frame on average. The number of solver steps between frames, the time step, and whether the 48 hours include a spin-up phase are not reported.

A solver with an accuracy similar to ScaleSplit-NO on the 256³ grid of our model would need a subgrid-scale model, because the DNS of this flow resolves its smallest eddies only on the 1024³ grid. We have not run such a solver.

Numerical solvers and learned models are also used in different ways. In a numerical solver, and especially in a DNS, the large-scale structures emerge from the computed interactions down to the smallest scales. Without resolving these small scales or modeling their effect with a subgrid-scale model, the solver cannot produce reliable large scales. The time step of the DNS is limited by the CFL condition on its fine grid. A learned model instead learns the evolution of the large scales directly from data on a coarse grid. Its time step is not limited by a CFL condition. A learned model can therefore run at a much lower spatial and temporal resolution than the DNS.

Following your constructive comments, we have added Table R3 and this comparison to the revised manuscript (Table 4 in Section 4.2, p. 10; Table 21 in Appendix D.4, p. 36; and Appendix D.5, pp. 36–37).

[R1.2] E. Perlman, R. Burns, Y. Li, and C. Meneveau, "Data exploration of turbulence simulations using a database cluster," Proceedings of the 2007 ACM/IEEE Conference on Supercomputing, 2007.

[R1.3] Y. Li et al., "A public turbulence database cluster and applications to study Lagrangian evolution of velocity increments in turbulence," Journal of Turbulence, 9, N31, 2008.

[R1.4] NVIDIA, "Multinode Multi-GPU: Using NVIDIA cuFFTMp FFTs at Scale," NVIDIA Technical Blog, 2022 (https://developer.nvidia.com/blog/multinode-multi-gpu-using-nvidia-cufftmp-ffts-at-scale/).

[R1.5] B. Burkhart et al., "The Catalogue for Astrophysical Turbulence Simulations (CATS)," The Astrophysical Journal, 905(1), 14, 2020.

[R1.6] R. Ohana et al., "The Well: a large-scale collection of diverse physics simulations for machine learning," Advances in Neural Information Processing Systems, 37, 2024.

Q3. Baseline capacity on JHTDB256

Response. Thank you for raising this point on the baseline capacity. We agree that it is important to verify that the conclusions do not depend on the reduced baseline sizes. We give the detailed answer in the common response (C1) and summarize the main point here. Most baselines on JHTDB256 were reduced to fit into 40 GB, because the same model needs more training memory as the resolution of the data increases. During training, the model stores the activations of every layer at every grid point. On GPUs with 96 GB of memory, we retrained all six baselines with larger configurations. Some of them improve, but all of them remain less accurate than ScaleSplit-NO (Table R2).

Regarding the FNO width, with width 30 and 20 modes on the 96 GB GPU, the NMSE of FNO is 92.0 × 10⁻³, close to 93.3 × 10⁻³ with width 12 (Table R2). The reduced width is therefore not the cause of the high error of FNO on JHTDB256. The FNO baseline processes the full 256³ field at once and had to be reduced to width 12 to fit into 40 GB. In contrast, by splitting the prediction between the Parent and the Child, ScaleSplit-NO lets each network process far fewer grid points. The same GPU memory can then hold FNOs with more parameters and greater expressive power. Table R4 shows the training memory of an FNO with widths of 12 and 32 for different grid sizes. The grid-independent part stays the same, while the grid-dependent part grows by about 8 times each time the grid size doubles along each axis. An FNO of width 32 needs more than 80 GiB on the 256³ grid but less than 1 GiB on a 32³ grid. We have added Table R4 to the appendix of the revised manuscript (Table 6 in Appendix B.2, p. 17).

Table R4. Training memory (GiB) of an FNO with four Fourier layers, 12 modes per axis, and batch size 1 for different grid sizes. Each entry is the grid-independent part (parameters and optimizer states) plus the grid-dependent part (inputs, targets, stored activations, and temporary FFT buffers).

GridWidth 12Width 32
32³0.09 + 0.090.63 + 0.20
64³0.09 + 0.570.63 + 1.43
128³0.09 + 4.380.63 + 11.26
256³0.09 + 34.890.63 + 81.65

Regarding the ranking of EddyFormer on the two datasets, with more subdomains and modes on the 96 GB GPU, the NMSE of EddyFormer drops from 40.3 × 10⁻³ to 15.4 × 10⁻³ (Table R2). The different ranking on the two datasets comes from how EddyFormer represents the field. EddyFormer is a Transformer-type architecture. It divides the domain into a small number of subdomains, represents the field in each subdomain by a fixed number of spectral modes, and treats each subdomain as a token for attention. We use the network configuration of the original work on both datasets. This configuration is sufficient for the 64³ fields of MHD64, but on JHTDB256 each subdomain becomes much larger and the modes of each subdomain are too few to represent the field accurately.

Thank you again for your helpful suggestions, which have strengthened our evaluation and improved the manuscript’s clarity.

AuthorsOpenReview ↗

Kind Follow-up on Our Rebuttal

Thank you again for your time and thoughtful comments. As the discussion period draws to a close, we would be grateful for any feedback you have on our rebuttal and additional results. We would be happy to answer any remaining questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Common response to all reviewers

We sincerely thank all the reviewers for their time, careful reading, thoughtful comments, and constructive suggestions. We are encouraged by their recognition of our approach as well motivated, principled, and intuitive, and by their appreciation of its ability to enable high-resolution turbulence prediction with modest training memory. In response, we have made the following updates to the manuscript:

  1. Additional experiments: We have added comparisons with larger and patch-trained baselines, coarse-conditioning analysis, and physics-based metrics.
  2. Inference efficiency optimization: We have optimized the inference speed of ScaleSplit-NO and added a discussion of estimated numerical solver runtimes.
  3. Additional references: We have added references and discussion connecting our approach to prior work, LES, and grid nesting.
  4. Clarity and presentation: Following the reviewers’ suggestions, we have included key results in the main paper and improved the labels in Figure 2.

All changes in the revised manuscript are marked in blue. Here, we summarize the shared comments regarding baseline capacity. For other comments, we have provided point-by-point responses to each reviewer.

C1. Larger baseline configurations (Reviewers rP4w, 5U56, TN6P)

Response. We thank the reviewers for raising this question about baseline capacity. The experiments in our original submission were conducted on A100 GPUs with 40 GB of memory. On JHTDB256, we therefore used smaller baseline configurations to fit within this budget. For full-field models, the memory needed to store intermediate activations typically grows with grid resolution, even when the number of parameters remains unchanged. Even with these smaller configurations, the baselines required 14.3–35.0 GiB of training memory, approximately 5–11 times the 3.0 GiB required by ScaleSplit-NO. As shown in the additional experiments below, ScaleSplit-NO remains more accurate than all tested enlarged baselines while requiring substantially less training memory.

To further address this concern, we retrained all six baselines with larger configurations on GPUs with 96 GB of memory, using 1/8 of the JHTDB256 training data. Table R1 lists the configuration changes, and Table R2 reports the test NMSE and training memory.

Table R1. Changes from the baseline configurations for the 40 GB GPU budget to those for the 96 GB GPU budget. Less or no activation checkpointing makes MG-TFNO and ReViT train faster. We used the saved time for more epochs, from 200 to 240 for MG-TFNO and from 300 to 450 for ReViT.

Method40 GB → 96 GB GPU
U-Netwidth 16 → 32
FNOwidth 12 → 30, modes 12 → 20
MG-TFNOwidth 40 → 64, padding 8 → 12, less activation checkpointing
EddyFormersubdomains per axis 8 → 12, modes per subdomain 13 → 15
P3DP3D-B → P3D-L (convolution channels ×2, Transformer width 128 → 256, attention heads 4 → 8)
ReViTwidth 48 → 96, no activation checkpointing

Table R2. One-step test NMSE (×10⁻³) and training memory (GiB) on JHTDB256 with 1/8 of the training data. The 40 GB GPU budget gives the baseline configurations of the submission, and the 96 GB GPU budget the enlarged configurations. Numbers in parentheses are ratios to ScaleSplit-NO, rounded to two significant figures. Training memory is measured as in the paper.

MethodNMSE (40 GB GPU)Memory (40 GB GPU)NMSE (96 GB GPU)Memory (96 GB GPU)
U-Net10.1 (8.0×)31.7 (10×)5.17 (4.1×)54.4 (18×)
FNO93.3 (74×)35.0 (11×)92.0 (73×)87.0 (29×)
MG-TFNO28.3 (23×)19.8 (6.5×)29.3 (23×)80.7 (27×)
EddyFormer40.3 (32×)26.4 (8.7×)15.4 (12×)87.8 (29×)
P3D2.65 (2.1×)25.6 (8.4×)2.27 (1.8×)52.3 (17×)
ReViT32.1 (26×)14.3 (4.7×)32.4 (26×)72.9 (24×)
ScaleSplit-NO1.253.0

As shown in Table R2, the enlarged baselines require 52 to 88 GiB of training memory. With this additional memory, U-Net, EddyFormer, and P3D improve. FNO, MG-TFNO, and ReViT show little or no improvement with the tested configurations. Nevertheless, ScaleSplit-NO remains the most accurate method. The strongest enlarged baseline, P3D-L, reaches an NMSE of 2.27 × 10⁻³ with 52.3 GiB, compared with 1.25 × 10⁻³ for ScaleSplit-NO with 3.0 GiB. The accuracy advantage of ScaleSplit-NO thus persists against these larger baseline configurations. These additional results further support the paper’s claim of higher accuracy with substantially less training memory in this setting. We have added Tables R1 and R2 to the appendix of the revised manuscript (Appendix B.2, Tables 8 and 7, p. 17).

We thank the reviewers again for their constructive comments and suggestions, which have helped us further strengthen the manuscript. We would appreciate any additional feedback and welcome continued discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.