acceptodds
Under review as a conference paper at ICLR 2027

HybridOSS: Orthogonal Scale Splitting for Hybrid Differentiable PDE Solvers

Abstract

Fine-resolution simulation of partial differential equations (PDEs) is important for resolving small-scale structures and cross-scale interactions, but becomes computationally expensive over long time horizons. Neural surrogates reduce this cost but can suffer from error accumulation, while hybrid models typically require the numerical solver and corrected state to operate at the same spatial resolution. We propose HybridOSS, an orthogonal scale splitting framework that decouples numerical-solver resolution from state resolution. HybridOSS decomposes the fine-resolution state into complementary resolved and unresolved components. A differentiable numerical solver advances the resolved state on a coarse grid, while a closure module corrects the resolved dynamics and a refinement module explicitly evolves the unresolved state. Complementary spectral projections constrain the learned updates to orthogonal subspaces while preserving cross-scale coupling. Our instantiated HybridOSS model reduces global prediction error by 31% and 69% over the strongest hybrid baseline on KS and Allen–Cahn, respectively, while improving both resolved- and unresolved-scale accuracy. On differentiable OpenFOAM for 3D incompressible Navier–Stokes turbulence, HybridOSS maintains lower error over long autoregressive rollouts and nearby unseen viscosity values while achieving approximately a inference speedup over the fine-grid solver.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
hybrid PDE solvers, differentiable physics, machine learning for science

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_o75GOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2(+1)
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents HybridOSS, a hybrid framework for fine-resolution PDE prediction using a coarse-grid numerical solver. The method decomposes the state into orthogonal resolved and unresolved spectral components. A differentiable numerical solver advances the resolved state on a coarse grid, while a learned closure module corrects the resolved dynamics and a refinement module updates the unresolved state. Spectral projections constrain the two learned corrections to complementary subspaces while allowing cross-scale interactions. The method is evaluated on Kuramoto–Sivashinsky, Allen–Cahn, and 3D Navier–Stokes equations. The reported results show improved prediction accuracy over the evaluated baselines, along with computational benefits in larger-scale simulations. Ablation studies further examine the contributions of the closure and refinement modules, spectral projection, and different coarsening factors.

Critical strengths

The motivation for decoupling the numerical solver resolution from the predicted state resolution is reasonable and well justified, providing a practical way to balance computational efficiency and fine-scale prediction accuracy. The paper also presents a series of experiments and ablation studies that support the effectiveness and feasibility of the proposed design. I consider these to be the main strengths of this work.

Critical weaknesses
  1. My main concern is the limited novelty of the proposed method. The first two contributions, including orthogonal scale decomposition, separate modeling of resolved and unresolved dynamics, and combining coarse-grid solvers with neural corrections, have been explored in prior work. The proposed framework appears to be more of a combination of existing techniques. I believe the authors should clarify what new technical contributions this combination offers compared with existing methods. What I mean is that the authors should more clearly explain why they chose this particular combination of components. I believe this is where the paper's novelty could be most clearly demonstrated.

Relevant references (including some already cited in the manuscript):

Marion & Temam (1989), Nonlinear Galerkin Methods. https://doi.org/10.1137/0726063

Devulder & Marion (1992), A Class of Numerical Algorithms for Large Time Integration: The Nonlinear Galerkin Methods. https://doi.org/10.1137/0729028

Hughes et al. (1998), The Variational Multiscale Method—A Paradigm for Computational Mechanics. https://doi.org/10.1016/S0045-7825(98)00079-6

Margenberg et al. (2022), A Neural Network Multigrid Solver for the Navier–Stokes Equations. https://arxiv.org/abs/2008.11520

Lai et al. (2024), Neural Downscaling for Complex Systems: From Large-scale to Small-scale by Neural Operator. https://arxiv.org/abs/2403.13016

  1. Following up on the previous point, the authors should use ablation studies to demonstrate why these components need to be combined in this particular way. The authors provide ablation studies for the Closure module, Refinement module, and orthogonal projection, but the role of cross-scale coupling is not directly examined. For example, removing the high-frequency input to the Closure module or the low-frequency input to the Refinement module would help determine whether these interactions are necessary. In addition, the current component ablations mainly focus on KS and AC. Extending these experiments to the 3D Navier–Stokes task would provide stronger evidence for why the proposed combination is effective.
Questions for authors

Please refer to the Critical Weaknesses section.

7 comments

We thank the reviewer for highlighting these connections.

W1: Clarifying the Novelty and Rationale of the Proposed Component Integration

1. Connection to prior work and our focus. We acknowledge that nonlinear Galerkin methods, their subsequent developments, and the variational multiscale method establish foundations for scale decomposition and cross-scale modeling. Our focus is maintaining the long-term evolution of a complete fine-grid state using a coarse-grid differentiable solver.

2. Distinction from related neural methods. DNN-MG feeds corrected fine-grid velocity information back into coarse-grid advancement through a restricted right-hand side, while Neural Downscaling maps large-scale temporal histories to small-scale states. HybridOSS explicitly propagates the unresolved state, learns its feedback through a separate Closure branch, and jointly trains complementary projected updates with a differentiable solver in the loop.

3. Why these components are combined. The design aims to reduce numerical cost while retaining coupled evolution across scales. Coarsening creates two tasks: correcting an existing numerical prediction on resolved scales and learning the evolution of unresolved scales without direct numerical advancement. Although one network can learn both tasks, two branches provide explicit responsibilities. Complementary projections restrict the frequency bands each branch can modify, while cross-scale inputs preserve information exchange between them. The Full-state control, projection ablation, and newly added cross-scale-input ablation provide complementary experimental support for this mechanism.

To the best of our knowledge, our work is the first to combine persistent unresolved-state evolution, cross-scale closure feedback, and complementary orthogonal projection constraints in a two-branch mechanism within a hybrid differentiable solver-in-the-loop framework, enabling coupled fine-grid evolution under coarse-grid numerical stepping.

W2: Validating Cross-Scale Coupling and Component Synergy through Extended Ablations

1. Cross-scale input ablations.
We added three ablations: masking the high-frequency input to Closure, masking the low-frequency input to Refinement, and masking both directions. Masking is applied before feature construction, so derived features are computed without direct access to the removed scale component. The cross-scale-input variants retain the same network architecture and training protocol. All results are reported as mean ± standard deviation across three independent training runs.

2. Results on KS and AC.
All three ablations increase the mean global error on both KS and AC. On AC, they also increase the mean resolved- and unresolved-scale errors. On KS, some individual scale-specific errors improve, but the global error still increases. These results support the benefit of cross-scale conditioning for full-state prediction.

ConfigurationKS GlobalKS ResolvedKS UnresolvedAC GlobalAC ResolvedAC Unresolved
w/o HF input to Closure0.1028±0.00300.0084±0.00060.9414±0.04750.0223±0.00650.0008±0.00060.2120±0.0118
w/o LF input to Refinement0.1047±0.00240.0147±0.00070.3509±0.04620.0202±0.00190.0006±0.00021.0095±0.0033
w/o cross-scale inputs0.1205±0.00290.0158±0.00070.7173±0.02140.0223±0.00330.0007±0.00041.0002±0.0002
HybridOSS(full)0.0858±0.00220.0097±0.00030.4756±0.11510.0173±0.00100.0004±2.95e-50.1709±0.0510

3. Extension to three-dimensional NS.
We further extended the component and cross-scale-input ablations to 3D NS. All variants were trained only at ν=10−5\nu=10^{-5} and evaluated on the same test trajectories using the mean relative L2L_2 error over 125-step autoregressive rollouts.

The complete model achieves the lowest mean error at all four test viscosities. At ν=10−5\nu=10^{-5}, its error is 0.29900.2990, compared with 0.48840.4884, 0.81150.8115, and 0.80240.8024 when masking the Closure high-frequency input, the Refinement low-frequency input, and both directions, respectively. Full-state and the projection ablation yield errors of 0.47270.4727 and 0.34650.3465. Removing either Closure or Refinement also increases the error. At the unseen viscosities ν=5×10−6, 5×10−5, 10−4\nu=5\times 10^{-6},\,5\times 10^{-5},\,10^{-4}, the complete model reduces the mean error relative to Full-state by approximately 43.4%43.4\%, 63.6%63.6\%, and 63.2%63.2\%, respectively, and by 12.6%12.6\% – 14.4%14.4\% relative to the projection ablation. Masking either direction of cross-scale input also increases the mean error at each unseen viscosity, indicating that the benefits extend beyond the training viscosity.

Configurationν=5×10⁻⁶ν=1×10⁻⁵ν=5×10⁻⁵ν=1×10⁻⁴
w/o Closure0.8457 ± 0.02710.8450 ± 0.02760.8637 ± 0.02760.9168 ± 0.0261
w/o Refinement0.4308 ± 0.09230.4482 ± 0.09440.5583 ± 0.10650.6919 ± 0.1183
w/o projection0.3636 ± 0.01600.3465 ± 0.01600.4856 ± 0.01000.6566 ± 0.0127
w/o HF input to Closure0.5050 ± 0.06480.4884 ± 0.07160.5396 ± 0.05680.6494 ± 0.0386
w/o LF input to Refinement0.8290 ± 0.04480.8115 ± 0.04720.8334 ± 0.03470.8876 ± 0.0280
w/o cross-scale inputs0.8176 ± 0.04830.8024 ± 0.05260.8164 ± 0.04260.8651 ± 0.0307
Full state0.5571 ± 0.01770.4727 ± 0.01241.1414 ± 0.01701.5580 ± 0.0247
HybridOSS(full)0.3151 ± 0.03370.2990 ± 0.03500.4157 ± 0.01690.5739 ± 0.0129

These results extend the evidence to three dimensions and support the complementary roles of branch separation, orthogonal projection, and cross-scale conditioning in the tested configurations. We will include these ablations and their variability across runs in the revised manuscript.

Thank you for your detailed responses and additional experiments. Most of my previous questions have been answered, and I now have a clearer and more positive view of the paper. However, I still have one important concern regarding the novelty of the method. I understand the differences between HybridOSS and previous methods, but I am still not sure why combining explicit unresolved-state evolution, a separate Closure module, and complementary orthogonal projections represents a distinct technical contribution. To me, this seems more like an effective combination of existing techniques rather than a new scientific insight. For example, I understand that the main motivation is to reduce computational cost while preserving cross-scale coupling. The first point is clear, and the paper provides convincing experimental evidence. However, for the second point, although the new ablations show that cross-scale inputs improve overall prediction accuracy, I am still unsure whether this means the model captures the underlying physical coupling between scales more accurately. The paper does not seem to provide a direct metric or analysis of these physical interactions. This is my main remaining concern, and I would be very willing to raise my score to 3 if the authors could clarify this point. I also noticed an interesting result in the KS ablations. Removing the high-frequency input to Closure improves the resolved error, while removing the low-frequency input to Refinement improves the unresolved error. Could this be related to the chaotic nature of KS, the coarsening factor, or error propagation between scales? I think this could be a useful case for further analysis, as it may help us better understand when cross-scale coupling is beneficial and when it becomes more difficult to learn.

Reviewer_2sdcOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes HybridOSS, a hybrid differentiable PDE solver that decouples the spatial resolution of numerical simulation from that of the predicted state. The method decomposes a fine-resolution state into orthogonal resolved and unresolved components using spectral projections. A coarse-grid numerical solver advances the resolved component, while a learned closure module corrects its dynamics and a separate refinement module evolves the unresolved component. The learned updates are constrained to complementary spectral subspaces while allowing cross-scale interactions, and the updated components are recombined for autoregressive prediction. The authors evaluate the framework on the Kuramoto–Sivashinsky, Allen–Cahn, and 3D incompressible Navier–Stokes equations, examining standard metrics including prediction accuracy, computational efficiency, component ablations, long-horizon behavior, and generalization across viscosity values.

Critical strengths

Practically motivated framework. PDE simulations by decoupling numerical-solver resolution from the resolution of the predicted state.

(Seemingly) Physically motivated architectural design. The framework combines coarse-grid numerical evolution with dedicated closure and refinement modules by using orthogonal spectral projections and complementary frequency subspaces.

Competitive predictive accuracy. HybridOSS achieves the lowest reported global prediction errors among the evaluated methods on KS and AC.

Informative ablation studies. Component-removal experiments and comparisons with unrestricted full-state correction provide empirical support for the proposed architectural design (although they do not fully isolate the advantages of orthogonal scale splitting.)

Promising long-horizon performance and computational efficiency. On the evaluated 3D NS benchmark, HybridOSS achieves approximately 19× inference acceleration relative to the fine-grid solver and maintains comparatively low prediction errors over 125 autoregressive steps despite training on three-step windows.

Critical weaknesses

Theoretical presentation

The presentation of the theoretical material needs to be tightened. Overall, the paper relies heavily on verbal explanations and lacks precise mathematical definitions of key concepts.

Insufficiently defined mathematical notions. The operators EfE_f and TfT_f are introduced without precise mathematical definitions. This ambiguity hinders understanding of the theoretical formulation and may also affect the reproducibility of the proposed method. Furthermore, the proof of Theorem 1 relies on the identities EfTf=MfE_f T_f=M_f and Tf(I−Mf)=0T_f(I-M_f)=0 without formally establishing them.

Limited theoretical contribution. Theorem 1 establishes a standard orthogonal spectral decomposition. Its theoretical contribution appears limited without additional results, such as unique advantages that the proposed method or modular components offer.

Unclear methodological motivation. Eqs. (10)–(12) introduce learned update rules motivated by the projected PDE dynamics in Eqs. (8)–(9), but the relationship between them is not precisely characterized or adequately explained.

Experimental evaluation

Unclear benefits of orthogonal scale splitting. Table 3 shows exceptionally strong performance at f=1f=1, where the unresolved subspace disappears. This raises questions about which components of the proposed architecture are primarily responsible for the performance improvements.

Limited experimental scope and scalability evaluation. The NS experiments compare the proposed method mainly against solver-only and full-state correction. Furthermore, the reference-grid resolutions are relatively modest, leaving the scalability of the method to finer spatial resolutions unclear.

Insufficient statistical reliability and validation details. Tables 1–3 omit standard deviations or confidence intervals, making it difficult to assess the reliability of the reported improvements. Moreover, the NS training data consist of a single trajectory divided into overlapping windows, and the validation procedure is not clearly described.

Questions for authors

Why does the refinement module in Eq. (11) receive the uncorrected resolved prediction Pf(v~n+1)P_f(\widetilde v^{n+1}) rather than the corrected prediction Pf(vn+1)P_f(v^{n+1})?

How precisely are Eqs. (10)–(12) related to the projected PDE dynamics in Eqs. (8)–(9)? Are they intended solely as physically motivated learned updates?

Why was f=16f=16 selected for the main KS experiments when smaller coarsening factors achieve substantially lower errors with nearly identical training times?

What explains the exceptionally strong performance at f=1f=1, where the unresolved subspace disappears?

13 comments

We thank the reviewer for the insightful comments, which have helped us improve the quality of our manuscript.

W1: Clarification of Mathematical Definitions and Theoretical Identities

1. Precise operator definitions.
In the revised manuscript, we will add the relevant background reference [1] and explicitly define the retained frequency set, the spectral truncation operator TfT_f, the embedding operator EfE_f, and the mask operator MfM_f. These definitions will specify the correspondence between fine- and coarse-grid frequency indices, the treatment of cutoff boundary modes, and the Fourier normalization convention, improving both mathematical clarity and reproducibility.

2. Explicit justification of the identities.
We will also provide brief derivations of the two identities used in Theorem 1. Extracting the coefficients at the retained frequencies and embedding them back into the fine-grid spectrum, with zeros elsewhere, is equivalent to applying the low-frequency mask. Therefore,

EfTf=Mf.E_fT_f=M_f.

Likewise, I−MfI-M_f sets all retained-frequency coefficients to zero, so subsequent truncation yields

Tf(I−Mf)=0.T_f(I-M_f)=0.

We will include the corresponding frequency-wise expressions to show explicitly how these identities follow from the operator definitions, completing this step of the proof.

Ref:[1] Rippel, O., Snoek, J., & Adams, R. P. (2015). Spectral Representations for Convolutional Neural Networks. Advances in Neural Information Processing Systems, 28.

W2:Theoretical Contribution and Novelty of the Proposed Framework

We agree that the orthogonal spectral decomposition in Theorem 1 is a standard property and does not, by itself, constitute a new theoretical contribution. Its role is to formalize the mathematical structure of the proposed architecture: the closure and refinement corrections are restricted to complementary frequency subspaces, namely,

Πf⊥Δuclosure=0,ΠfΔurefinement=0.\Pi_f^\perp\Delta u_{\mathrm{closure}}=0, \qquad \Pi_f\Delta u_{\mathrm{refinement}}=0.

These identities give the two branches distinct roles in the scales they modify, while allowing cross-scale coupling through their inputs.

Our main contribution is methodological: combining these structural constraints with a coarse-grid numerical solver and two coupled learned modules. We will revise the manuscript to distinguish these structural properties from the method’s empirical advantages. The Full-state control and the ablation without Πf⊥\Pi_f^\perp evaluate the roles of the branch design and projection constraint, respectively. Theorem 1 itself does not guarantee higher prediction accuracy or better long-term stability.

W3、Q2:Motivation and Theoretical Connection of the Proposed Update Rules

The key connection is that Eqs. (8)–(9) guide the design of the modules’ inputs and outputs, as well as the learned update rules in Eqs. (10)–(12). Eqs. (8)–(9) provide the physical motivation for the module decomposition and cross-scale coupling. Eqs. (10)–(12) translate this structure into learned updates over a finite time step: the coarse-grid solver advances the resolved dynamics, Closure learns corrections for cross-scale feedback and numerical integration errors, and Refinement predicts the increment of the unresolved component.We model it through Neural Network and ensure that the output subspace is strictly guaranteed by the projection operator.

Reviewer_V68yOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes HybridOSS, a hybrid differentiable PDE framework that maintains a fine-resolution state while advancing the numerical backbone on a coarser grid. The state is decomposed into complementary spectral subspaces. A learned closure corrects the resolved dynamics, while a separate refinement network advances an explicitly maintained unresolved state. Spectral projections constrain the two learned updates to their respective subspaces. The method is evaluated on the Kuramoto–Sivashinsky and Allen–Cahn equations, followed by a three-dimensional decaying-turbulence example using differentiable OpenFOAM. The paper reports improved prediction errors relative to several hybrid baselines, favorable results against an unrestricted full-state corrector, and approximately 19× inference acceleration relative to the reported solver-only configuration.

Critical strengths

1.The coarse/fine architecture is well motivated and supported by useful ablations. Separating resolved-scale correction from explicit unresolved-state evolution provides a coherent way to maintain a fine-resolution state while using a coarse numerical backbone. Table 2 includes a capacity-matched full-state corrector, against which HybridOSS reduces global error from 0.0975 to 0.0858 on KS and from 0.0252 to 0.0173 on AC. These comparisons provide meaningful evidence that the proposed organization of learned updates can be beneficial.

2.The OpenFOAM experiment provides a substantive test of autoregressive prediction. The model is trained on three-step windows from a single trajectory and evaluated without resetting over 125 steps on 40 held-out trajectories covering different initial conditions and viscosities (Section 4.4 and Appendix B.3). The reported improvements near the training viscosity support the practical potential of the approach. The paper also acknowledges the deterioration under larger viscosity shifts, which helps establish the scope of the results.

Critical weaknesses

1.The contribution of orthogonal scale splitting is not sufficiently separated from the corrector architecture and input features. Table 3 reports a KS error of 0.0047 at (f=1), where Appendix A gives (\Pi_f=I) and (\Pi_f^\perp=0), so the unresolved refinement contributes nothing. This is substantially below SITL’s 0.1241 in Table 1. The relationship between these settings needs to be explained: a large improvement is already reported for a configuration without a nontrivial scale split. Appendix B.4 describes derived physical features, multiscale dilated convolutions, and FiLM conditioning for HybridOSS, while the other hybrid baselines use a different ResNet architecture. I acknowledge the capacity-matched full-state control in Table 2, but it is unclear whether its input features, conditioning, receptive fields, and initialization are also matched. Consequently, the current evidence does not adequately distinguish the benefit of the projected two-branch design from that of the particular corrector implementation.

2.The practical accuracy–efficiency advantage is incompletely established. For NS, Appendix B.3 specifies a (64^3) reference simulation with (\Delta t=0.001) s, whereas the hybrid model advances with (\Delta t=0.008) s and uses (f=8). The exact temporal configuration and physical interval associated with a timed solver-only “step” in Figure 3 are not equally clear. This matters for interpreting and reproducing the reported 19× speedup. Figure 3 also reports 21× acceleration for the full-state control, so the additional value of the split architecture is its accuracy improvement at somewhat higher cost. A comparison with numerical configurations spanning resolution and time step, evaluated over a common physical interval, is needed to establish the practical trade-off. The selected KS operating point is also difficult to justify from Table 3: increasing (f) from 1 to 16 raises error from 0.0047 to 0.0858 while reducing training time only from 17.51 to 17.11 seconds per epoch. These limitations affect the central efficiency claim.

Questions for authors

1.Clarification of the controls and contribution attribution. What differs between the (f=1) experiment in Table 3 and SITL in Table 1 in terms of solver, training protocol, network architecture, and inputs? Does the full-state control receive the same derived features, conditioning, and initialization as HybridOSS? Please provide enough detail to determine whether the remaining comparison isolates the proposed splitting, together with variation across the three reported runs for the key ablation results.

2.Clarification of the timing comparison. Which grid, time step, internal substeps, and solver settings are used by the solver-only baseline in Figure 3? Does each timed operation advance the same physical interval across methods, and is differentiation overhead disabled during inference? These details are necessary to assess the reported acceleration.

3.Clarification of the Allen–Cahn implementation. Does the low-fidelity step of (\Delta t=1.5) apply the explicit reaction update in Eq. (28) once, or does it include internal reaction substeps? If applied once, its linearized amplification factor near the uniform equilibria (u=\pm1) is (1-2\Delta t=-2). This would make the uncorrected reaction update locally unstable, although it does not establish instability of the learned hybrid. Please clarify the implementation; if this is indeed a single update, evidence against a numerical baseline with stable reaction integration would materially affect my interpretation of the AC results.

6 comments

We sincerely thank the reviewer for the valuable comments and suggestions.

W1 and Q1: Clarification of the controls and contribution attribution.

1. Clarifying the f=1f=1–SITL comparison. SITL in Table 1 and HybridOSS at f=1f=1 use the same fine-grid numerical solver and training protocol. SITL uses the baseline ResNet without our derived physical features, whereas f=1f=1 retains our Closure network and these features. Since Πf=I\Pi_f=I and Πf⊥=0\Pi_f^\perp=0, Refinement contributes no update. The improvement at f=1f=1 therefore cannot be attributed to a scale split.

2. Full-state as a matched control for the two-branch design. Full-state is a SITL-style control using the same coarse-grid numerical backbone as HybridOSS and a single unrestricted fine-grid corrector. It matches the parameter budget, available state inputs and derived features, timestep FiLM conditioning, and zero-initialized output heads. It also uses the Refinement dilation set and matches the maximum receptive field of HybridOSS. The Full-state–HybridOSS comparison therefore provides a fair assessment of the proposed two-branch design, as both methods use the same coarse-grid solver and matched corrector settings, differing only in whether the learned correction is implemented through a single network or two branches.

3. Ablation evidence for the complementary roles of branching and projection.
Compared with Full-state, the unprojected two-branch variant reduces mean global error from 0.09750.0975 to 0.08820.0882 on KS and from 0.02520.0252 to 0.01900.0190 on AC, although its unresolved errors are higher. With the two-branch networks and inputs unchanged, restoring Πf⊥\Pi_f^\perp reduces unresolved error from 0.99680.9968 to 0.47560.4756 on KS and from 0.49830.4983 to 0.17090.1709 on AC. Together, these comparisons support the projected two-branch design. These results support the complementary roles of the two-branch organization and the orthogonal refinement projection in improving overall prediction accuracy and unresolved-scale modeling. As stated at the beginning of Section 4.3, Table 1 evaluates HybridOSS as a complete method, whose performance reflects both the corrector implementation and the projected two-branch formulation. The matched coarse-solver ablations provide more targeted evidence for the latter.

KS GlobalKS ResolvedKS UnresolvedAC GlobalAC ResolvedAC Unresolved
w/o both modules0.38590.17620.83500.53690.28982.0467
w/o closure0.3830±0.00370.1762±0.00141.4859±0.02560.5368±3.29e-50.2898±1.82e-70.9725±0.0033
w/o refinement0.1163±0.00430.0090±0.00040.8350±2.93e-50.0182±0.00010.0004±1.69e-52.0470±1.35e-8
w/o Πf⊥\Pi^{⊥}_{f}0.0882±0.00930.0067±0.00150.9968±0.03940.0190±0.00260.0007±2.95e-50.4983±0.094
Full-state0.0975±0.00160.0105±0.00030.5616±0.00610.0252±0.00030.0012±1.71e-50.4246±0.0332
HybridOSS(ours)0.0858±0.00220.0097±0.00030.4756±0.11510.0173±0.00100.0004±0.000020.1709±0.0510

W2 and Q2: Clarification of the timing comparison.

1. NS solver settings and a common physical interval.

For NS, the Solver-only baseline in Figure 3 uses the native OpenFOAM pimpleFoam solver on a 64364^3 grid with a time step of 0.008s0.008\mathrm{s}. Apart from the grid resolution and time-step size, the numerical settings, including the PIMPLE iteration settings, follow the official decayIsoTurb tutorial[1]. No additional temporal subcycling is used.

Full-state and HybridOSS use our differentiable implementation of pimpleFoam, on an 838^3 numerical grid while maintaining a 64364^3 prediction state. Both use an outer time step of 0.008s0.008\mathrm{s}. Thus, each timed update advances the same physical interval across Solver-only, Full-state, and HybridOSS. Automatic differentiation is disabled during inference for both hybrid methods. Apart from the numerical grid resolution, both hybrid methods use the same pimpleFoam solver settings as the Solver-only baseline.

The reference simulation uses a 64364^3 grid with Δt=0.001s\Delta t=0.001\mathrm{s} and provides the reference solution for error evaluation. Predictions and reference solutions are compared at matching physical times. The reported speedup is therefore measured against the Solver-only baseline at Δt=0.008s\Delta t=0.008\mathrm{s}, rather than against the reference simulation at Δt=0.001s\Delta t=0.001\mathrm{s}.

2. Scope of the efficiency evaluation and computational constraints.

We agree that a systematic comparison across numerical grid resolutions and time steps, evaluated over a common physical interval, would provide a more complete assessment of the practical accuracy–efficiency trade-off. Given the computational resources available during the rebuttal period, we have not been able to complete this full evaluation. We have therefore prioritized additional baselines and ablations at the existing 64364^3 prediction resolution.We added three ablations: masking the high-frequency input to Closure, masking the low-frequency input to Refinement, and masking both directions. These results extend the evidence to three dimensions and support the complementary roles of branch separation, orthogonal projection, and cross-scale conditioning in the tested configurations. We plan to extend the evaluation to higher resolutions and report prediction accuracy, runtime, and memory usage in future work.

Configurationν=5×10⁻⁶ν=1×10⁻⁵ν=5×10⁻⁵ν=1×10⁻⁴
w/o Closure0.8457 ± 0.02710.8450 ± 0.02760.8637 ± 0.02760.9168 ± 0.0261
w/o Refinement0.4308 ± 0.09230.4482 ± 0.09440.5583 ± 0.10650.6919 ± 0.1183
w/o projection0.3636 ± 0.01600.3465 ± 0.01600.4856 ± 0.01000.6566 ± 0.0127
Full state0.5571 ± 0.01770.4727 ± 0.01241.1414 ± 0.01701.5580 ± 0.0247
w/o HF input to Closure0.5050 ± 0.06480.4884 ± 0.07160.5396 ± 0.05680.6494 ± 0.0386
w/o LF input to Refinement0.8290 ± 0.04480.8115 ± 0.04720.8334 ± 0.03470.8876 ± 0.0280
w/o cross-scale inputs0.8176 ± 0.04830.8024 ± 0.05260.8164 ± 0.04260.8651 ± 0.0307
HybridOSS(full)0.3151 ± 0.03370.2990 ± 0.03500.4157 ± 0.01690.5739 ± 0.0129

3. Limited efficiency gains on KS.

KS is an inexpensive one-dimensional problem, and spatial coarsening provides only limited savings in the reported training time. We will make this limitation more explicit: as discussed in the Cost subsection of Section 4.2, our method does not offer an efficiency advantage on computationally inexpensive problems, but remains competitive on the more demanding 2D and 3D problems evaluated in this study.

Ref: [1] G. Comte-Bellot and S. Corrsin, "The use of a contraction to improve the isotropy of grid-generated turbulence," J. Fluid Mech. 25, 657 (1966).

Q3:Clarification of the Allen–Cahn implementation.

The original AC solver applies Eq. (28) once per step with Δt=1.5\Delta t=1.5, without internal reaction substeps. We agree that its amplification factor near u=±1u=\pm 1 is −2-2, making the uncorrected reaction update locally unstable.

To address this concern, we added a Solver-only baseline using the convex-splitting reaction update:

un+1−unΔt=un−(un+1)3.\frac{u^{n+1}-u^n}{\Delta t}=u^n-(u^{n+1})^3.

All other settings remain unchanged. This update has a unique real solution, and its amplification factor near u=±1u=\pm 1 is

GCS=1+Δt1+3Δt≈0.4545,G_{\mathrm{CS}}=\frac{1+\Delta t}{1+3\Delta t}\approx 0.4545,

so the reaction update is locally stable.

This replacement reduces the Solver-only global error from 0.48360.4836 to 0.24180.2418, while HybridOSS, retaining its original explicit reaction update, achieves 0.01730.0173. Thus, HybridOSS retains an accuracy advantage even against a baseline with locally stable reaction integration.

MethodGlobalResolvedUnresolved
Solver-only, explicit reaction0.48360.232358.1416
Solver-only, convex-splitting reaction0.24180.060911.7601
HybridOSS, explicit reaction0.0173±0.00100.0004±0.000020.1709±0.0510

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.