acceptodds
Under review as a conference paper at ICLR 2027

BalanceNO: Discovering What to Conserve for Coupled PDE Forecasting

27 views

Abstract

Accurate autoregressive prediction of coupled partial differential equations is essential for fast multiphysics simulation, where spatial transport coexists with interactions that redistribute conserved quantities among fields. Existing surrogates face a critical dilemma: unconstrained regression suffers from severe physical drift over long rollouts, while hard-coded conservation laws are overly rigid, failing to capture the dynamic redistribution of quantities among coupled fields. We introduce BalanceNO, an operator neural framework that reformulates coupled-field evolution as a transport–interaction balance governed by a linear invariant subspace identified from trajectory data. BalanceNO discovers and validates exact balance constraints solely from training trajectories. To strictly enforce them during updates, it integrates telescoping spatial fluxes with nullspace-projected local interactions—preserving field-coupled invariants without freezing individual field totals. By embedding selective conservation directly into its core discrete update, BalanceNO bypasses the need for complete governing equations, seamlessly accommodating weighted conservation laws and partially conservative dynamics. Evaluated across complex spatiotemporal systems combining local transport, global nonlocal interactions, and strongly coupled fields in heterogeneous, three-dimensional, and interface-evolving media, BalanceNO delivers accurate long-horizon rollouts alongside roundoff-level invariant preservation. The code for the experiments can be found in https://anonymous.4open.science/r/BNO-0BB3.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Deep learning, neural operators

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 4 reviews, above 27% of the others.
Reviewer_CQabOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces BalanceNO, which identifies conserved linear combinations of spatially integrated fields from training trajectories. The method enforces these invariants through conservative flux updates and nullspace-projected interactions, while allowing individual field totals to evolve. The authors evaluate autoregressive prediction errors, invariant drift, and errors in individual field totals on coupled flow–phase-field dynamics, rotating shallow-water flow, reaction–diffusion systems, and three-dimensional compressible flow.

Critical strengths
  • Explicit transport–interaction decomposition. The method separates prediction updates into conservative spatial transport and interactions between fields. Projecting the interaction term onto the nullspace of the identified conservation matrix preserves the conserved combinations while allowing individual field totals to change.

  • Evidence for the value of selective conservation. The experiments measure both invariant drift and field-total prediction error to distinguish selective conservation from overconstraint. On the shallow-water benchmark, BalanceNO maintains invariant drift near 10−810^{-8} and achieves a field-total error of 0.0083, compared with approximately 0.81 for the fieldwise-conservative controls.

  • Substantial experimental effort. The study covers three two-dimensional coupled PDE tasks and one three-dimensional flow task, with results reported over five random seeds. It includes comparisons under approximately matched parameter budgets and controlled ablations using a shared backbone.

Critical weaknesses
  1. Limited spatial and temporal diagnostics. The paper’s only figure is a schematic of the method. Despite evaluating 46-step NSCH and 22-step SWE rollouts, it provides no predicted-field visualizations, spatial error maps, or error-versus-time curves. This limits direct assessment of interface morphology, local structures, and error growth.

  2. Insufficient evidence for the added value of the transport–interaction decomposition. In the controlled ablation, the full method and the global-projection control share the same backbone and each contain 6,816 parameters. Their 40-step macro-averaged errors are 5.46% ± 4.27% and 6.16% ± 3.62%, respectively. The 95% confidence interval for the paired difference includes zero, so these results do not establish an advantage over the simpler global correction.

  3. Overconstrained fieldwise-conservative baselines. Domain-integrated momentum rotates under Coriolis forcing in SWE, while individual species totals change through reactions in RD. However, the fieldwise-conservative controls fix every channel’s total, yielding field-total errors of approximately 0.81 on SWE and 0.25 on RD. These comparisons primarily demonstrate that inappropriate constraints impair prediction; they do not establish the necessity of the proposed transport–interaction structure.

Questions for authors

see Weaknesses.

2 comments

We thank the reviewer for the careful review.

W1: Spatial and temporal diagnostics are missing.

R: We generated fixed-case field and error snapshots and error-vs-time curves. They show the NSCH phase, SWE height, RD species C and a CNS density slice at times 1, H/2, H and 2H. These will go in the revision.

W2: There is insufficient evidence for the transport–interaction decomposition.

R: Table 5 cannot show this. It is a small linear 16² diagnostic with 6,816 parameters, and its seed SD (≈\approx 4 pp) is as large as the differences between decoders. We therefore reran the comparison on NSCH at the paper's native setting (mean ±\pm std over five paired seeds, %).

NSCH (H / 2H = 46 / 92)

VariantMacro HMacro 2HFTDrift
BNO6.28 ± 1.6015.22 ± 5.780.51 ± 0.301.2e-8
GP-direct(Simple global projection)15.97 ± 8.8629.17 ± 23.637.29 ± 8.131.6e-8
GP-structured (Table 5 control)8.34 ± 2.8032.15 ± 19.541.55 ± 1.831.8e-8
BNO, identity filter17.31 ± 3.2559.81 ± 41.955.55 ± 1.571.2e-8
No projection15.91 ± 7.0340.14 ± 43.887.46 ± 7.086.9e-2

SWE (H / 2H = 22 / 44)

VariantMacro HMacro 2HFTDrift
BNO2.87 ± 0.433.09 ± 0.451.00 ± 0.422.8e-8
GP-direct4.61 ± 0.625.03 ± 1.132.17 ± 0.372.7e-8
GP-structured2.97 ± 0.323.04 ± 0.401.02 ± 0.382.7e-8
BNO, identity filter3.37 ± 0.9518.26 ± 30.591.23 ± 0.522.6e-8
No projection5.58 ± 0.535.92 ± 0.502.54 ± 0.333.1e-3

RD (H / 2H = 10 / 20)

VariantMacro HMacro 2HFTDrift
BNO1.41 ± 0.191.65 ± 0.250.23 ± 0.174.5e-8
GP-direct1.37 ± 0.161.54 ± 0.210.21 ± 0.084.8e-8
GP-structured1.42 ± 0.151.66 ± 0.210.19 ± 0.064.4e-8
BNO, identity filter1.29 ± 0.151.48 ± 0.180.20 ± 0.124.9e-8
No projection1.48 ± 0.141.77 ± 0.170.39 ± 0.092.1e-3

CNS 64³ (H / 2H = 6 / 12)

VariantMacro HMacro 2HFTDrift
BNO22.12 ± 2.3143.45 ± 3.280.004.3e-8
GP-direct33.97 ± 1.5364.35 ± 3.190.004.2e-8
GP-structured30.68 ± 1.4556.05 ± 3.760.004.6e-8
BNO, identity filter31.78 ± 1.3761.88 ± 2.960.003.4e-8
No projection35.39 ± 1.5963.65 ± 1.534.38 ± 0.914.0e-2

As you can see, against the simple global projection, our method performs exceptionally well on the NSCH and CNS(3d) problems, and also yields quite good results for the SWE and RD problems.

W3: The fieldwise-conservative baselines are overconstrained.

R: We agree that these baselines are negative controls, and we add two results. First, the penalty is intrinsic: any predictor that freezes all totals, even one given the true future fields, has a mean macro error of at least 61.3% on SWE and 23.1% on RD, whereas BNO reaches 8.5% and 4.2%. Second, a properly selective baseline still trails BNO. Selective NFV uses the same C, learned exchange in ker C, and parameters within 1% of BNO; under the compact protocol its macro error is 47.0% vs 8.5% on SWE and 6.9% vs 4.2% on RD, in 5/5 seeds. This separates our two contributions: choosing the right constraints, and the transport–interaction decoder.

Thank you for your response and clarifications. As my core concerns remain unresolved, I am still not convinced that the paper meets the bar for acceptance and will therefore maintain my current score.

Reviewer_9H34OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary
  • Identifies a subspace of conserved linear combinations from spatially integrated trajectory increments using SVD.
  • Combines telescoping fluxes with nullspace-projected interactions, preserving selected invariants while allowing individual field totals to evolve.
  • Evaluates forecasting accuracy and invariant drift on coupled 2D systems and 3D compressible Navier–Stokes
Critical strengths
  • The distinction between conserving every field individually and conserving the physically appropriate combinations is important and well motivated.
  • Conservation is enforced algebraically without tuning penalty weights, and the completeness result clarifies the expressiveness of the constrained update.
Critical weaknesses
  • Exact conservation is guaranteed only for the identified subspace. The consequences of enforcing an incorrect subspace under noise or insufficient excitation remain uncharacterized.
  • Evidence of improved accuracy over global projection with the same backbone is limited. The paired confidence interval in Appendix F includes zero, as the authors acknowledge.
Questions for authors
  • How do observation noise and the SVD singular-value gap affect subspace estimation error, and how does this error propagate through rollouts?
  • Under noisy observations, how do estimated invariants, oracle invariants, and soft constraints compare? When is hard enforcement beneficial?
  • Do the gains persist against conservative baselines supplied with the same invariant subspace and appropriate exchange/source mechanisms?
1 comment

We thank the reviewer for the careful review.

Q1 / W1: How do noise and the singular-value gap affect subspace error, and how does that error propagate through rollouts?

R: For the correct rank, the projector error is at most min⁡(1,2ε/δ)\min(1, 2\varepsilon/\delta), and the experiment follows this bound: on SWE, shrinking the gap by 10×10\times and 100×100\times raises the error from 1.4×10−41.4 \times 10^{-4} to 1.4×10−31.4 \times 10^{-3} and then 1.5×10−21.5 \times 10^{-2}. The true-invariant drift is at most ∥Q−Q^∥⋅∑∥Δmt∥\|Q - \hat{Q}\| \cdot \sum \|\Delta m_t\| plus the measured enforcement residual, and all 190190 frozen-model horizons satisfy this bound. At 1%1\% noise the fixed τ\tau returns rank 00. A noise-calibrated τ\tau restores the rank and cuts drift by 79–96%79\text{--}96\% in 2D and by 99.998%99.998\% on CNS, without improving accuracy (the rank-00 arm has slightly lower macro error).

Q2: How do estimated, oracle and soft constraints compare under noise, and when is hard enforcement beneficial?

R: We used the 3D exchange system with 1% noise on the discovery data.

Macro, 48 stepsTrue-invariant drift
Estimated hard4.54%5.5e-5
Oracle hard4.49%8.1e-8
Soft4.31%1.28e-3

The accuracy differences lie within the intervals. Hard enforcement keeps the drift 15–23× lower at no resolved accuracy cost, so it is beneficial when the physical consistency of the totals matters, but it does not guarantee lower spatial error. Enforcing a wrong law is harmful (5.41% vs 3.96% with the correct law), so correct identification comes first.

W2: Is there evidence against global projection with the same backbone?

R: Yes, we reran the comparison at native scale. Entries are % (mean ±\pm SD over 5 seeds); FT = field-total error at H; Drift = invariant drift; bold = best mean.

NSCH (H / 2H = 46 / 92)

VariantMacro HMacro 2HFTDrift
BNO6.28 ± 1.6015.22 ± 5.780.51 ± 0.301.2e-8
GP-direct(Simple global projection)15.97 ± 8.8629.17 ± 23.637.29 ± 8.131.6e-8
GP-structured (Appendix-F control)8.34 ± 2.8032.15 ± 19.541.55 ± 1.831.8e-8
BNO, identity filter17.31 ± 3.2559.81 ± 41.955.55 ± 1.571.2e-8
No projection15.91 ± 7.0340.14 ± 43.887.46 ± 7.086.9e-2

SWE (H / 2H = 22 / 44)

VariantMacro HMacro 2HFTDrift
BNO2.87 ± 0.433.09 ± 0.451.00 ± 0.422.8e-8
GP-direct4.61 ± 0.625.03 ± 1.132.17 ± 0.372.7e-8
GP-structured2.97 ± 0.323.04 ± 0.401.02 ± 0.382.7e-8
BNO, identity filter3.37 ± 0.9518.26 ± 30.591.23 ± 0.522.6e-8
No projection5.58 ± 0.535.92 ± 0.502.54 ± 0.333.1e-3

RD (H / 2H = 10 / 20)

VariantMacro HMacro 2HFTDrift
BNO1.41 ± 0.191.65 ± 0.250.23 ± 0.174.5e-8
GP-direct1.37 ± 0.161.54 ± 0.210.21 ± 0.084.8e-8
GP-structured1.42 ± 0.151.66 ± 0.210.19 ± 0.064.4e-8
BNO, identity filter1.29 ± 0.151.48 ± 0.180.20 ± 0.124.9e-8
No projection1.48 ± 0.141.77 ± 0.170.39 ± 0.092.1e-3

CNS 64³ (H / 2H = 6 / 12)

VariantMacro HMacro 2HFTDrift
BNO22.12 ± 2.3143.45 ± 3.280.004.3e-8
GP-direct33.97 ± 1.5364.35 ± 3.190.004.2e-8
GP-structured30.68 ± 1.4556.05 ± 3.760.004.6e-8
BNO, identity filter31.78 ± 1.3761.88 ± 2.960.003.4e-8
No projection35.39 ± 1.5963.65 ± 1.534.38 ± 0.914.0e-2

As you can see, against the simple global projection, our method performs exceptionally well on the NSCH and CNS(3d) problems, and also yields quite good results for the SWE and RD problems.

Q3: Do the gains hold against conservative baselines given the same subspace and an exchange mechanism?

R: Yes. Selective NFV uses the same C, learned exchange in kernel C, and parameters within 1% of BNO. Under the compact protocol BNO reduces macro error relative to it by 38–85% in 5/5 seeds. NFV is slightly better only on RD doubled-horizon totals (1.03% vs 1.05%).

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

This paper proposes BalanceNO, a neural operator framework for forecasting coupled PDE systems while preserving selected conservation laws. The method first identifies linear combinations of spatially integrated physical fields that remain invariant. It then constructs a balance-preserving evolution operator by decomposing state updates into conservative spatial fluxes and locally projected cross-field interactions. The proposed architecture guarantees exact preservation of the identified linear invariants under the stated conditions. Experiments on several coupled PDE benchmarks demonstrate improved long-horizon forecasting accuracy and conservation performance compared with existing neural operator baselines.

Critical strengths

The proposed framework provides a principled way to incorporate the identified conservation constraints directly into the prediction architecture, without relying on additional conservation penalties. The experimental results are also promising. BalanceNO achieves strong forecasting accuracy across several PDE benchmarks while maintaining near-exact conservation.

Critical weaknesses
  1. Insufficient justification of the Flux + Interaction decomposition.

While the effectiveness of selective conservation is well demonstrated, the additional benefits of the proposed Flux + Interaction decomposition remain unclear. Given the discovered invariant matrix CC, a standard neural operator with a simple global projection can enforce exactly the same global conservation constraints. Table 5 already considers this baseline, but the results are relatively close (5.46% vs. 6.16%), without establishing a clear advantage for the proposed decomposition. More systematic comparisons on the main PDE benchmarks would be valuable, particularly to clarify under which dynamical regimes the proposed structural inductive bias provides significant improvements and when its benefits are limited.

  1. Clarity and related work.

The paper would benefit from a clearer introduction to the coupled PDE setting, particularly the roles of spatial transport, cross-field interactions, and conserved quantities. Introducing a concrete example earlier in Section 3 would make the proposed framework more accessible. The distinction between globally identified conservation constraints and locally projected interactions also deserves further explanation.

In addition, the discussion of related work is incomplete, particularly regarding existing approaches for discovering and preserving unknown invariant quantities (e.g., FINDE, Kaiser et al., and ProbConserv). There are also unresolved citation and formatting issues. Finally, the SVD-based discovery relies on fixed thresholds and sufficiently informative trajectories, but the robustness of the identified invariant subspace is not thoroughly examined.

References

Matsubara and Yaguchi (2023). FINDE: Neural Differential Equations for Finding and Preserving Invariant Quantities. ICLR 2023.

Kaiser, Kutz, and Brunton (2018). Discovering Conservation Laws from Data for Control. IEEE Conference on Decision and Control, 2018.

Hansen et al. (2023). Learning Physical Models that Can Respect Conservation Laws. ICML 2023.

Questions for authors

Comparison with global projection on the main benchmarks. Table 5 shows that global projection performs comparably to the full BalanceNO decoder on the weighted-exchange benchmark. Could the authors extend this comparison to the main coupled PDE benchmarks, particularly the reaction–diffusion system where the conserved quantities involve multiple species? Using the same backbone and discovered invariant matrix, it would be helpful to compare forecasting accuracy, individual field-total errors, and long-horizon behavior.

When is the Flux + Interaction decomposition beneficial? Could the authors provide a more systematic characterization of the regimes in which the proposed decomposition offers a clear advantage over global projection, as well as cases where the difference is small? For example, a controlled experiment varying the relative strengths of spatial transport and cross-field interactions, or the degree of spatial heterogeneity, could help explain when the additional structural inductive bias is useful. Such an analysis would be valuable even if the proposed method does not consistently outperform global projection across all regimes.

Robustness of invariant discovery. Could the authors provide additional diagnostics of the SVD-based invariant identification procedure on the main benchmarks, such as the normalized singular-value spectra and their separation from the selected thresholds? A brief sensitivity analysis of the discovered rank with respect to the thresholds or the amount of trajectory excitation would help assess the reliability of the identified conservation subspace.

3 comments

We thank the reviewer for the careful review.

Q1 : Does the Flux + Interaction decomposition beat simple global projection on the main benchmarks, including field totals and long horizons?

R: We reran this comparison on NSCH at the paper's native setting. Values are mean ±\pm std over five paired seeds (%). NSCH (H / 2H = 46 / 92)

VariantMacro HMacro 2HFTDrift
BNO6.28 ± 1.6015.22 ± 5.780.51 ± 0.301.2e-8
GP-direct (simple global projection)15.97 ± 8.8629.17 ± 23.637.29 ± 8.131.6e-8
GP-structured (keeps the decomposition, global interaction correction)8.34 ± 2.8032.15 ± 19.541.55 ± 1.831.8e-8
BNO, identity filter17.31 ± 3.2559.81 ± 41.955.55 ± 1.571.2e-8
No projection15.91 ± 7.0340.14 ± 43.887.46 ± 7.086.9e-2

SWE (H / 2H = 22 / 44)

VariantMacro HMacro 2HFTDrift
BNO2.87 ± 0.433.09 ± 0.451.00 ± 0.422.8e-8
GP-direct4.61 ± 0.625.03 ± 1.132.17 ± 0.372.7e-8
GP-structured2.97 ± 0.323.04 ± 0.401.02 ± 0.382.7e-8
BNO, identity filter3.37 ± 0.9518.26 ± 30.591.23 ± 0.522.6e-8
No projection5.58 ± 0.535.92 ± 0.502.54 ± 0.333.1e-3

RD (H / 2H = 10 / 20)

VariantMacro HMacro 2HFTDrift
BNO1.41 ± 0.191.65 ± 0.250.23 ± 0.174.5e-8
GP-direct1.37 ± 0.161.54 ± 0.210.21 ± 0.084.8e-8
GP-structured1.42 ± 0.151.66 ± 0.210.19 ± 0.064.4e-8
BNO, identity filter1.29 ± 0.151.48 ± 0.180.20 ± 0.124.9e-8
No projection1.48 ± 0.141.77 ± 0.170.39 ± 0.092.1e-3

CNS 64³ (H / 2H = 6 / 12)

VariantMacro HMacro 2HFTDrift
BNO22.12 ± 2.3143.45 ± 3.280.004.3e-8
GP-direct33.97 ± 1.5364.35 ± 3.190.004.2e-8
GP-structured30.68 ± 1.4556.05 ± 3.760.004.6e-8
BNO, identity filter31.78 ± 1.3761.88 ± 2.960.003.4e-8
No projection35.39 ± 1.5963.65 ± 1.534.38 ± 0.914.0e-2

R: Yes on NSCH, SWE and CNS, where BNO lowers the macro error relative to simple global projection at both horizons and cuts the NSCH field-total error 14×; on RD the two decoders do not separate.

Q2: In which regimes does the decomposition help?

R: The gain depends on how strongly the interaction changes the non-conserved totals and on the rollout length. Relative to simple global projection, BNO reduces the macro error by 61% / 48% (H / 2H) on NSCH, where surface tension exchanges momentum between the flow and the phase field, and by 38% / 39% on SWE, where the Coriolis rotation exchanges the two momentum totals. On these two benchmarks it also cuts the field-total error 14× and 2×. On CNS the discovered C is full rank, so no selective exchange exists, yet the flux-form transport update alone still lowers the error by 35% / 32%. On RD, whose errors are already about 1.4%, the decoders do not separate. The controlled 3D A+B⇌CA+B\rightleftharpoons C study shows the same ordering across regimes: the reaction-dominated regime gains most (6.5% / 15.7% at 24 / 48 steps), the diffusion-dominated (1.4–3.8%) and heterogeneous (0.7–1.5%) regimes gain less, and the intervals include zero at its 5.3k-parameter scale. The small linear toy of Table 5 is too small to resolve any gain. The decomposition therefore matters most when the mean-changing interaction is large relative to transport and the rollouts are long.

W2: The presentation is unclear.

R: We will add an early A+B⇌CA+B\rightleftharpoons C example. Reaction moves species along (−1,−1,+1), A+C and B+C are conserved, and transport redistributes them. We will also state that C constrains global totals, while pointwise projection is only a decoder parameterization.

W3: Related work and formatting.

R: We will add FINDE (first integrals, ICLR'23), Kaiser et al. (Koopman-based discovery, CDC'18) and ProbConserv (Bayesian integral constraints, ICML'23). Our scope is narrower: linear combinations of field totals, enforced in a discrete PDE decoder. We will also fix the citation and formatting issues.

Q3 / W4: How robust is the SVD discovery (spectra, thresholds, excitation)?

R: We report normalized spectra for all main benchmarks. The discovered ranks are 1/1/3/5, and the active and null blocks are separated by about 6.7–8.3 orders of magnitude. The threshold τ\tau is at least 316× below the smallest active value, and the rank is unchanged over 6–8 decades of τ\tau. One or two trajectories suffice, and screening rejects a weak-excitation spurious direction on NSCH. At 1% noise the fixed τ\tau returns rank 0, so it fails without imposing a false law, and a noise-calibrated τ\tau recovers the rank.

Under 1% iid noise the fixed τ\tau returns rank 0, so it enforces no constraint rather than a false one. A noise-calibrated τ\tau restores the rank and cuts the true-invariant drift by 79–99.998%, but it does not improve accuracy: the rank-0 arm has the lowest macro error. Coherent noise, sparse sensors and joint weak excitation make discovery fail, and we now state these as limitations.

Thank you for the detailed rebuttal and additional experiments. I appreciate the new comparisons with global projection across the main benchmarks, as well as the analysis of different dynamical regimes. These results address an important part of my original concerns and provide stronger empirical support for the proposed framework.

However, I still have some reservations about the specific contribution of the Flux + Interaction decomposition. While BNO clearly outperforms simple global projection on several benchmarks, the comparison with GP-structured is less conclusive. The two methods perform similarly on SWE and RD, while the improvement on CNS does not directly support the selective interaction mechanism, since PC=0P_C=0 in this case. Moreover, the substantial performance degradation when removing the filter suggests that the filtering and transport parameterization also play an important role. I therefore remain somewhat unconvinced that the additional benefits of the proposed decomposition have been sufficiently isolated from these other design choices.

Overall, the rebuttal has improved my assessment of the empirical results, and I appreciate the authors' efforts to address the presentation and related-work issues. Nevertheless, my remaining concerns about the methodological contribution and robustness prevent me from recommending acceptance at this stage. I will therefore maintain my rating of 2 (Weak Rejection).

Reviewer_7HS9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses long-term forecasting of coupled PDE systems, where individual fields can exchange quantities even though certain combinations of their totals remain conserved. Enforcing conservation separately for every field can therefore impose the wrong physics, while unconstrained models can accumulate physical drift.

BalanceNO first estimates conserved linear combinations from training trajectories using singular value decomposition. It then builds these constraints into each prediction step: spatial fluxes account for transport, while interactions between fields are projected into the null space of the discovered conservation laws. This preserves the identified combinations without forcing each field total to remain fixed.

The authors evaluate the method on several coupled spatiotemporal systems and report improved long-horizon forecasts and near-roundoff preservation of the identified invariants. The key question for the review is whether those gains come from the proposed local structure itself and whether the inferred invariants remain reliable when training trajectories are noisy or unrepresentative.

Critical strengths
  1. The paper addresses the right conservation structure for coupled systems. Instead of treating every field as independently conserved, BalanceNO identifies conserved combinations of field totals while allowing individual fields to exchange quantities. The discrete update enforces these combinations at every step, and the paper clearly distinguishes exact conservation from claims about stability or identifying the true physical flux.
  2. The approach is supported by a useful range of experiments. Across the main 2D benchmarks, the authors report both long-rollout accuracy and invariant drift, with five-seed results and component-total diagnostics. They also disclose an initially unfavorable ablation against global projection and rerun the comparison after revising the decoder, which makes the evidence easier to assess fairly. The reported gains are promising, though the second strength should be weighed against whether the baselines isolate the contribution of the new architecture.
Critical weaknesses
  1. The experiments do not clearly establish that the proposed local transport–interaction decoder improves accuracy over a simpler global correction. The matched-backbone comparison is conducted on a small synthetic problem, not the main nonlinear benchmarks. The revised method wins in only three of five seeds, and the reported paired confidence interval includes zero. Moreover, the authors show that the revised decoder becomes algebraically equivalent to global projection when its low-pass filter is the identity. A matched-backbone comparison across the main tasks, with a separate filter ablation, is needed to determine what drives the gains.
  2. Invariant discovery is tested under favorable assumptions. The method assumes that the relevant conservation laws are linear in the supplied field totals. Training trajectories are designed to excite nonconserved directions, and in the weighted-conservation example the inputs are already transformed into stored densities. It remains unclear how reliably the method identifies the right constraints with noisy or sparsely observed data, weakly excited dynamics, or approximate conservation. This matters because a spurious discovered invariant would then be enforced at every forecast step. A related qualification is that the conservation-aware baselines deliberately conserve each channel separately. They are useful negative controls, but they do not establish superiority over a properly configured selective-conservation baseline.
Questions for authors

On the main nonlinear benchmarks, how does the same backbone perform with simple global invariant projection, no projection, and the proposed decoder? Can the low-pass filter be ablated separately? How sensitive is invariant identification to noise, sparse observations, weak excitation, and the singular-value threshold? What happens when an approximately conserved quantity is incorrectly treated as exact? Can a conservation-aware baseline be configured to conserve only the identified combinations, rather than every field separately? The 3D experiment uses two test trajectories per seed and six forecast steps, with all field totals conserved. Can the authors evaluate longer rollouts and a 3D case with genuine exchange between fields? Appendix G mentions a pilot-test fallback for two baseline configurations. Exactly which data were used for that selection, and were the final test trajectories independent of it?

2 comments

We thank the reviewer for the careful review.

Q1 / W1: Does the decoder beat a simple global projection on the main benchmarks (with no-projection also reported)? And what about the B=I equivalence?

R: We reran the comparison at native scale. GP-direct is the simple global projection (standard readout, then v − Q⟨v⟩). GP-structured is the Table 5 control: it keeps our transport–interaction structure and filter and replaces Pr by r − Q⟨r⟩. Entries are % (mean ±\pm SD over 5 seeds); FT = field-total error at H; Drift = invariant drift; bold = best mean.

NSCH (H / 2H = 46 / 92)

VariantMacro HMacro 2HFTDrift
BNO6.28 ± 1.6015.22 ± 5.780.51 ± 0.301.2e-8
GP-direct15.97 ± 8.8629.17 ± 23.637.29 ± 8.131.6e-8
GP-structured8.34 ± 2.8032.15 ± 19.541.55 ± 1.831.8e-8
BNO, identity filter17.31 ± 3.2559.81 ± 41.955.55 ± 1.571.2e-8
No projection15.91 ± 7.0340.14 ± 43.887.46 ± 7.086.9e-2

SWE (H / 2H = 22 / 44)

VariantMacro HMacro 2HFTDrift
BNO2.87 ± 0.433.09 ± 0.451.00 ± 0.422.8e-8
GP-direct4.61 ± 0.625.03 ± 1.132.17 ± 0.372.7e-8
GP-structured2.97 ± 0.323.04 ± 0.401.02 ± 0.382.7e-8
BNO, identity filter3.37 ± 0.9518.26 ± 30.591.23 ± 0.522.6e-8
No projection5.58 ± 0.535.92 ± 0.502.54 ± 0.333.1e-3

RD (H / 2H = 10 / 20)

VariantMacro HMacro 2HFTDrift
BNO1.41 ± 0.191.65 ± 0.250.23 ± 0.174.5e-8
GP-direct1.37 ± 0.161.54 ± 0.210.21 ± 0.084.8e-8
GP-structured1.42 ± 0.151.66 ± 0.210.19 ± 0.064.4e-8
BNO, identity filter1.29 ± 0.151.48 ± 0.180.20 ± 0.124.9e-8
No projection1.48 ± 0.141.77 ± 0.170.39 ± 0.092.1e-3

CNS 64³ (H / 2H = 6 / 12)

VariantMacro HMacro 2HFTDrift
BNO22.12 ± 2.3143.45 ± 3.280.004.3e-8
GP-direct33.97 ± 1.5364.35 ± 3.190.004.2e-8
GP-structured30.68 ± 1.4556.05 ± 3.760.004.6e-8
BNO, identity filter31.78 ± 1.3761.88 ± 2.960.003.4e-8
No projection35.39 ± 1.5963.65 ± 1.534.38 ± 0.914.0e-2

As you can see, our method performs exceptionally well on the NSCH and CNS(3d) problems, and also yields quite good results for the SWE and RD problems.

The B=I equivalence holds for the completed decoder T(J, a+Qr) + Pr ablated in Table 5. It does not hold for the main 2D decoder T(J, a) + Pr, which differs from GP-structured by Q(r−⟨r⟩) for any B. We will state this explicitly.

Why Table 5 was a near-tie: it is a linear 16² system with 6,816 parameters, whose seed SD ($\approx $4 pp) is as large as the decoder gap. At native scale the same pair separates:

SettingParamsBNOGP-structured
Table 5 toy (16², 40 steps)6,8165.46 ± 4.276.16 ± 3.62
NSCH, H = 4661,8216.28 ± 1.608.34 ± 2.80
NSCH, 2H = 9261,82115.22 ± 5.7832.15 ± 19.54

Once we employ a finer grid and a greater number of rollout steps for this problem, the gap widens and our method wins.

Q2: Can the filter be ablated separately?

R: Yes, we have added this method (BNO, identity filter) in the tables above. On NSCH and CBS-3d, the filter is essential: without it, BNO's error rises from 6.28% to 17.31% at H and from 15.22% to 59.81% at 2H, and it is then no more accurate than GP-direct (17.31% vs 15.97%). GP-direct has no such filter, so the BNO–GP-direct gap combines decoder and filter; the decoder-only effect, with the filter in both arms, is BNO vs GP-structured. The filter also helps on CNS (22.12% vs 31.78% at H, 43.45% vs 61.88% at 2H) and stabilizes long SWE rollouts (3.09% vs 18.26 ± 30.59% at 2H). On RD it is not needed: without it BNO is slightly better (1.29% vs 1.41% at H), but it still performs not bad. The filter is therefore an essential task-dependent component.

Q3: How sensitive is discovery to noise, sparse observations, weak excitation and the threshold?

R: On clean native data the discovered ranks are 1/1/3/5. The threshold τ=10−5\tau = 10^{-5} sits at least 316× below the smallest active singular value and at least 14,600× above the largest null one, and the rank is unchanged over 6–8 decades of τ\tau. One or two discovery trajectories keep the correct rank in all seeds; on NSCH, held-out screening rejects a spurious direction caused by weak excitation.

Under 1% iid noise the fixed τ returns rank 0, so it enforces no constraint rather than a false one. A noise-calibrated τ\tau restores the rank and cuts the true-invariant drift by 79–99.998%, but it does not improve accuracy: the rank-0 arm has the lowest macro error. Coherent noise, sparse sensors and joint weak excitation make discovery fail, and we now state these as limitations.

Q4: What happens when an approximately conserved quantity is treated as exact?

R: We tested a 3D A+B⇌CA+B\rightleftharpoons C system with a weak extra A⇌BA \rightleftharpoons B conversion, so that only one combination is exactly conserved. Enforcing the old rank-2 law freezes an evolving total: at 48 steps the macro error is 5.41% vs 3.96% with the correct law, and the slow-total error is 1.85% vs 0.28%. The default SVD at τ≤10−3\tau \leq 10^{-3} picks the correct rank 1 in all five seeds. This failure mode is not specific to BNO, since a global projector given the false law would inherit the same bias.

Q5 : Can a conservation-aware baseline conserve only the identified combinations?

R: Yes. We built a selective NFV with the same C, learned exchange in kernel C, and parameters within 1% of BNO. Under the compact protocol BNO reduces macro error relative to it by 82–85% on SWE and 38–69% on RD, in 5/5 seeds. The fieldwise baselines fail for an intrinsic reason: any predictor that freezes all field totals, even one given the true future fields, has a mean macro error of at least 61.3% on SWE and 23.1% on RD, whereas BNO reaches 8.5% and 4.2%.

Q6: Longer 3D rollouts, and genuine 3D exchange?

R: We have updated the table in the prior comment. For longer CNS rollouts we used 40 fresh trajectories and scored 6 and 12 steps on the same trajectories. Identity-filter BNO beats GP-direct by 12.7% and 20.6% (5/5 seeds, intervals exclude zero); the original-filter BNO degrades at 12 steps, which we report. All models produce some negative density or pressure cells, so conservation does not imply positivity.

For genuine 3D exchange we added an A+B⇌CA+B\rightleftharpoons C system on 16³ with three regimes, and compared BNO with the simple global projection (GP-direct). BNO then has lower mean macro and field-total error in all regime/horizon cells, with the largest gain of 15.7% in the reaction-dominated regime at 48 steps. At this 5.3k-parameter scale the macro intervals include zero.

Q7: For the Appendix G pilot fallback, which data were used, and were the final test trajectories independent of it?

R: The fallback affects only two baseline cells on NSCH (CoDA-NO and the clawNO-style control), and both were reverted to their defaults. No BNO configuration used test data. We will document this precisely in Appendix G.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.