acceptodds
Under review as a conference paper at ICLR 2027

Learning Multiplicative-Time Operators for Long-Horizon PDE Prediction

Abstract

Long-horizon prediction of dissipative PDEs challenges learned fixed-step solvers: solutions change scale while rollout length grows with physical time. We formulate prediction as learning a multiplicative-time evolution operator. For equations with known scaling covariance, normalising space and amplitude makes a fixed ratio of physical times correspond to an autonomous transition of the full state. Our multiplicative-time renormalisation-group evolution method (MTRG) learns this transition from early-time trajectories and composes it with conserved-mass correction to reach distant horizons at logarithmic depth. The basic method fixes the ratio; a ratio-conditioned extension represents a range of intervals. We establish the exact composition law and derive a conditional error recurrence separating approximation, discretisation and propagation. Tests on diffusion and Burgers equations evaluate accuracy across distant horizons and the contributions of conservation and composition consistency. Comparisons with analytic profiles identify regimes where evolving the full state remains useful beyond the supervision window, before the limiting shape becomes the better predictor. MTRG turns known scaling structure into a reusable learned transition for this finite-time evolution.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural operators, long-horizon prediction, scaling covariance, self-similar solutions

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_yLAZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper learns long-horizon evolution for PDEs with known scaling covariance by expressing the full solution in self-similar coordinates. Multiplying physical time by a fixed ratio then corresponds to applying an autonomous field-to-field map. MTRG-Fixed learns one residual transition, while MTRG-Variable conditions on the ratio and adds composition consistency; both apply conserved-mass correction during evaluation. The paper establishes the exact composition law and a conditional rollout-error recurrence separating approximation, reference mismatch, and propagation.

Training uses times 1–10, followed by evaluation at 10^2–10^8 using four to sixteen applications. Six analytic-reference benchmarks cover Burgers, porous-medium, fast-diffusion, and heat equations. Seven-seed comparisons and component studies show useful performance on several 1D/2D tasks and benefits from mass correction. Results are weaker on fast diffusion and the small 3D pilot. Analytic limiting profiles outperform all learned methods at 10^8 on the five diffusion benchmarks where they are evaluated; learned full-state evolution helps at some earlier horizons, especially for 2D porous medium.

Critical strengths
  1. A principled and reusable evolution target. The normalized full-state map follows exactly from scaling covariance, without requiring the input already to be self-similar. This gives a coherent way to learn finite-ratio transitions and reuse them at increasingly distant physical times. The distinction between exact operator structure and conditional learned accuracy is technically sound and consequential to the method's credibility.

  2. Informative and candid evaluation. Seven-seed results, complete horizon tables, mass/consistency ablations, and analytic-profile controls show both benefits and limitations. In particular, the authors explicitly report regimes where the attractor or another learned mechanism wins. This makes the positive 1D/2D findings more interpretable and avoids equating a large physical horizon with a uniformly difficult prediction problem.

Critical weaknesses
  1. The central transient-prediction advantage is insufficiently distinguished from low-dimensional profile tracking. Appendix C uses shifted Barenblatt profiles and Gaussian mixtures, with incompletely specified shifts and parameter distributions. Within a shifted Barenblatt family, the future is determined by mass, center, and effective age, whose normalized updates are known analytically. Table 11's 2D heat and porous-medium control errors decay approximately as T^(-1/2) and T^(-1/4), consistent with translation modes. This is a diagnostic observation, not proof of an implementation error. Matched-mass, centroid-aware, and age/width-fitted profile controls are needed, particularly for the sustained 2D porous-medium advantage. Without them or non-Barenblatt nonlinear tests, the evidence for a consequential full-state learning contribution is limited.

  2. The added computational value over established rescaling methods is not demonstrated. Classical dynamic rescaling already supplies the normalized clock; numerical RG likewise advances and rescales full fields. A relevant missing comparator is Braga et al. (2020), Numerical Renormalization Group Algorithms for Self-Similar Asymptotics of Partial Differential Equations. The physical-time timing comparison combines different rollout lengths with 32,768 versus 256 spatial points, as the authors disclose. The normalized learned comparisons evaluate adapted mechanisms using a common U-Net, rather than the complete source methods. A conventional solver in the same normalized coordinates is needed to establish practical accuracy/cost benefits from learning. This matters because the mathematical novelty is modest and the strongest contribution must come from empirical utility.

Questions for authors
  1. What are the exact initial-data and analytic-control formulas, including whether the Barenblatt shifts are temporal, spatial, or both, and how the 2D porous-medium center is handled? With identical exact-mass information, how much of the reported advantage remains against centroid-aware and fitted effective-age/width profile controls? Please also explain the absence of a Burgers diffusion-wave control.

  2. Do the gains persist for nonlinear diffusion initialized away from the Barenblatt family, preferably with mass and centroid controlled? A targeted perturbation test would help establish that the method predicts transient shape changes beyond a few similarity parameters.

  3. How does MTRG compare in error versus wall-clock cost with a conventional solver in self-similar coordinates or numerical RG on comparable spatial domains and resolutions? What training/setup cost must be amortized? This comparison would separate the benefit of learning from the coordinate and clock transformation.

  4. Can the authors provide the promised reporting archive and the data-generation, reference, and training configurations needed to reproduce the experiments? In particular, how large are quadrature/truncation and exact-mass-versus-grid-reference discrepancies, and do the principal control comparisons persist under spatial refinement? These details matter for interpreting the small reported errors and the special-family controls.

3 comments
AuthorsOpenReview ↗

W1 / Q2 — full-state prediction versus profile tracking & submitted controls, centre and age; formulas.

We thank the reviewer for a precise diagnosis. We ran the two tests it asks for: nonlinear diffusion away from the Barenblatt family with matched information, and a conventional solver in the same coordinates. They are supplementary experiments (separately implemented pipeline; five training seeds unless noted); they neither replace nor reproduce the submitted tables. They support a narrower empirical claim than the submission's, and the revised draft states it.

W1 / Q2 — full-state prediction versus profile tracking. New experiment (supplementary; Appendix D.2): 2D porous medium with initial data that are sums of 2–4 anisotropic compact bumps; 1,024 training trajectories supervised only at t = 1 → √10 → 10; 64 independent test trajectories; five training seeds. Every method receives the same 128² initial cell averages, whose cell-area-weighted sum is the exact mass. We compare with six initial-only fitted-profile controls (Table S12). The strongest at every horizon is the elliptical quadratic-pressure solution family fitted to mass, centroid and covariance: a known-form solution of the equation, though not the exact evolution of a multi-bump field. Mean relative L² error:

T√10 (in window)10 (in window)10√1010²10³10⁸
MTRG-Fixed1.44 ± 0.18 %0.88 ± 0.12 %0.71 ± 0.12 %0.65 ± 0.11 %0.65 ± 0.10 %0.74 ± 0.14 %
strongest fitted profile10.2 %3.9 %1.6 %0.64 %0.11 %≈ 0.01 %
finite-volume solver, same grid0.21 %0.20 %0.19 %0.19 %0.20 %0.24 %
test cases where MTRG beats the profile95 %91 %75 %34 %0 %0 %

At the two supervised horizons and at T = 10√10, the first tested time beyond the window, MTRG is 7×, 4.5× and 2.2× more accurate than the strongest control, in all five seeds. At T = 10² the difference is unresolved. From T ≈ 3·10² the profile is more accurate and remains so, and the same-grid solver is 3–7× more accurate than MTRG at every tested horizon (all 16 horizons: Fig. S1, Table S5). The supported claim is finite-time predictive value beyond the six tested controls, not a statement about every low-dimensional representation and not sustained late-time superiority. Out of distribution (Table S6), the advantage holds for 5–6-bump data through T = 10²; for flat-topped data as a whole it is established only at √10.

W1 / Q1 — submitted controls, centre and age; formulas. We agree that the submitted comparison (Table 11) does not isolate full-state utility against centre-, age- or covariance-aware controls. In the new experiment the mean-error crossover location depends on the control: about 3·10⁷ for a centred limiting profile that does not track the centroid, 3·10³–10⁴ for a centroid- and age-matched isotropic profile, and 10²–3·10² for the elliptical family. These are where the mean errors cross; the change of ranking is resolved under our rule only from 10⁸, 10⁴ and 3·10² respectively. Untracked translation can thus produce a prolonged apparent gain of the kind the reviewer inferred. We do not claim that this explains the submitted Table 11 numbers, and we cannot add detail on the Barenblatt shifts or the 2D centre handling of the submitted runs beyond what the manuscript states. Within a correctly specified analytic family a fitted profile describes its own trajectories without full-state learning, so Barenblatt-family benchmarks cannot separate the two. The revised draft says so, removes the sentence about a sustained interval of advantage, and adds the horizon-limited comparison with fitted controls (Section 5.2, Table S5). Initial-data formulas, parameter distributions, centre and age definitions and fitting objectives of the new study are in Appendix D.2 and Table S12.

AuthorsOpenReview ↗

W2 / Q3 — value over rescaling methods and solvers.

W2 / Q3 — value over rescaling methods and solvers. The coordinates, logarithmic clock and renormalization step are classical; the revised draft credits dynamic rescaling and numerical RG (Braga et al., 2020). We did not implement that algorithm and claim no advantage over it, nor over the complete source methods behind our adapted learned comparators. Against a conventional solver in the same coordinates from the same observations (Table S3), we find a task-dependent online accuracy–latency trade-off, not a general saving:

  • 2D porous medium: to T = 10² at batch 64, MTRG takes 0.054 s; a 128² finite-volume solve takes 2.6 s and is 3.4× more accurate; a 64² solve takes 0.74 s and is 2.1× less accurate.
  • 1D Burgers, same 256 points: a second-order solver reaches 0.13–0.28 % against 1.1–1.4 % for MTRG from T = 10²; we claim no speed advantage.
  • Linear heat (separate study, one trained model, 16-step rollout to 10⁸): a conservative solver with mass projection takes 7–8 ms per batch against 54–91 ms for MTRG, with a largest horizon-mean error of 0.09 % against 0.52 %. The solver is cheaper and more accurate.

Timings are recorded for research implementations on one GPU type (FP32 network, FP64 solver), with no optimized solver. Offline cost counted: 4.3 GPU-hours of training references and 90 s of training for one model; validation and test references, pilot runs and development time are not counted. As conditional arithmetic on implementations of unequal accuracy, this equals the measured online difference after ≈ 3.9·10⁵ porous-medium trajectories to T = 10² at batch 64. The revised draft states that the timings do not establish a general computational advantage. It adds one motivation, stated as untested: inference never evaluates the differential operator, so training could in principle use observed trajectories when the scaling law, state, time labels and invariant are available.

AuthorsOpenReview ↗

Q1 — Burgers diffusion-wave control & Q4 — numerical reliability and reporting materials.

Q1 — Burgers diffusion-wave control. The submitted comparison omitted it; we added limiting, least-squares-fitted and moment-matched diffusion waves (Table S7). At T = 10² MTRG has the lower cohort mean (1.2 % against 20.8 %, 9.35 % and 16.4 %), but the moment-matched wave is more accurate on 86 % of trajectories, and all three waves overtake MTRG from T = 10⁶. The Cole–Hopf formula agrees with the reference to rounding on these data, so this benchmark shows no advantage over an analytic solution.

Q4 — numerical reliability and reporting materials. For the new porous-medium references (Table S11): N = 512 differs from N = 1024 by ≤ 2.4·10⁻⁴ for T ≤ 10², and from the exact elliptical solution by 1–2·10⁻⁴ on average through T = 10⁸; halving the time step changed one pilot family by ≈ 10⁻¹⁰ (T ≤ 10²). These are measured discrepancies, not error bounds. Our decision threshold is three times the larger grid-refinement discrepancy (raw discrepancy ≤ 8.8·10⁻⁴, so 0.14–0.26 percentage points); this applies to the porous-medium comparisons reported above as resolved; lower means and crossover locations are descriptive. The mass label equals the cell-area-weighted sum of the observation to rounding, and no mass reaches the outer ring of the computational box; we did not vary the box size, and these checks do not extend to the submitted benchmarks. We have prepared a reporting archive for the new studies: saved per-seed and per-trajectory errors, recorded timings, configurations, recorded summaries, and scripts that regenerate the supplementary tables and the figure from these metrics and explicitly identified recorded summaries. It does not retrain models, regenerate trajectories, remeasure timings or re-execute field-level checks. We are not able to provide the seed-level archive for the submitted tables that the original reproducibility statement describes; the revised statement describes only what is available.

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

The paper proposes a method for long-horizon prediction of dissipative PDEs by rescaling space and amplitude according to known scaling laws and reparameterizing time logarithmically. This allows a learned operator to advance the solution through geometrically increasing times. The authors introduce two variants: MTRG-Fixed, which learns a fixed-ratio transition with conserved-mass correction, and MTRG-Variable, which conditions on the time ratio and adds a composition-consistency loss. They evaluate the methods on heat, nonlinear diffusion, and viscous Burgers equations, comparing predictions beyond the training window against learned baselines and known asymptotic profiles.

Critical strengths
  1. A useful extension of known rescaling methods: The method learns how the full solution evolves toward its limiting shape, improving on predictions based only on that shape in some experiments. Its application to several equations suggests a reusable framework for problems with known scaling laws.

  2. Demonstrated savings in long-term prediction for the solution of the viscous Burgers equation at (T=10^4). The method uses eight network calls rather than 1,250, while achieving lower prediction error, demonstrating the computational benefit of using the equation’s scaling structure.

Critical weaknesses

The broader applicability remains insufficiently demonstrated. The framework requires known scaling laws, and the experiments focus on diffusion and Burgers equations using analytic solution families. Although these include nonlinear problems, they provide limited evidence for more general dynamics or initial conditions. It remains unclear how the approach would accommodate boundaries, forcing, or approximate rather than exact scaling. A clearer account of its applicability, supported by a less restricted benchmark, would help establish the contribution’s broader significance.

Questions for authors
  1. Does MTRG retain its advantage on initial conditions beyond the selected analytic families, when all methods receive the same mass information?

  2. Could you clarify how the framework would be applied to equations with scaling laws different from the tested examples? When the scaling exponents are unknown or the symmetry holds only approximately, which theoretical guarantees would remain valid?

3 comments
AuthorsOpenReview ↗

Q1 — initial conditions beyond the analytic families, with the same mass information.

We thank the reviewer. Both questions are answered below, the first with a new experiment. The new results are supplementary experiments (separately implemented pipeline; five training seeds unless noted); they neither replace nor reproduce the submitted tables.

Q1 — initial conditions beyond the analytic families, with the same mass information. New experiment (supplementary; Appendix D.2): 2D porous medium with initial data that are sums of 2–4 compactly supported anisotropic bumps, outside the Barenblatt family; references from a finite-volume solver checked against an exact non-radial solution; supervision only on t ∈ [1, 10]; five training seeds. Every method receives the same initial cell averages, whose cell-area-weighted sum is the mass. The comparison is with the strongest of six initial-only fitted profiles, an elliptical exact-solution family fitted to mass, centroid and covariance.

  • Independent test trajectories from the training distribution (64; Fig. S1, Table S5). MTRG is 7×, 4.5× and 2.2× more accurate at T = √10, 10 and 10√10 (1.44 %, 0.88 %, 0.71 % against 10.2 %, 3.9 %, 1.6 %), in all five seeds; the first two horizons are supervised, the third is not. The difference is unresolved at T = 10², and the fitted profile is more accurate from T ≈ 3·10².
  • Out-of-distribution, 5–6 bumps (32; Table S6). MTRG is more accurate on all cases through 10√10 and on 78 % at 10².
  • Out-of-distribution, flat-topped data (32). As a whole the advantage is established only at √10. Split after seeing the results into one- and two-component cases, MTRG is ahead for two components at the first three horizons and behind for a single ellipse at every horizon; this split is descriptive, not a confirmed result.
  • Transfer. A model trained only on Barenblatt data does not transfer to bump data (21 ± 8 % at √10). This is a limit of that trained model's coverage.

So with equal mass information the advantage extends beyond the analytic families, at finite times. At long horizons a fitted profile is the better predictor, and a conventional solver at the same resolution is more accurate than MTRG throughout (Table S3). Sign-changing data and other exponents or dimensions were not tested in this new study.

AuthorsOpenReview ↗

Q2 — other scaling laws, unknown exponents, approximate symmetry.

Q2 — other scaling laws, unknown exponents, approximate symmetry.

  • Different known exponents. The construction does not depend on the particular exponents of our examples. For an equation with exact scaling covariance and a conserved invariant, the correctly normalized dynamics are autonomous and the composition law holds. Learned accuracy remains conditional on approximation, coverage and error propagation, as in the manuscript's recurrence.
  • Incorrect exponent. Normalizing with a relative exponent error δ loses autonomy: the true profile drifts in the mis-normalized frame, by about 1.6|δ| (heat) and 1.15|δ| (porous medium) in relative L² per decade against a fixed profile. This is a kinematic statement, not a bound on network error. In our study, with a normalization that keeps mass, re-trained models (three seeds per exponent error) show that a 1 % exponent error raises the T = 10⁸ error from 0.8 % to 10–12 % (heat) and from 1.4 % to 7.8 % (porous medium) (Table S9). The number of calls is independent of the exponents: for a fixed ratio μ and a target on the geometric landing grid it is log_μ(T/t₀).
  • Unknown exponent. For exactly self-similar, noise-free data, fitting V(t) = C(t + t₀)^{2β} to in-window snapshots of the second moment recovers β to within 2·10⁻⁴, whereas a naive log–log slope is off by about 42 %. With 0.1 % additive noise our simple estimator fails for heat (+15 %, with 389 of 1,024 trajectories giving no valid fit) (Table S10). This is a feasibility check, not an identification pipeline, and no model was trained with a noisy estimate.
  • Approximate symmetry. We have neither a bound nor an experiment. None of the guarantees above is established for this case; that is a limit of the present evidence.

Boundaries and forcing. The theory is for whole-space, unforced problems, and the computations use documented finite computational domains. A fixed physical boundary moves in similarity coordinates, and forcing can break covariance or change the mass, so neither the autonomous normalized evolution nor the constant-mass projection follows automatically. Scale-compatible boundaries or forcing might permit a related construction, but each needs its own covariance and invariant argument, and none is validated here. The revised draft adds a limitations paragraph with these points (Section 6).

Thank you for your response. Although Figure S1 shows that the strongest fitted profile wins for T above approximately 300, as you acknowledge, the experiment does not isolate full-state utility, and your solver timings demonstrate only a task-specific trade-off for the 2D porous medium problem. The additional experiments however, demonstrate broader applicability and highlight the value of the framework. Threrefore, I will keep my score unchaneged.

Reviewer_3eH4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Multiplicative-Time Renormalisation-Group Evolution (MTRG), a framework for long-horizon prediction of dissipative partial differential equations exhibiting known scaling covariance. By transforming the physical solution into self-similar coordinates, the authors construct an autonomous evolution operator that maps a normalised state at time (t) to its state at time (\mu t). Repeated application of this operator reduces the number of learned transitions from linear to logarithmic dependence on the physical prediction horizon.

Critical strengths

The most compelling contribution is the formulation of long-horizon PDE prediction as learning an autonomous multiplicative-time evolution operator. By exploiting known scaling covariance, the framework explicitly connects self-similar PDE dynamics with reusable learned state transitions. This provides a principled alternative to repeatedly applying a fixed physical-time update and reduces composition depth to (O(\log T)). Besides, the method evolves the full normalised state rather than directly replacing the solution with its limiting self-similar profile. This distinction is relevant for resolving finite-time transients, where asymptotic approximations may still be inaccurate. This strength is consequential because it identifies a precise mathematical structure that can inform neural evolution operator design beyond a particular network architecture.

Critical weaknesses

Insufficient isolation of methodological novelty. The main ingredients of MTRG, including self-similar normalisation, logarithmic-time evolution, semigroup composition, and conservation correction, are closely related to established techniques. Although the proposed integration is mathematically motivated, the paper does not sufficiently isolate its independent contribution relative to existing self-similar and semigroup-based operator-learning approaches. Potentially confounded baseline comparisons. The physical-time comparison uses substantially different spatial resolutions (32,768 versus 256 grid points) and temporal stepping strategies, making it difficult to attribute the observed accuracy and efficiency improvements specifically to the proposed operator-learning formulation. Limited benchmark diversity and generalisation evidence. The experiments mainly consider analytically tractable, scaling-covariant dissipative PDEs with relatively restricted initial-condition families, including Gaussian mixtures and Barenblatt profiles. May consider more complex PDE benchmark.

Questions for authors

refer to weakness

3 comments
AuthorsOpenReview ↗

W1 — isolating the contribution.

We thank the reviewer. Each weakness is answered below with a targeted additional experiment where one was needed; all are reported in the revised draft (Appendix D). They are supplementary experiments (separately implemented pipeline; five training seeds unless noted); they neither replace nor reproduce the submitted tables.

W1 — isolating the contribution. Self-similar normalization, the logarithmic clock, semigroup composition and conservation corrections are established, and the revised draft credits them. MTRG's contribution is both methodological and empirical. Methodologically, it is a learning formulation built on known scaling covariance: MTRG-Fixed learns one fixed-ratio transition of the full normalized state from an early window and reuses it at every later scale, and MTRG-Variable conditions on the ratio with composition-consistency training. Empirically, the new experiments (all on Fixed) show what this buys and where it stops:

  • Representation at a fixed ratio (Table S1). With the ratio fixed at √10 and the same training trajectories, backbone and number of updates, we compared a physical-state model with a time input, a normalized-state model with a time input, and MTRG-Fixed (normalized, no time input), five seeds each, on 64 fresh held-out non-Barenblatt porous-medium cases. All three fit the supervised window (0.59 %, 0.85 % and 0.85 % at T = 10; differences unresolved under the rule fixed in advance, which is not equivalence). Beyond it, at T = 10√10 and 10², the errors are 28 ± 6 % and 65 ± 22 %; 2.8 ± 2.9 % and 7.2 ± 7.4 %; and 0.69 % and 0.64 %. The physical-state model differs from both normalized models under that rule. The time-input variant is worse than MTRG-Fixed in all five seeds, but that difference is unresolved at these two horizons and resolved from T = 10^3.5. This supports the implemented scaling-aware representation as a package under early-window supervision. It does not test normalization alone or the clock, and the physical baseline is untuned, uses a 416² grid against 128² with about 5.4× the compute per update, and, like the time-input variant, saw only two source times.
  • Full-state transition versus fitted profiles (Fig. S1, Table S5). On held-out non-Barenblatt data with identical initial information, MTRG-Fixed is 7×, 4.5× and 2.2× more accurate than the strongest of six fitted-profile controls at T = √10, 10 (both supervised) and 10√10; the difference is unresolved at 10² and the profile is better afterwards. This is finite-time utility beyond the tested controls, not superiority over every low-dimensional or neural alternative.
  • Mass projection (Table S2). Adding the projection at inference to unchanged projection-free weights lowers the T = 10⁸ error by 31–75 % (heat) and 25–47 % (non-Barenblatt porous medium, three models); four analytic porous-medium models improve by 29–56 % and a fifth, already diverging, gets worse. On held-out non-Barenblatt data this effect is unresolved through T = 10², so it does not explain the early advantage over profiles. Models trained with the projection remain 2–18× more accurate under the same inference rule, a difference that includes training effects; they fail without it, because their raw output amplitude is left to the projection (raw output mass 5–1,700× the label). The intervention identifies the effect of the implemented recurrent projection; it does not separate a pure mass correction from rescaling of the field, and we do not infer from it how the submitted models behave.
  • Semigroup variant (Table S8). Our Semigroup baseline adapts the Deep-OSG mechanism to our backbone and clock; it is not that complete system. Over five seeds there is no consistent ordering on heat, where the submission gives Fixed the lower mean; on analytic porous medium Fixed has the lower mean but the ranking is unresolved and training workloads differ. We claim no ranking.
AuthorsOpenReview ↗

W2 — confounded physical-time comparison.

We agree. Tables 1–2 compare complete systems with 32,768 versus 256 grid points and different stepping. The revised draft keeps their numbers and changes their interpretation: they do not isolate the clock and do not establish general computational superiority. We do not have a clean clock-only ablation. The fixed-ratio comparison above uses the same multiplier in every arm and different grids, so it is neither grid-controlled nor additive-versus-multiplicative. A same-grid additive-clock network on 1D Burgers (256 points, three seeds) is far less accurate from T = 10², but it is an untuned baseline, already worse inside its training window, and it extrapolates its time input (Table S4). What we can report is a task-specific trade-off (Table S3): on 2D porous medium the learned rollout is 42–48× faster at batch 64 than a same-resolution finite-volume solver that is 3.1–3.4× more accurate, in research implementations on one GPU type. On 1D Burgers and linear heat a conventional solver is more accurate, and we claim no speed advantage there, nor any general advantage over classical solvers, numerical RG or complete physical-time learning systems.

AuthorsOpenReview ↗

W3 — benchmark diversity and generalization.

We added a numerically generated, strongly non-Barenblatt porous-medium benchmark: a 2–4-bump training family, 64 independent test cases and two test-only out-of-distribution cohorts (Fig. S1, Tables S5–S6). The finite-time advantage over fitted profiles carries over to 5–6-bump data through T = 10². For flat-topped data as a whole it is established only at √10, and a model trained only on Barenblatt data does not transfer to bump data. This broadens the initial conditions for one equation; it does not add a PDE class. The method assumes an exactly scaling-covariant equation with a known invariant and accurate exponents (a 1 % exponent error raises the T = 10⁸ error from about 1 % to 8–12 %, Table S9). The revised draft states this scope and adds a limitations paragraph.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.