acceptodds
Under review as a conference paper at ICLR 2027

SCOPE: Observation-Conditioned Full-Target Prediction for Sparse PDE Inference

Abstract

Recovering complete physical fields from sparse observations is challenging because the measurements may not uniquely determine the underlying state. Diffusion-based PDE solvers address this problem through iterative sampling whereas neural operators provide deterministic one-pass predictions. We propose SCOPE (Sparse-Context Observability-aware Predictive Embeddings) to recover complete PDE fields from sparse observations by coupling full-field latent prediction with physical reconstruction. A shared decoder reconstructs fields from both predicted and complete-view representations so that representation learning is guided by both physical recovery and latent matching. We derive a quadratic risk decomposition at fixed teacher-decoder pairs showing why optimal latent prediction need not yield optimal field reconstruction. We also establish sufficient conditions for decoder improvements on complete inputs to transfer to recovery from partial observations. Experiments across five PDE settings show that SCOPE outperforms mask-aware neural operators on all ten forward and inverse tasks and achieves lower errors than those reported for diffusion-based solvers including DiffusionPDE and FunDPS. Decoder-only adaptation further improves recovery without retraining the backbone while retaining deterministic single-pass inference.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Scientific machine learning, partial differential equations, sparse observations, joint-embedding prediction, representation learning, field reconstruction

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_U4QCOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper trains an AE to do forward/inverse solving. In training AE, the latent is required to match teacher.

Critical strengths

This paper is easy to follow. The demonstration is clear. The theorems are intuitive and easy to understand.

Critical weaknesses

The theorem part is trivial. The main contribution of this paper shoud lie in the empirical part. However, the experiements are not convincing. See questions.

Questions for authors
  1. FO (without teacher) is also a mask deterministic AE. Why it already largely surpasses all mask AE baselines? It there any techniques inside mask AE? Why exactly the same formulation differs so much in performance?
  2. The training budgets and model capacity of SCOPE is much larger than the ones of baselines, this makes comparision not fair.
  3. In the ablation, it seems that teacher matching (Table. 7) brings suble improvement, and it is even worse than AE in certain cases. Does it mean teacher matching is not very useful?
7 comments
AuthorsOpenReview ↗

Weakness 1

The theorem part is trivial. The main contribution of this paper shoud lie in the empirical part. However, the experiements are not convincing. See questions.

Response:

We sincerely thank you for this candid and helpful assessment. We appreciate your emphasis on the empirical evidence. In response, we have completed the approximately 40M comparison to strengthen the empirical evaluation. Below, we summarize that evidence and clarify the explanatory role of the analysis alongside the requested experiments.

[D1] Expanded empirical evidence.

The expanded evaluation includes FNO40M, Transolver40M, CNO40M and DeepONet40M on all five PDEs in both directions. Under the evaluated configurations, we observe lower error for Full-native than for FNO, Transolver, CNO and DeepONet on all ten endpoints.

For ease of reference, we provide the complete table and computational-resource interpretation in our response to Q2. The pretrained FGV and native-size decoder refits provide separate controls for the component-attribution questions raised in the reviews.

[D2] What the theory is intended to establish.

We would like to clarify the intended role of the fixed-reference squared-risk analysis. It distinguishes observation uncertainty, decoder incompatibility and predictor excess risk to explain why optimal latent prediction need not yield optimal physical reconstruction. Consistent with this, full-target prediction without main-stage grounding (FJV) has higher error than Full on all ten endpoints of Table 7, for example 11.6476% versus 8.7023% on Helmholtz inverse.

The weighting analysis examines the target metric under its stated reference assumptions. In the corresponding experiment, uniform weighting gives 9.5024% Poisson inverse error and a mask-dependent equal-group weighting gives 10.7760% (our response to Reviewer VrNy, V-W2).

The readout-transfer result gives sufficient conditions connecting complete-view and sparse-view decoder behavior. In Table 10, a decoder fitted on complete-view representations (DP-10M) lowers the NS (BCs) inverse error from 0.6491% to 0.4576% but not the Darcy inverse BER (1.5564% to 1.6468%), whereas a decoder fitted on the sparse path (SP-10M) lowers both, to 0.4524% and 1.5503%.

The contribution of these results lies in their relevance to the design, and the observed benefits of the full training procedure are assessed through the empirical controls.

Manuscript changes. Guided by your feedback, the revised manuscript foregrounds the full approximately 40M comparison (Table 1 and Section 4.2, page 8, lines 414–416). Section 3.3 now opens by stating the three questions examined by the analysis and the experiment that addresses each of them (page 6, lines 301–303), which connects each empirical conclusion to its corresponding control and makes explicit that the measured performance is supported by experiments.

AuthorsOpenReview ↗

Question 1

FO (without teacher) is also a mask deterministic AE. Why it already largely surpasses all mask AE baselines? It there any techniques inside mask AE? Why exactly the same formulation differs so much in performance?

Response:

Thank you for this perceptive question about the teacher-free model. We appreciate the need to explain why models with a common reconstruction formulation can perform differently. The original models share the masked-reconstruction task, while their architectures, capacities and training configurations differ. We clarify those differences below.

[D1] What FO contains.

For clarity, FO uses the SCOPE inference architecture (Appendix E.5): a patch-token Transformer encoder that receives the observed values together with their masks, a mask conditioner that encodes how much of each patch is observed, a latent predictor and the native decoder, totaling 40,040,834 parameters. We train it from random initialization using only the physical field-reconstruction loss, without complete-view pretraining or a teacher.

We use the new approximately 40M comparison to discuss external performance at a more closely aligned scale.

[D2] Evidence from the expanded comparison.

To help contextualize FO's performance, we compare it with the approximately 40M FNO and Transolver baselines. Their forward/inverse results are shown below, with the FO values taken from Table 7:

PDEFOFNO40MTransolver40M
Poisson1.832/11.2463.205/15.0972.605/12.718
Helmholtz1.831/10.7073.228/14.8172.244/12.698
Darcy (inverse: BER)2.163/1.5832.626/3.5942.444/2.641
NS2.627/7.3293.975/11.1626.072/13.771
NS (BCs)1.983/0.7082.317/0.7362.615/1.156

We report the values as percentages. FO uses no technique beyond the architecture and field supervision described in [D1].

To make the interpretation clearer, the revised manuscript reports the networks, inputs and training budgets of the compared models (Tables 2, 18, 20 and 21).

Manuscript changes. Following your helpful question, the revised manuscript states FO's architecture, initialization and loss in the footnote of Table 1 (page 8, lines 410–412) and in the “Field-only comparison” paragraph of Section 4.3 (page 9, lines 456–457), and lists FO as a same-architecture recipe reference beside the approximately 40M baselines in Table 22 (Appendix K.1, page 36). The approximately 40M results serve as the external comparison, and latent matching is discussed separately through Full–FGV (Section 4.3, page 9, lines 460–463, and Appendix K.2).

AuthorsOpenReview ↗

Question 2

The training budgets and model capacity of SCOPE is much larger than ...

Response:

We sincerely appreciate your concern about the fairness of the capacity and resource comparison. In response, we have expanded the comparison to four approximately 40M baselines—FNO, Transolver, CNO and DeepONet—evaluated against SCOPE Full-native across all five PDEs and both recovery directions. We summarize these results and the resource accounting below.

[D1] Completed approximately 40M comparison.

To keep the evaluation protocol consistent, all methods use the same test records, normalization and observation masks. Each entry reports forward/inverse error in percent, with BER for Darcy inverse and the whole-grid metric for NS (BCs).

ModelParameters (M)PoissonHelmholtzDarcyNSNS (BCs)
SCOPE Full-native40.041.680/9.5021.636/8.7022.174/1.5562.581/6.9732.115/0.649
FNO40M39.403.205/15.0973.228/14.8172.626/3.5943.975/11.1622.317/0.736
Transolver40M40.192.605/12.7182.244/12.6982.444/2.6416.072/13.7712.615/1.156
CNO40M39.9211.314/24.20113.810/25.25315.084/2.95512.104/17.25414.616/7.385
DeepONet40M40.0510.375/27.7659.949/35.2805.597/8.01221.121/32.43726.614/22.046

[D2] Conclusions supported by the expanded comparison.

Under the evaluated training configurations, we observe lower error for Full-native than for FNO40M, Transolver40M, CNO40M and DeepONet40M on all ten forward/inverse endpoints.

For the inverse tasks on Poisson, Helmholtz and NS, we obtain 9.502%, 8.702% and 6.973%, respectively, compared with 12.718%, 12.698% and 11.162% for the strongest approximately 40M baseline on the corresponding endpoints.

This ordering is stable. The table below reports the K=500 errors as mean ± standard deviation over five test-mask seeds:

EndpointFullFNO40MTransolver40MCNO40MDeepONet40M
Poisson F1.648 ± 0.0193.137 ± 0.0472.585 ± 0.02011.312 ± 0.03110.368 ± 0.008
Poisson I9.510 ± 0.01015.102 ± 0.01112.724 ± 0.01124.215 ± 0.01427.805 ± 0.032
Helmholtz F1.639 ± 0.0103.229 ± 0.0122.244 ± 0.00913.822 ± 0.0099.954 ± 0.007
Helmholtz I8.706 ± 0.00214.810 ± 0.00612.695 ± 0.00525.253 ± 0.00135.275 ± 0.004
Darcy F2.172 ± 0.0052.614 ± 0.0092.441 ± 0.00615.082 ± 0.0025.593 ± 0.002
Darcy I (BER)1.558 ± 0.0013.594 ± 0.0012.643 ± 0.0022.955 ± 0.0018.014 ± 0.002
NS F2.576 ± 0.0083.960 ± 0.0116.065 ± 0.01312.106 ± 0.01321.110 ± 0.009
NS I6.979 ± 0.00411.171 ± 0.00813.784 ± 0.01717.264 ± 0.00832.434 ± 0.002
NS (BCs) F2.094 ± 0.0182.290 ± 0.0152.656 ± 0.03314.607 ± 0.03326.788 ± 0.119
NS (BCs) I0.649 ± 0.0000.740 ± 0.0051.151 ± 0.0037.351 ± 0.02522.052 ± 0.010

Full has the lowest mean error on all ten endpoints, and its standard deviation across mask seeds is at most 0.019. In addition, the paired per-record bootstrap intervals of all 40 Full-minus-baseline differences lie below zero (table in our response to Reviewer MiBT, M-W2 / M-Q2).

To keep the comparison balanced, we retain all task-specific results in the complete table.

[D3] Computational-resource accounting.

Following your emphasis on computational resources, we also compiled per-method training-resource records covering main-training GPU-hours and hardware. We report parameter-scale alignment and resource accounting separately, so the accuracy measurements can be interpreted under their actual training configurations.

Full's main training takes 45.0, 41.7, 36.3, 28.6 and 12.7 GPU-hours on Poisson, Helmholtz, Darcy, NS and NS (BCs). Transolver40M is trained for 41.4, 34.2, 43.6, 41.6 and 22.1 GPU-hours, CNO40M for 27.8, 44.9, 58.1, 52.1 and 5.7 GPU-hours, and FNO40M for 10.3, 10.1, 10.1, 10.4 and 2.8 GPU-hours. On Darcy and NS, Transolver40M and CNO40M are trained longer than Full, and Full has lower error on all four endpoints of these PDEs. The training hours and training-loss trajectories of Full and Transolver40M are given in our response to Reviewer MiBT (M-W2 / M-Q2). The training configurations and the main-training GPU-hours of every method are given in Tables 20 and 21 of the revised manuscript (Appendix K.1, page 35).

Manuscript changes. In response to your feedback, the revised manuscript makes the complete approximately 40M comparison central to the evaluation (Table 1 and Section 4.2, page 8, lines 414–416) and provides its training configurations and resource accounting in a new Appendix K.1 (page 35, Tables 20 and 21). The ten-endpoint comparison with the four baseline families is reported with its paired intervals and test-mask seeds in Tables 22–24 (page 36), with the evaluation conditions made clear.

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

This paper proposes a latent matching method for solving sparsely observed PDE inverse problems. Its loss has 4 parts: physics-space matching, latent-space matching, autoencoder loss, and regularization. It works very well in single reconstruction.

Critical strengths
  • Novel architectural design and bringing JEPA's latent matching idea to PDE representations.
  • Theory proves that equal-weighted field reconstruction is better than mask-dependent weighting, which is interesting.
  • SotA performance on five PDE reconstruction settings with only one evaluation.
Critical weaknesses
  • The introduction is hard to understand. For example, on line 47, "latent" is first introduced without any context. It's not clear what latent refers to on the first page.
  • More ablation studies can greatly help the soundness of the paper. (i) How's Full vs F+G+V? aka drop latent matching loss and see how much gain is from latent matching. (ii) The paper's theory shows the advantage of equal full-field weighting. There should be a study comparing uniform weighting vs mask-dependent weighting. (ii) The native-to-SP comparison has a flaw: SP replaces the 0.17M native decoder with a 5M decoder and requires more training time. Thus the gain is unclear from the unfair comparison.
  • Lots of missing values in the main Table 1.
Questions for authors
  • For Figure 3, why is the T-bar surprisingly of equal length across different cases? How's the T-bar calculated?
  • What happens if the number of observed points changes? Can you plot error vs # of observations? In an extremely sparse case, the solution may not be unique. Could uncertainty be handled with SCOPE?
  • Does this method generalize to unseen distribution? e.g. it is trained with coeff from a certain GRF family and tested on another set of GRF hyperparameters.
11 comments
AuthorsOpenReview ↗

Weakness 1

The introduction is hard to understand. For example, on line 47, "latent" is first introduced without any context. It's not clear what latent refers to on the first page.

Response:

We sincerely thank you for pointing to the first use of this term. The revised introduction now defines it at its first use (page 1, lines 46–49). Here, “latent” denotes a learned internal feature representation of the paired physical fields; it is not an additional physical variable. We explain the roles of the representations below.

[D1] The representations used by SCOPE.

SymbolMeaningRole
z_MRepresentation predicted from sparse measurements and their observation maskDecoded at inference
z̄Stop-gradient complete-field target produced by the EMA encoderLatent supervision during training
z_FOnline encoder's complete-view representationGrounding and variance regularization during training

To clarify the information available to each pathway, complete paired fields are available during training. At inference, we use only the sparse-context pathway, without the teacher or complete-view inputs. Grounding trains the online complete-view representation and shared decoder; the EMA target remains stop-gradient and is not trained directly through this loss.

[D2] First-use explanation in the revised introduction.

“A latent representation is a learned internal encoding of the paired fields. SCOPE predicts a complete-field teacher representation from sparse measurements and masks, while a shared decoder connects latent prediction to physical reconstruction.”

Manuscript changes. Following your helpful comment, the revised manuscript inserts this explanation before the location you identified (Introduction, page 1, lines 46–49), states the roles of z_M, z̄ and z_F in Section 3.1 (page 5, lines 245–247), and marks the training-only quantities in the caption of Figure 2 (page 5, lines 240–241). These edits make the existing information access clearer while preserving the method itself.

AuthorsOpenReview ↗

Weakness 2

More ablation studies can greatly help the soundness of the paper. (i) How's Full vs F+G+V? aka drop latent matching loss and see how much gain is from latent matching. (ii) The paper's theory shows the advantage of equal full-field weighting. There should be a study comparing uniform weighting vs mask-dependent weighting. (ii) The native-to-SP comparison has a flaw: SP replaces the 0.17M native decoder with a 5M decoder and requires more training time. Thus the gain is unclear from the unfair comparison.

Response:

We are very grateful for these thoughtful and concrete ablation suggestions. They identify distinct questions about latent matching, target weighting and decoder adaptation that deserve separate controls. In response, we completed the grounded no-latent control, an equal-group latent-target comparison and native-size decoder refits. We address each of your requests below.

[D1] Full versus F+G+V.

Following the control you suggested, FGV is the same model as the original Full, with pretraining, architecture, G, V and main schedule unchanged. The only difference is that J is forced to zero throughout main training.

For Poisson, we obtain forward/inverse errors of 1.6804%/9.5024% for Full and 1.9121%/11.0441% for FGV. For Helmholtz, we obtain 1.6361%/8.7023% for Full and 1.7906%/10.3872% for FGV. Full has lower error in all four comparisons, by 1.5417 and 1.6849 percentage points for inverse recovery.

We interpret this as the contribution of latent matching on these two PDEs.

[D2] Uniform versus mask-dependent weighting.

To address your weighting question, we compare uniform weighting over latent coordinates with equal total weights for the mask-defined observed and unobserved token groups. We reweight only J, keeping the physical field loss F unchanged.

PDE/runUniform inverse (%)Equal-group inverse (%)
Poisson9.502410.7760
NS, run 16.97327.0424
NS, run 26.96646.9040

We observe lower error with uniform weighting on Poisson, by 1.27 percentage points. On NS, the two comparisons show no consistent difference: the equal-group minus uniform differences are +0.069 and −0.062 percentage points, with a mean of +0.003.

The theoretical comparison concerns latent-coordinate weighting under its reference assumptions; the equal-group weighting above is one such mask-dependent weighting.

[D3] Native-size decoder refitting.

For the decoder control you requested, we refitted the 171,010-parameter native head on the same frozen Full parent, using either continued or reinitialized weights.

We obtain Poisson inverse errors of 9.5024% originally, 9.4582% after continuation, 9.4659% after reinitialization and 9.4260% with SP-10M. The corresponding cylinder whole-grid inverse errors are 0.6491%, 0.6193%, 0.9982% and 0.4524%. For comparison, DP-10M gives 0.4576% on cylinder.

The tested small-head refits recover part of the non-cylinder improvement, and DP-10M also achieves much of the large-head cylinder improvement. These controls help distinguish the total adaptation benefit from the narrower SP–DP pathway difference. We would also like to clarify that the principal SP-10M head has 10,010,658 parameters; the 5M head is a separate setting.

Manuscript changes. Following your suggestions, the revised manuscript reports FGV in a new Appendix K.2 (page 37, Table 26), cross-referenced from Table 7; provides the weighting definition and comparison in a new Appendix K.4 (page 38, Table 28), referenced from Section 3.3 (page 6, lines 301–303); and reports both native-size refits in a new Appendix K.3 (pages 37–38, Table 27), cross-referenced from Tables 8 and 10. Section 4.3 summarizes the FGV and refit results (page 9, lines 460–463 and 477–479), so that task-dependent component effects, latent-target weighting and readout-capacity/optimization effects are distinguished more clearly.

AuthorsOpenReview ↗

Weakness 3

Lots of missing values in the main Table 1.

Response:

Thank you for drawing our attention to the missing entries. The dashes occur in the literature-reference rows and denote settings not reported there; the common-protocol comparison has complete entries. We explain this distinction below, and the revised Table 1 makes it explicit.

[D1] Meaning of the missing entries.

For clarity, all 27 dashes in the submitted table lie in the four literature-reference rows: PRISMA-50, Equi-FunDPS, PISD and DDIS. According to the submitted caption and source notes, these entries correspond to settings not reported in the cited work.

The common-protocol rows contain all evaluations, and the new approximately 40M comparison also covers all five PDEs in both directions. We leave the unreported entries as dashes, without interpreting them as zero or substituting development scores.

[D2] Revised organization.

To make this distinction explicit, the revised Table 1 lists the models trained and evaluated under our common protocol in an upper panel and the literature values retained under their source protocols in a lower panel. The latter are presented as contextual references.

Manuscript changes. The revised Table 1 (page 8) is reorganized accordingly: the complete approximately 40M comparison is placed with the common-protocol results in the upper panel, and the caption (lines 379–384) states that a dash denotes a setting not reported in the cited work. The evaluated models are annotated with their parameter counts, and the caption points to their training configurations and budgets (Tables 2, 18, 20 and 21).

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

SCOPE reconstructs complete physical fields from sparse observations for forward and inverse PDE problems. Complete paired fields are used to train a model for each setting (one for each). A student predicts a teacher’s complete-field representation from masked observations while a shared decoder reconstructs fields from the student prediction and encoder representations of complete inputs. The backbone can be frozen so that it fits a larger decoder. Inference is deterministic without sampling or test-time optimization.

Under squared loss and fixed teacher–decoder pair, the theory shows that optimal latent prediction need not be optimal field reconstruction. It also gives sufficient conditions for decoder improvements to transfer from complete observations to sparse ones.

Across Darcy, Poisson, Helmholtz and Navier–Stokes settings native model outperforms the tested mask-aware operators and improves Poisson and Helmholtz inverse recovery over its field only baseline. Replication across training seeds is not achieved; operator baselines use smaller models and less training, ablations do not isolate latent prediction contribution.

Critical strengths
  • Full training recipe achieves significant improvements over the same architecture field-only baseline: Poisson inverse error drops from 11% to 9% and Helmholtz inverse error from 10% to 8%. Full improves all ten endpoints over FJV (which shares its pretraining and architecture) as well, since full view reconstruction is supported in the proposed recipe (Table 7).
  • The DP–SP comparisons keep the backbone, decoder architecture and training budget fixed. SP has lower primary mean errors at every matched head size (controlled evidence for fitting the decoder on the sparse-context representations used at inference) (Tables 8 and 10).
  • The fixed-snapshot quadratic analysis separates observation uncertainty, decoder incompatibility and predictor excess risk. It shows why optimal latent prediction need not be optimal field reconstruction under squared loss through a nonlinear decoder. The paper explicitly separates this reference analysis from its training objective (Section 3.3).
Critical weaknesses
  • The contribution of latent prediction is not isolated. Full versus FO changes pretraining and several auxiliary objectives together, Full versus FJV tests grounding within the predictive recipe. A pretrained grounded control without the latent loss is missing. The strong combined-recipe results thus cannot tell us what latent prediction itself contributes (Table 7).
  • The main operator comparison has very different resources: roughly 40M parameters and 500 epochs for SCOPE vs. approximately 5M parameters and 100 epochs for the operators. Batch size difference means roughly 20 times as many updates to the main stage for SCOPE. Without convergence or resource controlled comparisons it is not clear where the large performance gap comes from (Tables 2,6 and 18).
  • The native-to-adapted comparison changes decoder capacity, optimization and acquisition specialization. SP-10M replaces a 171k-parameter decoder with a 10M-parameter decoder and adds 100 epochs; its non-cylinder gains are less than 1% relative whereas DP already gets most of the larger cylinder gain. A native-size decoder refit would show what enlargement is worth vs additional fitting (Tables 6, 8 and 10).
Questions for authors
  • Could you compare Full with a control that retains the same pretraining, grounding, variance regularization, architecture, and training schedule but removes the latent loss, recomputing the auxiliary cap as prescribed? Do the Poisson and Helmholtz inverse gains persist relative to this control?
  • What convergence evidence supports the competitiveness of the main operator baselines? Could you report how the strongest baseline, Transolver, changes with increased training budget or capacity, including error versus measured training resources?
  • How much of the SP–Full improvement remains when a decoder with the original 171k-parameter architecture is refitted on the frozen backbone using the same sparse-path objective and 100-epoch schedule?
6 comments
AuthorsOpenReview ↗

Weakness 1 & Question 1

The contribution of latent prediction is not isolated. Full versus FO changes pretraining and several auxiliary objectives together, Full versus FJV tests grounding within the predictive recipe. A pretrained grounded control without the latent loss is missing. The strong combined-recipe results thus cannot tell us what latent prediction itself contributes (Table 7).

and

Could you compare Full with a control that retains the same pretraining, grounding, variance regularization, architecture, and training schedule but removes the latent loss, recomputing the auxiliary cap as prescribed? Do the Poisson and Helmholtz inverse gains persist relative to this control?

Response:

We sincerely thank you for this perceptive observation and for formulating the precise control needed to clarify the contribution of latent matching. In response, we have completed FGV, a pretrained grounded model that removes latent matching while retaining the other specified components, and both the Poisson and the Helmholtz inverse gains persist relative to this control. Below, we give the configuration and clarify how this control examines J within the grounded training recipe, separately from the original Full–FO comparison.

[D1] Control definition.

To follow the control you proposed, FGV is the same model as the original Full: it retains Full's complete-view pretraining, online architecture, 171,010-parameter native decoder, grounding, variance regularization and main-training schedule. The only difference is that the latent-matching term J is forced to zero from the first main step and throughout main training, which gives the objective F + α_t(0.25G + 0.01V). The auxiliary cap follows the same rule as in Full.

[D2] What the new comparison establishes.

For the two PDEs you highlighted, we report the forward and inverse errors below:

PDEForward: FullForward: FGVFull − FGVInverse: FullInverse: FGVFull − FGV
Poisson1.68041.9121−0.23179.502411.0441−1.5417
Helmholtz1.63611.7906−0.15458.702310.3872−1.6849

Errors are relative L2 in percent; Full is the model reported in Table 7.

We observe lower error for Full than for FGV on both PDEs and in both directions: by 0.2317 and 0.1545 percentage points for forward recovery, and by 1.5417 and 1.6849 percentage points for inverse recovery. The Poisson and Helmholtz inverse gains therefore persist relative to the requested control, and we interpret this control as evidence that latent matching contributes on Poisson and Helmholtz beyond pretraining, grounding and variance regularization.

This control measures the contribution attributable to J, while the complete recipe's performance is assessed through its own comparisons.

Manuscript changes. Following your helpful suggestion, the revised manuscript reports the FGV results in a new “Latent matching” paragraph of Section 4.3 (page 9, 460–463), and gives FGV, its definition and the complete comparison in a new Appendix K.2 (page 37, Table 26), cross-referenced from Table 7 (page 26). Appendix K.2 states the control conditions accompanying the Poisson and Helmholtz results, and Section 4.3 now discusses whole-recipe performance (Full–FO), grounding within predictive training (Full–FJV) and the incremental contribution of J (Full–FGV) in separate paragraphs.

AuthorsOpenReview ↗

Weakness 2 & Question 2, Part 1 of 2

The main operator comparison has very different resources: ....

and

What convergence evidence supports the competitiveness of the main operator baselines? Could you report how the strongest baseline, Transolver, changes with increased training budget or capacity, including error versus measured training resources?

Response:

Thank you very much for highlighting this important concern about capacity and computational resources, and for your request for budget and convergence evidence for Transolver. In response, we have expanded the comparison to four approximately 40M baselines across all five PDEs and both recovery directions, and we report the recorded training hours and training-loss trajectory of Transolver40M alongside Full. We summarize the results below.

[D1] Complete approximately 40M comparison.

For this comparison, we use the same test records, normalization, observation masks and evaluation protocol for all methods. We report forward/inverse errors in percent in each entry, using BER for Darcy inverse and the whole-grid metric for NS (BCs).

ModelParams (M)PoissonHelmholtzDarcyNSNS (BCs)
Full-native40.041.680/9.5021.636/8.7022.174/1.5562.581/6.9732.115/0.649
FNO40M39.403.205/15.0973.228/14.8172.626/3.5943.975/11.1622.317/0.736
Transolver40M40.192.605/12.7182.244/12.6982.444/2.6416.072/13.7712.615/1.156
CNO40M39.9211.314/24.20113.810/25.25315.084/2.95512.104/17.25414.616/7.385
DeepONet40M40.0510.375/27.7659.949/35.2805.597/8.01221.121/32.43726.614/22.046

[D2] Conclusions supported by these results.

Under these evaluated configurations, we observe lower error for Full-native than for FNO40M, Transolver40M, CNO40M and DeepONet40M on all ten endpoints.

The paired per-record differences confirm this on every endpoint. The table below reports Full minus each baseline at uniform K=500, with a 95% bootstrap interval from 10,000 resamples of the test records:

EndpointFull− FNO40M− Transolver40M− CNO40M− DeepONet40M
Poisson F1.680−1.525 [−1.606, −1.446]−0.924 [−0.976, −0.874]−9.634 [−9.920, −9.353]−8.695 [−8.917, −8.484]
Poisson I9.502−5.594 [−5.682, −5.509]−3.216 [−3.293, −3.142]−14.699 [−14.840, −14.560]−18.262 [−18.611, −17.921]
Helmholtz F1.636−1.592 [−1.618, −1.566]−0.608 [−0.620, −0.596]−12.174 [−12.273, −12.075]−8.313 [−8.381, −8.249]
Helmholtz I8.702−6.115 [−6.144, −6.086]−3.996 [−4.022, −3.970]−16.551 [−16.598, −16.502]−26.578 [−26.675, −26.484]
Darcy F2.174−0.452 [−0.464, −0.441]−0.270 [−0.278, −0.263]−12.910 [−12.994, −12.824]−3.423 [−3.441, −3.405]
Darcy I (BER)1.556−2.038 [−2.058, −2.018]−1.085 [−1.098, −1.072]−1.399 [−1.412, −1.386]−6.456 [−6.518, −6.394]
NS F2.581−1.393 [−1.421, −1.366]−3.490 [−3.529, −3.451]−9.523 [−9.588, −9.457]−18.539 [−18.804, −18.274]
NS I6.973−4.189 [−4.225, −4.153]−6.798 [−6.852, −6.743]−10.281 [−10.348, −10.213]−25.464 [−25.804, −25.111]
NS (BCs) F (whole grid)2.115−0.202 [−0.357, −0.070]−0.501 [−0.678, −0.265]−12.501 [−13.169, −11.857]−24.498 [−28.022, −21.993]
NS (BCs) I (whole grid)0.649−0.087 [−0.130, −0.049]−0.507 [−0.535, −0.483]−6.736 [−7.218, −6.273]−21.397 [−21.801, −21.011]

All 40 intervals lie below zero. Across five test-mask seeds, Full also has the lowest mean error on all ten endpoints (table in U-Q2).

We interpret these results as support for competitive recovery at approximately aligned parameter scale under the evaluated configurations.

[D3] Training hours of Transolver40M.

Capacity alignment addresses the scale concern. For the training budget, we follow your question on Transolver and report its training hours at approximately 40M parameters alongside Full. For clarity, the entries below are the recorded training times in GPU-hours on each PDE; the entries for Full are its main training.

ConfigurationParams (M)PoissonHelmholtzDarcyNSNS (BCs)
Transolver40M40.1941.434.243.641.622.1
SCOPE Full-native40.0445.041.736.328.612.7

Transolver40M is trained longer than Full on Darcy, NS and NS (BCs) (43.6 versus 36.3, 41.6 versus 28.6 and 22.1 versus 12.7 GPU-hours), and on these PDEs its forward/inverse errors are 2.444/2.641, 6.072/13.771 and 2.615/1.156, against 2.174/1.556, 2.581/6.973 and 2.115/0.649 for Full. Full has lower error than Transolver40M on all ten endpoints.

AuthorsOpenReview ↗

Weakness 2 & Question 2, Part 2 of 2

[D4] Convergence evidence.

Regarding convergence, the table below reports the training field loss of Full and Transolver40M on every PDE at each 10% of its own training time, together with its total training hours and the change over the final 10%:

PDEModelTotal h10%20%30%40%50%60%70%80%90%100%90%→100%
PoissonFull45.00.10760.09710.09250.08850.08370.07930.07490.07110.06830.0672−1.7%
PoissonTransolver40M41.40.20400.14330.12670.11570.10370.09540.08760.08160.07800.0771−1.1%
HelmholtzFull41.70.10490.09450.08820.08410.08010.07660.07260.06900.06640.0654−1.5%
HelmholtzTransolver40M34.20.27820.17590.14610.12990.11620.10590.09710.08600.07890.0758−4.0%
DarcyFull36.30.07970.07470.07200.07040.06900.06740.06530.06290.06030.0590−2.2%
DarcyTransolver40M43.60.16500.11790.10280.09250.08080.07430.07020.06660.06300.0616−2.2%
NSFull28.60.09690.08890.08550.08280.07980.07660.07290.06910.06600.0648−1.9%
NSTransolver40M41.60.23140.17700.14620.12840.11520.10560.09900.09410.09090.0899−1.1%
NS (BCs)Full12.70.04280.02830.02280.01900.01590.01270.00980.00800.00700.0067−3.8%
NS (BCs)Transolver40M22.10.14510.09450.06200.04980.03960.03320.02730.02250.01830.0166−9.2%

Convergence can be observed from the table: the training loss of both models flattens toward the end of training.

Manuscript changes. Guided by your feedback, the revised manuscript makes the complete approximately 40M comparison central to the evaluation: the four approximately 40M baselines are added to Table 1 (page 8) and described in Sections 4.1 and 4.2 (page 7, lines 368–371; page 8, lines 414–416). A new Appendix K.1 (pages 35–37) provides the configurations of the approximately 40M baselines (Table 20), the training-budget table (Table 21), the complete comparison with its paired intervals and test-mask seeds (Tables 22–24) and the training-loss table above (Table 25), with the comparison conditions stated explicitly.

Comments on the paper

2 comments

Dear Area Chair and reviewers,

Thank you for your careful reading and thoughtful suggestions. Your comments have guided the additional comparisons and explanations reported below. In response, our main new evidence is a complete comparison with four approximately 40M baselines: FNO, Transolver, CNO and DeepONet, evaluated on all five PDEs in both recovery directions.

In these evaluated training configurations, we observe lower error for Full-native than for FNO40M, Transolver40M, CNO40M and DeepONet40M on all ten endpoints. The paired per-record bootstrap intervals of all 40 Full-minus-baseline differences lie below zero, and Full remains lowest on all ten endpoints across five test-mask seeds.

Following your helpful emphasis on computational resources, we conduct this comparison under fairer resource conditions: the baselines are matched to Full-native in parameter scale, and the training budget and training-loss trajectories of Full and Transolver40M are reported on every PDE.

We have also completed the grounded no-latent control, native-size decoder refits, latent target-weighting comparison, observation-count evaluations and frozen-model GRF-hyperparameter tests requested in the reviews, the last two for Full and all approximately 40M baselines. We are grateful for these suggestions, which help us examine component attribution and generalization separately from the external-baseline comparison.

Below, we provide the completed evidence and indicate where each corresponding change appears in the revised manuscript, in which all revisions are marked in red. Guided by your feedback, the revision makes the approximately 40M comparison central (Table 1 and Section 4.2, page 8), collects the additional experiments in a new Appendix K (pages 35–40), and presents the performance and component conclusions with their supporting controls. We sincerely appreciate your constructive feedback.

I would like to thank their authors for their rebuttal. Dear Reviewers MiBT and U4QC, as we are approaching the end of the author-reviewers discussion period, I would appreciate it if you could take a moment to acknowledge the rebuttal and see if it addresses your concerns or no. Thank you for your time and service.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.