acceptodds
Under review as a conference paper at ICLR 2027

Spherical Laplace Neural Operators with Spatio-Temporal Modeling

Abstract

Neural operators provide efficient surrogate models for dynamical partial differential equations (PDEs). Existing spherical neural operators have advanced the modeling of global spatial interactions through spherical spectral representations, while transient temporal dynamics remain implicit in learned discrete-time mappings. To address this limitation, we introduce the Spherical Laplace Neural Operator (SLNO), which combines spherical Fourier modeling in space with Laplace-domain modeling in time. By conditioning temporal transfer functions on the Laplace–Beltrami spectrum, SLNO explicitly represents scale-dependent decay and oscillatory dynamics through learnable poles and channel-valued residues. We further derive analytic initial-state and trajectory-driven responses to construct and propagate internal temporal features within nonlinear neural operator blocks. Extensive Experiments on controlled linear fast–slow system, spherical PDEs and global weather forecasting demonstrate the accurate and stable long-horizon prediction of SLNO. These results demonstrate the potential of explicit spatio-temporal spectral modeling for learning long-term spherical dynamics.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Spherical Neural Operator, Laplace Transform, Partial Differential Equations

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

This paper proposes the Spherical Laplace Neural Operator (SLNO), which combines spatial spectral modeling with temporal responses through scale-conditioned pole–residue representations and internal latent trajectories. The method targets long-horizon rollout prediction of spherical dynamics and is evaluated on a controlled fast–slow system, shallow-water equations, and partial-autoregressive weather forecasting.

Critical strengths

The paper addresses the practically relevant problem of error accumulation in long-horizon prediction of spherical dynamical systems.

The method connects spherical spatial scales with learnable temporal responses, offering an interesting direction for spatiotemporal modeling.

The experiments cover controlled dynamics, spherical shallow-water equations, and weather data, providing some diversity in evaluation tasks.

Critical weaknesses

Insufficient discussion of existing temporal modeling methods. The paper presents explicit temporal structure as key to improving long-horizon rollout prediction, but does not sufficiently analyze how existing Laplace-based, state-space, and memory-based methods model temporal dynamics or what limitations remain. Consequently, the specific research gap addressed by this work is not clearly established.

Duplicated text and an incomplete cross-reference. On page 5, lines 250–251, the description of time broadcasting in the initial block is repeated verbatim, including the malformed junction “block.matching.” On page 4, line 205, the text refers to an appendix for the shared spectral parameterization without identifying the relevant section, Appendix A.2. These issues indicate insufficient proofreading of the final manuscript.

Limited baseline coverage, with relevant temporal modeling baselines missing. The controlled-system and shallow-water experiments compare only against SFNO and GSNO, while the weather experiment compares only against GSNO. Temporal modeling methods directly relevant to the paper’s motivation are not included. The current experiments therefore do not adequately establish the added value of the proposed temporal structure over existing approaches.

Missing ablations on the contribution of explicit temporal modeling. The paper presents explicit temporal structure as key to improving long-horizon rollout prediction, but does not validate its contribution through ablations that remove or replace this structure. Comparisons between the full model and the baselines cannot distinguish gains from explicit temporal modeling from those due to other architectural differences. The link between the central design and the reported long-horizon improvements therefore remains insufficiently validated.

Questions for authors

Could the authors provide ablations that remove or replace the explicit temporal structure, with parameter counts and computational budgets matched as closely as possible, to establish whether this design drives the improvements in long-horizon rollout prediction?

lease provide a fully autoregressive experiment over all dynamic input variables: use the ground-truth state only at initialization, and recursively feed back model predictions thereafter, without supplying ERA5 ground truth at subsequent times. Please compare SLNO and the baselines under this same protocol to assess their long-horizon performance without ground-truth auxiliary inputs.

4 comments

We sincerely thank the reviewer for the careful review and for highlighting the practical importance of long-horizon error accumulation, the connection between spherical spatial scales and learnable temporal responses, and the range of tasks in our evaluation.

Weakness 1 — Related work on temporal modeling

We thank the reviewer for this important comment. We have expanded Section 2 of the revised manuscript. LNO represents temporal dynamics through learned pole--residue responses, while structured state-space and memory-based neural operators propagate temporal information through latent states or observed/input sequences. These approaches establish important mechanisms for temporal modeling, but they do not specifically address the representation studied in SLNO: explicitly conditioning transient temporal responses on spherical spectral scale within a single-state spherical forecasting operator.

In contrast, SLNO generates an internal latent trajectory from the current spherical state, conditions its pole--residue responses directly on the Laplace--Beltrami spectrum, and propagates these scale-dependent temporal responses through successive operator blocks before reading out the forecast endpoint. Thus, the contribution is not temporal modeling alone, but the coupling of spherical spectral structure with explicit within-step temporal responses under a single-state forecasting interface.

We have revised the Related Work section to make this distinction and the resulting research gap explicit. We also added direct temporal baselines and component ablations, reported in our responses to Weaknesses 3--4 and Question 1, to empirically test this distinction. All corresponding revisions are highlighted in red.

Weakness 2 — Duplicated text and incomplete cross-reference

We thank the reviewer for pointing out these issues. In Section 4.3 of the revised manuscript, we removed the repeated description of time broadcasting in the initial block and corrected the sentence containing the editing error. Section 4.1 now explicitly points to Appendix A.2 for the shared spectral parameterization.

Weakness 3 — Temporal modeling baselines

We thank the reviewer for this suggestion. We have added a temporal convolutional baseline (TCN) on the spherical shallow-water equations (SWE). TCN uses the same spherical harmonic transforms, four operator blocks, and 25 latent time nodes as SLNO. It maps the current state to latent features at those time nodes. In each subsequent block, a learned degree-dependent map mixes the spherical harmonic channels, followed by a three-tap causal convolution (Bai et al., 2018) applied independently to each channel along the latent time axis. Both models were trained for 100 epochs with three paired seeds (666–668) and evaluated on the same 150 initial conditions per seed. SLNO and TCN have 1,459,749 and 1,456,369 real trainable degrees of freedom, respectively, a difference of 0.23%. The table reports mean joint relative L2L_2 rollout error in physical units at each forecast horizon, averaged over the three paired training seeds.

ModelMean 1 h errorMean 5 h errorMean 10 h errorMean 20 h error
SLNO0.00064220.00222490.00234940.0026888
TCN0.00067570.00314180.00629480.2363048
SLNO error reduction4.97%29.18%62.68%98.86%

On SWE, SLNO has lower 20 h rollout error than TCN for each of the three paired training seeds. The reduction in mean error grows from 4.97% at 1 h to 98.86% at 20 h. Appendix F.5 also reports a comparison with LNO on the same SWE evaluation set, where SLNO reduces mean 20 h error by 96.2%.

Appendix F.5 of the revised manuscript describes the TCN architecture and the shared SWE evaluation protocol. Table 12 reports the real trainable degrees of freedom and mean rollout errors for SLNO, LNO, and TCN at 1, 5, 10, and 20 h.

Weakness 4 and Question 1 — Temporal ablations and computational cost

We appreciate the reviewer’s suggestion to test the contribution of explicit temporal structure at comparable parameter counts. On SWE, we evaluate two controls. The first retains the pole–residue response and passes only the latent endpoint between blocks, with each subsequent block treating that endpoint as a constant input. The second replaces the temporal response with a shared, degree-conditioned static spectral map. We match the training schedule at 100 epochs for each configuration. The endpoint-only control has the same parameter count as SLNO, and the static spectral control has 0.38% fewer trainable degrees of freedom. We evaluate all configurations on the same 150 initial conditions per training seed and report mean joint relative L2L_2 rollout error over seeds 666–668. We report measured inference time and peak GPU memory separately below.

SWE configurationReal trainable degrees of freedom1 h5 h10 h20 h
SLNO1,459,7490.00064220.00222490.00234940.0026888
SLNO (endpoints only)1,459,7490.00064080.00255510.00581480.3187674
SLNO (static spectral map)1,454,1930.00212770.00402860.00612215.7741058

The endpoint-only control has the same number of trainable degrees of freedom as SLNO and a similar mean 1 h error. At 20 h, passing the full latent trajectory reduces mean error by 99.16% relative to passing only the endpoint. The static spectral control has 0.38% fewer trainable degrees of freedom than SLNO. Relative to this control, SLNO reduces mean 20 h error by 99.95%. Appendix F.6 describes both controls and reports the rollout errors in Table 13.

We also measured inference cost in a separate profiling run using the seed-666 checkpoints. We used the same NVIDIA TITAN X (Pascal), a 256 × 512 grid, batch size 1, and 50 matched initial conditions. Timing covers 20 model forward passes with autoregressive feedback, with GPU synchronization and gradients disabled.

SWE configurationTime per 20-step rolloutPeak allocated GPU memory
SLNO3.0622 s899.9 MiB
SLNO (endpoints only)0.2387 s297.3 MiB
SLNO (static spectral map)0.1919 s263.5 MiB

The endpoint-only and static spectral controls reduce measured rollout time by 92.2% and 93.7%, respectively. With 100 training epochs for each configuration and comparable parameter counts, SLNO achieves lower mean 20 h error than both controls and incurs higher measured inference cost. Appendix F.7 gives the profiling protocol and the results of this separate run.

Question 2 — Weather evaluation protocol

We thank the reviewer for raising the distinction between conditional and fully autoregressive forecasting. Following the GSNO WeatherBench study (Tang et al., 2026), we use six forecast targets. GSNO, SFNO, and SLNO are trained for 20 epochs on the same data under the same training protocol. After initialization, each model feeds back its predictions for the six targets, uses ERA5 values for the other 18 dynamic variables at each forecast step, and keeps the three static fields fixed. We now describe this setting explicitly as conditional autoregressive forecasting in Section 5.3 and Appendix G.1.

Selected lead times from Table 3 of the revised paper. Lower wRMSE and higher ACC are better.

ModelMetricDay 1Day 10Day 20Day 30
GSNOwRMSE2.8867123.3827353.4592703.455369
SFNOwRMSE3.0217573.6001213.6961923.719947
SLNOwRMSE2.8615723.3336583.3882833.391095
GSNOACC0.8409060.7859320.7826920.782854
SFNOACC0.8220610.7492940.7434800.739562
SLNOACC0.8427860.7884940.7882960.787944

SLNO achieves lower wRMSE and higher ACC than both GSNO and SFNO at all eight lead times reported in the revised paper. At days 20 and 30, SLNO reduces wRMSE by 2.05% and 1.86% relative to GSNO, and by 8.33% and 8.84% relative to SFNO, respectively. These results characterize long-range forecast skill under the six-target conditional rollout protocol with observed ERA5 auxiliary inputs.

Fully autoregressive evaluation. To further address the concern, we also evaluated SLNO, GSNO, and SFNO on the same 512 test initial conditions. Observed ERA5 dynamic fields are used only at initialization. At each subsequent step, each model predicts and feeds back all 24 dynamic variables, while the three static fields remain fixed. We compute wRMSE and ACC over all 24 dynamic variables.

ModelScoreDay 1Day 5Day 10Day 20Day 30
GSNOwRMSE ↓3.4358725.6360476.8337917.9686198.696467
SFNOwRMSE ↓3.5620595.6812126.5402907.6070519.746844
SLNOwRMSE ↓3.3554465.5629506.7695687.4891977.732783
GSNOACC ↑0.8459530.5415850.3457750.2577200.229404
SFNOACC ↑0.8324290.5114120.3491600.2583480.179869
SLNOACC ↑0.8531620.5602500.3592980.2684060.233431

In this fully autoregressive setting, SLNO achieves the lowest wRMSE at days 1, 5, 20, and 30, while SFNO obtains a slightly lower wRMSE at day 10. SLNO achieves the highest ACC at every reported lead time. At day 30, SLNO reduces wRMSE by 11.08% relative to GSNO and 20.66% relative to SFNO.

Both protocols and the corresponding results have been added to the revised manuscript and are highlighted in red.


We thank the reviewer again for the constructive comments. All corresponding modifications and added results are highlighted in red in the revised manuscript.

Dear Reviewer KswF,

We hope this message finds you well. We have added new experiments and clarifications in response to your constructive comments. As the discussion phase is nearing its end, we would greatly appreciate any further feedback or remaining concerns you may have.

Thank you again for your time and consideration.

Best regards,

The Authors

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

The paper presents a spherical neural operator with learnable, scale-dependent poles and residues. By processing internally generated latent trajectories through an initial-state block and a trajectory-driven block, the proposed method achieves improved long-horizon performance on controlled dynamics and shallow-water forecasting, with additional conditional weather experiments.

Critical strengths
  1. Scale-conditioned temporal responses provide a coherent mechanism for coupling spatial scales with decay and oscillation.
  2. The fast–slow benchmark distinguishes one-step accuracy from long-horizon prediction quality.
  3. SLNO achieves consistently lower long-horizon shallow-water errors.
Critical weaknesses
  1. The paper lacks direct comparisons with related Laplace-based operators, as well as ablation studies on spectral sharing and latent-trajectory processing, leaving the individual contributions of these components unclear.
  2. Because observed ERA5 values for 21 auxiliary variables are supplied throughout the rollout, the results provide limited evidence regarding autonomous long-range forecasting capability.
  3. Although parameter counts are matched across models, the absence of runtime and memory measurements makes it difficult to assess the overhead introduced by SLNO's 25-node latent trajectory.
Questions for authors
  1. Do the authors have ablation results isolating the contributions of the scale-conditioned pole–residue parameterization and the latent-trajectory processing? Such evidence would materially strengthen the case for the paper's central claim.
  2. Could the authors report inference time and peak memory for SLNO and the parameter-matched baselines under identical settings, particularly given the use of 25 latent temporal nodes?
  3. Since the pole–residue temporal modeling builds on prior Laplace neural operators, could the authors explain why no Laplace-based baseline adapted to spherical data is included, and provide evidence that the gains stem from SLNO's scale-conditioned formulation and latent-trajectory processing rather than from Laplace modeling alone?
5 comments

We sincerely thank the reviewer for the careful review and for recognizing the scale-conditioned temporal responses, the distinction between one-step accuracy and long-horizon forecasting in the fast–slow benchmark, and the lower long-horizon errors on shallow water.

Weakness 1 and Question 3 — Spherical Laplace baseline and component ablations

We thank the reviewer for raising this important point. In the original submission, we focused the SWE comparison on native spherical neural operators, namely SFNO and GSNO, because our primary goal was to compare SLNO against existing spherical spectral operators under the same spherical forecasting setting. LNO (Cao et al., 2024), in contrast, was originally developed as a Laplace-domain operator for temporal dynamics rather than as a spherical neural operator, so a direct comparison requires adapting its pole--residue response to spherical harmonic coordinates.

To further address this concern, we have added LNO, a four-block spherical adaptation that applies a constant-input endpoint response in spherical harmonic coordinates. The table reports mean L2L_2 rollout error over the same 150 initial conditions per training seed, averaged across seeds 666–668:

Model1 h5 h10 h20 h
LNO0.00059720.00249050.00395780.0715542
SLNO0.00064220.00222490.00234940.0026888

LNO achieves the lower mean error at 1 h. SLNO achieves lower mean error at 5, 10, and 20 h, with reductions of 40.6% at 10 h and 96.2% at 20 h. Appendix F.5 of the revised manuscript describes LNO and reports the comparison in Table 12. This comparison shows that Laplace-domain pole--residue modeling alone does not account for SLNO's long-horizon improvement.

The LNO comparison measures the overall difference between the two architectures. To assess individual design choices within SLNO, the controls in our response to Question 1 and Appendix F.6 (Table 13) separately examine degree-dependent poles, cross-degree spectral sharing, and latent trajectory propagation.

Weakness 2 — Conditional weather forecasting

We follow the GSNO WeatherBench study (Tang et al., 2026) in selecting the six forecast targets. GSNO, SFNO, and SLNO are trained for 20 epochs on the same data under the same training protocol. We evaluate all three models on the same 512 initial conditions for up to 30 days. After initialization, each model feeds back its six predictions, uses ERA5 values for the other 18 dynamic variables at each step, and keeps the three static fields fixed. Section 5.3, the caption of Table 3, and Appendix G.1 of the revised manuscript specify the ERA5 inputs used at each forecast step. The table below summarizes the day-20 and day-30 results under the same conditional rollout protocol:

ModelDay 20 wRMSE ↓Day 20 ACC ↑Day 30 wRMSE ↓Day 30 ACC ↑
GSNO3.4592700.7826923.4553690.782854
SFNO3.6961920.7434803.7199470.739562
SLNO3.3882830.7882963.3910950.787944

At days 20 and 30, SLNO reduces wRMSE by 2.05% and 1.86% relative to GSNO, and by 8.33% and 8.84% relative to SFNO, respectively.

Fully autoregressive evaluation. To further address the concern, we also evaluated SLNO, GSNO, and SFNO on the same 512 test initial conditions. Observed ERA5 dynamic fields are used only at initialization. At each subsequent step, each model predicts and feeds back all 24 dynamic variables, while the three static fields remain fixed. We compute wRMSE and ACC over all 24 dynamic variables.

ModelScoreDay 1Day 5Day 10Day 20Day 30
GSNOwRMSE ↓3.4358725.6360476.8337917.9686198.696467
SFNOwRMSE ↓3.5620595.6812126.5402907.6070519.746844
SLNOwRMSE ↓3.3554465.5629506.7695687.4891977.732783
GSNOACC ↑0.8459530.5415850.3457750.2577200.229404
SFNOACC ↑0.8324290.5114120.3491600.2583480.179869
SLNOACC ↑0.8531620.5602500.3592980.2684060.233431

In this fully autoregressive setting, SLNO achieves the lowest wRMSE at days 1, 5, 20, and 30, while SFNO obtains a slightly lower wRMSE at day 10. SLNO achieves the highest ACC at every reported lead time. At day 30, SLNO reduces wRMSE by 11.08% relative to GSNO and 20.66% relative to SFNO.

Both protocols and the corresponding results have been added to the revised manuscript and are highlighted in red.

Question 1 — Isolating scale-conditioned responses and trajectory processing (Also addressing Weakness 1)

We thank the reviewer for this valuable suggestion. We agree that directly isolating these components is important for establishing the source of SLNO's long-horizon improvement. We therefore added matched SWE controls for degree-dependent poles, latent-trajectory propagation, and cross-degree spectral sharing. Each row below averages training seeds 666–668, using the same 150 initial conditions per seed for all variants:

Variant1 h5 h10 h20 h
SLNO0.00064220.00222490.00234940.0026888
Poles shared across degrees0.00065740.00271000.00327270.0048440
Endpoint-only propagation0.00064080.00255510.00581480.3187674
Untied spectral parameters0.00059580.00242420.00434540.0161870
  • Degree-dependent poles. The shared-pole control keeps the full trajectory, degree-dependent response matrices, and the same 1,459,749 real trainable degrees of freedom. Allowing the poles to vary with spherical harmonic degree reduces the mean 20 h error by 44.49%.
  • Trajectory propagation. The endpoint-only control retains the pole–residue responses and the same number of trainable parameters, while passing a single endpoint state between blocks. The mean 1 h errors are similar (0.0006422 for SLNO and 0.0006408 for the endpoint-only control). At 20 h, propagating the full trajectory reduces the mean error from 0.3187674 to 0.0026888, a 99.16% reduction across the three matched training seeds. The comparison shows that passing the full latent trajectory between blocks improves long-horizon rollout accuracy on SWE.
  • Cross-degree parameter sharing. The untied control uses independently learned spectral parameters across degrees, increasing the trainable real degrees of freedom from 1,459,749 to 7,031,341. The shared parameterization reduces the mean 20 h error by 83.39% on SWE.

Together with the LNO baseline above, these ablations provide complementary evidence that SLNO's long-horizon improvement arises from the combination of scale-conditioned temporal responses, latent-trajectory propagation, and shared spectral structure.

These results and the corresponding control definitions have been added to Appendix F.6 (Table 13) and are highlighted in red in the revised manuscript.

Weakness 3 and Question 2 — Inference time and peak memory

We appreciate the reviewer’s suggestion to quantify the inference cost of the 25-node latent trajectory. On SWE, we profiled GSNO, SFNO, and SLNO on the same NVIDIA TITAN X (Pascal), at 256 × 512 resolution, with batch size 1 and 50 matched initial conditions. We timed 20 model forward passes with autoregressive feedback, using CUDA synchronization and with gradients disabled.

ModelTrainable parametersTime per 20-step rolloutPeak allocated GPU memory
GSNO1,487,9810.205 s269.3 MiB
SFNO1,459,0810.194 s279.3 MiB
SLNO1,459,7493.184 s899.9 MiB

Under these matched settings, SLNO requires 15.5$\times$ the rollout time of GSNO and 16.4$\times$ that of SFNO, while its peak allocated GPU memory during inference is 3.34 times that of GSNO and 3.22 times that of SFNO. These measurements quantify the computational overhead associated with explicitly constructing and processing the latent temporal trajectory. We now report this accuracy--computation trade-off explicitly in the revised manuscript. Appendix F.7 provides the complete profiling protocol and the three-model comparison in Table 14, together with additional profiling of the endpoint-only and static spectral controls. All corresponding additions are highlighted in red.


We thank the reviewer again for the constructive comments. All corresponding modifications and added results are highlighted in red in the revised manuscript.

Dear Reviewer Qwvf,

We hope this message finds you well. We have added new experiments and clarifications in response to your constructive comments. As the discussion phase is nearing its end, we would greatly appreciate any further feedback or remaining concerns you may have.

Thank you again for your time and consideration.

Best regards,

The Authors

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

The authors propose a new SLNO architecture (Spherical Laplace Neural Operator), designed to improve long-term predictions of time-dependent PDEs on a sphere. While standard neural operator approaches for such problems aim at learning a forward map that predicts the state at the next time step from the current state, primarily by mixing its spatial modes, SLNO also introduces temporal variability of these modes through oscillations and exponential decay. This is achieved through learnable poles and residues in the Laplace domain, which are used to construct and process latent temporal trajectories. The paper demonstrates improved long-horizon prediction accuracy and stability in autoregressive rollouts for a manufactured fast–slow spherical dynamics problem, the spherical shallow-water equations, and global weather forecasting.

Critical strengths
  1. Experiments demonstrate lower long-horizon prediction errors for SLNO compared to other spherical neural operators, namely SFNO and GSNO. While SLNO does not necessarily provide the most accurate one-step predictions, its 20-step error is substantially lower for the controlled fast–slow dynamics problem. Similarly, for the shallow-water equations, the error after 20 hours is approximately 93% lower than GSNO's and 98.5% lower than SFNO's. The improvements are more modest for global weather forecasting, but SLNO still outperforms both models across the evaluated forecast horizons.

  2. The introduction of explicit temporal dynamics through learnable oscillatory and exponential responses provides an interesting extension of existing spherical neural operators, which primarily focus on spatial interactions. By conditioning these temporal responses on the spherical Laplace–Beltrami spectrum, the architecture allows different spatial modes to exhibit different temporal behaviors. Furthermore, these temporal dynamics are modeled through internally generated latent trajectories, without requiring additional supervision at intermediate time steps.

Critical weaknesses
  1. While the introduction of explicit temporal dynamics through learnable poles and residues is interesting, the paper does not provide sufficient ablation studies to demonstrate which components of the proposed architecture are responsible for the improved long-horizon predictions. In particular, it would be interesting to investigate the sensitivity of the model to the number of internal latent temporal steps, the number of poles, and the parameterization of these poles. Moreover, the paper doesn't provide information about the possible slowdown of rollout time induced by the prediction of the intermediary latent steps, training time, and memory cost of SLNO when compared to the other two models.

  2. While the authors demonstrate improved long-horizon predictions, the global weather forecasting experiments are only partially autoregressive. Indeed, out of the 27 atmospheric variables used as inputs, only six are predicted by the model, while the remaining 21 are provided from the ground-truth ERA5 data at every forecast step. This significantly limits the accumulation of errors and casts doubt on whether the improved stability demonstrated in the simpler experiments would translate to fully autoregressive weather forecasting. Furthermore, the improvements in this experiment are relatively modest, with an approximately 2% reduction in error compared to GSNO after 20 and 30 days. Given that long-term stability is the main motivation of the proposed architecture, fully autoregressive weather forecasting experiments would considerably strengthen the paper.

Questions for authors
  1. Could the authors provide some more thorough ablation studies around the architecture, in particular the number of latent trajectory blocks, to understand their importance for the long rollout stability ?

  2. Can the authors provide more information about the runtime needed for rollout in comparison with the other models, as well as training time, and memory consumption during training and at inference ?

  3. Could the authors train a full autoregressive global weathe forecasting on the 27 channels considered, without resorting to re-injecting the ERA5 data for 21 of them ?

6 comments

We sincerely thank the reviewer for the careful and constructive assessment. We particularly appreciate the reviewer’s recognition of SLNO’s improved long-horizon forecasting performance and of the value of introducing scale-dependent temporal responses without requiring supervision at intermediate time steps.

Weakness 1 — Component ablations and computational cost

We thank the reviewer for this important suggestion. To address this concern, we have added SWE ablations of trajectory propagation, temporal node count, pole count, and degree-dependent pole parameterization. All comparisons use training seeds 666–668 and 150 matched initial conditions per seed. We report the computational costs in our response to Question 2.

  • Trajectory propagation. With the same parameter count and similar mean 1 h errors, SLNO reduces the mean 20 h error by 99.16% relative to passing only the endpoint between blocks. This comparison tests the contribution of propagating the latent trajectory.
  • Latent time nodes. At a fixed latent interval and parameter count, retaining 25 nodes reduces the mean 20 h error by 85.54% relative to 5 nodes and by 77.07% relative to 13 nodes. This comparison tests the temporal resolution used to represent the trajectory.
  • Pole count. The 5-pole configuration reduces the mean 20 h error by 88.87% relative to 3 poles. This comparison tests the number of poles in the temporal response.
  • Degree-dependent poles. Allowing the poles to depend on spherical degree reduces the mean 20 h error by 44.49% relative to shared poles. Both variants retain degree-dependent response matrices, so the comparison isolates degree dependence in the poles.
  • Explicit temporal response. SLNO reduces the mean 20 h error by 99.95% relative to the static spectral control at a comparable parameter count. This comparison tests the overall contribution of replacing a static spectral mapping with explicit temporal processing.
  • Cross-degree parameter sharing. SLNO reduces the mean 20 h error by 83.39% relative to the untied variant, which has more trainable parameters. This comparison tests the shared spectral parameterization.

Table A. SWE ablations across rollout horizons. For each horizon, we first average the errors over 150 matched initial conditions per training seed. The table reports the mean and population standard deviation of the three seed means. The reference SLNO uses 25 latent time nodes and five poles. Each percentage above is computed relative to the corresponding control's mean 20 h error. The “Ablations of temporal and spectral structure” subsection of the revised appendix (F.6) reports the same means and describes the controls.

SWE variant1 h5 h10 h20 h
SLNO0.0006422 ± 0.00000530.0022249 ± 0.00007440.0023494 ± 0.00018710.0026888 ± 0.0001300
Endpoint-only0.0006408 ± 0.00000610.0025551 ± 0.00037390.0058148 ± 0.00367630.3187674 ± 0.4437776
5 nodes0.0006348 ± 0.00000100.0023782 ± 0.00011220.0034777 ± 0.00051560.0186013 ± 0.0198614
13 nodes0.0006380 ± 0.00000320.0023567 ± 0.00013680.0029403 ± 0.00034400.0117256 ± 0.0110880
3 poles0.0006397 ± 0.00000470.0023740 ± 0.00011650.0038543 ± 0.00100990.0241551 ± 0.0286723
Shared poles0.0006574 ± 0.00001230.0027100 ± 0.00034040.0032727 ± 0.00044900.0048440 ± 0.0015685
Static spectral control0.0021277 ± 0.00197520.0040286 ± 0.00130810.0061221 ± 0.00381885.7741058 ± 8.1605243
Untied0.0005958 ± 0.00000060.0024242 ± 0.00005480.0043454 ± 0.00065010.0161870 ± 0.0027682

Weakness 2 and Question 3 — Conditional and fully autoregressive weather forecasting

We thank the reviewer for raising this important point. We agree that the weather experiment should be clearly distinguished from fully autonomous forecasting of the complete atmospheric state.

We use the same six forecast targets as the GSNO WeatherBench study (Tang et al., 2026). In our comparison, GSNO, SFNO, and SLNO are trained for 20 epochs on the same data under the same training protocol and evaluated from the same 512 initial conditions. At each forecast step after initialization, each model feeds back its predictions for the six targets, uses ERA5 values for the remaining 18 dynamic variables at the corresponding lead time, and keeps the three static fields fixed. Section 5.3, the caption of Table 3, and Appendix G.1 of the revised manuscript specify this conditional rollout protocol.

ModelDay 20 wRMSE ↓Day 20 ACC ↑Day 30 wRMSE ↓Day 30 ACC ↑
GSNO3.4592700.7826923.4553690.782854
SFNO3.6961920.7434803.7199470.739562
SLNO3.3882830.7882963.3910950.787944

At days 20 and 30, SLNO reduces wRMSE by 2.05% and 1.86% relative to GSNO, and by 8.33% and 8.84% relative to SFNO, respectively. SLNO also achieves higher ACC at both lead times. These results quantify long-range forecast skill when the six target variables are predicted recursively and the remaining dynamic inputs are supplied by ERA5.

Fully autoregressive evaluation. To further address the concern, we also evaluated SLNO, GSNO, and SFNO on the same 512 test initial conditions. Observed ERA5 dynamic fields are used only at initialization. At each subsequent step, each model predicts and feeds back all 24 dynamic variables, while the three static fields remain fixed. We compute wRMSE and ACC over all 24 dynamic variables.

ModelScoreDay 1Day 5Day 10Day 20Day 30
GSNOwRMSE ↓3.4358725.6360476.8337917.9686198.696467
SFNOwRMSE ↓3.5620595.6812126.5402907.6070519.746844
SLNOwRMSE ↓3.3554465.5629506.7695687.4891977.732783
GSNOACC ↑0.8459530.5415850.3457750.2577200.229404
SFNOACC ↑0.8324290.5114120.3491600.2583480.179869
SLNOACC ↑0.8531620.5602500.3592980.2684060.233431

In this fully autoregressive setting, SLNO achieves the lowest wRMSE at days 1, 5, 20, and 30, while SFNO obtains a slightly lower wRMSE at day 10. SLNO achieves the highest ACC at every reported lead time. At day 30, SLNO reduces wRMSE by 11.08% relative to GSNO and 20.66% relative to SFNO.

Both protocols and the corresponding results have been added to the revised manuscript and are highlighted in red.

Question 1 — Number of latent trajectory blocks

We varied the total number of operator blocks from two to four on the native SWE grid. For SLNO, one initial-state block is followed by one, two, or three trajectory blocks. We kept the SLNO channel width and the 25 latent time nodes fixed. For each model and depth, we used the same 150 initial conditions per training seed and averaged the results over seeds 666, 667, and 668.

SWE depth ablation (Appendix F.8, Table 15 of the revised manuscript). We report the L2L_2 rollout error at 20 h, averaged over the three training seeds. Bold indicates the lowest mean error in each column.

Model2 blocks3 blocks4 blocks
SLNO0.0035330.0093200.002689
GSNO0.1067190.0028380.040243
SFNO1.479112 × 10⁶0.3859550.175332

At two and four blocks, SLNO reduces the mean error relative to GSNO at the same depth by 96.69% and 93.32%, respectively. At three blocks, GSNO has the lowest mean error. The higher mean for three-block SLNO is driven mainly by seed 668, whose error is 0.021818, compared with 0.002928 and 0.003214 for seeds 666 and 667. The lowest mean error across the tested models and depths is 0.002689, obtained by SLNO with four blocks.

Question 2 — Rollout time, training time, and memory

We agree that the additional latent-time computation introduces a non-negligible computational cost and that this trade-off should be reported explicitly. On SWE, we measured pure 20-step rollout on the same NVIDIA TITAN X (Pascal), at 256 × 512 with batch size 1 and 50 matched initial conditions. The timer covers model forward passes and autoregressive feedback, with CUDA synchronization and gradients disabled. The 100-epoch training time includes validation and checkpoint writes. The training memory column gives peak allocated GPU memory during training epochs.

ModelTraining time (100 epochs)Training peak allocated GPU memoryTime per 20-step rolloutInference peak allocated GPU memory
GSNO60.69 min490.3 MiB0.205 s269.3 MiB
SFNO60.53 min511.7 MiB0.194 s279.3 MiB
SLNO336.17 min1787.7 MiB3.184 s899.9 MiB

Over 100 epochs on SWE, SLNO takes about 5.5 times as long to train as either baseline. Its peak allocated GPU memory during training is 3.65 times that of GSNO and 3.49 times that of SFNO. Appendix F.7 gives the measurement protocol for the training and rollout results in the table.


We again thank the reviewer for the constructive comments. All corresponding modifications and added results are highlighted in red in the revised manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

Dear Reviewers and Area Chair,

Thank you very much for your time and consideration. We thank the reviewers for their constructive feedback and for recognizing the importance of long-horizon spherical forecasting and SLNO’s scale-conditioned temporal modeling. In response, we made the following major additions and revisions:

  1. Additional temporal baselines. We added spherical LNO and TCN comparisons under matched SWE settings. SLNO substantially improves long-horizon accuracy over both baselines.

  2. Ablations and sensitivity analyses. We added controls for latent-trajectory propagation, explicit temporal processing, degree-dependent poles, spectral parameter sharing, temporal node count, pole count, and block depth. These results more directly isolate the contributions of SLNO’s temporal and spectral design.

  3. Fully autoregressive weather forecasting. Beyond the original conditional setting, we added a 24-variable fully autoregressive 30-day evaluation. SLNO achieves the best day-30 performance, reducing error by 11.08% relative to GSNO and 20.66% relative to SFNO.

  4. Computational analysis. We added matched measurements of training time, rollout time, and GPU memory, together with implementation profiling and optimization. We now explicitly report the remaining computational overhead as a limitation and future direction.

  5. Clarified positioning and presentation. We expanded the related-work discussion, clarified the weather and ablation protocols, and corrected the presentation issues identified by the reviewers. All corresponding revisions and new results are highlighted in red.

We thank the reviewers again for helping us strengthen the manuscript. We hope this summary will be helpful for the remaining discussion. As the first discussion phase is nearing its end, we would greatly appreciate any remaining feedback or concerns.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.