acceptodds
Under review as a conference paper at ICLR 2027

Modelling Slow and Fast Variables for Subseasonal-to-Seasonal Forecasting

Abstract

Deep learning models for weather prediction have reached impressive accuracy. Yet, a significant performance and reliability gap remains at subseasonal-to-seasonal (S2S) forecasting horizons, i.e., 15 days to 2 months. To reach high predictive skill in this regime, accurate estimates are required not only for fast atmospheric dynamics but also for slowly changing boundary conditions on the Earth's surface (for instance, sea surface temperature). This makes S2S dynamics a coupled multi-scale system in which not all variables evolve at the same pace. We address this challenge by extending the Earth System Foundation Model (ESFM), a transformer-based model of weather and climate, such that it is able to explicitly handle the different update frequencies. We introduce a Temporal Attention Aggregator (TAA) taking as input samples from arbitrary past timesteps, and devise a mixed-lead-time (mixed-LT) approach encouraging the model to focus on selected variables' timesteps for each time scale. Our extended TAA model for S2S forecasting (TAAM-S2S), trained with adapted block rollout finetuning, improves the skill of the standard ESFM baseline for slow variables, while matching its accuracy for fast variables at S2S time scales. Moreover, TAAM-S2S outperforms DLESyM, AIFS v1.1 and Aurora v1.5 in terms of ensemble metrics for slow variables. It also reaches competitive performance in a rare-event case study, predicting droughts in Sichuan in 2022 and in Amazonia in 2023.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Subseasonal-to-Seasonal, Subseasonal Weather Forecasting

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_VBu3OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the different evolution rates of fast and slow variables in subseasonal-to-seasonal forecasting. It extends ESFM with a Temporal Attention Aggregator (TAA), mixed-lead-time training and inference, residual prediction for slow variables, and block rollout finetuning. The reported results show improved forecasting performance for several slow variables and competitive results against existing models on selected variables.

Critical strengths
  1. The method combines temporal attention aggregation with mixed lead times to accommodate the different evolution rates of fast and slow variables within one model, while retaining frequent forecast outputs. This is a practically useful direction.

  2. The experiments show improvements for several slow variables and evaluate the model using pointwise errors, anomaly correlations, and spectral metrics, providing supporting evidence for the approach.

Critical weaknesses
  1. The final model combines TAA and mixed lead times with residual prediction, latent-channel reassignment, and block rollout finetuning. The paper examines several of these choices, but it is still hard to tell how much of the improvement comes from TAA and mixed-LT themselves. Comparisons with the same input history, initialization, and training budget would make their contributions clearer. In particular, comparing TAA with a simpler way of combining the same past inputs would help show whether the learned attention is needed.

  2. The metrics and spectral analysis are useful, but there are few direct comparisons of the predicted fields. The appendix shows some failure cases, while the drought examples mainly track regional averages, which can hide differences in where and how anomalies develop. Maps comparing the final model, baselines, and ERA5 at several lead times would make these differences easier to assess. Showing slow variables alongside related atmospheric fields would also help check whether their evolution is consistent, beyond improvements in their individual scores.

Questions for authors
  1. Under matched inputs, initialization, and training budgets, how much does each of TAA and mixed lead times contribute? Can targeted ablations verify their independent effects?

  2. Can the authors compare forecast fields, reference fields, and spatial errors for representative fast and slow variables at different lead times, to show whether the metric improvements correspond to more realistic spatial structures and temporal evolution?

4 comments

Thank you kindly for the insightful review.

Adding ablation studies

To address your first concern and question, we added a section with ablation studies to the appendix. We compared the base mixed-LT 6/48h model with its versions without: TAA, slow variables latent level, residual prediction of slow variables, and the mixed-LT approach. All of the models are initialized from the same 48h baseline model and trained for the same number of epochs. In this particular case of S2S predictions of the mixed-LT 6/48h model, we observe that the majority of gain comes from the mixed-LT approach, as well as from changing the target of slow variables from absolute values to residuals. As reported in our response to reviewer 6XoP, since TAA does not significantly increase the number of parameters, one can reasonably expect that the processor compensates for any lack of representation with variable-specific patch embedding. However, it is important to note that TAA is promising to offer more flexibility, never degrades the metrics, and improves them for some variables (e.g., LSE of 6-hourly total precipitation).

In addition, TAA is more robust to perturbations in the inputs than the baseline variable-specific patch embedding. This robustness was quantified in the mixed-LT model described in Appendix G and trained with five lead times and the corresponding six past inputs (t_0, t_0 - 6h, t_0-48h, t_0-96h, t_0-192h, t_0-384h). The equivalent model with variable-specific patch embedding was trained under the same configuration. At inference time with delta_1=6h and delta_2=48h, inputs t_0-48h, t_0-96h and t_0-384h were replaced with Gaussian noise, leaving only t_0, t_0-6h, and t_0-192h as informative inputs. On fast variables, RMSE after two weeks increases by 10 to 80% with perturbed inputs and the variable-specific patch embedding, while our TAA module shows no noticeable degradation of the error. On slow variables, variable-specific patch embedding suffers from an error increase of 50 to 900%, while our TAA module limits the error increase to 7 to 64%. Sea surface temperature is the exception, as it is the most sensitive variable with TAA, increasing by up to 330%, which is nonetheless still lower than the 680% increase of variable-specific patch embedding. Therefore, although TAA shows moderate improvement over a linear combination of inputs in the baseline variable-specific patch embedding in the setup matching the training conditions, TAA is significantly more robust to perturbed inputs.

Qualitative results

Following your suggestions, in the revised version of the article, we also provide an additional section in the appendix, where we present predictions and error maps of baselines and the TAAM-S2S model. To cover as many variables as possible, we gave examples of mean sea-level pressure and 6-hourly total precipitation (fast surface variables), V-component of wind at 1000hPa and geopotential at 850hPa (atmospheric variables), and sea surface temperature and volumetric soil water layer 1 (slow variables). As challenging as it is to draw conclusions from prediction maps corresponding to one rollout, the figures show that TAAM-S2S doesn’t lose performance of 48h baseline on slow variables, and that they both supersede the 6h baseline under this regime.

Thank you for the response and additional experiments. Several of my concerns have been addressed. Two questions remain:

  1. Do the ablations in Appendix M use identical historical inputs and include the final block-rollout fine-tuning?
  2. T2m performs substantially worse than climatology. Could you explain this gap and compare with FuXi-S2S under the same evaluation protocol to clarify the model’s S2S advantage?

We are happy that our answers helped to clarify some ambiguities.

Ad 1. All ablations in Appendix M use identical inputs, but none of them include block rollout finetuning. The latter is not added, as it introduces substantial computational cost to each training run and may confound the ablation under study. All of the models have been compared under fair conditions. In particular, we note that the base mixed-LT 6/48h was also not rollout-finetuned.

Ad 2. As you accurately noticed, 2m-temperature performs worse than climatology on S2S timescales in RMSE metric but remains below the theoretical threshold of 2⋅climatology\sqrt{2} \cdot \text{climatology}. As RMSE measures point-wise accuracy, it rewards oversmoothed predictions that minimize single-pixel errors. Climatology maps are smoother than those created by TAAM-S2S, meaning that our model is penalized for creating more realistic predictions. To support this claim, we attach the mean LSE (Logarithmic Spectral Error) over 48 test-time predictions for the 15th and 30th day of rollout for TAAM-S2S compared with Climatology. Furthermore, we break down the score to report LSE for small- and large-scale features separately (split at k=20k=20, which corresponds to 18∘18^\circ east-west and 9∘9^\circ north-south), showing that TAAM-S2S captures the variance of both significantly better than climatology.

lead time (features)+15 days (all)+15 days (large-scale)+15 days (small-scale)+30 days (all)+30 days (large-scale)+30 days (small-scale)
TAAM-S2S0.144 ± 0.0200.186 ± 0.0340.115 ± 0.0220.146 ± 0.0270.183 ± 0.0520.119 ± 0.021
Climatology0.266 ± 0.0540.311 ± 0.0750.238 ± 0.0500.268 ± 0.0530.313 ± 0.0710.239 ± 0.051

We ran a comparison with FuXi-S2S on 12 separate rollouts, but it must be noted that both models work on slightly different data. FuXi-S2S inputs and outputs are daily averages, whereas our model processes instantaneous values. To create daily means using TAAM-S2S, for each forecast, we started four rollouts at 00, 06, 12, 18 UTC on the initial date. Then we averaged those predictions to obtain mean values for each pixel. We used the same protocol to obtain the ground truth we are comparing with. This incurs a slight discrepancy with FuXi-S2S as its outputs are daily averages over 24 hours. However, we do not expect it to revert our conclusions.

MetricModelWeek 1Week 2Week 3Week 4Week 5Week 6
RMSE [K]TAAM-S2S1.574 ± 0.1582.590 ± 0.4062.794 ± 0.3482.877 ± 0.2592.851 ± 0.3182.810 ± 0.423
RMSE [K]FuXi-S2S1.360 ± 0.1042.715 ± 0.3573.191 ± 0.3113.352 ± 0.3573.509 ± 0.5603.480 ± 0.636
ACCTAAM-S2S0.753 ± 0.0320.318 ± 0.1260.178 ± 0.1130.124 ± 0.0670.109 ± 0.1140.106 ± 0.155
ACCFuXi-S2S0.833 ± 0.0230.395 ± 0.0880.189 ± 0.0910.134 ± 0.1010.095 ± 0.1300.112 ± 0.124
LSETAAM-S2S0.118 ± 0.0080.171 ± 0.0140.184 ± 0.0220.189 ± 0.0240.182 ± 0.0190.178 ± 0.016
LSEFuXi-S2S0.250 ± 0.0130.257 ± 0.0220.273 ± 0.0180.289 ± 0.0160.307 ± 0.0340.315 ± 0.027

We can see that TAAM-S2S achieves a lower RMSE than FuXi-S2S from week 2 onwards and keeps significantly better LSE throughout the whole predicted period. On the other hand, FuXi-S2S achieves slightly better ACC, although, with the exception of week 1, the scores are within one standard deviation.

With the discussion period closing soon, are there any remaining concerns that we could address?

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

The paper proposes a forecasting model that treats fast-changing atmospheric variables and slow-changing surface variables differently. The Temporal Attention Aggregator (Figure 1) learns how much weight to give past observations. The mixed-lead-time procedure (Figure 2) predicts fast variables every six hours while updating slow variables over 48-hour intervals. The experiments compare the proposed model with ESFM baselines (Figure 3) and external forecasting models (Figure 6). The paper also evaluates soil moisture forecasts during two drought events (Figure 8 and Table 3).

Overall, the approach is well motivated. Atmospheric variables, sea surface temperature, and soil moisture change at different speeds, so using different prediction intervals makes sense. Table 1 shows the trade-off: longer intervals can improve forecasts for slow variables but reduce accuracy for fast variables. However, it is unclear how much of the improvement comes from temporal attention and mixed lead time forecasting. The final model also includes other training changes. Appendix H, Table 7 shows that the 48-hour residual baseline already performs similarly on several slow variables before final rollout fine-tuning. For example, both this baseline and the mixed LT residual model achieve a sea surface temperature ACC of 0.47 for weeks 3-6. Comparing the models after equivalent fine-tuning would help establish the additional benefit of the proposed method The precipitation results raise a separate concern. Appendices E and J report unrealistic spikes after predictions are converted back to physical units. The authors should explain how often these spikes occur and how severe they are. This problem also weakens the claim that the method preserves performance on fast variables. Each main evaluation uses only 12 forecast starting dates, one per month (Sections 4.2 and 5.1). The authors should report uncertainty estimates for the differences between models to show how consistently the proposed method improves performance across these dates. The ensemble results also require qualification. Figure 7 and Section 5.1 show that no evaluated model outperforms climatology for sea ice or soil-water layer 1 beyond week 3. Therefore, outperforming the other models does not necessarily mean providing useful skill relative to climatology at these longer horizons. The drought results in Figure 8 and Table 3 are encouraging. However, Table 3 combines errors across lead times, making performance during weeks 3-6 unclear. Reporting these weeks separately and comparing against persistence, keeping the initial conditions unchanged, would better demonstrate the model’s longer-range forecasting ability.

Critical strengths

The paper addresses an important forecasting problem and offers a practical way to combine different update intervals within one model. The authors investigate several design choices, including lead times (Table 1), residual prediction (Table 2), and spectral loss weighting (Figure 5). They evaluate both prediction accuracy and spatial structure, which provides a broader assessment than a single error measure. The drought examples in Figure 8 and Table 3 also demonstrate a potentially useful application.

Critical weaknesses

The main weakness is that the contribution of temporal attention and mixed lead time forecasting remains unclear, particularly given the competitive residual baseline in Table 7. The precipitation failures reported in Appendices E and J weaken the claim that fast-variable performance is preserved. The small number of forecast starting dates limits confidence in the comparisons, while Figure 7 shows that relative improvements do not always translate into better performance than climatology. The drought results in Table 3 also lack persistence comparisons and separate weeks 3-6 scores.

Questions for authors

Questions: Q1: How does the final model compare with the 48-hour residual baseline in Table 7 after similar rollout fine-tuning? Q2: Does the TAA module in Figure 1 outperform a simpler method that combines the same historical inputs? Q3: How frequent and severe are the precipitation spikes reported in Appendices E and J, and can they be corrected? Q4: Were the settings examined in Tables 1-2 and Figure 5 selected using only validation data? How uncertain are the reported improvements? Q5: Do the drought improvements in Table 3 remain when compared with persistence and evaluated only over weeks 3-6?

Suggestions: The authors should prioritize a fair comparison with the strongest residual baseline from Table 7 and, if feasible, one experiment isolating temporal attention or mixed lead time forecasting. They should explain and quantify the precipitation failure reported in Appendices E and J, clarify how model settings were selected, and report uncertainty in the main comparisons. Adding persistence and weeks 3-6 results to Table 3 would strengthen the drought analysis. The discussion of Figure 7 should also distinguish outperforming another model from outperforming climatology. The baseline comparison and the explanation of the precipitation failure would have the greatest influence on my assessment

2 comments

Thank you kindly for your in-depth review.

Clarification on comparison between 48h baseline and TAAM-S2S

First, we would like to issue a small clarification relating to Q1. In Table 7, you are referring to neither model is block rollout finetuned. There, we are comparing baseline models trained for 100 epochs to perform either 6h or 48h prediction with a model finetuned for 10 epochs using the mixed-LT scheme (starting from the processor and decoder initialized from the 48h baseline). The goal here is to show exactly what you noticed, which is that we don’t lose 48h step performance by training also on 6h steps. We added a sentence of clarification to the corresponding paragraph.

Comparison of TAA with simpler linear approach

To address Q2, and your comment about the ambiguity of how much improvement comes from TAA and mixed-lead-time forecasting, we added a section in the appendix with ablation studies, where we compared mixed-LT 6/48h and models trained with the same training scheme, but without one component. As reported in our response to reviewer 6XoP, we study ablating: TAA, slow variables latent level, residual prediction of slow variables, mixed-LT approach. In our experiments, we show that for the selection of 6h and 48h lead times, the gain in metrics comes mostly from the mixed-LT approach and predicting slow variables as residuals. TAA never degrades, sometimes still improves, and is promising to offer more flexibility. It is also more robust to perturbations in the inputs than the baseline variable-specific patch embedding, as described in the answer to reviewer VBu3.

Analysis of unphysical precipitation spikes

Following your suggestions, we conducted an extensive analysis of the root cause, characteristics, and frequency of the precipitation spikes. As we show, these are a side effect of trying to optimize the energy spectra in the logarithmic space, where a small change can have a drastic effect when transferred back to the physical units. We also propose a solution involving capping the precipitation values at the maximum ever noted per pixel. With this modification, spikes disappear while the metrics of the other variables are mostly unaffected. All the results presented in the updated paper now apply this capping solution at inference time.

Extending the number of runs for each experiment

The settings examined in Tables 1-2 and Figure 5 didn’t use validation data at all. They only used test split data, which corresponded to the period from 2020 onwards for our models. To strengthen our improvement claims, we extend all of the experiments (Tables 1-3, 5-7; Figures 3-6, 10-16, 20-29) to be based on 48 forecasts (4 initial dates per month in 2021) instead of 12. All of the conclusions drawn stay the same. To account for the randomness and forecasts’ dependence on initial conditions, each table in the submitted manuscript reports the standard deviation of the metrics. When applicable, we extended our interpretations of the results to emphasize that the improvements reported as significant are separated by at least one standard deviation (e.g., ocean variables in Table 1).

Comparison with climatology

You accurately noticed that Figure 7 and Section 5.1 show that no evaluated model outperforms climatology for sea ice or soil-water layer 1 beyond week 3. We noted this fact in L407. This is an obviously challenging task in which climatology is a surprisingly difficult baseline to beat. Nevertheless, despite this fact, we think that every improvement is a step forward in the field of S2S forecasts.

Breaking down the drought prediction results into periods

As suggested, in the updated version of the paper, we have added persistence and extended Table 3 (reproduced below) to contain separate results for weeks 1-2, 3-6, and from week 7 onwards. After this change, we can see that TAAM-S2S outperforms the competitors and persistence on both soil water layers during both droughts at lead times of 3-6 and 7+ weeks, except for volumetric soil water layer 1 in the Amazonia event at 3-6 weeks, where Aurora scores slightly higher. Overall, we can see that the drought results hold.

RegionModelSWVL1 ( Week 1-2)SWVL2 ( Week 1-2)SWVL1 ( Week 3-6)SWVL2 ( Week 3-6)SWVL1 ( Week 7+)SWVL2 ( Week 7+)
Sichuan 2022TAAM-S2S+0.452+0.525+0.249+0.447+0.046+0.155
AIFS 1.0/1.1+0.782+0.736+0.020+0.036-0.126-0.189
Aurora 1.5+0.681–+0.079–+0.039–
Persistence+0.070+0.024-0.070-0.176-0.179-0.291
Amazonia 2023TAAM-S2S+0.264+0.588+0.309+0.372+0.259+0.361
AIFS 1.0/1.1+0.340+0.610+0.228+0.313+0.168+0.211
Aurora 1.5+0.277–+0.381–+0.123–
Persistence+0.203+0.573+0.141+0.160+0.006+0.113

As the discussion period deadline approaches, do you have any further concerns that we didn’t address? We are happy to clarify any ambiguities if there are any left.

Reviewer_6XoPOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper offers a deep learning model for subseasonal-to-seasonal forecasting by extending a foundational model, ESFM, through two techniques. First, the Temporal Attention Aggregator (TAA) takes as input samples from arbitrary past timesteps; second, the mixed-lead-time (mixed-LT) approach encourages the model to focus on selected variables’ timesteps for each time scale. The paper compares the performance across versions of the foundational model and some other deep learning and reanalysis model (ERA5) baseline approaches.

Critical strengths
  • The offered method solves the error accumulation problem during autoregressive prediction.
  • The paper offers a use case and shows the applicability of the model in a real-life case.
  • The paper offers results for a long horizon and various time windows.
Critical weaknesses
  • When the model rolls out the predictions to fill slow variables, the model’s own 6h predictions of slow variables are thrown away. This also limits the model's reaction to sudden variable changes, such as under a passing storm. More importantly, the fast variables will fit on slow predictions, which are highly smoothed due to linear interpolation.
  • The Mixed-LT is simple and can be applicable to other models that can predict the long horizons. Is mixed-lt INFERENCE applicable to other methods? And when it is applied to the base models in Figure 6, do we expect TAAM-S2S to still have the best performance?
  • The graphs are hard to read and need further analysis. For example, why wasn't the model good at LSE of the 10 m u-component of wind in Figure 6, while it was good at RMSE?
  • The attention weights are not surprising; due to the interpolation, it is expected that the t_0 - 48h receives more weight.

I will improve my score if the authors address my concerns.

Questions for authors
  • The authors used TAA in place of variable-specific patch embeddings. Do you compare both approaches?
3 comments

Thank you kindly for your thorough review addressing important aspects of our work.

Justification of slow variables interpolation

Ad W1. As you correctly stated during the mixed-LT rollout, the model uses linear interpolation of slow variables predicted in the 48h step to do a sequence of 6h steps. However, slow variables are characterized by having little change between timesteps. The whole idea of modeling them differently comes from the observation that they need a separate treatment, as they do not change fast enough for the model to efficiently capture the trends with 6h timesteps. In their case, linear interpolation between 48h steps is an accurate representation, as shown by the table below reporting the RMSE of the linear interpolation of each slow variable across 2021 (normalized by the variable range to account for differing magnitudes): NRMSE=RMSEymax−yminNRMSE = \frac{RMSE}{y_{max} - y_{min}}

VariableNRMSE
sst0.00210
ci0.00435
swvl_10.00700
swvl_20.00276
swvl_30.000694
swvl_40.0000830

Such an approach also ensures continuity in slow variable values between inference blocks.

Application of mixed-LT to other models

Ad W2. Mixed-LT inference is applicable to other models, but the prerequisite is that the model is trained on multiple lead times. In particular, one can apply mixed-LT training to achieve this goal, but this is not a strict requirement. If other models, like AIFS and Aurora, were trained using multiple lead times, we would expect them to outperform TAAM-S2S with the mixed-LT inference scheme, as they are larger, more capable models, trained using extensive resources and working at a finer resolution. We added this point to the conclusions and limitations section.

Clarification on the metrics discrepancy

Ad W3. The metrics we present are complementary (L266) and require to be analysed together, which we tried to do throughout the article. Referring to the given example of Figure 6, a model can achieve low RMSE by oversmoothing prediction maps, thereby limiting the risk of high single-pixel error by outputting cautious averages. However, we also care about having accurate power spectra of the predictions. Thus, we also look into LSE. In the example model, oversmoothing predictions achieve low RMSE, but this behaviour is accompanied by high LSE, as the small-scale power spectrum is far from the ground truth.

TAA module attention weights

Ad W4. We would like to point out a different analysis of the attention weights than you reported, as t0−48t_0 - 48h receives less weight than t0−6t_0 - 6h, and t0t_0 in the majority of cases. In case we misunderstood your point, could you please elaborate on the reasons for timestep t0−48t_0 -48h to receive more weight than δ1=6h\delta_1=6h due to interpolation? We note that those weights were calculated for a single 6h and 48h steps, without any rollout.

Comparison of TAA with variable-specific patch embeddings

Ad Q1. We have added a new ablation section to the appendix of the revised paper containing, among others, the comparison between TAA and variable-specific patch embeddings which the question concerned. We compare the mixed-LT 6/48h and models trained with the same training scheme, but without one component. We study ablating: TAA, slow variables latent level, residual prediction of slow variables, mixed-LT approach. All of the models are initialized from the same 48h baseline model and trained for the same number of epochs. In our experiments, we show that for the selection of 6h and 48h lead times, the gain in metrics comes mostly from the mixed-LT approach and predicting slow variables as residuals. In this specific case, TAA provides only marginal gain over variable-specific patch embedding, where the interaction between time steps is learned by a linear combination. Since TAA does not significantly increase the number of parameters, one can reasonably expect that the processor compensates for any lack of representation with variable-specific patch embedding. However, it is important to note that TAA never degrades the metrics and improves them for some variables (e.g., LSE of 6-hourly total precipitation). In addition, TAA is more robust to perturbations in the inputs than the baseline variable-specific patch embedding, as described in the answer to reviewer VBu3.

I thank the authors for their detailed rebuttal, which addresses most of my concerns; therefore, I raised my score to 3. However, some of my concerns remain:

  • The abrupt changes through a passing storm and how the model adapt this change.
  • Discarding the model's 6h predictions
  • W2 is answered candidly, and I appreciate the addition to the limitations section.

We are glad we could address your concerns and appreciate you raising the score.

Following up on the remaining points. In ERA5, the variables we deem slow behave relatively smoothly even through extreme events. In ERA5, SST and CI come from the daily analysis and change only at 00 UTC (we interpolate them linearly to 6h granularity, as described in Appendix B). Therefore, the sub-daily information we can obtain there is limited, which is another reason justifying discarding the 6h steps for those variables.

We also analyzed how those variables change during storms. From what we found, for the strongest 50 typhoons in the training period, 48h interpolation underestimates the cold wake by only 0.09K for a day (median) and up to a maximum of about 1K (the 95th percentile of the approximation error is 0.51 K). Therefore, when training on ERA5, sampling SST at 48h resolution does not incur a significant loss of accuracy.

Addressing your concern about the model discarding its own 6h predictions of slow variables during inference. Figure 3 and Appendix E show that predicting slow variables at longer lead times is more beneficial, especially on the S2S timescale. Therefore, 6h slow variables forecasts could only potentially be used instead of introducing interpolation between 48h predictions. Since interpolating slow variables between 48h steps is a good approximation, switching to 6h slow-variable steps would only introduce compounding errors and break the continuity between the last 6h step in the inference block and the 48h step prediction of the same variables.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.