acceptodds
Under review as a conference paper at ICLR 2027

WeatherJEPA: Joint Embedding of Observational and Atmospheric State Representations for Weather Forecasting

Abstract

Accurate weather forecasting depends not only on learning atmospheric dynamics, but also on determining how atmospheric information should be represented. Many leading data-driven weather models predict future gridded atmospheric fields from reanalysis-based training data, whereas observation-driven systems infer atmospheric representations directly from heterogeneous measurements. These approaches reveal a fundamental representation trade-off: atmospheric state representations provide coherent and physically structured information but inherit assumptions from assimilation systems, while observational representations preserve direct measurements but must discover atmospheric structure from incomplete evidence. We introduce WeatherJEPA, a joint embedding framework that aligns complementary observational and atmospheric-state representations into a shared latent space. The atmospheric-state representation provides a structured view during training, while the observational representation defines the deployment pathway. Rather than reconstructing one representation from the other, WeatherJEPA learns their relationship through joint embedding, grounding, and temporal prediction objectives. Frozen-state readouts and ten-day forecasts show that initial accessibility does not consistently predict forecast ranking. Under matched observation-driven comparisons, WeatherJEPA achieves the lowest aggregate probe errors for both physical fields and in-situ observations. Across six representative station, radiosonde, and satellite targets, it reduces Day-10 forecast error by 4.4–12.7% relative to the strongest competing learned method for each target. On ICOADS marine zonal wind observations, its Day 1–10 mean MAE is 4.4% lower than IFS-HRES.

Primary area:
learning on time series and dynamical systems
Keywords:
weather forecasting, joint embedding predictive architectures, representation learning

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_6Ep1OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

WeatherJEPA learns a shared atmospheric representation from observations and reanalysis data. At deployment, it forecasts from observations without reanalysis inputs. Training objectives align representations, preserve weather information, and support prediction over time.

The authors evaluate current-state decoding, 10-day forecasts, and training-objective ablations. Against controlled learned baselines, day-10 errors fall by 4.4-12.7% across six selected targets. Results also show that better current-state decoding does not always imply better forecasts.

Critical strengths
  1. Weather JEPA uses reanalysis to guide representation learning while allowing forecasts to start from observations alone. Its improvements extend across station, radiosonde, and satellite targets, and across forecast lead times. This breadth strengthens the empirical case beyond a single aggregate score or selected target.

  2. Current-state decoding rankings do not consistently match forecast rankings, and rankings change with forecast duration. This shows why decoding tests alone are insufficient. I believe this is a lesson relevant to broader representation learning research.

Critical weaknesses

The authors use controlled adaptations of three competing methods, but provide limited details about their training. Appendix A.2 states that the baselines follow the original architectural and training designs and use the same observation data and targets. However, their model sizes and training budgets are not reported. The reanalysis records used by each baseline in this study are also unclear. Appendix A.4 describes sequential adaptation across forecast leads for WeatherJEPA, but does not state whether the baselines use this procedure.

These choices can affect forecast accuracy. Without these details, I cannot assess how much of the reported improvement reflects the proposed method rather than differences in training. Since these comparisons provide the main evidence for gains over competing learned methods, this limits my confidence in those gains.

Questions for authors
  1. Appendix A.2 states that the baselines follow the original architectural and training designs. What changes were made for this study? Please give each baseline’s model size, training budget, and checkpoint selection procedure.

  2. Appendix A.1 states that the longer ERA5 archive ends in 2017. What is its starting year? Which reanalysis records were used to train each baseline in this study?

  3. Appendix A.4 describes sequential adaptation across forecast leads and separate observation decoders for WeatherJEPA. Were these procedures also used for the baselines? If not, what forecasting and readout training did each baseline use?

  4. Section 5.1 identifies the WeatherJEPA state as the observation-only deployment state. Which pre-forecast state was tested for each baseline? Please specify the architecture and training settings of the lightweight readouts used for the matched comparison.

Reviewer_2PTaOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper propose WeatherJEPA, which combines observation and ERA5 to learn the latent representation for weather forecasting. The training includes Joint Embedding, Grounding and Temporal Prediction. During deployment, the model usable to predict future weather using observation data. Experiments shows that WeatherJEPA outperforms existing baselines on multiple observation forecasting tasks.

Critical strengths
  1. Reasonably combine different data sources, including ERA5 and observation.
  2. The experiment results are competitive. The paper compared their method with different baselines on multiple observation forecasting tasks and observation time. WeatherJEPA outperforms other leaned baseline in many tasks.
Critical weaknesses
  1. Limited clarity of experimental setup and baseline implementation. For example, Figure 3 and Table 1(b) reports relative error without the absolute errors, making the results less intuitive. In Table 2, both the main text and caption does not explain the evaluation procedure, makes it hard for readers to understand the content of the ablation study. Other than these, the runner-up ranking in Figure 5(b) does not have much correlation with the author proposed method. The architectures of EϕE_\phi, Πobs\Pi_{\mathrm{obs}}, Πdense\Pi_{\mathrm{dense}}, KK, and Φψ\Phi_\psi are not described in sufficient detail.
Questions for authors
  1. Why using two different dynamics predictors? What is the difference between their architecture and model capacity? Is it possible to use the trained KK for forecasting?
  2. What is the motivation for choosing future observation forecasting as the primary task rather than physical-field forecasting, since we now both have the ERA5 data and observation data? Could the authors clarify the practical advantages of observation forecasting over physical-field forecasting? Is observation forecasting more useful than physical-field forecasting?
  3. In Figure 6, IFS-HRES performs better at several shorter forecast leads, while WeatherJEPA outperforms it from Day 8. What explains this change in relative performance as the forecast horizon increases?
Reviewer_KxovOpenReview ↗
Rating
2
Confidence
1
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

This paper introduces WeatherJEPA, which uses reanalysis data to learn representations better suited to weather forecasting while requiring only observations at inference time. The method aligns observations and reanalysis through joint embedding, uses reconstruction objectives to preserve meteorological information, and introduces a temporal prediction objective to encourage representations that support forecasting. The experiments examine the information that can be decoded from different representations, evaluate forecast performance over a ten-day horizon, and analyze the roles of the reconstruction and temporal prediction objectives.

Critical strengths
  1. The paper addresses a clear and meaningful problem. Observations provide incomplete coverage, while reanalysis offers a more comprehensive view of atmospheric structure. Learning from both while retaining observation-only inference is a worthwhile goal. The paper presents a fairly complete implementation and demonstrates consistent improvements across multiple observational targets and forecast lead times.
  2. Beyond reporting forecast errors, the paper attempts to understand what the model learns. For example, better decoding of the current state does not necessarily translate into better forecasts, and different reconstruction targets affect the information retained in the representations. These analyses add value beyond performance gains and help clarify what makes a representation useful for weather forecasting.
Critical weaknesses
  1. The baseline implementations and training settings are not sufficiently clear. The comparisons use the authors’ adaptations of Aardvark, AIFS-DOP, and Earth-o1. Although the observational inputs and evaluation targets are standardized, the paper does not adequately document model sizes, training budgets, or how each method uses the longer reanalysis record. It is therefore difficult to determine how much of the improvement comes from the proposed representation learning approach and how much comes from differences in training and adaptation.
  2. The added value of joint embedding is not fully established. The method combines observation reconstruction, reanalysis reconstruction, cross-view alignment, and temporal prediction, but does not directly test whether simpler alternatives—such as joint reconstruction or direct reanalysis supervision of observation representations—could achieve comparable results. In particular, removing reanalysis reconstruction barely affects the decoding error for current observations. This leaves unclear where and how the information supplied by reanalysis contributes to the final forecasts.
Questions for authors
  1. Could the authors provide a clearer account of each baseline’s model size, training budget, reanalysis data coverage, and procedures for training the forecasting model and output heads? How much of WeatherJEPA’s advantage remains when these factors are matched as closely as possible?
  2. How does removing the temporal prediction or reanalysis reconstruction objective affect actual Day 1–10 forecast performance? Could these results establish whether the representation-level improvements in Table 2 translate into better forecasts, particularly at longer lead times?
Reviewer_igG1OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

WeatherJEPA jointly uses heterogeneous observations and ERA5 fields during training to learn a shared atmospheric representation. At forecast time, it constructs the initial state from observations alone and evolves it in latent space. The experiments show lower multi-day forecast errors than the compared learned methods on the selected observation targets.

Critical strengths
  1. The paper addresses a practical question in observation-driven forecasting. Using reanalysis as guidance during training while initializing forecasts from observations alone is a worthwhile direction.

  2. The experiments cover station, radiosonde, and satellite observations. Comparing initial-state readouts with multi-day forecasts also helps reveal differences between learned state representations.

Critical weaknesses
  1. The paper does not explain clearly enough how the model learns an atmospheric state from sparse observations. As I understand it, the goal is to use paired observations and reanalysis during training to establish a relationship between their latent features, so that a physically meaningful state can later be constructed from observations alone. This step infers spatially continuous atmospheric information from incomplete measurements and therefore shares some goals with data assimilation. However, the paper mainly describes the training objectives. It gives less explanation of how observation features correspond to dense-field features, how the model handles missing or unevenly distributed observations, and what physical information remains in the state at inference. I encourage the authors to explain this process more clearly and examine the learned state through dense-field reconstruction or forecasts, performance in observation-sparse regions, and visual examples.

  2. The experimental comparisons do not yet show clearly how each design contributes to the final forecast performance. The paper reimplements several baselines and uses representation probes to study the training objectives, but it remains difficult to tell how much of the gain comes from the shared representation and temporal modeling, as opposed to additional training objectives, reanalysis data, or training settings. I suggest more direct forecast comparisons with data, model size, and training budget matched as closely as possible. For example, the authors could compare observation-only training with training that uses reanalysis guidance, compare joint embedding with direct supervision from reanalysis, and remove the temporal and grounding objectives separately while reporting multi-day observation forecast errors. These comparisons would make the contribution of each design easier to assess.

Questions for authors
  1. How does the model construct a state corresponding to a dense atmospheric field from sparse and unevenly distributed observations? Can dense-field reconstruction or forecasts, results in observation-sparse regions, and visual examples show what physical information the inference-time state retains?

  2. With data and training budgets matched as closely as possible, how much do joint embedding, reanalysis guidance, and the individual training objectives each contribute to multi-day observation forecasts? Can the authors provide these controls and clarify the baseline training conditions?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.