acceptodds
Withdrawn from ICLR 2027

WindWeaver: Along the Wind and Across Scales for Generative Weather Super-Resolution

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Fine-resolution weather sequences support applications sensitive to local conditions and their evolution. Generative super-resolution must recover unresolved fine-scale states and connect their changes across gaps in coarse observations. Coarse wind offers approximate cross-time correspondence while leaving multivariate fine-scale evolution undetermined. We introduce WindWeaver, which uses atmospheric transport to structure the computation of generative updates. Coarse wind organizes cross-time feature exchange, while the network learns how the exchanged information contributes to multivariate reconstruction. Wind-derived coordinates retrieve features of the current candidate sequence at each flow evaluation. Separate readouts predict coarse block means and within-block detail, while a locally correlated source initializes unresolved variation. These interfaces support velocity-based conditional flow matching (CFM) and a task-adapted, denoiser-based adaptive flow matching (AFM) host. Across East Asian ERA5, same-product CERRA, and paired ERA5-to-CERRA reconstruction, WindWeaver improves probabilistic reconstruction of wind, temperature, cloud, and radiation in both hosts. Controlled CFM experiments establish net communication gains over branch removal in state and six-hour increment scores. Matched static and reversed correspondence show that the coordinate rule matters. Increment benefits extend across every reconstructed variable group. These results support using approximate atmospheric correspondence to guide generative feature exchange for reconstructing multivariate weather states and their temporal changes.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Flow Matching, Spatiotemporal Super-Resolution, Weather Downscaling, Physics-Guided Machine Learning, Conditional Generative Modeling

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_sU57OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The work focuses on spatio-temporal super-resolution for weather forecasting using re-analysis dataset at different resolutions. They proposed a generative model trained on paired grid-coarsened data of six-hourly observations and fine-grid data at hourly of and three-hourly resolution. Their method is presented as an extension of Conditional Flow Matching and Adaptive Flow Matching for weather forecast tasks. In the paper, they present the importance of the super-resolution tasks being modelled in both axes: temporal and spatial, proposing a method that addresses both problems jointly. The authors highlights its contribution to the baseline by including a better informant and guidance encoding, wind-guided local operators, modelling a complementary weather source, and a residual mixer model. Experiments are validated on different resolutions according to the different datasets East Asian ERA 5, ERA5-CERRA and CERRA, for which they report the CRPS and relative gain percentage performance from each baseline.

Critical strengths

None

Critical weaknesses

Summary of concerns

The paper reads as preliminary work. It is hard to follow, and I have several major concerns about the experiments and the clarity of the method.

Experiments

Missing baselines and references. Besides the extended CFM and AFM variants, the paper should compare against additional baselines. PISR (Xu et al., 2026b) is already cited in the related work and addresses both the spatial and temporal aspects of the problem, so it should also appear in the experiments. Other relevant works include:

  1. A generative framework for probabilistic, spatiotemporally coherent downscaling of climate simulations (Schmidt et al., 2024), which should be included as a baseline.
  2. FlowConGRU (Hader et al.), which should also be included as a baseline.
  3. Fourier Neural Operators for arbitrary-resolution climate data downscaling (Yang et al., 2023), which also addresses the spatiotemporal setting on ERA5 and should therefore be discussed in the related work and included as a baseline.

The related work should also cover physics-aware (hybrid/grey-box) models for climate and weather, such as Neural General Circulation Models (Kochkov et al., 2023) and Variational Grey-Box Dynamics Matching (Singh et al., 2026), in addition to the already cited ClimPhyDM from the same author.

Evaluation. The evaluation relies mostly on CRPS and compares against only two baselines, which falls short of a standard climate/weather evaluation protocol. At a minimum, I would expect:

  • Deterministic forecast skill (latitude-weighted RMSE and Anomaly Correlation Coefficient) across multiple lead times, with climatology and persistence as reference forecasts.
  • Spectral analysis, such as power spectra or spectral power ratios.
  • Spread-to-skill ratio alongside CRPS, to assess the ensemble variance and calibration.
  • Evaluation on extreme events, to assess generalization out of distribution.

Clarity of the paper

The introduction and the application task are relevant and well motivated. However, after the introduction the paper becomes hard to follow, and the method section has several technical problems. The notation is often undefined or introduced without explanation, and it is heavier than necessary. Examples:

  • In Eq. (2), the quantities u\mathbf{u} are not defined, and it is unclear what they represent or how they differ.
  • PH\mathcal{P}_H is introduced but later defined as an identity matrix, which makes the notation redundant, and HH is never defined.
  • Fs,tF_{s,t} (Eq. 3) is not defined anywhere in the paper.

More importantly, it is unclear which functions are learned and which are fixed. It is also unclear whether time and spatial coordinates are inputs to the learned model and, if so, how they are encoded.

The contributions are hard to identify from the method as presented. The background on AFM, the most relevant prior work, is not presented clearly, which makes it difficult to see what the paper extends from it and what is new.

Questions for authors

The work should require a substantial amount of work adressing the points remarked in the critical weakness:

  1. Improve the submitted text and method clarity
  2. Include the new baselines asked
  3. Extend the evaluation pipeline to include all the metrics required to assess the weather forecast model
Reviewer_Km8NOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces WindWeaver to address the challenge of reconstructing fine-resolution, multivariate weather fields with missing temporal states from coarse weather observations. The core idea is to guide cross-time feature retrieval with the coarse wind fields to enable the reconstructed weather states exchanging relevant information during generative refinement. Key contributions include wind-guided feature communication, prediction of coarse-scale means and fine-scale details, and a structured stochastic source, adapted into both CFM and AFM frameworks. Experiments on ERA5 and CERRA show improvements in probabilistic reconstruction and six-hour change prediction over internal baselines and ablations.

Critical strengths
  1. The method uses coarse wind to guide the model where related information is in neighboring time frames, and a learned mixer decides how that information updates each variable and rereads it at every generation step. So, the model gets physical guidance at small extra cost and still learns how variables like temperature or cloud change.
  2. Experiments show evidence that the wind is what really helps. The authors show this by Statice and Reverse to keep the architecture and change the coordinates. Both perform worse than No-exchange, while correct wind improves results. This is important because it shows that the gains are due to using the right wind-based coordinates not capacity.
Critical weaknesses
  1. There are no comparison to SOTA methods; all results are measured against the authors’ own Base model; no other weather super-resolution or temporal interpolation methods like diffusion-based downscaling (CorrDiff), GAN-based methods (Leinonen et al., 2021) or recent methods not cited in the paper such as spateGAN-ERA5 [1], AirCast-SR [2] are tested. It is unclear whether WindWeaver is competitive under matched conditions.
  2. It’s not clear if the physical wind is really needed in the first place. The results show that correct wind outperforms wrong wind, but not whether a learned motion (letting the network predict its own offsets) or a wider spatiotemporal attention would do just as well. The paper’s main claim is that physics is what helps. Without this control, we are not sure if gain can come from any cross-time spatial lookup rather than from the wind. [1] Glawion, Luca, et al. "Global spatio-temporal ERA5 precipitation downscaling to km and sub-hourly scale using generative AI." npj Climate and Atmospheric Science 8.1 (2025): 219. [2] Luitel, Somnath, et al. "AirCast-SR: A Foundation Model for Kilometer-Scale Atmospheric Super-Resolution via Latent Consistency Diffusion." arXiv preprint arXiv:2605.26130 (2026).
Questions for authors
  1. Does WindWeaver outperform strong existing reconstruction approaches such as methods mentioned in weakness 1?
  2. Would a variant when the network predicts its own displacement offsets instead of using wind-derived G^l (Y) match WindWeaver? What about wider spatiotemporal attention across neighboring frames?
  3. How does performance change when wind estimates are inaccurate (for example with added noise, time shift, or missing one/two pressure level)? Reverse correspondence scores worse than No-exchange, which shows inaccurate winds could actively hurt. At what error level WindWeaver becomes worse than No-exchange?
  4. Can you report power rank histograms and some individual samples rather than only ensemble means? It is not clear if WindWeaver generates realistic fine-scale variability or mainly improves the ensemble mean.
Reviewer_qvz1OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This manuscript addresses high spatial resolution weather downscaling, particularly for wind. The paper coined the term "atmospheric transport", and claim the construct can be used to downscale the wind. CFM (conditional flow matching) and AFM (adaptive flow matching) are utilized. ERA5 and CERRA datasets are used. Standard metrics such as CRPS are used.

Critical strengths

The capability to downscale wind has great application value.

Critical weaknesses

The write-up is hard to understand. I have the feeling that the manuscript is written by an LLM. Perhaps such a writing style is effective for an LLM to communicate with another LLM. Unfortunately this is not how humans communicate. A few examples:

  • "Motion correspondence connects features across location and time" (line 151)
  • "The source combines local persistence with independent random inputs at each weather time" (line 247)
  • "Training uses 2020-2021" (line 282)
  • "CFM controls separate the branch's net benefit form the choice of correspondence" (line 317)

The manuscript should be revised, both in terms of clarity and writing style.

Questions for authors
  1. What is the definition of motion correspondence? If it's a standard terminology, a reference should be provided, along with a brief description. If it's a term the author(s) coined, a detailed explanation should be provided.
  2. Title of section 3.5, what is a host? Or it should be a branch?
  3. What is the rationale of design specified in eqn (6)?

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

Related papers

Discussion (0)

Sign in to comment.