acceptodds
Under review as a conference paper at ICLR 2027

WindINR: Latent-State INR for Fast Local Wind Query and Correction in Complex Terrain

Abstract

Many downstream decisions in complex terrain require fast wind estimates at a small number of user-specified locations and heights for a given forecast valid time, rather than another dense forecast field on a fixed grid. We present WindINR, a latent-state implicit neural representation framework for continuous high-resolution local wind query and sparse-observation correction. WindINR maps static terrain descriptors, a low-resolution background field, and continuous query coordinates to a high-resolution wind state through a latent-conditioned decoder. To enable rapid inference-time correction, WindINR separates reusable representation learning from sample-specific latent-state correction. During training, a privileged encoder infers a reference latent state from high-resolution supervision, a deployable latent predictor estimates an initial latent state from inference-time inputs alone, and their discrepancies are summarized into a dataset-adaptive Gaussian prior over latent corrections. At inference time, within the WindINR module, network weights remain fixed and only the latent state is updated by minimizing a regularized correction objective using sparse observations and their uncertainty. We train WindINR on the large-scale FuXi-CFD dataset spanning diverse terrain and evaluate it across complementary observation settings, including controlled OSSEs, simulated helicopter trajectories, and real UAV measurements over Tromsø. Across these settings, assimilating sparse observations consistently improves local wind-field predictions, with WindINR outperforming several baselines, including WindSeer. Moreover, by performing online correction only in a low-dimensional latent space, WindINR achieves a 14 speedup over full-network fine-tuning. The resulting representation remains continuously queryable at arbitrary spatial coordinates, making WindINR a promising and efficient correction framework that provides a practical interface between kilometer-scale background products, sparse local observations, and wind queries in complex terrain.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
latent data assimilation, wind field estimation, ai for fluid dynamics

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_vggnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

WindINR uses a terrain- and background-conditioned implicit decoder for continuous local wind queries. A privileged encoder learns reference latents, a deployable predictor estimates initial latents, and training residuals define a Gaussian correction prior. Sparse observations update latent controls with fixed network weights. Experiments include 1,881 held-out CFD cases, an eight-case simulated flight pilot and three protocols from one real UAV episode. The main simulation reports 13.26% improvement over the uncorrected background at 1,024 observations and roughly 14× faster updates than the specified full-fine-tuning recipe.

Critical strengths
  1. A practical representation with substantial paired simulation evidence. Direct queries at observation locations avoid a separate interpolation operator, and freezing network weights makes online correction compact. Across 1,881 cases, the same checkpoint and solver improve the background at all four observation counts. At 1,024 observations, mean component RMSE falls from 0.4051 to 0.3514 m/s, with 99.68% of cases improving. This is stronger evidence than selected qualitative examples.
  2. Useful effort to separate fitting from verification. The random-observation study reserves query points, and the mobile pilot uses only observations already available and excludes queries within 50 m of them. The spatial-block real test has more than 100 m separation. The appendix also correctly limits the 14× figure to update time under fixed solver budgets, rather than end-to-end airborne latency.
Critical weaknesses
  1. The claimed prior contribution is not yet convincingly isolated. Latent assimilation with INRs already exists, including LAINR [1]; the main methodological increment is the conditional representation and its reference-to-prediction residual prior. Yet Table 6 uses a diagonal covariance inflated by 64 and a 3×3 spatial control grid. This does not demonstrate learned cross-channel covariance, and the inflation still controls regularization strength. The adaptive-prior gain over the isotropic variant is modest: 0.3574 to 0.3514 m/s at 1,024 observations (about 1.7%). The isotropic tuning protocol, mapping of global residual statistics to spatial controls, and held-out calibration are not sufficiently documented to attribute that gain to the claimed statistical insight.
  2. The empirical comparisons do not yet support the broader superiority and transfer claims. The WindSeer row is a local implementation inspired by the original, but Appendix B.5 does not establish fidelity to its masked volumetric CNN [2]. The original accepts variable sparse inputs, so excluding it from the mobile test because these local checkpoints use fixed counts is an implementation limitation. The FuXi-CFD number is imported from its published supplement rather than established as a paired rerun. Case-disjoint splitting is documented, but terrain-disjoint splitting is not. The real evidence is one episode: only 12 spatial-block holdouts, a 0.67% survey gain, and a stopped-transect suffix entirely within 50 m of observed locations. These caveats substantially narrow the demonstrated generalization.

[1] Li, Dong and Zhang. Latent assimilation with implicit neural representations for unknown dynamics. JCP 2024, 506, 112953. [2] Achermann et al. WindSeer: real-time volumetric wind prediction over complex terrain aboard a small uncrewed aerial vehicle. Nature Communications 2024.

Questions for authors
  1. How are the residual covariance and mean lifted to the 3×3 controls, and how were the scale 64, observation covariance and isotropic-prior strength chosen? A validation-tuned isotropic control and a scale sensitivity check on the same cases would determine whether the adaptive prior adds more than regularization tuning.
  2. What is the exact architecture behind the WindSeer label, and which original components are retained? Please distinguish a faithful adaptation from a new supervised observation encoder, and provide a variable-density comparator on the mobile protocol. Was the FuXi-CFD reference scored on identical cases, heights and query locations?
  3. Can the split manifest verify that terrain regions, including overlapping patches and alternative wind cases over the same terrain, are separated? For the real episode, please provide the uncorrected background score, sensor/error treatment and protocol-level uncertainty. The existing spatial-block test is relevant, but repeated nearby samples do not constitute independent terrain or event validation.
Reviewer_TqkhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes WindINR, a conditional implicit neural representation that maps static terrain descriptors, a low-resolution wind background, a continuous 3D query coordinate and a sample-specific latent state to the three wind components (u, v, w). Training has three stages: (1) a decoder is learned jointly with a privileged encoder that infers a reference latent from high-resolution support points; (2) with the decoder frozen, a deployable predictor learns to estimate an initial latent from terrain and background alone, using reconstruction plus alignment to the reference latent; (3) a Gaussian prior on latent corrections is estimated from the training-set discrepancies between reference and predicted latents. At inference, all network weights are frozen and sparse observations update only the latent by minimizing a 3D-Var-style objective (prior term plus observation misfit evaluated through the decoder), solved with L-BFGS. The stated contributions are (i) a latent-state conditional INR for terrain-aware local wind estimation with arbitrary-coordinate query, and (ii) a training and inference scheme for fast observation-guided latent correction using a dataset-adaptive prior. Validation consists of OSSEs on 1,881 held-out FuXi-CFD cases with 256–2,048 noisy random observations (13.26% mean component RMSE reduction at 1,024 observations, better than partial and full fine-tuning, a WindSeer-inspired supervised baseline, residual IDW and an isotropic-prior ablation); a simulated UAV-ahead-of-helicopter pilot on 8 cases (15.17% reduction over the background, about 14× faster than full fine-tuning); and a real-UAV evaluation at Tromsø with three protocols of 12–24 held-out points.

Critical strengths

● One frozen checkpoint serves any observation count, and the correction is consistently positive (beats the raw field in 99.68% of cases at 1,024 obs). ● Paired per-case statistics, a causal trajectory protocol with 50 m exclusion, and full per-case curves.

Critical weaknesses

● W1. “The deployed prior is diagonal, scaled ×64 with 0.1 shrinkage, and no selection procedure is given. The Iso. The ablation gap is only 0.8–1.8%, its weight is unreported, and it is absent outside the OSSE. ● W2. Eq. 5 uses one global latent, but experiments update a 3×3 spatial control grid (~1,152 dims, more than the 32–768 observed values) through an undescribed readout and prior. ● W3. FuXi-CFD numbers are copied from its paper, not rerun. “WindSeer” is a small MLP on WindINR’s own background, with unexplained val/test RMSE gap (~0.046). Full FT is worse than no correction at ≤1,024 obs. ● W4. The 14× speedup compares 8 L-BFGS vs. 100 Adam steps (Partial FT, few parameters, is still ~11× slower). The Tromsø table lacks the uncorrected background and CIs; the survey win is 0.67% on 19 points.

Questions for authors

● Q1. How were the covariance scale (64), shrinkage, 3×3 grid, and the Iso. regularization weight selected? ● Q2. How do the 3×3 controls feed the decoder, and how is their prior derived from the global training discrepancies Δz? ● Q3. Why is the supervised baseline’s validation RMSE (Table 8) ~0.046 higher than its test RMSE (Table 2)? ● Q4. What is the uncorrected background RMSE for each Tromsø protocol, and what observation error R_n was used?

Reviewer_jt5YOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This manuscript proposes a method to downscale wind at selected locations, without generating high spatial resolution wind solutions. The method is based on implicit neural representation. Large scale numerical experiment is used to validation the effectiveness of the proposed method.

Critical strengths

The paper proposes an interesting way to investigate downscaling of wind in weather modeling. Instead of fine resolution, the proposed method only focus on selected locations. The utilization of conditional implicit neural representation is a very interesting concept.

Critical weaknesses

The proposed method is not trained end-to-end. Instead it is constructed by training one component at a time. To fully validate the proposed architecture and training method are sound, a more structured ablation study is needed, to attribute the contribution of each (separately trained) component to the overall accuracy.

The author(s) provided substantial numerical results. But most of them are synthetic, for example, the helicopter-drone data is a synthetic data. The only observed data is from Tromsø region, however the results do not demonstrate the effectiveness of the proposed method (see questions for more details).

Questions for authors
  1. What is the rationale of the proposed structure show in Fig 2? How to tell which component is critical for the accuracy of the proposed model? My suggestion is to conduct a comprehensive set of ablation studies to investigate the accuracy contribution of different components. Such a study can also illustrate which component should be trained more thoroughly.

  2. Table 1, second column (under Raw mean). I am puzzled why all numbers are practically the same 0.4051.

  3. Fig 3b. What are the percentage RMSE? Are the results from both u and v components?

  4. Table 4. my reading is that this testcase is almost trivial. At which level is the wind data collected and forecasted? If it's at higher level, the wind will be dominated by the global model output, and has little to do with the local topography. By comparing the results between "Constant", and WindINR results, my feeling is this is the case. More results on wind velocity and wind direction will be needed for readers to better understand.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.