Pith. sign in

REVIEW 3 major objections 5 minor 46 references

This paper claims that a transformer conditioned on the latest in-situ observations can correct GFS marine wind forecasts, cutting RMSE by 45% at one hour and 13% at 48 hours across the Atlantic.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 18:44 UTC pith:25QCZBWK

load-bearing objection A useful, honest post-processing paper whose headline 45% gain is inflated by persistence-like inputs; the 13% at 48h is the more credible claim. the 3 major comments →

arxiv 2512.03606 v2 pith:25QCZBWK submitted 2025-12-03 cs.LG

Observation-driven correction of numerical weather prediction for marine winds

Matteo Peduto , Qidong Yang , Jonathan Giezendanner , Devis Tuia , Sherrie Wang This is my paper
classification cs.LG
keywords marine wind forecastingNWP correctiontransformerself-attentioncross-attentionin-situ observationsICOADSGFS post-processing
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to establish that marine wind forecasts from a global numerical model can be meaningfully improved by a learned, observation-informed correction — without retraining the dynamical model. The proposed transformer, ORCA, ingests the most recent ship, buoy, tide-gauge, and coastal-station measurements, pairs them with the corresponding GFS forecast, and outputs corrected winds at any requested coordinates. Across the Atlantic, it reduces GFS 10-meter wind RMSE at every lead time up to 48 hours, with a 45% improvement at 1 hour and 13% at 48 hours, and beats the ERA5 reanalysis up to 12 hours. The largest and most consistent gains appear where observations are densest — coastlines and shipping routes — while sparse mid-ocean regions show smaller, sometimes negative corrections. A sympathetic reader would care because the approach is a low-latency post-processing layer that could slot into operational forecasting pipelines rather than replacing NWP.

Core claim

ORCA reformulates marine wind forecasting as a correction task: instead of predicting winds from scratch, it learns state-dependent errors of the GFS forecast and adjusts them using the latest in-situ observations. The model represents each observation–forecast pair as a token, processes all such tokens with self-attention, and uses cross-attention to let arbitrary target locations query these tokens, producing corrected wind components in a single forward pass. On ICOADS data over the Atlantic, the model reduces GFS wind-speed RMSE by 45% at 1-hour lead time and by 13% at 48 hours, improves over ERA5 up to 12 hours, and shows the most consistent gains along coastlines and shipping corridors

What carries the argument

The central mechanism is a set-based transformer with two attention stages: self-attention over the current set of {observation, GFS forecast} pairs, which learns which past pairs matter, and cross-attention from target-location queries to those encoded pairs, which transfers the correction signal to arbitrary coordinates. Masking handles the irregular, time-varying observation sets; cyclical day-of-year and hour-of-day embeddings encode time; and latitude-longitude coordinates are mapped through spherical harmonics into a sinusoidal representation network so the model can be queried at any point on the ocean without grid interpolation. This design is what lets one model serve both site-spec

Load-bearing premise

The load-bearing premise is that the most recent observations are useful for correcting a forecast at any location regardless of distance; the paper adopts this from a correlation analysis but never tests the mid-ocean case where no recent observation exists nearby.

What would settle it

Run ORCA and a persistence baseline (latest observation used directly as the 1-hour forecast) on fixed platforms and compare—if ORCA does not beat persistence, the headline short-lead gain is inherited from the observations, not learned. Separately, evaluate a mid-ocean grid cell with no ICOADS observation within ~500 km and 6 hours; the paper's claims imply the model should still beat GFS there, which the reported spatial maps only partially support.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • GFS 10-m marine wind RMSE can be reduced at every lead time up to 48 h — 45% at 1 h, 13% at 48 h — so NWP post-processing with recent observations yields large near-term gains and persistent but smaller longer-term gains.
  • The corrector beats ERA5 reanalysis for leads up to 12 h, meaning observation-driven corrections can temporarily exceed a retrospectively optimized global field.
  • Single-pass inference at arbitrary coordinates generates corrected basin-scale fields in under 5 minutes on one GPU, making operational, low-latency forecast updates feasible.
  • Correction quality tracks observational density: fixed platforms and coasts gain most, drifting buoys and mid-ocean areas least, and some sparse-region cases degrade relative to GFS.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The 1-hour 45% gain is plausibly dominated by persistence — the latest observation is itself a near-perfect 1-hour forecast for a fixed platform — so the paper has not yet separated learned correction from observation carry-over; a persistence baseline would settle this.
  • Because the input set is not spatially capped in the paper's description, the same architecture could be tested with distance-limited or age-limited observation pools to find where the temporal-proximity assumption breaks down.
  • If near-real-time vessel weather reports were added to the pipeline, the tokenized design could densify coverage in exactly the mid-ocean regions where current gains are weakest.
  • The framework is not ocean-specific: the same pair-conditioning and arbitrary-coordinate inference could be retrained for other basins or variables (waves, currents) wherever paired forecast–observation histories exist.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes ORCA, a transformer-based post-processing method that corrects GFS 10-m wind forecasts over the North Atlantic by conditioning on recent ICOADS in-situ observations. The model is trained to predict wind at arbitrary target coordinates using self-attention over historical observation-forecast pairs and cross-attention from target tokens, with masking for irregular observation sets. Evaluation on a chronological held-out test period reports RMSE reductions relative to GFS at all lead times up to 48 h, with a 45.1% reduction at 1 h and 13.1% at 48 h (Table 2), and the paper demonstrates single-pass inference on a 0.25-degree Atlantic grid.

Significance. If the headline result holds, the paper makes a useful contribution to marine wind post-processing. The formulation of NWP correction as a set-based, observation-conditioned transformer is sensible, the chronological split is a reasonable leakage control, and the availability of code and data is a strength. The reported RMSE numbers are internally consistent, and Table A1 documents a set of negative architectural ablations. However, the central quantitative claims are currently difficult to interpret because the 1-h improvement is confounded with persistence of the latest observation, and the basin-scale claim is not quantitatively evaluated in data-sparse mid-ocean regions where the model is most likely to be used. The paper would be strengthened by adding a persistence baseline, ablating recent co-located observations, and evaluating at unobserved grid points.

major comments (3)
  1. [Sec. 4.1, Sec. 5.2, Table 1, Table 2] The headline 45.1% RMSE reduction at 1-h lead time (Table 2) is confounded with persistence. The conditioning set is defined as 'the most recent observations ... regardless of spatial distance' (Sec. 4.1), and test targets are 'all the observations at some future lead time' (Sec. 5.2). Table 1 shows that roughly two-thirds of ICOADS records come from hourly reporting fixed platforms (moored buoys, C-MAN, coastal stations, tide gauges). For a 1-h target at such a platform, the input set generally contains a co-located observation from the immediately preceding hour, and a cross-attention model can copy that value. No persistence baseline and no ablation removing or randomizing recent co-located observations are reported, so the 45% gain cannot be attributed to a learned NWP correction. Please add a persistence-of-latest-observation baseline and stratify results by whether a co-located pre
  2. [Sec. 6.3, Fig. A4, Sec. 6.2] The paper claims to improve GFS marine winds 'at all lead times' over the Atlantic, but the quantitative evaluation is only at ICOADS observation locations. No quantitative skill is reported at arbitrary mid-ocean grid points. The case studies in Sec. 6.3 explicitly acknowledge that in a mid-ocean region with sparse prior observations, 'adjustments farther west introduce distortions that result in lower accuracy compared to the GFS forecast.' Fig. A4 shows that degradation is present in some data-sparse cells, with low R² values (0.017-0.064) for the density-improvement relationship. To support the basin-scale contribution, the authors should evaluate corrected fields at unobserved grid points against an independent gridded reference (e.g., ERA5) and report skill as a function of distance and time to the nearest input observation.
  3. [Sec. 4.1, Sec. 4.2.3] The design decision to use 'the most recent observations ... regardless of spatial distance' is justified only by the correlation analysis in Fig. A1, but the actual input selection rule is not specified. The manuscript does not state how many observations are retained, whether there is a radius or cap, how subsampling is done, or how masks are constructed in practice. This matters for reproducibility and for the interpretation of the cross-attention mechanism: the model's behavior in data-sparse regions depends on the composition of the input set. Please specify the exact observation-selection rule used in training and inference, and report sensitivity to the number of input observations and to the time window.
minor comments (5)
  1. [Sec. 5.2, Fig. A3] The text states that the split is chronological, but the Fig. A3 caption says the subsets are spatially separated and 'ensuring spatial diversity and minimal overlap between regions.' Please clarify whether the split is purely temporal or also spatial, and adjust the caption accordingly.
  2. [Table A1] The 'Performance gain' column is not quantified. Report the actual RMSE differences (or lack thereof) for each architectural variation, and specify which lead time or averaged lead times are used.
  3. [Fig. 4b] The y-axis label 'Mean Reduction Error to GFS' is ambiguous. Define the sign convention explicitly (negative values indicate improvement) in the caption.
  4. [Sec. 6.1] The comparison to ERA5 is to a reanalysis, not a forecast. The statement that the model 'outperforms ERA5 up to 12h' should be phrased more carefully, since ERA5 benefits from later observations and is not a direct forecast baseline.
  5. [Fig. A2] The schematic of GFS cycle alignment is dense and hard to read. Consider simplifying the figure or adding a worked example with concrete timestamps.

Circularity Check

0 steps flagged

No significant circularity: the reported improvements are empirical held-out RMSE comparisons, not reductions to model inputs or self-citations.

full rationale

The paper's derivation chain is a supervised post-processing setup: a transformer takes recent observation-GFS pairs as input and is trained to minimize vector error against future ICOADS observations, with a chronological train/validation/test split (Sec. 5.2). Table 2 reports held-out RMSE against GFS and ERA5. The target values are not used to define the model, fit its parameters, or select the correction formula, so the 45%/13% improvement is not true by construction. The 1-h result may be influenced by temporal autocorrelation for hourly fixed platforms, and the paper does not quantify mid-ocean/no-nearby-observation skill; these are evaluation-attribution concerns, not circularity. Self-citations to Yang et al. (2024) for the correction framing and for preferred architecture are present, but the central result does not depend on accepting those citations: it rests on this paper's own held-out experiments. No uniqueness theorem, fitted parameter renamed as prediction, or ansatz-smuggled-via-citation was found. Per instructions, self-citation without a load-bearing reduction does not raise the circularity score.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The ledger contains no invented physical entities. The paper's contribution is an empirical correction function; its free parameters are the learned network weights plus several under-specified design choices (context length, token set size, encoder capacity). The key domain assumptions are that ICOADS is a trustworthy reference, that GFS error patterns are stationary, that temporal proximity dominates spatial distance, and that observation-location metrics represent basin-wide skill.

free parameters (4)
  • Neural network weights and biases = Learned from 27.8M training samples
    The correction function is entirely data-driven; no closed-form or physically parameterized error model is provided, so the central result depends on these fitted values.
  • Input time context length = One previous time step (bΔt=1)
    Ablation with two previous hours gave no gain (Table A1); the model sees only the immediately preceding observation set, which is the main driver of 1-h skill.
  • Observation token set size / cap / radius = Not stated
    No cap, subsampling, or radius is reported for the 'most recent observations' used as input; this affects attention behavior and reproducibility in dense versus sparse regions.
  • Geographic encoder capacity = Spherical-harmonic degree and SirenNet dimensions not reported
    Positional encoding follows Rußwurm et al. (2023), but the specific degree and widths are omitted, so arbitrary-coordinate generalization cannot be reproduced exactly.
axioms (4)
  • domain assumption ICOADS observations are accurate, quality-controlled ground truth for 10-m marine winds at their reported locations and times.
    All training labels, validation, and test references are ICOADS reports (Section 3.1); observation error and QC limitations are not modeled, and no independent higher-quality reference is used.
  • domain assumption GFS error statistics are stationary enough that a model trained on 2015-2022 data corrects 2023-2024 forecasts.
    The chronological split (Section 5.2) assumes no significant distribution shift in GFS biases or observation sampling over the test period; no shift analysis is provided.
  • domain assumption Temporal proximity dominates spatial proximity for predictive correlation, so spatial distance can be ignored in selecting input observations.
    Section 4.1 and Figure A1 justify this design choice from pairwise correlations, but the paper does not show that the learned attention can compensate for the absence of nearby observations in mid-ocean regions.
  • domain assumption Quantitative evaluation at ICOADS observation locations is representative of Atlantic-basin performance.
    Numbers in Table 2 are computed only where in-situ observations exist (Section 6.1); the basin-scale gridded inference in Section 6.3 is presented qualitatively through case studies.

pith-pipeline@v1.3.0-alltime-deepseek · 15470 in / 11992 out tokens · 117654 ms · 2026-08-03T18:44:12.155957+00:00 · methodology

0 comments
read the original abstract

Accurate marine wind forecasts are essential for safe navigation, ship routing, and energy operations, yet they remain challenging because observations over the ocean are sparse, heterogeneous, and temporally variable. We present an observation-informed correction approach for global numerical weather prediction (NWP) of marine winds. Rather than forecasting winds directly, we learn local correction patterns by assimilating the latest in-situ observations to adjust the Global Forecast System (GFS) output. We propose ORCA (Observation-informed Real-time Correction with Attention), a transformer-based deep learning architecture that (i) handles irregular and time-varying observation sets through masking and set-based attention mechanisms, (ii) conditions predictions on recent observation--forecast pairs via cross-attention, and (iii) employs cyclical time embeddings and coordinate-aware location representations to enable single-pass inference at arbitrary spatial coordinates. We evaluate ORCA over the Atlantic Ocean using observations from the International Comprehensive Ocean-Atmosphere Data Set (ICOADS) as reference. ORCA reduces GFS 10-meter wind error at all lead times up to 48 hours, achieving 45% improvement at 1-hour lead time and 13% improvement at 48-hour lead time. Spatial analyses reveal the most persistent improvements along coastlines and shipping routes, where observations are most abundant. The tokenized architecture naturally accommodates heterogeneous observing platforms (ships, buoys, tide gauges, and coastal stations) and produces both site-specific predictions and basin-scale gridded products in a single forward pass. These results demonstrate a practical, low-latency post-processing approach that complements NWP by learning to correct systematic forecast errors.

Figures

Figures reproduced from arXiv: 2512.03606 by Devis Tuia, Jonathan Giezendanner, Matteo Peduto, Qidong Yang, Sherrie Wang.

Figure 1
Figure 1. Figure 1: Comparison of spatial and statistical wind patterns from GFS forecasts, in-situ [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Spatial density of ICOADS observations by platform type (2015–2024). Higher [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Schematic representation of the spatio–temporal learning framework. Each [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of model performance across forecast lead times and observation [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Spatial distribution of wind-speed prediction errors for the machine-learning [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of our model wind fields, GFS field, and their differences across [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Case studies illustrating localized wind-field corrections across three represen [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

46 extracted references · 3 canonical work pages

  1. [1]

    atlas2011ccmp APACrefauthors Atlas, R. \ . APACrefauthors \ 2011 . A cross-calibrated, multiplatform ocean surface wind velocity product for meteorological and oceanographic applications A cross-calibrated, multiplatform ocean surface wind velocity product for meteorological and oceanographic applications . Bulletin of the American Meteorological Society ...

    2011

  2. [2]

    APACrefauthors \ 2023

    awe_icoads_decline APACrefauthors AWE International . APACrefauthors \ 2023 . New Insights from Old Ship Data: Challenges in Maintaining Global Marine Observations. New insights from old ship data: Challenges in maintaining global marine observations. https://www.awe.international/article/1841509/new-insights-old-ship-data . Accessed: October 2025

    arXiv 2023

  3. [3]

    bi2023pangu APACrefauthors Bi, K. \ . APACrefauthors \ 2023 . Accurate medium-range global weather forecasting with 3D neural networks Accurate medium-range global weather forecasting with 3d neural networks . Nature 619 533--538

    2023

  4. [4]

    APACrefauthors \ 2017

    copernicus_climate_change_service APACrefauthors Copernicus Climate Change Service (C3S) . APACrefauthors \ 2017 . ERA5: Fifth generation of ECMWF atmospheric reanalyses of the global climate. Era5: Fifth generation of ecmwf atmospheric reanalyses of the global climate. https://cds.climate.copernicus.eu . Copernicus Climate Change Service

    2017

  5. [5]

    dong2024marine_gnn APACrefauthors Dong, D. \ . APACrefauthors \ 2024 . Short-Term Marine Wind Speed Forecasting Based on Dynamic Graph Embedding and Spatiotemporal Information Short-term marine wind speed forecasting based on dynamic graph embedding and spatiotemporal information . Journal of Marine Science and Engineering 12 3 502

    2024

  6. [6]

    APACrefauthors \ 2023

    ecmwf_aeolus_impact APACrefauthors ECMWF . APACrefauthors \ 2023 . The NWP impact of Aeolus Level-2B winds at ECMWF The nwp impact of aeolus level-2b winds at ecmwf \ . ECMWF Technical Memo . APACrefURL https://www.ecmwf.int/en/elibrary/81546-nwp-impact-aeolus-level-2b-winds-ecmwf APACrefURL

    2023

  7. [7]

    APACrefauthors \ 2024

    ECMWF APACrefauthors European Centre for Medium-Range Weather Forecasts (ECMWF) . APACrefauthors \ 2024 . ECMWF Reanalysis Documentation. Ecmwf reanalysis documentation. https://www.ecmwf.int/en/forecasts/datasets/reanalysis-datasets/era5

    2024

  8. [8]

    \ Lowry, D

    glahn1972mos APACrefauthors Glahn, H. \ Lowry, D. APACrefauthors \ 1972 . The use of Model Output Statistics (MOS) in objective weather forecasting The use of model output statistics (mos) in objective weather forecasting . Journal of Applied Meteorology 11 8 1203--1211

    1972

  9. [9]

    , Raftery, A E

    gneiting2005emos APACrefauthors Gneiting, T. , Raftery, A E. , Westveld, A H. \ Goldman, T. APACrefauthors \ 2005 . Calibrated probabilistic forecasting using ensemble model output statistics and minimum CRPS estimation Calibrated probabilistic forecasting using ensemble model output statistics and minimum crps estimation . Monthly Weather Review 133 5 1098--1118

    2005

  10. [10]

    APACrefauthors \ 2023

    metnet3_blog APACrefauthors Google Research & DeepMind . APACrefauthors \ 2023 . MetNet-3: A state-of-the-art neural weather model available in Google products. Metnet-3: A state-of-the-art neural weather model available in google products. Blog post . APACrefURL https://research.google/blog/metnet-3-a-state-of-the-art-neural-weather-model-available-in-go...

    2023

  11. [11]

    han2023offshore_dnn APACrefauthors Han, X. \ . APACrefauthors \ 2023 . Enhanced Offshore Wind Speed Forecasts along the U.S. East Coast: A Deep Learning Framework Enhanced offshore wind speed forecasts along the u.s. east coast: A deep learning framework . Open Research Europe 3 31

    2023

  12. [12]

    , Bell, B

    hersbach2020era5 APACrefauthors Hersbach, H. , Bell, B. , Berrisford, P. , Hirahara, S. , Hor \'a nyi, A. , Mu \ n oz-Sabater, J. Th \'e paut, J N. APACrefauthors \ 2020 . The ERA5 global reanalysis The era5 global reanalysis . Quarterly Journal of the Royal Meteorological Society 146 730 1999--2049 . APACrefDOI doi:10.1002/qj.3803 APACrefDOI

    doi:10.1002/qj.3803 2020

  13. [13]

    jaegle2021perceiverio APACrefauthors Jaegle, A. \ . APACrefauthors \ 2021 . Perceiver IO: A general architecture for structured inputs and outputs Perceiver io: A general architecture for structured inputs and outputs . ICML. Icml

    2021

  14. [14]

    , Dehghani, M

    kazemnejad2019positional APACrefauthors Kazemnejad, A. , Dehghani, M. , Bahri, D. \ Tay, Y. APACrefauthors \ 2019 . Positional Encoding in Transformers Positional encoding in transformers . arXiv preprint arXiv:1908.06982 . APACrefURL https://arxiv.org/abs/1908.06982 APACrefURL

    Pith/arXiv arXiv 2019

  15. [15]

    , Hinrichs, M

    researchgate_ais_weather APACrefauthors Kumpf, A. , Hinrichs, M. \ Isermann, J. APACrefauthors \ 2021 . Improving Meteorological Models Using Ships’ Weather Data Communicated via AIS Improving meteorological models using ships’ weather data communicated via ais . Journal of Marine Science and Engineering . APACrefURL https://www.researchgate.net/publicati...

    arXiv 2021

  16. [16]

    lam2023graphcast APACrefauthors Lam, R T. \ . APACrefauthors \ 2023 . Learning skillful medium-range global weather forecasting Learning skillful medium-range global weather forecasting . Science 382 6669 1416--1421

    2023

  17. [17]

    lee2019set APACrefauthors Lee, J. \ . APACrefauthors \ 2019 . Set Transformer: A framework for attention-based permutation-invariant neural networks Set transformer: A framework for attention-based permutation-invariant neural networks . ICML. Icml

    2019

  18. [18]

    lin2021cat APACrefauthors Lin, C. , Li, S. , Zhang, Y. \ Ding, Z. APACrefauthors \ 2021 . CAT: Cross Attention in Vision Transformer Cat: Cross attention in vision transformer . Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops Proceedings of the ieee/cvf international conference on computer vision workshops \ ( \ 0--0). AP...

    Pith/arXiv arXiv 2021

  19. [19]

    liu2023agn_gnn APACrefauthors Liu, J. \ . APACrefauthors \ 2023 . Adaptive graph-learning convolutional network for multi-node offshore wind forecasting Adaptive graph-learning convolutional network for multi-node offshore wind forecasting . Journal of Marine Science and Engineering 11 4 879

    2023

  20. [20]

    APACrefauthors \ 2024

    climatedataguide_icoads APACrefauthors National Center for Atmospheric Research . APACrefauthors \ 2024 . ICOADS: Surface marine weather observations. Icoads: Surface marine weather observations. https://climatedataguide.ucar.edu/climate-data/icoads-surface-marine-weather-observations . Accessed: October 2025

    2024

  21. [21]

    APACrefauthors \ 2020

    ncep_gfs APACrefauthors National Centers for Environmental Prediction (NCEP) . APACrefauthors \ 2020 . NCEP Global Forecast System (GFS) Documentation Ncep global forecast system (gfs) documentation \ . NOAA/National Weather Service . https://www.nco.ncep.noaa.gov/pmb/products/gfs/

    2020

  22. [22]

    APACrefauthors \ 2025

    NOAA_ML4BC_2025 APACrefauthors NOAA/NWS/NCEP . APACrefauthors \ 2025 . A Machine Learning-Based Bias Correction Method for 2-m Temperature of the GFSv16 A machine learning-based bias correction method for 2-m temperature of the gfsv16 \ . NOAA NCEP Office Note 520 . APACrefURL https://library.oarcloud.noaa.gov/noaa_documents.lib/NWS/NCEP/NCEP_office_notes...

    2025

  23. [23]

    pathak2022fourcastnet APACrefauthors Pathak, J. \ . APACrefauthors \ 2022 . FourCastNet: A global data-driven high-resolution weather model using adaptive Fourier neural operators Fourcastnet: A global data-driven high-resolution weather model using adaptive fourier neural operators . arXiv:2202.11214

    Pith/arXiv arXiv 2022

  24. [24]

    metnet3_arxiv APACrefauthors Pathak, J. \ . APACrefauthors \ 2023 . Deep Learning for Day Forecasts from Sparse Observations Deep learning for day forecasts from sparse observations . arXiv:2306.06079

    Pith/arXiv arXiv 2023

  25. [25]

    , Kubovčík, M

    Pospichal2022Solar APACrefauthors Pospíchal, J. , Kubovčík, M. \ Dirgov\' a -Lupt\' a kov\' a , I. APACrefauthors \ 2022 . Solar Irradiance Forecasting with Transformer Model Solar irradiance forecasting with transformer model . Applied Sciences 12 17 8852 . APACrefURL https://www.mdpi.com/2076-3417/12/17/8852 APACrefURL APACrefDOI doi:10.3390/app12178852...

    work page doi:10.3390/app12178852 2022

  26. [26]

    , Gneiting, T

    raftery2005bma APACrefauthors Raftery, A E. , Gneiting, T. , Balabdaoui, F. \ Polakowski, M. APACrefauthors \ 2005 . Using Bayesian model averaging to calibrate forecast ensembles Using bayesian model averaging to calibrate forecast ensembles . Monthly Weather Review 133 5 1155--1174

    2005

  27. [27]

    APACrefauthors \ 2023

    raschka2023selfattention APACrefauthors Raschka, S. APACrefauthors \ 2023 . A Tour of Self-Attention A tour of self-attention . arXiv preprint arXiv:2308.13086 . APACrefURL https://arxiv.org/abs/2308.13086 APACrefURL

    Pith/arXiv arXiv 2023

  28. [28]

    \ Lerch, S

    rasp2018nnpost APACrefauthors Rasp, S. \ Lerch, S. APACrefauthors \ 2018 . Neural networks for postprocessing ensemble weather forecasts Neural networks for postprocessing ensemble weather forecasts . Monthly Weather Review 146 11 3885--3900

    2018

  29. [29]

    APACrefauthors \ 2024

    ccmp_remss APACrefauthors Remote Sensing Systems . APACrefauthors \ 2024 . CCMP Wind Vector Analysis Product. Ccmp wind vector analysis product. Product documentation . APACrefURL https://www.remss.com/measurements/ccmp/ APACrefURL

    2024

  30. [30]

    rennie2021aeolus APACrefauthors Rennie, M P. \ . APACrefauthors \ 2021 . The impact of Aeolus wind retrievals on ECMWF global weather forecasts The impact of aeolus wind retrievals on ecmwf global weather forecasts . Quarterly Journal of the Royal Meteorological Society 147 740 3555--3586

    2021

  31. [31]

    , Klemmer, K

    russwurm2023locationencoder APACrefauthors Ru wurm, M. , Klemmer, K. , Rolf, E. , Zbinden, R. \ Tuia, D. APACrefauthors \ 2023 . Geographic Location Encoding with Spherical Harmonics and Sinusoidal Representation Networks Geographic location encoding with spherical harmonics and sinusoidal representation networks . arXiv:2310.06743

    Pith/arXiv arXiv 2023

  32. [32]

    \ Hamill, T M

    scheuerer2020qrforests APACrefauthors Scheuerer, M. \ Hamill, T M. APACrefauthors \ 2020 . Postprocessing of ensemble temperature forecasts via quantile regression forests Postprocessing of ensemble temperature forecasts via quantile regression forests . Monthly Weather Review . Early ML post-processing example

    2020

  33. [33]

    sitzmann2020siren APACrefauthors Sitzmann, V. \ . APACrefauthors \ 2020 . Implicit Neural Representations with Periodic Activation Functions Implicit neural representations with periodic activation functions . NeurIPS. Neurips

    2020

  34. [34]

    , Zuo, X

    Su2025systematic APACrefauthors Su, L. , Zuo, X. , Li, R. , Wang, X. , Zhao, H. \ Huang, B. APACrefauthors \ 2025 . A systematic review for transformer-based long-term series forecasting A systematic review for transformer-based long-term series forecasting . Artificial Intelligence Review 58 80 . APACrefDOI doi:10.1007/s10462-024-11044-2 APACrefDOI

    work page doi:10.1007/s10462-024-11044-2 2025

  35. [35]

    sun2023gfs_precip APACrefauthors Sun, D. \ . APACrefauthors \ 2023 . Deep Learning Improves GFS Wintertime Precipitation Forecasts over the Southeastern United States Deep learning improves gfs wintertime precipitation forecasts over the southeastern united states . Geophysical Research Letters 50 22 e2023GL104406

    2023

  36. [36]

    tancik2020fourier APACrefauthors Tancik, M. \ . APACrefauthors \ 2020 . Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains Fourier features let networks learn high frequency functions in low dimensional domains . NeurIPS. Neurips

    2020

  37. [37]

    Army Engineer Research and Development Center

    erdc_ais_weather APACrefauthors U.S. Army Engineer Research and Development Center . APACrefauthors \ 2024 . Sharing Ships’ Weather Data via AIS: Concept and Results from Multiyear Observations. Sharing ships’ weather data via ais: Concept and results from multiyear observations. https://www.erdc.usace.army.mil/Media/Publication-Notices/Article/3836011/sh...

    arXiv 2024

  38. [38]

    , Shazeer, N

    vaswani2017attention APACrefauthors Vaswani, A. , Shazeer, N. , Parmar, N. , Uszkoreit, J. , Jones, L. , Gomez, A N. Polosukhin, I. APACrefauthors \ 2017 . Attention Is All You Need Attention is all you need . Advances in Neural Information Processing Systems Advances in neural information processing systems \ ( \ 5998--6008)

    2017

  39. [39]

    APACrefauthors \ 2011

    wilks2011statistical APACrefauthors Wilks, D S. APACrefauthors \ 2011 . Statistical Methods in the Atmospheric Sciences Statistical methods in the atmospheric sciences \ ( 3rd \ ). Academic Press

    2011

  40. [40]

    , Giezendanner, J

    yang2024local_off_grid APACrefauthors Yang, Q. , Giezendanner, J. , Civitarese, D S. , Jakubik, J. , Schmitt, E. , Chandra, A. Wang, S. APACrefauthors \ 2024 . Local Off-Grid Weather Forecasting with Multi-Modal Earth Observation Data Local off-grid weather forecasting with multi-modal earth observation data . arXiv preprint arXiv:2410.12938 . APACrefURL ...

    Pith/arXiv arXiv 2024

  41. [41]

    , Brodie, J

    airu_wrf APACrefauthors Ye, F. , Brodie, J. , Miles, T. \ Ezzat, A A. APACrefauthors \ 2023 . AIRU-WRF : A physics-guided spatio-temporal wind forecasting model for U.S. Mid-Atlantic offshore AIRU-WRF : A physics-guided spatio-temporal wind forecasting model for u.s. mid-atlantic offshore . arXiv:2303.02246. arxiv:2303.02246

    Pith/arXiv arXiv 2023

  42. [42]

    , Zhang, X

    deepmide_2024 APACrefauthors Ye, F. , Zhang, X. , Stein, M. \ Ezzat, A A. APACrefauthors \ 2024 . DeepMIDE: A Multivariate Spatio-Temporal Method for Ultra-Scale Offshore Wind Energy Forecasting Deepmide: A multivariate spatio-temporal method for ultra-scale offshore wind energy forecasting . arXiv:2410.20166

    arXiv 2024

  43. [43]

    , Zhang, L

    Zhang2024_wind_climatology APACrefauthors Zhang, W. , Zhang, L. \ Liu, J. APACrefauthors \ 2024 . Satellite‑based re‑examination of changes in terrestrial near‑surface wind speed in the last 30 years Satellite‑based re‑examination of changes in terrestrial near‑surface wind speed in the last 30 years . Frontiers in Earth Science 12 . APACrefURL https://do...

    arXiv 2024

  44. [44]

    \ Hasager, C B

    zuo2023aeolus APACrefauthors Zuo, H. \ Hasager, C B. APACrefauthors \ 2023 . The impact of Aeolus winds on near-surface wind forecasts over tropical ocean and high-latitude regions The impact of aeolus winds on near-surface wind forecasts over tropical ocean and high-latitude regions . Atmospheric Measurement Techniques 16 3901--3913 . APACrefDOI doi:10.5...

    work page doi:10.5194/amt-16-3901-2023 2023

  45. [45]

    apacite url apacite =6pt Acknowledgments. 6pt 1sp \@dates Received \@recvdate\@empty\@rcvaccrule \@recvdate \@revisedate\@empty ; revised \@revisedate; \@accptdate\@empty \@revisedate\@empty; accepted \@accptdate \@pubdate\@empty. ; published \@pubdate. -2pt \@authaddrs @list\@empty =.15in @list 1sp @list =9pt plus 2pt minus 6pt \@sluginfo width 4pc =3000...

    2001

  46. [46]

    gc" journal option for G-Cubed Nov 3, 2003 M Kelly, fixed noindent in subsubsubsection titles and for all sections in rog option Oct 2, 2003 M Kelly, added

    \@ifstar \@figbox \@figbox \@figbox#1#2#3 to !#1! #3 [#1][c] !#2!#3 \@tempdima#2 \@tempdima by2 \@tempdima by- \@tempdima by- \@height\@tempdima\@depth\@tempdima\@width @ to @ #3 Bib ??? ??? ??? =0 =0 = @figure=0 @table=0 #1 --#1 -24pt -2ex #1 0= #1 to 0 #1 I NDEX T ERMS: #1 #1 Citation: #1 Feb 9, 2009 Changed name and references to name from agu2001 to a...

    arXiv 2009