pith. sign in

arxiv: 1906.08066 · v1 · pith:Y4H334IBnew · submitted 2019-06-19 · ⚛️ physics.ao-ph

Predicting the Ocean Currents using Deep Learning

Cihan Bayindir This is my paper

Pith reviewed 2026-05-25 20:02 UTC · model grok-4.3

classification ⚛️ physics.ao-ph
keywords ocean currentsLSTMdeep learningcurrent predictionMassachusetts Bayu and v velocitiesspectral properties
0
0 comments X

The pith

LSTM networks can predict ocean current speeds in two horizontal directions from historical measurements.

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

The paper trains a Long Short Term Memory network on NOAA data collected in Massachusetts Bay from late 2002 to early 2003. It demonstrates that the model can forecast the east-west and north-south current components, called u and v. The work shows how the size and resolution of the training set change both the size of the forecast errors and the frequency content of the predicted time series. If these forecasts hold, they could support calculations of tidal energy output and estimates of forces on underwater structures. The results also indicate that higher-resolution data can extend the useful prediction horizon in both time and space.

Core claim

An LSTM network applied to the 2002-2003 Massachusetts Bay current records produces forecasts of the two horizontal velocity components u and v whose accuracy and spectral character depend on the temporal and spatial resolution of the training data.

What carries the argument

The Long Short Term Memory (LSTM) network that processes sequences of ocean current measurements to output future values of u and v.

If this is right

  • Forecast skill for u and v improves or degrades directly with the temporal or spatial density of the training records.
  • The frequency spectrum of the LSTM output can be made to match the measured spectrum more closely by selecting appropriate training intervals.
  • Prediction horizons in time and distance become longer when the data resolution is increased.
  • The same trained model supplies inputs for tidal-energy estimates and for calculations of current-induced forces on marine structures.

Where Pith is reading between the lines

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

  • The method could be tested on real-time buoy streams to see whether forecasts remain stable when new measurements arrive continuously.
  • Pairing the LSTM output with a simple physical circulation model might reduce drift over multi-day horizons.
  • If the spectral fidelity holds across sites, the approach could help locate regions where chaotic currents block surface waves.

Load-bearing premise

The accuracy achieved on the single 2002-2003 Massachusetts Bay training period will continue when the same model is applied to other periods or other coastal sites.

What would settle it

Running the trained LSTM on current measurements from a later year or a different bay and finding that the predicted u and v series deviate substantially from the recorded speeds or lose the correct frequency peaks.

Figures

Figures reproduced from arXiv: 1906.08066 by Cihan Bayindir.

Figure 1
Figure 1. Figure 1: FIG. 1: LSTM layer architecture [33] [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2: LSTM gates [33] [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3: LSTM components and formula [33]. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: We compare the predicted and observed time series and depict the root-mean-square (rms) error in Fig. 6. As [PITH_FULL_IMAGE:figures/full_fig_p003_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4: Ocean current speed time series a) u (m/s) b) v (m/s). [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5: Observed and predicted time series of the first component of the current velocity (u) obtained using the initial 95 % [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6: a) Comparisons between the observed and predicted time series of the first component of the current velocity (u) [PITH_FULL_IMAGE:figures/full_fig_p004_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIG. 7: Observed and predicted time series of the first component of the current velocity (u) obtained using the initial 95 % [PITH_FULL_IMAGE:figures/full_fig_p005_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: FIG. 8: a) Comparisons between the observed and predicted time series of the first component of the current velocity (u) [PITH_FULL_IMAGE:figures/full_fig_p005_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: FIG. 9: Observed and predicted time series of the second component of the current velocity (v) obtained using the initial 95 [PITH_FULL_IMAGE:figures/full_fig_p006_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: FIG. 10: a) Comparisons between the observed and predicted time series of the second component of the current velocity (v) [PITH_FULL_IMAGE:figures/full_fig_p006_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: FIG. 11: Observed and predicted time series of the second component of the current velocity (v) obtained using the initial 95 [PITH_FULL_IMAGE:figures/full_fig_p006_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: FIG. 12: a) Comparisons between the observed and predicted time series of the second component of the current velocity (v) [PITH_FULL_IMAGE:figures/full_fig_p007_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: FIG. 13: a) Comparisons between the observed and predicted time series of the first component of the current velocity (u) [PITH_FULL_IMAGE:figures/full_fig_p007_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: FIG. 14: Comparison of the Fourier spectra of the observed and predicted time series of the first component of the current [PITH_FULL_IMAGE:figures/full_fig_p007_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: FIG. 15: a) Comparisons between the observed and predicted time series of the first component of the current velocity (u) [PITH_FULL_IMAGE:figures/full_fig_p008_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: FIG. 16: Comparison of the Fourier spectra of the observed and predicted time series of the first component of the current [PITH_FULL_IMAGE:figures/full_fig_p008_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: FIG. 17: a) Comparisons between the observed and predicted time series of the second component of the current velocity (v) [PITH_FULL_IMAGE:figures/full_fig_p009_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: FIG. 18: a) Comparisons between the observed and predicted time series of the second component of the current velocity (v) [PITH_FULL_IMAGE:figures/full_fig_p009_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: FIG. 19: Comparison of the Fourier spectra of the observed and predicted time series of the second component of the current [PITH_FULL_IMAGE:figures/full_fig_p009_19.png] view at source ↗
read the original abstract

In this paper, we analyze the predictability of the ocean currents using deep learning. More specifically, we apply the Long Short Term Memory (LSTM) deep learning network to a data set collected by the National Oceanic and Atmospheric Administration (NOAA) in Massachusetts Bay between November 2002-February 2003. We show that the current speed in two horizontal directions, namely u and v, can be predicted using the LSTM. We discuss the effect of training data set on the prediction error and on the spectral properties of predictions. Depending on the temporal or the spatial resolution of the data, the prediction times and distances can vary, and in some cases, they can be very beneficial for the prediction of the ocean current parameters. Our results can find many important applications including but are not limited to predicting the tidal energy variation, controlling the current induced vibrations of marine structures and estimation of the wave blocking point by the chaotic oceanic current and circulation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The manuscript applies Long Short-Term Memory (LSTM) networks to predict the u and v components of ocean currents from NOAA buoy measurements collected in Massachusetts Bay between November 2002 and February 2003. It examines how training dataset size and temporal/spatial resolution affect prediction error and the spectral properties of the forecasts, and suggests applications including tidal energy variation, marine structure vibrations, and wave blocking estimation.

Significance. If the reported LSTM forecasts were shown to outperform standard baselines on temporally or spatially disjoint test data while preserving spectral content, the work would establish a viable data-driven method for short-term ocean current prediction. The focus on resolution-dependent prediction horizons and spectral fidelity would then provide concrete guidance on data requirements for operational forecasting in ocean engineering and renewable energy contexts.

major comments (3)
  1. [Abstract] Abstract: The central claim that 'the current speed in two horizontal directions, namely u and v, can be predicted using the LSTM' is presented without any quantitative error statistics (RMSE, correlation, etc.), baseline comparisons, or description of the train-test split, rendering the result an unverified assertion rather than a demonstrated finding.
  2. [Experiments/Results] Experiments/Results sections: All reported predictions and spectral comparisons are generated from the single 2002-2003 Massachusetts Bay dataset with no held-out seasons, locations, or cross-validation; without persistence, AR(1), or linear-regression baselines, any apparent accuracy cannot be attributed to learned dynamics rather than short-term autocorrelation already present in the buoy time series.
  3. [Methods] Methods: No model hyperparameters, regularization strategy, optimizer settings, or loss function are specified, and no numerical values for prediction error as a function of training-set size or resolution are tabulated, preventing assessment or reproduction of the claimed dependence of prediction times/distances on data resolution.
minor comments (1)
  1. [Abstract] The terms 'prediction times and distances' are used in the abstract without explicit definition or units, and the manuscript does not clarify whether these refer to forecast lead time, spatial extrapolation distance, or both.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major comment below and indicate the revisions planned.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that 'the current speed in two horizontal directions, namely u and v, can be predicted using the LSTM' is presented without any quantitative error statistics (RMSE, correlation, etc.), baseline comparisons, or description of the train-test split, rendering the result an unverified assertion rather than a demonstrated finding.

    Authors: We agree that the abstract should include quantitative support for the central claim. In the revised manuscript we will update the abstract to report key error statistics (RMSE and correlation for u and v) and briefly note the temporal train-test split employed in the study. revision: yes

  2. Referee: [Experiments/Results] Experiments/Results sections: All reported predictions and spectral comparisons are generated from the single 2002-2003 Massachusetts Bay dataset with no held-out seasons, locations, or cross-validation; without persistence, AR(1), or linear-regression baselines, any apparent accuracy cannot be attributed to learned dynamics rather than short-term autocorrelation already present in the buoy time series.

    Authors: The manuscript is an exploratory study of how training-set size and resolution affect error and spectral fidelity on this specific dataset. We will add persistence and AR(1) baseline comparisons in the revised results to help attribute performance beyond autocorrelation. The original analysis used a temporal split within the available record; we will clarify this and add time-series cross-validation where feasible. However, the study contains only one buoy record from a single season and location. revision: partial

  3. Referee: [Methods] Methods: No model hyperparameters, regularization strategy, optimizer settings, or loss function are specified, and no numerical values for prediction error as a function of training-set size or resolution are tabulated, preventing assessment or reproduction of the claimed dependence of prediction times/distances on data resolution.

    Authors: We will expand the Methods section to specify the LSTM architecture (layers, hidden units), all hyperparameters, regularization, optimizer, and loss function. We will also add a table reporting numerical prediction errors as a function of training-set size and resolution to support reproducibility and the claimed dependence. revision: yes

standing simulated objections not resolved
  • The study is restricted to a single NOAA buoy record from one location and four-month period; results on temporally or spatially disjoint test data from other seasons or locations cannot be provided without new measurements.

Circularity Check

0 steps flagged

No circularity: LSTM predictions are generated by trained network on data, not algebraically equivalent to inputs.

full rationale

The paper applies a standard LSTM architecture to time-series buoy data from a single NOAA deployment. The claimed predictions of u and v components are produced by forward passes of the trained network on held-out segments of that data; they are not obtained by algebraic rearrangement of the training inputs, by re-using fitted parameters as the output, or by any self-referential definition. No uniqueness theorem, ansatz, or prior self-citation is invoked to force the result. The central claim therefore remains an empirical statement about model performance rather than a tautology. Minor self-citations, if present, are not load-bearing for the prediction step itself.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The work rests on standard neural-network assumptions and the representativeness of a single four-month buoy record; no new physical entities are postulated.

free parameters (1)
  • LSTM architecture hyperparameters
    Number of layers, hidden units, learning rate, and sequence length are chosen to fit the training data.
axioms (1)
  • domain assumption The measured current time series contains learnable temporal patterns that an LSTM can exploit for multi-step forecasting.
    Invoked when the authors state that u and v can be predicted from the data.

pith-pipeline@v0.9.0 · 5678 in / 1158 out tokens · 27144 ms · 2026-05-25T20:02:08.145750+00:00 · methodology

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

34 extracted references · 34 canonical work pages

  1. [1]

    Richardson, Benjamin Franklin and Timothy Folgers first printed chart of the Gulf Stream, Science, 207, 643, (1980)

    P. Richardson, Benjamin Franklin and Timothy Folgers first printed chart of the Gulf Stream, Science, 207, 643, (1980)

    work page 1980

  2. [2]

    R. G. Peterson and L. Stramma and G. Kortum, Early concepts and charts of ocean circulation, Prog. Oceanogr., 37, 1, (1996)

    work page 1996

  3. [3]

    Munk, On the wind driven ocean circulation, J

    W.H. Munk, On the wind driven ocean circulation, J. Meteor., 7, 79 (1950)

    work page 1950

  4. [4]

    H. U. Sverdrup, Wind Driven Currents in a Baroclinic Ocean; With Application to the Equatorial Currents of the Eastern Pacific, Proc. Natl. Acad. Sci. USA, 33, 318, (1947)

    work page 1947

  5. [5]

    B. B. Parker, Tidal Hydrodynamics, John Wiley and Sons, New York, (1991)

    work page 1991

  6. [6]

    Stommel, A survey of ocean current theory, Deep Sea Res., 4, 149, (1957)

    H. Stommel, A survey of ocean current theory, Deep Sea Res., 4, 149, (1957)

    work page 1957

  7. [7]

    Tomczak and J.S

    M. Tomczak and J.S. Godfrey, Regional Oceanography: an Introduction, Pergamon Press, Oxford, 1994

    work page 1994

  8. [8]

    Wust, Schichtung und Zirkulation des Atlantisches Ozeans

    G. Wust, Schichtung und Zirkulation des Atlantisches Ozeans. Die Stratosphare, Wissenschaftliche Ergebnisse der Deutschen Atlantischen Expedition auf dem Forschungs-und Vermessungsschiff, Meteor. 19251927, 6, 109 (1935)

    work page 1935

  9. [9]

    H. U. Sverdrup and M.W. Johnson and R.H. Fleming, The Oceans: Their Physics, Chemistry and General Biology, Prentice Hall, New Jersey, 1942

    work page 1942

  10. [10]

    Siedler and J

    G. Siedler and J. Church and J. Gould, Ocean Circulation and Climate: Observing and Modelling the Global Ocean, Academic Press, New York, 2001

    work page 2001

  11. [11]

    O˘ guz and E.¨Ozsoy and M

    T. O˘ guz and E.¨Ozsoy and M. A. Latif and H. I. Sur and ¨U. ¨Unlata, Modeling of hydraulically controlled exchange flow in the Bosphorus Strait, J. Phys. Oceanogr., 20, 945, (1990)

    work page 1990

  12. [12]

    Jarosz and W

    E. Jarosz and W. J. Teague and J. W. Book and S ¸. Be¸ siktepe, On flow variability in the Bosphorus Strait, J. Geophys. Res., 116, C08038, (2011)

    work page 2011

  13. [13]

    Bayındır, Shapes and statistics of the rogue waves generated by chaotic ocean current, Proc

    C. Bayındır, Shapes and statistics of the rogue waves generated by chaotic ocean current, Proc. of the 26th ISOPE, (2016)

    work page 2016

  14. [14]

    Bayındır, Rogue waves of the Kundu-Eckhaus equation in a chaotic wavefield, Phys

    C. Bayındır, Rogue waves of the Kundu-Eckhaus equation in a chaotic wavefield, Phys. Rev. E., 93, 032201, (2016)

    work page 2016

  15. [15]

    Bayındır, Rogue wave spectra of the Kundu-Eckhaus equation, Phys

    C. Bayındır, Rogue wave spectra of the Kundu-Eckhaus equation, Phys. Rev. E., 93, 062215, (2016)

    work page 2016

  16. [16]

    Bayındır, Early detection of rogue waves by the wavelet transforms, Phys

    C. Bayındır, Early detection of rogue waves by the wavelet transforms, Phys. Lett. A, 380, 156, (2016)

    work page 2016

  17. [17]

    Magagna and A

    D. Magagna and A. Uihlein, Ocean energy development in Europe: current status and future perspectives, Int. J. Mar. Energy, 11, 84, (2015)

    work page 2015

  18. [18]

    Uihlein and D

    A. Uihlein and D. Magagna, Wave and tidal current energy A review of the current state of research beyond technology, Renew. Sust. Energ. Rev., 58, 1070, (2016)

    work page 2016

  19. [19]

    Lynn, Electricity from wave and tide: an introduction to marine energy, John Wiley and Sons, Chichester, UK, 2013

    P.A. Lynn, Electricity from wave and tide: an introduction to marine energy, John Wiley and Sons, Chichester, UK, 2013

    work page 2013

  20. [20]

    Roulston and J

    M.S. Roulston and J. Ellepola and J. von Hardenberg and L.A. Smith, Forecasting wave height probabilities with numerical weather prediction models, Ocean. Eng., 32, 1841, (2005)

    work page 2005

  21. [21]

    Reikard and P

    G. Reikard and P. Pinson and J-R. Bidlot, Forecasting ocean wave energy: the ECMWF wave model and time series methods, Ocean Eng., 38, 1089, (2011)

    work page 2011

  22. [22]

    Bayındır, Compressive spectral method for simulation of the nonlinear gravity waves, Scien

    C. Bayındır, Compressive spectral method for simulation of the nonlinear gravity waves, Scien. Rep., 22100 (2016)

    work page 2016

  23. [23]

    Malmberg and U

    A. Malmberg and U. Holst and J. Holst, Forecasting near-surface ocean winds with Kalman filter techniques, Ocean Eng., 32, 273, (2005)

    work page 2005

  24. [24]

    Ho and J

    P.C. Ho and J. Z. Yim, Wave height forecasting by the transfer function model, Ocean Eng., 33, 1230, (2006)

    work page 2006

  25. [25]

    Londhe and V

    S.N. Londhe and V. Panchang, One-day wave forecasts based on artificial neural networks, J. Atmos. Ocean. Technol., 23, 1593, (2006)

    work page 2006

  26. [26]

    Gaur and M.C

    S. Gaur and M.C. Deo, Real-time wave forecasting using genetic programming, Ocean. Eng., 35, 1166, (2008)

    work page 2008

  27. [27]

    S. C. James and Y. Zhang and F. O’Donncha, A machine learning framework to forecast wave conditions, Coast. Eng., 137 , 1, (2018)

    work page 2018

  28. [28]

    Jiang and Y

    Y. Jiang and Y. Gou and T. Zhang and K. Wang and C. Hu, A Machine Learning Approach to Argo Data Analysis in a Thermocline, Sensors, 17, 2225, (2017)

    work page 2017

  29. [29]

    Hollinger and A

    G. Hollinger and A. Pereira and V. Ortenzi and G. Sukhatme, Towards improved prediction of ocean processes using statistical machine learning, Proc. Robotics: Science and Systems Workshop on Robotics for Environmental Monitoring (RSS), Sydney, Australia, (2012)

    work page 2012

  30. [30]

    Sarkar and M

    D. Sarkar and M. A. Osborne and T. A. Adcock, Spatiotemporal prediction of tidal currents using Gaussian processes, J. Geophy. Res.: Oceans, (2019). doi:10.1029/2018jc014471

    work page doi:10.1029/2018jc014471 2019

  31. [31]

    Bolton and L

    T. Bolton and L. Zanna, Applications of Deep Learning to Ocean Data Inference and Subgrid Parameterization, J. Adv. Model. Earth Sys., 11, 376, (2019). 11

    work page 2019

  32. [32]

    D. J. C. MacKay, Information theory, inference and learning algorithms, Cambridge University Press, Cambridge, UK , 2003

    work page 2003

  33. [33]

    Hochreiter and J

    S. Hochreiter and J. Schmidhuber, Long Short-Term Memory, Neur. Comp., 9, 1735, (1997)

    work page 1997

  34. [34]

    Greff and R

    K. Greff and R. K. Srivastava and J. Koutnik and B. R. Steunebrink and J. Schmidhuber, LSTM: A search space odyssey, IEEE Trans. on Neur. Netw. and Learn. Sys., 28, 2222, (2015)

    work page 2015