climate-weather/ai produced the result/Scientific Reports 2024 · v2
Neural network forecasts freak waves minutes ahead from a single buoy's record
Researchers trained a neural network on archived buoy measurements of sea surface height to judge, from the preceding twenty minutes, whether an unusually large wave was about to arrive. The network's classifications are the study's result.
spectrum · one line per step, placed by what the step does · bright lines used AI
Prediction of freak waves from buoy measurements
Scientific Reports, 2024
doi:10.1038/s41598-024-66315-3 · record aix-00285 v2 · checked 2026-11-05
- AI was for
- Forecasting, Classification
- Model family
- Recurrent neural network
- Checked by
- Held-out3757 tested, 2902 worked
- Code
- available
The finding the paper is about came from the AI.
What this research was about
Out at sea, waves come in groups of varying size, and occasionally one rises far above its neighbours. Oceanographers call these rogue or freak waves, and they are usually defined by comparison: a wave whose height is more than roughly twice the significant wave height, which is a standard measure of the typical size of the larger waves in a sea state. Such waves are rare, short-lived and dangerous to ships and offshore structures. Because they appear and vanish within a minute or two, and because the sea surface at any one point is a tangle of overlapping waves, knowing in advance that one is coming is hard.
The researchers asked whether a short record of the sea surface at a single location holds enough information to warn of a rogue wave a few minutes before it arrives. They drew on the archive of the Coastal Data Information Program at the Scripps Institution of Oceanography, whose buoys record how far the surface rises and falls, and which holds more than 20 billion such measurements. From it they extracted half-hour windows containing rogue waves under three different definitions, and paired each with a window from the same buoy containing none.
Where AI came in
The forecast itself was made by the AI. The team rephrased the problem as a sorting task: given a 20-minute stretch of buoy measurements ending one to five minutes before the moment of interest, decide which of two classes the stretch belongs to, one followed by a rogue wave and one not. They trained a recurrent neural network built from LSTM layers, a design suited to sequences because it carries information forward as it reads through a signal. Roughly two thirds of the labelled windows were used for fitting, with portions kept back for checking and for final testing. The networks were built in TensorFlow and trained on graphics processors.
On the held-out test data for the strictest definition of a rogue wave, with one minute of warning, 2,902 of 3,757 rogue-wave windows were sorted correctly, and overall accuracy was 76 percent, falling to 73 percent when the warning time was stretched to five minutes. The team also retrained with every record from one buoy removed, then asked the network about that unseen location: 251 of 331 rogue waves at Buoy 067 were sorted correctly, with 75 percent accuracy at two further withheld buoys. The scripts and trained networks are publicly available.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
The authors trained an LSTM network to decide, from 20 min of single-point buoy records of sea surface elevation, whether a rogue wave would follow within a set advance warning time. The training data were drawn from a CDIP buoy archive containing more than 20 billion sea surface elevation samples, from which 30-min windows containing rogue waves were extracted and paired with control windows from the same buoy without rogue waves. On the held-out testing split of the strictest data set (H/Hs>2.2) with a 1-min advance warning time, 2902 of 3757 rogue-wave samples were classified correctly and total accuracy was 76 percent, falling to 73 percent at a 5-min advance warning time. When all records from a buoy were withheld from training, 251 of the 331 rogue waves at Buoy 067 were classified correctly, and accuracy was 75% for two further withheld buoys.
How AI was used
Buoy records of sea surface elevation were quality controlled, scanned for waves satisfying three rogue wave definitions, normalised by significant wave height and shifted so the rogue wave falls at minute twenty-five; each rogue-wave window was paired with a randomly selected window from the same buoy containing no rogue wave, giving balanced two-class data sets. From each sample a 20-min recording ending one to five minutes before the rogue wave was taken, recasting forecasting as a two-class time series classification problem. A recurrent network with LSTM layers alternating with batch normalisation, followed by dropout and a fully connected layer with a sigmoid activation, was fitted to a 64 percent training split using the Adam optimizer with a cosine-decay schedule from an initial learning rate of 0.001, with 16 percent used for validation and 20 percent held out for testing; convolutional and transformer architectures were tried first, and hyperparameters were checked against tuning algorithms and random grid searches. Separate networks were trained for each advance warning time, for each rogue wave definition, and for runs in which all records from one buoy were excluded from training and then predicted; the balanced-data network was also applied to a test set enriched with additional non-rogue-wave samples. Networks were built in TensorFlow 2.9.1 and trained on NVIDIA Quadro P1000 and A100 GPUs.
The shape of the work
Structural · the record, drawn
Assemble buoy surface-elevation records
Obtaining raw data, whether by measurement, download or retrieval.
The data set is provided and maintained by the Coastal Data Information Program (CDIP), Scripps Institution of Oceanography.where the paper describes this · verbatim
Quality control, rogue wave extraction and class labelling
Cleaning, filtering, normalising or labelling data already obtained.
For each rogue wave sample, a time series without a rogue wave from the same buoy is randomly selected and stored.where the paper describes this · verbatim
Train LSTM classifier
Fitting model parameters, including fine-tuning an existing model. The AI stood in for new capability.
the data sets are randomly split into training (64 %), validation (16 %), and testing (20 %) datawhere the paper describes this · verbatim
Classify held-out windows at varying advance warning times
Running a trained model over new data to predict, classify or score. The AI stood in for new capability.
To yield a definitive prediction, the class with the higher probability is selected as the forecast.where the paper describes this · verbatim
Retrain with one buoy withheld and predict at that location
Testing outputs against ground truth. The AI stood in for new capability.
all measurements from CDIP Buoy 067 are removed from the data set Awhere the paper describes this · verbatim
Evaluate on imbalanced test set emulating the ocean
Testing outputs against ground truth. The AI stood in for new capability.
the testing portion of data set A is enriched by 2.8 million non-rogue-wave sampleswhere the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The paper's result is the forecast itself: the LSTM network's classifications of buoy records are what the study reports, so the finding depends entirely on the model.
the forecasting problem has been rephrased into a time series classification taskwhere the paper describes this · verbatim
which is about 77% out of all rogue-wave samples (2902 out of 3757)where the paper describes this · verbatim
All scripts, including the trained neural networks are publicly available at https://github.com/tbreunung/Freak-wave-forecastingwhere the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- Version of LSTM classification networkWhich version of the model was used is not stated.
About this article
Record aix-00285, version 2, checked by a person on 2026-11-05. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error
Cite this record: AIxSci (2026). Neural network forecasts freak waves minutes ahead from a single buoy's record. Record aix-00285, version 2, checked 2026-11-05. https://aixsci.org/articles/aix-00285/