climate-weather/ai produced the result/PLoS ONE 2026 · v2
Neural network forecasts daily sea surface temperature up to 90 days ahead
Researchers built WDFormer, a trained forecasting model that splits a year of daily sea surface temperature readings into a slow trend and faster wiggles, handles each separately, then merges the predictions for horizons from one week to three months.
spectrum · one line per step, placed by what the step does · bright lines used AI
SST subseasonal-to-seasonal forecasting: A heterogeneous three-path fusion model with wavelet decomposition
PLoS ONE, 2026
doi:10.1371/journal.pone.0351923 · record aix-00292 v2 · checked 2026-11-05
- AI was for
- Forecasting
- Model family
- Transformer, Multilayer perceptron, Linear model
- Checked by
- Held-out
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
Sea surface temperature is the temperature of the top layer of the ocean. It matters far beyond the water: it shapes rainfall, storms and fisheries, and it is the quantity used to track El Nino and La Nina, the warm and cool phases of a recurring pattern in the tropical Pacific. Forecasting it weeks to months ahead sits in an awkward middle ground. That range is longer than ordinary weather prediction, where the atmosphere's day-to-day memory still helps, but shorter than the slow seasonal cycle. A simple guess that tomorrow looks like today works well for a few days and then fades. Beyond that, the signal is a mix of a slow drift and faster fluctuations on top of it.
The researchers set out to forecast daily sea surface temperature at five buoy monitoring sites in the El Nino region of the Pacific, at horizons of 7, 15, 30, 60 and 90 days, using only past temperature records as input.
Where AI came in
The forecasts themselves are the output of a trained machine learning model, so every number reported comes from software rather than from a physics-based ocean simulation. The team fed the model a 365-day window of daily satellite-and-buoy temperature records for a small patch of grid cells around each site. Before the learning began, a wavelet transform, a standard mathematical tool for splitting a wiggly signal into slow and fast parts, separated each window into a smooth trend and a leftover residual. Three branches then made predictions in parallel: small, simple networks handled the trend and the raw undecomposed series, while a transformer, the architecture behind modern language models, handled the residual. A learned weighting combined the three.
The model was trained from scratch on data up to 2014, tuned on data to 2021 and tested on the years that followed, so it was judged on dates it had not seen. Five other deep learning forecasters were trained the same way for comparison, alongside two reference yardsticks that need no learning: assuming no change from today, and using the long-term average for the calendar day. The authors also removed the decomposition step to see how much it contributed, and varied how far back the input window reached.
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 study trains WDFormer, a daily sea surface temperature forecasting model that splits a 365-day input window into a low-frequency trend and a higher-frequency residual with a db4 wavelet transform, sends each component plus the undecomposed series through separate MLP or Transformer branches, and combines the three predictions with a learnable gate. Models were trained and tested on NOAA OISST V2 daily data for 4x4 grids around five ENSO-region buoy stations, with chronological train, validation and test splits, and forecast horizons of 7, 15, 30, 60 and 90 days. Aggregated across the five stations at the 90-day horizon, WDFormer reached an RMSE of 0.521 (95% CI 0.515-0.528) against 0.617 for iTransformer, 0.620 for DLinear, 0.633 for TimeMixer, 0.639 for PatchTST and 0.661 for Crossformer, and reduced average RMSE relative to persistence by 10.7%, 23.2% and 31.9% at 30, 60 and 90 days; persistence had lower RMSE than WDFormer at 7 days (0.245 versus 0.275), and Crossformer sometimes had lower error at 7-15 day horizons. Ablation of the decomposition module raised average RMSE at one station from 0.460 to 0.559.
How AI was used
Daily OISST V2 SST series were extracted for 4x4 grids around five ENSO monitoring buoy stations, giving 16 features per time step, and split chronologically into training, validation and test sets at a 7:2:1 ratio. Each 365-day input window was decomposed with a Daubechies 4 wavelet at a tunable decomposition level into an approximation-based trend series and a residual; the trend and the original undecomposed series were each downsampled, linearly embedded and passed through lightweight MLP predictors, while the residual was downsampled, embedded and passed through a stack of Transformer encoder layers. The three branch predictions were merged by a learnable scalar sigmoid gate. The model and five deep learning baselines (iTransformer, PatchTST, Crossformer, DLinear, TimeMixer) were trained by the authors with MSE loss on the same 365-day look-back window, then run over the held-out test period to produce deterministic daily point forecasts at 7, 15, 30, 60 and 90 day horizons. Scoring used RMSE and MAE with non-parametric bootstrap resampling (2,000 replicates, paired indices across models), plus persistence and daily climatology reference baselines on matched target windows, four architecture ablations, a look-back window sweep over 90, 180, 365, 540 and 1,100 days, and ONI-based stratification of 90-day samples by ENSO phase and El Nino event stage.
The shape of the work
Structural · the record, drawn
Assemble daily SST observations
Obtaining raw data, whether by measurement, download or retrieval.
We use the NOAA Optimum Interpolation Sea Surface Temperature (OISST) V2 dataset, which provides globally gridded daily mean SST at 0.25° spatial resolution.where the paper describes this · verbatim
Build look-back windows and chronological splits
Cleaning, filtering, normalising or labelling data already obtained.
The dataset is split chronologically into training, validation, and test sets at a 7:2:1 ratio.where the paper describes this · verbatim
Wavelet decomposition into trend and residual
Encoding data into features, descriptors, embeddings or graphs.
A wavelet decomposition module separates 𝐗 into trend and residual components.where the paper describes this · verbatim
Train WDFormer and deep learning baselines
Fitting model parameters, including fine-tuning an existing model.
All models use the same 365-day look-back window and are trained with MSE loss.where the paper describes this · verbatim
Forecast daily SST at 7-90 day horizons
Running a trained model over new data to predict, classify or score. The AI stood in for statistical model.
All experiments use the past 365 days of daily observations as input to forecast future daily SSTwhere the paper describes this · verbatim
Score forecasts, bootstrap and stratify by ENSO phase
Testing outputs against ground truth.
We use Root Mean Square Error (RMSE) and Mean Absolute Error (MAE)where 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 results are the forecasts produced by the proposed learned model and its learned baselines; all reported skill numbers come from model output
Lightweight MLP-based branches process the original and low-frequency components, while a Transformer-based branch models the residual dynamics.where the paper describes this · verbatim
This corresponds approximately to training from 1991-01-01 to 2014-10-19, validation from 2014-10-20 to 2021-08-08, and testing from 2021-08-09 to 2024-12-31where the paper describes this · verbatim
The processed SST datasets used to support the findings of this study are openly available in Zenodowhere the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- CodeWhether the code is available is not stated.
- Trained model weightsWhether the trained model is available is not stated.
- ComputeThe hardware or time used is not stated.
- How many were testedThe paper gives no count of what was tested.
- Version of WDFormerWhich version of the model was used is not stated.
- Version of iTransformerWhich version of the model was used is not stated.
- Version of PatchTSTWhich version of the model was used is not stated.
- Version of CrossformerWhich version of the model was used is not stated.
- Version of DLinearWhich version of the model was used is not stated.
- Version of TimeMixerWhich version of the model was used is not stated.
- What step 4 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00292, 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 daily sea surface temperature up to 90 days ahead. Record aix-00292, version 2, checked 2026-11-05. https://aixsci.org/articles/aix-00292/