~/aixsci
weekly updates

climate-weather/ai produced the result/Science Advances 2026 · v2

Deep learning forecast of El Niño suggests Indian and Atlantic oceans extend warning time

Researchers trained a neural network, GL-Geoformer, to forecast tropical ocean temperatures and winds 24 months ahead. Switching off the Indian and Atlantic oceans in the model's inputs shortened its useful El Niño forecasts from 17 months to 11.

1. Assemble and preprocess tropical ocean-atmosphere fields2. Pretrain GL-Geoformer on CMIP6 simulations3. Transfer training with coupling-enhanced loss4. Generate rolling hindcasts on reanalysis5. Score prediction skill against reanalysis and dynamical models6. Run basin-masking ablation experiments7. Run forced-response and pacemaker integrations8. Attribute skill changes and teleconnection mechanisms

spectrum · one line per step, placed by what the step does · bright lines used AI

Tropical basin interactions reduce spring predictability barrier of ENSO in a deep learning model
Science Advances, 2026

doi:10.1126/sciadv.aeb0901 · record aix-00283 v2 · checked 2026-11-05

ai-resultrole of AI
AI was for
Forecasting, Simulation surrogate
Model family
Transformer, Convolutional neural network
Checked by
Held-out
Code
not reported

The finding the paper is about came from the AI.

read as

The science is explained before the AI appears. Switch to field specialist to go straight to the method.

Assumes the discipline and goes straight to the method.

The diagram, the record and what the paper did not report are identical in both modes. Only the framing changes, never the evidence.

Introduction by AIxSci · plain language

What this research was about

El Niño and La Niña are the warm and cool phases of a see-saw in the tropical Pacific called ENSO, and they shift rainfall and temperature across much of the world. Forecasting them months in advance is valuable but awkward. Skill tends to collapse for forecasts that start in the northern spring, a long-standing problem known as the spring predictability barrier. The usual tool is a dynamical model, which solves equations for the ocean and atmosphere step by step. Such models are costly to run and their forecasts of ENSO typically fade within a year. The tropical Pacific also does not sit alone: the Indian and Atlantic oceans warm and cool too, and the three basins pull on one another through winds.

The researchers set out to build a neural network that predicts the whole tropical ocean rather than a single index, and then to use that network as a laboratory. By removing one ocean basin at a time from its inputs, they could ask how much each basin contributed to the forecast.

Where AI came in

GL-Geoformer is a transformer, the same broad design used in language models, here reading maps instead of text, with a convolutional front end that chops each map into patches. It was given nine consecutive months of 13 fields, two components of surface wind stress plus ocean temperature at 11 depths across the global tropics, and produced 24 months ahead by feeding its own output back in. It was first trained on 29 CMIP6 climate simulations, then retrained using all 30 plus older reanalysis data, and finally run as ten copies from monthly observed starting conditions for 1980 to 2023. Correlation with observed Niño3.4 stayed above 0.5 beyond 17 months.

The model also stood in for the dynamical simulations climate scientists normally use to probe cause and effect. The team zeroed out Indian or Atlantic anomalies and re-ran it, which cut the useful lead time to 12, 15 and 11 months depending on which basins were removed. They also ran pacemaker experiments inside the network, holding one basin's temperatures to observations while blanking the other, and these reproduced observed lead-lag relationships with the following winter's Niño3.4. For the 2020 to 2021 La Niña, started in April 2020, the Pacific-only setup missed the event's size while the full version did not. Every such claim is model behaviour.

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

GL-Geoformer is a Vision Transformer encoder-decoder trained on CMIP6 historical simulations and SODA reanalysis to predict 13 monthly variables (two surface wind stress components and ocean temperature at 11 depths) across the global tropics for 24 months ahead. Tested on GODAS reanalysis for 1980 to 2023 with a 10-member ensemble, it exceeded a 0.5 Pearson correlation for the Niño3.4 index beyond 17 months, and beyond 16 months for predictions initialised in February to May. Ablation experiments that zeroed Indian and Atlantic Ocean fields reduced the effective lead time from 17 months to 12 months (Indian removed), 15 months (Atlantic removed) and 11 months (both removed), and pacemaker integrations reproduced the observed DMI and ATL3 lead-lag relationships with the following winter Niño3.4 index. In a hindcast of the 2020 to 2021 La Niña initialised in April 2020, the Pacific-only configuration did not reproduce the event amplitude while the configuration including the Indian and Atlantic Oceans did.

How AI was used

A transformer model, GL-Geoformer, was built with a convolutional preprocessing module producing 160-dimensional patch embeddings with temporal and spatial positional encodings, followed by a three-submodule encoder and decoder using four-head spatiotemporal self-attention. It takes nine consecutive months of 13 standardised anomaly fields (zonal and meridional wind stress and ocean temperature at 11 depths in the upper 200 m, over 0°–360°, 32°S–32°N) and produces 24 months of the same variables by recursively feeding its own predictions back into the decoder. Input data were deseasonalised and detrended, interpolated to a 2° zonal by 0.5°/1° meridional grid, with land and missing points set to zero and all fields standardised by the global mean standard deviation. Pretraining used 29 CMIP6 historical simulations with one held for validation and roughly 10,000 training and 250 validation samples drawn per epoch, under a loss combining RMSE over all predicted fields with an RMSE term on Niño3.4 SST anomalies; a transfer training stage then used all 30 CMIP6 models with pre-1980 SODA reanalysis for validation and a coupling-enhanced loss constraining spatiotemporal covariance between wind and temperature fields. Training used Adam with warm-up schedules, peak learning rates of 1 × 10−3 for pretraining and 1 × 10−5 for transfer learning, and early stopping on validation RMSE. The model was then run from monthly GODAS initial conditions for 1980 to 2023 as a 10-member seed ensemble, re-run with Indian and/or Atlantic Ocean anomalies removed from the predictors and the rolling prediction for ablation attribution, and integrated for 1980 to 2023 in an ocean-only run forced by observed wind stress and in two pacemaker runs restoring Indian or Atlantic SST anomalies to observations while zeroing the other non-Pacific basin.

The shape of the work

Structural · the record, drawn

PREPARATIONTRAININGTRAININGINFERENCEVALIDATIONINFERENCESIMULATIONINTERPRETATION12345678AIAIAIAIAIAssemble andpreprocesstropical ocean-a…PretrainGL-Geoformer onCMIP6 simulationsTransfer trainingwithcoupling-enhance…Generate rollinghindcasts onreanalysisScore predictionskill againstreanalysis and d…Run basin-maskingablationexperimentsRunforced-responseand pacemaker in…Attribute skillchanges andteleconnection m…↤ simulation↤ simulation
AI stepNo AI↤ what the AI stood in for

Assemble and preprocess tropical ocean-atmosphere fields

Cleaning, filtering, normalising or labelling data already obtained.

climatological seasonal cycle and long-term linear trends are removed to eliminate the effects of global warmingwhere the paper describes this · verbatim
in the paper

Pretrain GL-Geoformer on CMIP6 simulations

Fitting model parameters, including fine-tuning an existing model.

we select 29 CMIP6 model simulations for training and 1 additional model simulation for validationwhere the paper describes this · verbatim
in the paper

Transfer training with coupling-enhanced loss

Fitting model parameters, including fine-tuning an existing model.

the GL-Geoformer is trained on simulations from all 30 CMIP6 models, and SODA reanalysis before 1980 are used for validationwhere the paper describes this · verbatim
in the paper

Generate rolling hindcasts on reanalysis

Running a trained model over new data to predict, classify or score. The AI stood in for simulation.

the GL-Geoformer is tested using GODAS reanalysis spanning the period from 1980 to 2023, with the model initialized each monthwhere the paper describes this · verbatim
in the paper

Score prediction skill against reanalysis and dynamical models

Testing outputs against ground truth.

All-months PCC and RMSE skills for 3-month running mean Niño3.4 indices as a function of prediction lead monthswhere the paper describes this · verbatim
in the paper

Run basin-masking ablation experiments

Running a trained model over new data to predict, classify or score.

a series of ablation experiments are conducted using the GL-Geoformerwhere the paper describes this · verbatim
in the paper

Run forced-response and pacemaker integrations

Numerical or physics simulation, including where a learned surrogate replaces it. The AI stood in for simulation.

we conduct three experiments within the GL-Geoformer that emulate the forced-response and pacemaker experiments commonly used in dynamical modelingwhere the paper describes this · verbatim
in the paper

Attribute skill changes and teleconnection mechanisms

Extracting understanding from model behaviour.

we extend this analysis to evaluate the individual versus synergistic contributions of the Indian and Atlantic Oceanswhere the paper describes this · verbatim
in the paper

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.

~Role of AI
AI produced the resultour reading

Every reported prediction skill, ablation attribution and pacemaker response is produced by the trained GL-Geoformer; the paper's claims about basin interactions and the spring predictability barrier exist only as behaviour of the model.

+What the AI was for
the model architecture adopts a Vision Transformer framework, which is structured on an encoder-decoder configurationwhere the paper describes this · verbatim
+Model families
+How it was taught
+Models named
GL-Geoformer · Trained from scratchin the paper
+How results were checked
Held-outin the paper
we apply it to an independent test set using Global Ocean Data Assimilation Systemwhere the paper describes this · verbatim
−Code · weights · data
code not reportedweights not reporteddata not reportednot reported
−Compute
not reportednot reported

What this paper did not report

Technical · absence is published deliberately

Reported as not stated: 9 items
  • CodeWhether the code is available is not stated.
  • Trained model weightsWhether the trained model is available is not stated.
  • DataWhether the data are 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 GL-GeoformerWhich version of the model was used is not stated.
  • What step 2 replacedThe paper gives no basis for what the AI stood in for.
  • What step 3 replacedThe paper gives no basis for what the AI stood in for.
  • What step 6 replacedThe paper gives no basis for what the AI stood in for.

About this article

Record aix-00283, 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). Deep learning forecast of El Niño suggests Indian and Atlantic oceans extend warning time. Record aix-00283, version 2, checked 2026-11-05. https://aixsci.org/articles/aix-00283/