climate-weather/ai produced the result/Nature Geoscience 2026 · v2
Neural network turns hourly satellite heat images into Gulf Stream current maps
Researchers trained a neural network on output from a fine-scale ocean simulation, then applied it to hourly infrared images of the Gulf Stream from a weather satellite to produce maps of surface currents and their swirling and spreading.
spectrum · one line per step, placed by what the step does · bright lines used AI
An unprecedented view of ocean currents from geostationary satellites
Nature Geoscience, 2026
doi:10.1038/s41561-026-01943-0 · record aix-00284 v2 · checked 2026-11-05
- AI was for
- Structure determination
- Model family
- Convolutional neural network, Multilayer perceptron
- Checked by
- Experimental
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
The ocean's surface is in constant motion, and much of that motion happens in small, short-lived features: eddies, filaments and fronts only a few kilometres across that last hours to days. These stir heat, carbon and nutrients between the surface and the depths. Measuring them is awkward. Ships and drifting buoys sample single points or tracks. Satellites that measure the height of the sea surface, from which currents can be worked out, cover wide areas but give a coarse picture that smooths out the smallest features. So the flow at scales of a few kilometres and a few hours has largely been out of reach.
There is, however, another signal. Satellites parked over a fixed point of the Earth (geostationary satellites) take infrared pictures every hour, and those pictures show the temperature pattern of the sea surface. Currents drag that pattern about, so the way warm and cold water is arranged, and how the arrangement changes from hour to hour, carries information about the flow. The researchers set out to turn that sequence of temperature images into the velocity field itself.
Where AI came in
The link between temperature pictures and currents was learned rather than derived. The team trained a U-Net, a type of image-processing network that compresses an image and then rebuilds it at full size, keeping direct connections between the two halves so fine detail survives. Its lessons came from a high-resolution computer simulation of the Atlantic, where both the surface temperature and the true currents are known. The network was fed three consecutive hourly snapshots of how sharply temperature varied across the surface, and asked to produce the velocity field at the middle hour. Tidal wobbles were filtered out of the targets first, and the training also rewarded getting the balance of energy across scales right.
The trained network was then applied directly to hourly infrared images from the GOES-East satellite over the Gulf Stream, with cloudy patches masked out afterwards. Everything reported about the real ocean, the surface velocities and the derived swirling (vorticity), stretching (strain) and spreading (divergence), comes out of the network; the record describes no other route to these fields. It stands in, in effect, for a physical equation nobody writes down: the step from a moving temperature pattern to the current that moved it. The outputs were then checked against ship-based current measurements, drifting buoys and existing satellite products.
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
GOFLOW is a U-Net trained on one year of hourly output from the 1/48° MITgcm LLC4320 simulation over Atlantic sub-domains to map three consecutive hourly snapshots of the logarithm of the sea surface temperature gradient magnitude onto the surface velocity field at the central time. Applied to hourly GOES-East infrared brightness temperature imagery of the Gulf Stream, it produces hourly surface velocity fields together with their vorticity, strain and divergence. The authors compared the output with AVISO geostrophic currents, SWOT L3 Expert products, shipboard ADCP velocities at 20–25 m depth and surface drifters, and report a kinetic energy spectral slope of about k−3 with a drop in variance near 10-km scales. Strain–vorticity joint PDFs and 1D PDFs computed over a 12.8° × 6.6° Gulf Stream region show skewness towards high strain, positive vorticity and negative divergence associated with scales smaller than 30 km.
How AI was used
A U-Net, a fully convolutional encoder–decoder network with skip connections, was trained from scratch on hourly MITgcm LLC4320 output extracted over 5.3° × 5.3° Atlantic sub-domains between 20° N and 45° N, using 256 × 256-point patches. Inputs were sequences of three consecutive hourly fields of the logarithm of the temperature gradient magnitude; targets were the model surface velocities after an 18-hour low-pass Butterworth filter removed semi-diurnal tidal signals. The loss combined an L1 velocity term with a log-spectral term computed on two-dimensional kinetic energy spectra after Tukey windowing, weighted by λ; λ was scanned over ten values between 0.05 and 0.9, training proceeded as 100 epochs at λ = 0 followed by 100 epochs at λ = 0.2 with cosine annealing and warm restarts, and patch sizes of 256, 128 and 64 were compared on a held-out test set. Because the network is fully convolutional, the trained model was applied directly to full GOES-East ABI Band 14 brightness temperature images without tiling, with velocities masked after inference where clouds were present. The authors also ran two exploratory alternatives: a regularised linear inversion of a discrete forced advection–diffusion equation, and an implicit Neural Field representation of each velocity component with the advection–diffusion residual as a loss term.
The shape of the work
Structural · the record, drawn
Extract training sub-domains from ocean model
Obtaining raw data, whether by measurement, download or retrieval.
we extract 5.3° × 5.3° sub-domains within the 20° N–45° N latitude band in the Atlantic Oceanwhere the paper describes this · verbatim
Build gradient inputs and filter target velocities
Encoding data into features, descriptors, embeddings or graphs.
we pre-processed the target LLC4320 velocity fields by applying an 18-hour low-pass Butterworth filterwhere the paper describes this · verbatim
Train U-Net with velocity and spectral loss
Fitting model parameters, including fine-tuning an existing model. The AI stood in for conventional algorithm.
The network is trained by minimizing a composite loss function designed to ensure physical consistency across scaleswhere the paper describes this · verbatim
Acquire geostationary infrared imagery
Obtaining raw data, whether by measurement, download or retrieval.
For inference, we use Level 1b hourly infrared brightness temperature data (Band 14, 11.2 μm) from the GOES-East Advanced Baseline Imagerwhere the paper describes this · verbatim
Infer hourly surface velocity fields from satellite imagery
Running a trained model over new data to predict, classify or score. The AI stood in for conventional algorithm.
Three consecutive hourly snapshots are used to predict the velocity field at the central time.where the paper describes this · verbatim
Evaluate on held-out model domains and select configuration
Testing outputs against ground truth. Its result feeds back into an earlier step.
we performed comprehensive validation on held-out LLC4320 domainswhere the paper describes this · verbatim
Compare with in situ and other satellite products
Testing outputs against ground truth.
we compare its predicted zonal and meridional surface velocities with independent in situ ship-based Acoustic Doppler Current Profiler (ADCP) measurementswhere the paper describes this · verbatim
Compute submesoscale kinematic statistics
Extracting understanding from model behaviour.
GOFLOW statistics are computed in a 12.8o × 6.6o region on 13 April 2023 from 16:30 UTC to 20:30 UTC.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 reported surface velocity, vorticity and divergence fields and all submesoscale statistics are outputs of the trained U-Net; no non-learned route to the result is reported.
We employ a standard U-Net architecture, an encoder–decoder network with skip connectionswhere the paper describes this · verbatim
Shipboard ADCP observations: 300-kHz acoustic Doppler current profiler data from multiple research cruiseswhere 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.
- 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 GOFLOW U-NetWhich version of the model was used is not stated.
- Version of Neural Field (implicit physics-informed velocity representation)Which version of the model was used is not stated.
About this article
Record aix-00284, 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 turns hourly satellite heat images into Gulf Stream current maps. Record aix-00284, version 2, checked 2026-11-05. https://aixsci.org/articles/aix-00284/