← News
Science Breakthroughs Science Breakthroughs Knowledge

A satellite can see the ocean’s surface—but not its currents. AI may help fill the gap

A satellite can see the ocean’s surface—but not its currents. AI may help fill the gap
50 Meters Model depth
1.5 Km Grid spacing
0.049 M/S Speed error
29% speed error reduction
0.93 vorticity correlation
0.66 divergence correlation

At roughly 53 meters below the surface, a neural network inferred ocean speed from sea-surface height with an error of 0.049 meters per second—29% lower than the conventional geostrophic estimate. More strikingly, it recovered swirling motion with a correlation of 0.93, compared with 0.75 for that standard method, and detected divergence: the spreading and converging flow that geostrophy rules out by definition.

That matters because a satellite measures the shape of the ocean surface, not the moving water beneath it. For decades, oceanographers have turned surface height into currents using a clean physical shortcut: where the sea surface slopes, pressure pushes water downhill, while Earth’s rotation deflects it sideways. The shortcut works remarkably well for broad, slowly changing currents. But new satellites can resolve structures only a few kilometers across, where the ocean’s motion is messier—and the old shortcut begins to miss important parts of the picture.

Yaswanth Reddy Yaradoddi, Jinbo Wang and Elyes Serghine trained a neural network to make that translation in a high-resolution ocean simulation. Their result is a proof of concept, not yet a satellite-ready current product: the model learned from one simulated region, and it used sea-surface temperature and wind stress as well as sea-surface height. Still, it suggests that the fine detail now visible to the Surface Water and Ocean Topography mission (SWOT) might become more than a map of surface bumps. It could help reveal the currents, eddies and smaller-scale exchanges that shape ocean circulation.

The Science

Sea-surface height (SSH) is the height of the ocean surface relative to a reference level. Differences in height create pressure gradients. Under geostrophic balance—the condition in which pressure’s push is balanced by the deflection caused by Earth’s rotation—those gradients point to the speed and direction of currents. This is the foundation of many satellite-derived maps of ocean circulation.

The method has a blind spot. Geostrophy assumes that the flow changes slowly and that the effects of rotation dominate. At small scales, currents can curve sharply, accelerate, respond to wind, or be disturbed by internal waves moving through the ocean. These motions are called ageostrophic when they depart from geostrophic balance. A snapshot of SSH can contain traces of several kinds of motion at once, without revealing which one made which pattern.

SWOT, launched to measure ocean-surface height across broad swaths, brings this limitation into focus. Its unusually fine spatial detail makes it possible to observe shorter-scale features that older altimetry could not resolve well. But having a sharper picture of surface height does not automatically make the current beneath it easier to infer. A single image does not directly provide the time evolution or depth structure needed to disentangle every wave and flow.

The team framed the task as a supervised learning problem: show a model examples of ocean conditions and their associated velocities, then ask it to predict velocity from observable inputs. The model was a U-Net, a type of image-processing neural network that combines broad spatial patterns with fine local detail. It learned from LLC4320, a high-resolution global ocean simulation, using a regional subset of the Cape Basin southwest of the Agulhas retroflection, off southern Africa. The simulation includes tides and atmospheric forcing; its grid spacing is about 1.5 to 1.7 kilometers.

The input fields were SSH, sea-surface temperature (SST) and wind stress—the push exerted by wind on the ocean. The target was horizontal velocity, meaning the east-west and north-south components of flow. The main target sat at 52.7 meters depth, where direct wind influence is weaker than at the surface; the authors also tested surface velocity. They compared raw SSH with its horizontal gradients and tested whether adding SST or wind stress helped.

A crucial design choice was how to test whether the model had learned persistent relationships rather than simply recognizing near-duplicate snapshots. The ocean changes slowly, so randomly scattering neighboring hours into training and testing would make the test artificially easy. Instead, the researchers divided the record into consecutive training, validation and test periods, with 48-hour gaps between them. They trained on the earlier period and evaluated on a later one, using the validation period to select model settings.

The evaluation kept training and test hours apart

Duration of each temporal block used to train, tune and evaluate the model. Blocks were separated by 48-hour gaps.

The evaluation kept training and test hours apart
LabelValue
Training7,137 hours
Validation971 hours
Test1,992 hours
Figure 1: Snapshot of the fields used in this study at 10:00 UTC on 10 April 2012, on the 2 km analysis grid (x eastward from 11.8∘E, y northward from 45.0∘S; latitude on the right). (a) Sea-surface height anomaly minus its domain mean; (b) SST; (c) wind stress (vectors, every 16th grid point) over its magnitude; (d) velocity at 53 m depth (vectors, every 12th grid point) over speed. Contours show η every 0.05 m on all panels (dashed where below the domain mean). Dashed lines in (a) mark the six fixed 128×128 km subdomains; validation and test patches are these tiles, whereas training samples are drawn at random positions in the domain (Section 3.6). This snapshot is from the training period.
Figure 1: Snapshot of the fields used in this study at 10:00 UTC on 10 April 2012, on the 2 km analysis grid (x eastward from 11.8∘E, y northward from 45.0∘S; latitude on the right). (a) Sea-surface height anomaly minus its domain mean; (b) SST; (c) wind stress (vectors, every 16th grid point) over its magnitude; (d) velocity at 53 m depth (vectors, every 12th grid point) over speed. Contours show η every 0.05 m on all panels (dashed where below the domain mean). Dashed lines in (a) mark the six fixed 128×128 km subdomains; validation and test patches are these tiles, whereas training samples are drawn at random positions in the domain (Section 3.6). This snapshot is from the training period. Source: Yaswanth Reddy Yaradoddi, Jinbo Wang

What They Found

At 52.7 meters, the best model’s speed root-mean-square error (RMSE)—a measure of typical prediction error that penalizes larger misses—was 0.049 meters per second. That was 29% lower than the geostrophic approximation’s error, according to the paper. The simulation’s median speed at that depth was 0.22 meters per second, so the improvement concerns currents that are often much slower than a brisk walk, but that move enormous volumes of water across ocean basins.

The model also captured a more demanding property: relative vorticity, a measure of local rotation in a flow. Its predicted vorticity correlated with the simulated truth at 0.93, while the geostrophic estimate reached 0.75. Correlation does not mean every whirlpool was perfectly placed or every value was exact; it indicates that the predicted pattern tracked the simulation’s variations closely.

The model better captures ocean rotation

Reported correlations between predicted and simulated relative vorticity at 53 meters. Higher correlation means closer tracking of spatial variations; it does not mean exact values.

The model better captures ocean rotation
LabelValue
Geostrophic estimate0.75 correlation
U-Net prediction0.93 correlation

Another diagnostic was divergence: whether water locally spreads out or converges. Geostrophic velocity is nondivergent by construction, so it cannot represent that feature at all. The neural network recovered divergence with a reported correlation of 0.66. This is significant because divergence can carry information about small-scale, wave-driven motion, even though it is also a challenging signal to infer.

The results depended on what the network was given and how it was trained. The authors report four especially useful lessons. First, SSH gradients worked better than raw SSH. Second, wind stress was the most useful added environmental input for predicting total velocity. Third, SST’s benefit varied with season. Fourth, penalizing errors in vorticity and divergence as well as velocity improved the predictions. In other words, the network did better when it was asked not only to match the current at each pixel, but also to respect patterns in how neighboring currents rotate and spread.

Figure 5: Relative vorticity ζ/f (top) and horizontal divergence δ/f (bottom) on a representative test patch (128×128 km; 50 km scale bar in (d)). Left: LLC4320 truth; center: f-plane geostrophic estimate from SSH, whose divergence is zero by construction; right: final model. f0<0, so ζ/f>0 (red) is cyclonic. The correlation with the truth for this patch is given in each panel. The patch is the one whose rζ for the final model (seed-0 run) equals the test-block median.
Figure 5: Relative vorticity ζ/f (top) and horizontal divergence δ/f (bottom) on a representative test patch (128×128 km; 50 km scale bar in (d)). Left: LLC4320 truth; center: f-plane geostrophic estimate from SSH, whose divergence is zero by construction; right: final model. f0<0, so ζ/f>0 (red) is cyclonic. The correlation with the truth for this patch is given in each panel. The patch is the one whose rζ for the final model (seed-0 run) equals the test-block median. Source: Yaswanth Reddy Yaradoddi, Jinbo Wang

The frequency content of the inputs helps explain why this is not a simple conversion. In the simulation, 88% of SSH variance at a fixed point fell in the semidiurnal tidal band, with periods around 11.5 to 13 hours. Yet 92% of velocity variance at 53 meters came from periods longer than two days. The surface-height signal and the subsurface current were therefore dominated by different timescales.

The surface-height signal and subsurface current run on different clocks

Reported variance fractions at fixed points in the simulation. Categories are distinct frequency bands; they are not a complete partition of variance.

The surface-height signal and subsurface current run on different clocks
LabelValue
SSH: semidiurnal band88 % variance
SSH: periods >2 days11 % variance
Velocity at 53 m: periods >2 days92 % variance
Velocity at 53 m: semidiurnal band2 % variance
Velocity at 53 m: near-inertial band1.5 % variance

That mismatch is part of the puzzle the model must solve. A network that merely learned to map each visible bump onto a nearby velocity pattern would struggle when the bump reflects a tide or wave rather than a slowly evolving current. The model instead uses spatial patterns and the additional environmental fields to infer statistical relationships in the simulation. The authors also found that training choices mattered: their paper reports that a basic U-Net did not outperform geostrophy until choices such as data augmentation, predicting the departure from geostrophy, and derivative-based loss terms helped it generalize.

The team assessed more than pixel-by-pixel errors. It examined the joint distribution of vorticity and strain (the rate at which a flow stretches or shears nearby water) and compared the spectral kinetic-energy flux of the true, geostrophic and predicted currents. Spectral energy flux describes how kinetic energy transfers between motions of different sizes. The authors report that geostrophy misrepresented this cross-scale transfer, while their model reproduced it more faithfully at both 53 meters and the surface.

Figure 7: Kinetic-energy spectrum E(K) (top) and spectral kinetic-energy flux Π(K) (bottom) as functions of wavelength 2π/K, averaged over the test period: (a, b) at 53 m; (c, d) at the surface. Π>0 is a forward cascade toward small scales, Π<0 an inverse cascade toward large scales. The geostrophic estimate is computed from the same SSH at both depths. Shading: one standard deviation over three seeds.
Figure 7: Kinetic-energy spectrum E(K) (top) and spectral kinetic-energy flux Π(K) (bottom) as functions of wavelength 2π/K, averaged over the test period: (a, b) at 53 m; (c, d) at the surface. Π>0 is a forward cascade toward small scales, Π<0 an inverse cascade toward large scales. The geostrophic estimate is computed from the same SSH at both depths. Shading: one standard deviation over three seeds. Source: Yaswanth Reddy Yaradoddi, Jinbo Wang

That kind of test matters because a current map can look plausible and still get the ocean’s dynamics wrong. If a model distorts how energy moves from large eddies to smaller features, it may be less useful for studying mixing, turbulence or the life cycle of eddies. The paper’s evaluation asks whether the predictions preserve some of those relationships, not just whether each arrow points roughly the right way.

Why This Changes Things

Ocean currents transport heat, carbon, nutrients and pollutants. Their smaller-scale structure helps determine where these materials mix into deeper water, where plankton encounter nutrients, and how coastal and open-ocean ecosystems respond. Better current estimates could also improve the interpretation of satellite observations used to study marine heatwaves, eddies and the ocean’s role in climate. The study does not measure those downstream benefits directly, but it addresses a basic bottleneck: turning a two-dimensional view of the sea surface into a more complete picture of moving water.

The most promising idea is not that machine learning replaces ocean physics. The network learns from a simulation built around physical equations, and its predictions still depend on that simulation’s representation of the ocean. Rather, it can act as a learned closure: a way to use the surrounding spatial and environmental context to account for effects that a simple balance formula leaves out. Where geostrophy imposes a fixed answer, a trained model can learn when real or simulated flows tend to depart from that answer.

This opens a route for using SWOT’s finer-scale measurements. Traditional altimetry has been especially effective for broad currents and mesoscale eddies—features tens to hundreds of kilometers across. At smaller scales, geostrophic estimates become less reliable, but those are precisely the scales that wide-swath measurements now expose. A model trained on high-resolution ocean simulations could, in principle, help translate those fine patterns into velocity estimates and diagnostics that are difficult to derive from SSH alone.

But the word “could” matters. This paper trains and evaluates on a single simulated region, not on a global collection of real SWOT observations. Its inputs include wind stress and SST, so it does not show that SSH by itself uniquely determines the full velocity field. Nor does it demonstrate that the model transfers unchanged to a different season, ocean basin, satellite error profile or region with different wave and current regimes. As the authors note, two different velocity fields can be associated with similar SSH snapshots; machine learning can use prior examples to choose a likely answer, but it cannot make an intrinsically ambiguous snapshot unambiguous.

Figure 7: Kinetic-energy spectrum E(K) (top) and spectral kinetic-energy flux Π(K) (bottom) as functions of wavelength 2π/K, averaged over the test period: (a, b) at 53 m; (c, d) at the surface. Π>0 is a forward cascade toward small scales, Π<0 an inverse cascade toward large scales. The geostrophic estimate is computed from the same SSH at both depths. Shading: one standard deviation over three seeds.
Figure 7: Kinetic-energy spectrum E(K) (top) and spectral kinetic-energy flux Π(K) (bottom) as functions of wavelength 2π/K, averaged over the test period: (a, b) at 53 m; (c, d) at the surface. Π>0 is a forward cascade toward small scales, Π<0 an inverse cascade toward large scales. The geostrophic estimate is computed from the same SSH at both depths. Shading: one standard deviation over three seeds. Source: Yaswanth Reddy Yaradoddi, Jinbo Wang

What’s Next

The next test is whether the learned relationships survive contact with observations. SWOT measurements include noise and sampling patterns unlike the idealized, complete fields used in this training setup. A model intended for practical use would need to be tested on those actual measurement conditions, ideally against independent current observations such as drifters, while accounting for the fact that drifters do not sample every place or every kind of motion evenly.

Geographic and seasonal generalization also need careful scrutiny. The Cape Basin is an energetic setting, but one region cannot represent the range of latitudes, currents, tides and stratification found across the global ocean. Wind stress emerged as a valuable input for total velocity, and SST’s usefulness shifted by season; both findings suggest that a model may need reliable environmental inputs and region-aware validation rather than a single universal mapping.

Uncertainty is another missing piece for any operational product. Because the same SSH pattern can arise from different dynamics, a useful system should communicate where its prediction is well supported and where several velocities remain plausible. Probabilistic outputs, tests across multiple simulations, and comparisons with independent observations could help distinguish a confident reconstruction from a guess shaped by the training set.

The paper’s result is therefore a measured step, not a finished ocean-current atlas. Yet it shows what becomes possible when high-resolution simulations are used not just to forecast the ocean, but to teach a model how surface patterns relate to subsurface motion. If that lesson transfers to SWOT, satellites may eventually help us see not only the changing shape of the sea, but more of the intricate circulation hidden beneath it.