A simulated Atlantic shows how AI could read a current from drifting floats

One way to measure an ocean current is to listen electrically to a telephone cable. NOAA derives daily Florida Current transport estimates from voltage measured across a submarine telecommunications cable between Florida and Grand Bahama.[4]

Moving seawater carries charged particles through Earth’s magnetic field, producing an electrical signal. The cable detects a voltage related to the flow crossing it. Scientists remove magnetic interference and calibrate the estimate against measurements made in the water.[3]

It is a striking example of a general scientific bargain: measure something accessible, establish its relationship to something harder to observe, and preserve the conditions under which that relationship holds.

The larger circulation is the Atlantic Meridional Overturning Circulation, or AMOC. It carries warm water northwards near the surface and colder water southwards at depth, redistributing heat through the Atlantic.[5]

An Ocean Science study published on 18 December 2025 investigates another such bargain. Virtual Argo floats helped reconstruct the broad interior flow governed by pressure gradients and Earth’s rotation: the geostrophic component. The simulation explained up to 80% of its variability, with mean absolute error below one Sverdrup—a million cubic metres per second. Application to real observations remains future work.[1]

The sensor is drifting

A conventional Argo float spends much of its cycle drifting about a kilometre beneath the surface. Roughly every ten days it descends to two kilometres, then rises while measuring the water around it. At the surface it transmits its position and observations by satellite before sinking again. Temperature, salinity and pressure form the programme’s basic measurements.[2]

Temperature and salinity help determine water density. Differences in density structure across the basin create pressure gradients, providing physical clues to the flow.[1]

The resulting information is rich in vertical structure. It also arrives from instruments that keep moving. Imagine trying to infer traffic across a city from devices that change streets between reports: the places observed, their spacing and the number of observations all vary. A fixed spreadsheet of permanent sensor locations would describe the wrong problem.

The researchers placed virtual profiles according to real Argo sampling patterns from 2012–2023, drawing their water properties from the VIKING20X simulation. A graph neural network assembled the changing set of profiles into a stable representation, alongside boundary-current and wind information. The target lay at 26.5° north. Seasonal estimates were strongest; performance weakened for year-to-year variation.[1]

The graph exchanges information between groups of profiles.[1]

Its attraction is that a missing or relocated float need not shift every subsequent input into the wrong position. The model can work with a collection rather than demanding a permanently numbered row of instruments.

A useful signal inside an existing measurement

Argo makes its observations openly available. Rapid releases support users who need timely information, while a later version receives more extensive checks against nearby floats and research-ship observations. The same instrument therefore supplies measurements whose appropriate use depends partly on their stage of processing.[2]

This matters for any proposed circulation estimate. A computational method would inherit the coverage and quality of its inputs. Adding software does not create observations in a stretch of ocean where none were collected; it supplies a way to connect the observations that exist.

Removing Argo profiles substantially reduced skill. Deeper virtual profiles brought no consistent gain when unfamiliar deep-ocean states were absent from training.[1]

That combination makes the result useful as a question about observing design. Which measurements contain information about a particular process? At what timescale does that information become reliable? Greater depth and a larger dataset need not be interchangeable with the right observations.

The cable and the float solve different measurement problems. Their shared lesson is that an instrument can be more informative than its most obvious reading. The next challenge is to establish whether the relationship learned inside a simulated Atlantic survives the untidy, sparsely observed real one.

Report an error or suggest a correction