The data
- Currents
- Monthly averages of the OSCAR surface currents, which are derived from satellites, over 2011 to 2014, on a grid of 1° cells.
- Wind
- Monthly averages of QuikSCAT satellite winds at 10 m, 1999 to 2009, with gaps near coasts and sea ice filled from the nearest open water.
- Drifters
- The Global Drifter Program’s buoys, 2012 to 2014, one position a day. Some lost their drogue and rode at the surface, catching more wind; the model uses that difference.
- Land
- Natural Earth coastlines.
The model
A small neural network reads ten numbers about a spot in the ocean: where it is, that month’s average current and wind, the time of year, and how much the object catches the wind. It predicts how far a drifter moves over the next day, east and north, together with how unsure it is. On the globe, every fragment is a cloud of copies, each moved by that guess plus its own share of the uncertainty.
Loading the model’s numbers…
Its limits
- Climatology, not live. Currents and winds are monthly averages over a few years. No particular storm or eddy is in them.
- A coarse grid. Currents sit on 1° cells, about 111 km at the equator, and winds on 2° cells, so coasts, straits and anything smaller are blurred.
- Windage is a proxy. The drifter record only covers objects that slip downwind at about 0.1 % to 1 % of the wind speed, so a bottle riding high and a flat bag look the same to the model.
- Fragmentation timing is uncertain. No study has timed how long a floating item takes to break into smaller pieces. Drift’s rates are assumptions built from lab, beach and seafloor measurements.
- Flat water. Sinking, beaching and gathering in gyres follow simple rules with cited rates; nothing moves in depth.

