Satellites see the ocean's surface. They measure sea surface temperature, sea surface height, surface chlorophyll, surface wind stress. Everything below the surface is invisible. But the ocean is three-dimensional, and what happens at 500 meters depth — the movement of intermediate water masses, the transport of heat into the deep ocean, the formation of mode waters that set the climate's memory on decadal timescales — cannot be observed from space.
The conventional approach is data assimilation: combine sparse in-situ measurements from Argo floats and ship transects with physics-based ocean models to estimate the subsurface state. This works, but it produces a single best estimate — one temperature field, one salinity field, one velocity field at each depth. The estimate is wrong everywhere (the ocean is undersampled), but it does not say where it is wrong or by how much.
Souza and colleagues use score-based diffusion models — the same generative framework behind image synthesis — to produce not one subsurface estimate but a distribution of plausible subsurface states consistent with the surface observations. Given a satellite snapshot of sea surface temperature and height, the model generates many possible three-dimensional ocean states, each physically plausible, each different. The spread of the ensemble is the uncertainty.
What makes this more than a clever application is that the uncertainty behaves physically. Skill degrades systematically as resolution drops or inference depth increases. Near the surface, where the observed data constrains the reconstruction tightly, the ensemble is narrow. At depth, where surface information has been diluted by mixing and advection, the ensemble spreads. The model has learned something about the physics of vertical information propagation — how much the surface tells you about the deep ocean depends on the dynamics that connect them, and the model captures that dependence without being told the physics explicitly.
The through-claim is about the information content of a boundary. The surface of the ocean is a boundary condition for the three-dimensional flow beneath it. How much that boundary tells you about the interior depends on the coupling between them — strong coupling means the surface constrains the interior tightly, weak coupling means the surface is nearly independent of what lies below. The diffusion model's uncertainty estimates are, in effect, a learned map of this coupling strength. Where uncertainty is low, the surface and the interior are tightly linked. Where uncertainty is high, the deep ocean has its own dynamics that the surface cannot see. The invisible column reveals itself not through observation but through the structure of what observation fails to constrain.