Surface in.
Intelligence
within.
A window into the ocean we cannot see.
Built from the signals we can.
One connected framework.
From satellite observations to 1,000
metres.
Explore the labeled architecture below.
Conceptual system view Drag to rotate
Every connection.
A deeper understanding.
Seven days of surface observations become one compact representation. The decoder brings it back into focus, at fifteen depths.
The surface tells
more than a story.
Twelve features. Each chosen for the physical information it carries into the ocean below.
one fused surface state
Regrid to 0.25° Normalize Stack 7 days
Select a feature to see why it belongs in the model.
Built to learn.
Designed to be tested.
A reconstruction is only useful if it holds up beneath the surface. Training and independent evaluation are part of the same design.
Learn the structure.
Then the variability.
Climatology pre-training
Monthly GLORYS means establish seasonal stratification and the mean vertical temperature structure.
Daily fine-tuning
Seven-day rolling inputs teach the full network to resolve daily changes, eddies, and monsoon dynamics.
Independent by design.
ARGO profiles are
reserved for evaluation and never enter training.
Focus where it matters.
Depth-weighted lossThe thermocline receives twice the weight: its steep temperature gradients make reconstruction particularly challenging.
Measure at every depth.
RMSE / Bias / R² / Pearson r
Across seasons, the Bay of Bengal & Arabian SeaThis page presents the proposed OceanEmbed system. Diagrams illustrate the architecture; they do not represent a trained model or measured performance.