01
SURFACE OBSERVATIONS
Daily SST, SSS, SSH / SLA, currents, and winds define the observed surface state.
FROM SURFACESIGNALSTO SUBSURFACETEMPERATURE
OceanEmbed reconstructs the vertical temperature structure of the North Indian Ocean from surface ocean observations alone.
The model learns the relationship between daily surface ocean conditions and the subsurface temperature field represented by the training target. Evaluation is performed depth by depth and against independent ARGO observations to determine where the reconstruction is reliable and where uncertainty increases.
MODEL PIPELINE
A complete workflow from surface data to validated subsurface temperature.
02
DATA HARMONIZATION
Input fields are quality-checked and placed on the standardized 0.25° OceanEmbed grid.
03
OCEAN EMBEDDING
The multi-variable surface state is converted into a learned representation.
04
DEPTH-WISE RECONSTRUCTION
The model reconstructs temperature at multiple depths from the ocean embedding.
05
GLORYS REFERENCE
Reconstructed fields are evaluated against the GLORYS reanalysis reference target.
06
ARGO VALIDATION
Independent ARGO observations are used to assess real-world consistency.
WHAT THE MODEL SEES
The model receives surface ocean-state information and reconstructs temperature through the upper ocean and into deeper layers. The two regions’ final models use different surface inputs.
SURFACE INPUT
01SSTBOTH REGIONS
02SSHBOTH REGIONS
03WIND STRESS CURLARABIAN SEA ONLY
04MLDARABIAN SEA ONLY
05SSSARABIAN SEA ONLY
06EDDY VORTICITYARABIAN SEA ONLY
Bay of Bengal’s final model uses SST + SSH with 5-region clustering; Arabian Sea’s adds four more surface fields with 10-region clustering — both with bias correction.
DAILY
SURFACE
STATE→
SURFACE
STATE→
RECONSTRUCTION TARGET
SUBSURFACE
TEMPERATURE
TEMPERATURE
0 m5 m10 m20 m30 m50 m75 m100 m125 m150 m200 m300 m500 m700 m1000 m
EMBEDDING ARCHITECTURE
OceanEmbed converts the multi-variable surface ocean state into a learned representation that can be used to reconstruct temperature at multiple depths.
01INPUT FIELDS→
02FEATURE ENCODING→
03OCEAN EMBEDDING→
04DEPTH-AWARE DECODER→
05TEMPERATURE PROFILE
TECHNICAL NOTE Architecture-specific details are reported from the trained model configuration.
HOW THE MODEL IS EVALUATED
Evaluation is performed independently of training, using temperature fields and profiles withheld from the fitting process.
RMSE
Root Mean Square Error
Magnitude of reconstruction error between predicted and reference temperature.
BAY OF BENGAL0.637°C
ARABIAN SEA0.834°C
CORRELATION
Pearson Correlation
How closely predicted temperature variations follow the reference field.
BAY OF BENGAL0.997
ARABIAN SEA0.988
BIAS
Mean Prediction Error
Whether the model systematically overestimates or underestimates temperature.
BAY OF BENGAL-0.385°C
ARABIAN SEA+0.066°C
Avg. of published example predictions (BoB n=3, AS n=2) — not the full validation sample RMSE/Correlation above are drawn from.
PERFORMANCE THROUGH THE WATER COLUMN
Validated RMSE by depth, against withheld ARGO profiles. Error peaks in the thermocline and falls again in deep water for both regions.
BAY OF BENGAL ARABIAN SEA
STANDARD DEPTHS 0 m5 m10 m20 m30 m50 m75 m100 m125 m150 m200 m300 m500 m700 m1000 m
RMSE & CORRELATION BY DEPTH
A depth-selectable readout of validated RMSE and correlation for both regions.
DEPTH SELECTOR
BAY OF BENGAL · 0 m
RMSE0.261°C
CORRELATION0.949
ARABIAN SEA · 0 m
RMSE1.034°C
CORRELATION0.808
INDEPENDENT ARGO VALIDATION
ARGO observations provide an independent reference for assessing whether reconstructed subsurface temperature fields remain physically consistent outside the training target.
MODEL
PREDICTION
+PREDICTION
ARGO
PROFILE
↓PROFILE
DEPTH-MATCHED
COMPARISON
↓COMPARISON
RMSE / BIAS /
CORRELATION
CORRELATION
30°N5°N45°E105°E
BAY OF BENGAL30,019 ARGO MEASUREMENTS
RMSE 0.637°CCORRELATION 0.997
ARABIAN SEA2,129,593 ARGO MEASUREMENTS
RMSE 0.834°CCORRELATION 0.988
WHERE THE MODEL PERFORMS
Regional comparisons distinguish reconstruction behavior across the two proof-of-concept regions.
ACTIVE REGION
BAY OF BENGAL
Final model: SST, SSH · 5-region clustering + bias correction.
RMSE0.637°C
BIAS-0.385°CAVG. OF EXAMPLES
CORRELATION0.997
ARGO VALIDATION COUNT30,019
UNDERSTANDING MODEL ERROR
Reconstruction skill is not uniform with depth or location. Validation against ARGO shows:
01Both regions reconstruct the near-surface layer (0–10 m) most reliably — Bay of Bengal RMSE holds at 0.24–0.26°C, Arabian Sea at 0.62–1.03°C.
02Error peaks in the thermocline, not at depth: Bay of Bengal peaks at 100 m (RMSE 1.464°C), Arabian Sea at 75 m (RMSE 1.351°C) — where vertical gradients are steepest.
03Deep water is easier, not harder, to reconstruct: below 500 m, Bay of Bengal falls to 0.163–0.202°C and Arabian Sea to 0.431–0.540°C.
04Arabian Sea error runs 2–5× higher than Bay of Bengal at nearly every depth — why its model adds four predictors beyond SST/SSH and clusters into 10 regions instead of 5.
05Bay of Bengal runs cold (-0.385°C avg.) and Arabian Sea runs close to neutral (+0.066°C avg.), averaged across the 5 published example predictions on /solution — not yet a full-validation-sample aggregate.
DATA USED FOR EVALUATION
The evaluation design separates the model training target, independent validation, and surface input fields.
| ROLE | DATASET | PURPOSE |
|---|---|---|
| MODEL TRAINING TARGET | GLORYS Global Ocean Reanalysis | Reference subsurface temperature field |
| INDEPENDENT VALIDATION | Gridded ARGO / INCOIS | Independent in-situ validation |
| SURFACE INPUT | Satellite SST | Model input |
| SURFACE INPUT | Satellite / observation SSS | Model input |
| SURFACE INPUT | Satellite / observation SSH / SLA | Model input |
| SURFACE INPUT | Ocean current product | Model input |
| SURFACE INPUT | Atmospheric analysis | Model input |
OCEANEMBED MODEL STATUS
Two regional models
trained and validated
TRAININGCOMPLETE — 2 REGIONAL MODELS
EVALUATIONCOMPLETE — VS. GLORYS + ARGO
ARGO VALIDATIONCOMPLETE — 2,159,612 MEASUREMENTS
DEPLOYMENT5 VALIDATED PREDICTIONS LIVE ON /SOLUTION