Credits: Gilles Louppe
Understanding the Ocean
from Observations Using
Deep Learning
PRESENTED BY Victor Mangeleer1–4
PHD SUPERVISORS Marilaure Grégoire1,3 and Gilles Louppe2,3 HOST SUPERVISOR Laure Zanna4 IN AFFILIATION WITH Modelling for Aquatic Systems1, Science with AI Laboratory2,
University of Liège3, New York University4.

What can we predict from space?

Progress

Since our last meeting, here is what we worked on:

  • Cleaning, standardizing and copying the datasets on Torch, with GLORYS coarsened by a factor .
  • A fixed validation, automatic stopping and resuming, and faster updates*.
  • We finally finished training the encoders, and all their latents are already saved on Torch ( GB).
* Sea surface temperature encoder on one H200, 32 days per update: 2.96 s with the setup of the first trainings, 0.99 s with torch.compile, channels last and no activation checkpointing.

What the loss of the encoders compares

Each encoder is trained with a Gromov-Monge embedding loss: for every pair of days in a batch, the distance between their latents must match the distance between their fields. Now, the encoders have been changed:

  • no mask as input;
  • no date as input;
  • the convolutions are periodic along the longitude (it helped in our experiments).

Figure A. Two days and are encoded into and , and the loss compares their distances.

Key result. One loss for every dataset, which only compares distances between days.

Credits: Lee, Geometry-preserving encoder/decoder in latent generative models, 2026.

A fixed validation

The validation now uses the same days at every evaluation, spread over – ( days for the observations, for GLORYS). We keep the best validation model, stop when the loss stops improving, and resume a job automatically when it is cut. We then evaluated each encoder on the whole validation and on as many training days: the results are in the next figure.

Figure B. Validation loss over training loss, now and before fixing the validation: closer to than before.

Key result. The validation loss now reflects much more faithfully what the encoders can do.

Training of the seven encoders

Each encoder sees days per update ( for GLORYS), on one GPU for the observations and on two GPUs for GLORYS and the chlorophyll*. All encoders finished, and their best validation models encoded every training and validation day.

* NVIDIA RTX Pro 6000 GPUs, with 102 GB of memory each; GLORYS ran on two NVIDIA B200.

Figure C. Training (running median) and validation losses of GLORYS and the chlorophyll: both still decrease.

Figure D. Same for temperature, salinity and wind (solid: ascending passes, dashed: descending): salinity drifts apart.

Salinity and wind have two encoders each, one for the ascending passes of the satellites and one for the descending passes, hence two validation curves.

Summary

Key results

  1. One loss for every dataset, on distances only.
  2. A validation that reflects what the encoders can do.

Problems to solve

  1. Salinity overfits its training data.
  2. A retraining with a more frequent validation stopped too early, so we kept the first models.

What's next

  • Train the processor on the latents, with the observations and the states outside the observation windows as inputs.

Sources