Task description

Overview

Data Challenge 1 (DC1) is an open benchmark for emulating global ocean reanalyses at the surface level. Participants train a neural emulator on historical ocean data and submit 10-day surface-level predictions of the global ocean state. Predictions are evaluated against a suite of independent in-situ and satellite observations covering the period 1 January 2024 – 1 January 2025.

The fundamental goal of DC1 is to assess whether neural networks can faithfully reproduce the time evolution of a global ocean reanalysis (e.g. GLORYS12) using only 2-D surface fields, without requiring the full 3-D water column.

DC1 is part of the PPR Océan & Climat (Projet Prioritaire de Recherche), a national research program launched by the French government and managed by CNRS and Ifremer to improve understanding of the ocean and climate.

Goal

Given any set of input data (e.g. reanalysis fields, satellite observations, in-situ profiles), produce daily global ocean state emulations at 0.25° × 0.25° horizontal resolution for lead times $t = 0, 1, \ldots, 9$ days. Five physical variables must be predicted at the surface level only:

CF standard name

Short name

Description

sea_surface_height_above_geoid

zos

Sea surface height

sea_water_potential_temperature

thetao

Sea surface temperature

sea_water_salinity

so

Sea surface salinity

eastward_sea_water_velocity

uo

Surface eastward current

northward_sea_water_velocity

vo

Surface northward current

All variables are 2-D fields with dimensions (time, lat, lon). Unlike DC2, no depth dimension is required: the evaluation pipeline automatically extracts the surface level (~0.5 m depth) from any submission that includes a depth axis.

Evaluation setup

Predictions are launched every 7 days (evaluation interval) throughout the benchmark year. Each forecast covers 10 days of lead time. The evaluation pipeline:

  1. Downloads or reads the submitted emulation for each initialization date.

  2. Interpolates predicted surface fields to the space-time locations of each observation dataset.

  3. Computes RMSD (and other metrics) between the interpolated prediction and the observations.

  4. Aggregates scores per variable and lead time and publishes them on the leaderboard.

Spatial domain

The target grid covers the global ocean at the surface only:

  • Latitude: −78° to +90° (step 0.25°, 672 points)

  • Longitude: −180° to +180° (step 0.25°, 1 440 points)

  • Depth: surface only (~0.49 m)

This is the key distinction from DC2 which evaluates on 21 depth levels from ~0.5 m to ~5 275 m.

Reference model — GloNet

The baseline against which all submissions are compared is GloNet (Global Neural Ocean Forecasting System), a deep-learning model developed by Mercator Ocean International within the PPR Océan & Climat framework. GloNet produces daily global forecasts at 0.25° resolution. For DC1, only the surface level of GloNet’s output is used for evaluation, serving as the benchmark score on the leaderboard.