Skip to content

DepthDif

DepthDif is a research codebase for reconstructing dense subsurface temperature or salinity fields from sparse EN4/ARGO observations. The maintained pixel-space model uses three ordered surface conditions—SST, SSS, and ADT—plus sparse profile support, ocean masks, coordinates, and date context.

!!! important "Two checkpoint interfaces" Current repository configs use [sst, sss, adt]. The downloadable public depthdif_v1.ckpt uses the earlier single-OSTIA interface. Configs and checkpoints from those interfaces are not interchangeable.

Supported workflows

  • Export and download the GeoTIFF/Zarr training dataset.
  • Train temperature, salinity, or joint pixel-space diffusion models.
  • Pretrain with deterministic surface-plus-depth-offset synthetic targets.
  • Run public regional inference for one ISO week.
  • Export global, paired-variable, temporal, paper-metric, and spectral products.
  • View generated products in the spatial, temporal, comparison, and spectral analysis pages.

Analysis

Explore generated DepthDif products

Open the analysis landing page for spatial and temporal globes, error dashboards, model comparisons, and wavenumber spectra.

Open Analysis

Start here

  • Quick start: install, train, and run inference.
  • Model: tensor contract, conditioning, diffusion, and outputs.
  • Training: current presets, scenarios, resume behavior, and monitors.
  • Data overview: active products and model-facing dataset.
  • Inference: local, global, temporal, and paper workflows.
  • Public package: stable one-week API and legacy checkpoint.
  • CLI reference: maintained commands and module entry points.
  • Development: contributor workflow and repository structure.
  • Tests: validation coverage and commands.

The documentation describes current code and committed configuration. Historical experiment notes and speculative roadmaps are intentionally not part of the site.