PatchPODiff: Uncertainty-Aware Super-Resolution of Sea Surface Temperature Forecasts

PatchPODiff is an uncertainty-aware diffusion model for generating high-resolution sea surface temperature (SST) fields from coarse ensemble forecasts. It extends PODiff, accepted at ICML 2026, to climate and ocean forecasting applications on a high-resolution ROMS grid.

This demo focuses on SST downscaling for the Western Australia ocean region. It allows users to compare the original forecast input with the PatchPODiff downscaled prediction and its ensemble-based uncertainty.

What you can explore

1. ECMWF SST input — interpolated ensemble-mean forecast field
2. PatchPODiff high-resolution SST — downscaled high-resolution ensemble-mean prediction on the ROMS grid
3. Uncertainty map — pixel-wise variability across downscaled ensemble members for the selected forecast day
4. 2-degree regional zoom — local SST structure around a selected latitude/longitude point
5. 90-day location time series — ECMWF input vs PatchPODiff downscaled SST at the selected grid point
6. p90 exceedance probability — percentage of ensemble members exceeding the daily 90th-percentile SST threshold for the selected forecast day

This is a research demo using preprocessed ECMWF-style inputs for the first 90 days of 2025. It is designed to illustrate AI-based climate downscaling, uncertainty-aware ocean prediction, and high-resolution SST reconstruction.

Available latitude range: -34.33 to -22.58
Available longitude range: 108.51 to 116.28