[Task] Climate Downscaling

[LingFH]

Task Description

We aim to optimize supervised climate downscaling from CMIP6 historical simulations to ERA5 reanalysis at higher spatial resolution. The goal is to reduce daily precipitation RMSE on the test set by at least 10% compared to the baseline model, while keeping inference speed within ±20% of baseline. Success is defined by improvements in RMSE, MAE, and spatial Pearson correlation on the held-out test period. Background: CMIP6 outputs are coarse-resolution and contain systematic biases; ERA5 provides higher-resolution reanalysis data suitable as a target for supervised bias correction and downscaling.

Baseline Repository Link (Must be Public)

https://github.com/bthrasher/daily_BCSD

Baseline reproduction (minimal)

commit/tag: main

Dataset (Must be Public)

https://cds.climate.copernicus.eu/datasets/reanalysis-era5-single-levels?tab=download

Results and Evaluation Metrics

RMSE, MAE, spatial Pearson correlation

Preconditions (required)

• The baseline I provide comes from public code hosted on an open-source platform such as GitHub.
• The task I use/optimize is based on a public dataset.
• If my request involves private code/data, I will contact interndiscovery@pjlab.org.cn by email instead.