CORDEXBench: A benchmarking dataset for AI-based regional climate downscaling.

Introduction to CORDEXBench: A Benchmarking Dataset for Regional Climate Downscaling

CORDEXBench is a standardized benchmarking dataset designed to evaluate empirical-statistical downscaling (ESD) and deep learning-based regional climate emulators. It supports rigorous model evaluation across multiple climate zones and experimental setups. This dataset spans three distinct geographic domains:

New Zealand (NZ) - 0.11° resolution

Europe - 0.11° resolution

South Africa - 0.25° resolution

Each region includes carefully structured training and testing data derived from dynamically downscaled Global Climate Models (GCMs), enabling systematic analysis of emulator performance in both historical and future climates.

The dataset provides two core training experiments:

ESD Pseudo-Reality (1961–1980)A 20-year historical training period using a single GCM (e.g., ACCESS-CM2 for NZ), designed to mimic ESD training. 

Emulator Hist+Future (1961–1980 + 2081–2100)A more comprehensive 40-year training period combining historical and future climates. This experiment supports evaluation of extrapolative skill, including transferability across GCMs.

Both setups will be tested with and without topography as a predictor. 

For each training setup, the dataset enables evaluation across multiple test periods and inference conditions:

Historical (1981–2000): For both perfect and imperfect inference.

Mid-century (2041–2060) and End-century (2081–2100): To assess extrapolation to future climates, including hard transferability scenarios using unseen GCMs.

Evaluation Types

The dataset supports several benchmarking configurations:

PP cross-validation: Same GCM used in training and testing.

Imperfect inference: Same GCM but different realizations or noise.

Transferability testing: Inference using a different GCM than the training set.

Change signal evaluation: Assessment of climate change response in future periods.

Data Structure Example: New Zealand Domain

Each domain follows a consistent file structure, with subdirectories for training and testing data, and further divisions by period, GCM, and evaluation type. Predictors include both dynamic variables (e.g., temperature, precipitation) and optional static fields (e.g., topography).

NZ_Domain/
├── train/
│ ├── ESD_pseudo-reality/
│ │ ├── predictors/
│ │ └── target/
│ ├── Emulator_hist_future/
│ │ ├── predictors/
│ │ └── target/
├── test/
│ ├── historical/
│ ├── mid_century/
│ └── end_century/

A More Detailied Overview of the Experiments

📘 Table 1: Evaluation using ESD “pseudo-reality” (T1)

Training Setup
Inference Set
Evaluation Type
Notes
Eval
Required

ESD “pseudo-reality”Period: 1961–1980Static fields: Yes/No
historical (1981–2000) 
PP cross-validation
Same GCM used in training, perfectly
Error, Clim
X

historical 1981–2000
Imperfect cross validation
Same GCM, but imperfectly
Error, Clim

2041-2060 + 2081-2100
Extrapolation
Same GCM, but perfectly
change signal
X

2041-2060 + 2081-2100
Extrapolation
Same GCM but imperfectly
change signal

📗 Table 2: Evaluation using Emulator (T2)

Training Setup
Inference Set
Evaluation Type
Notes
Eval
Required

Emulator hist + future

period: 1961-1980 + 2081-2100

Static fields: Yes/No

historical (1981–2000) 
PP cross-validation
Same GCM used in training, perfectly
Error, Clim
X

historical 1981–2000
Imperfect cross validation
Same GCM, but imperfectly
Error, Clim
X

2041-2060 + 2081-2100
Extrapolation
Same GCM, but perfectly
change signal
X

2041-2060 + 2081-2100
Extrapolation / Hard Transferibility
Different GCM, but perfectly
change signal
X

2041-2060 + 2081-2100
Extrapolation / Hard Transferibility
Different GCM, but imperfectly
change signal
X

# NZ Domain/├── train/│   ├── ESD_pseudo-reality/│   │   ├── predictors/│   │   │   ├── ACCESS-CM2_1961-1980.nc│   │   │   └── static.nc│   │   └── target/│   │       └── pr_tasmax_ACCESS-CM2_1961-1980.nc││   ├── Emulator_hist_future/│   │   ├── predictors/│   │   │   ├── ACCESS-CM2_1961-1980_2080-2099.nc│   │   │   └── static.nc│   │   └── target/│   │       └── pr_tasmax_ACCESS-CM2_1961-1980_2080-2099.nc│├── test/│   ├── historical/│   │   ├── predictors/│   │   │   ├── perfect/│   │   │   │   ├── ACCESS-CM2_1981-2000.nc│   │   │   │   └── EC-Earth3_1981-2000.nc│   │   │   └── imperfect/│   │   │       ├── ACCESS-CM2_1981-2000.nc│   │   │       └── EC-Earth3_1981-2000.nc│   │   └── target/│   │       ├── pr_tasmax_ACCESS-CM2_1981-2000.nc│   │       └── pr_tasmax_EC-Earth3_1981-2000.nc││   ├── mid_century/│   │   ├── predictors/│   │   │   ├── perfect/│   │   │   │   ├── ACCESS-CM2_2040-2059.nc│   │   │   │   └── ...│   │   │   └── imperfect/│   │   │       ├── ACCESS-CM2_2040-2059.nc│   │   │       └── ...│   │   └── target/│   │       ├── pr_tasmax_ACCESS-CM2_2040-2059.nc│   │       └── ...││   └── end_century/│       ├── predictors/│       │   ├── perfect/│       │   │   ├── ACCESS-CM2_2080-2099.nc│       │   │   └── ...│       │   └── imperfect/│       │       ├── ACCESS-CM2_2080-2099.nc│       │       └── ...│       └── target/│           ├── pr_tasmax_ACCESS-CM2_2080-2099.nc│           └── ...

Data Preprocessing for the NZ domain

For information (roughly) on how the data was preprocessed, please see the following repository: https://github.com/nram812/CORDEXBench-nzdomain-preprocessing