CORDEX-ML-Bench

A Standardized Benchmarking Dataset for Data-Driven Regional Climate Downscaling

📋 Overview

CORDEX-ML-Bench provides a standardized framework for evaluating machine learning models for regional climate downscaling. Participants will train models on Regional Climate Model (RCM) simulations and test their performance across different time periods and spatial domains. This Zenodo repository only provides the training data, other relevant information will be provided below. 

Key Features

• RCM simulations spanning three distinct geographic domains. Allowing for testing and evaluation of emulators in multiple domains. Currently, this repository is in a testing phase, and not all the data has been uploaded. 
• Evaluation framework across historical and future periods. CORDEX-ML-Bench provides tools for training and evaluating ESD and RCM emulators. These can be found here📦 WCRP-CORDEX/ml-benchmark.

🌍 Geographic Domains

An illustration of the predictor (16 x 16)  and target domains (128 x 128)  are shown above. Noting that the predictor domains are slightly larger than the target domains. 

Predictor Variables (16 total, 16 x 16)

• Atmospheric variables at 850 hPa, 700 hPa, 500 hPa :
  • u - zonal wind component
  • v - meridional wind component
  • q - specific humidity
  • t - temperature
  • z - geopotential height
• Static field: Orography (topography)

🔧 Getting Started

Loading Data

For code examples and tutorials on loading the training data, refer to the official repository:

📦 WCRP-CORDEX/ml-benchmark

Data Preprocessing

Region-specific preprocessing information:

• NZ Domain: nram812/CORDEXBench-nzdomain-preprocessing
• ALPS Domain: jgonzalezab/CORDEXBench-alpsdomain-preprocessing

🎯 Training Framework

Perfect Framework Training

The benchmark investigates training in the "perfect" framework:

• X (predictors): Coarsened RCM fields
• y (targets): High-resolution truth from RCM

Evaluation is supported in both perfect (coarsened RCM) and imperfect (applied directly to the GCM) configurations.

🧪 Experimental Setups

1️⃣ ESD Pseudo-Reality (1961–1980)

Training Period: 20 years (1961–1980)
GCM Examples: ACCESS-CM2 (NZ), CNRM-CM5 (ALPS)

This experiment mimics traditional Empirical Statistical Downscaling (ESD) training.

Configurations:

• ✅ With orography
• ❌ Without orography

2️⃣ Emulator Historical + Future (1961–1980 + 2081–2100)

Training Period: 40 years combined
Periods: Historical (1961–1980) + Future (2081–2100)
GCM Examples: ACCESS-CM2 (NZ), CNRM-CM5 (ALPS)

This experiment evaluates model extrapolation skill and transferability across different GCMs.

Configurations:

• ✅ With orography
• ❌ Without orography

📊 Evaluation Framework

Test Periods

• Historical: 1981–2000
• Mid-Century: 2041–2060
• End-Century: 2081–2100

Evaluation Types

• PP Cross-Validation: Same GCM used in training and testing
• Imperfect Inference: Same GCM but different realizations or noise
• Transferability Testing: Different GCM than training set
• Change Signal Evaluation: Climate change response in future periods

📋 Evaluation Tables

Table 1: ESD "Pseudo-Reality" Evaluation (T1)

Table 2: Emulator Evaluation (T2)

📁 Data Structure

New Zealand Domain

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/
    │   │   └── imperfect/
    │   └── target/
    │
    └── end_century/
        ├── predictors/
        │   ├── perfect/
        │   └── imperfect/
        └── target/

ALPS Domain

ALPS_Domain/
├── train/
│   ├── ESD_pseudo-reality/
│   │   ├── predictors/
│   │   │   ├── CNRM-CM5_1961-1980.nc
│   │   │   └── static.nc
│   │   └── target/
│   │       └── pr_tasmax_CNRM-CM5_1961-1980.nc
│   │
│   └── Emulator_hist_future/
│       ├── predictors/
│       │   ├── CNRM-CM5_1961-1980_2080-2099.nc
│       │   └── static.nc
│       └── target/
│           └── pr_tasmax_CNRM-CM5_1961-1980_2080-2099.nc
│
└── test/
    ├── historical/
    │   ├── predictors/
    │   │   ├── perfect/
    │   │   │   ├── CNRM-CM5_1981-2000.nc
    │   │   │   └── MPI-ESM-LR_1981-2000.nc
    │   │   └── imperfect/
    │   │       ├── CNRM-CM5_1981-2000.nc
    │   │       └── MPI-ESM-LR_1981-2000.nc
    │   └── target/
    │       ├── pr_tasmax_CNRM-CM5_1981-2000.nc
    │       └── pr_tasmax_MPI-ESM-LR_1981-2000.nc
    │
    ├── mid_century/
    │   ├── predictors/
    │   │   ├── perfect/
    │   │   └── imperfect/
    │   └── target/
    │
    └── end_century/
        ├── predictors/
        │   ├── perfect/
        │   └── imperfect/
        └── target/

🎓 Key Terminology

• Perfect Inference: Using the same GCM realization that was used in training
• Imperfect Inference: Using different realizations or added noise from the same GCM
• PP Cross-Validation: Perfect-Prognosis evaluation with the same GCM
• Hard Transferability: Testing on a completely different GCM than used in training
• Change Signal: The difference between future and historical climate periods

📚 Resources

• Main Repository: WCRP-CORDEX/ml-benchmark
• NZ Preprocessing: CORDEXBench-nzdomain-preprocessing
• ALPS Preprocessing: CORDEXBench-alpsdomain-preprocessing

Last Updated: November 2025