ICCS Summer School 2026 — Temperature Downscaling with a U-Net

This repository contains the ICCS Summer School 2026 practical on spatial downscaling of ERA5 2-m temperature over Southeast Asia using a U-Net in PyTorch. The notebooks can be run with Google Colab (on a GPU-based runtime) without local installations. Participants are strongly encouraged to use the VS Code plugin for Google Colab.

Contents

• Learning Objectives
• Teaching Material
• Preparation and Prerequisites
• Data
• CDS API Key Handling
• Installation and Setup
• License
• Contribution Guidelines and Support

Learning Objectives

By the end of this practical, participants should be able to:

• Explain spatial downscaling and why it matters for climate applications.
• Describe the U-Net architecture for image-to-image regression.
• Prepare geospatial fields for CNN training in PyTorch.
• Train and evaluate a U-Net for 1° to 0.25° temperature downscaling.
• Explore extension ideas beyond the baseline model.

Teaching Material

Exercises

• notebooks/unet_temperature_downscaling_exercise.ipynb
  • Participant notebook used during the session.
  • Contains guided TODO-style implementation steps.

Worked Solution

• notebooks/unet_temperature_downscaling_solution.ipynb
  • Complete worked solution for debrief and post-session review.

Recommended teaching order:

1. Start with exercise notebook.
2. Use solution notebook for walkthrough and recap.

Running on Google Colab

You can run both notebooks directly in Colab from this repository:

• Exercise notebook: notebooks/unet_temperature_downscaling_exercise.ipynb
• Solution notebook: notebooks/unet_temperature_downscaling_solution.ipynb

The "Open in Colab" badge at the top of this README launches the repository in Colab.

Preparation and Prerequisites

To get the most out of this practical, participants should have basic familiarity with:

Mathematics and Machine Learning

• Basic regression concepts.
• Matrix and tensor representations of data.
• High-level neural network concepts (layers, training, validation).

Python

• Core Python syntax and functions.
• Basic use of numpy, matplotlib, and Jupyter notebooks.
• Basic PyTorch familiarity is helpful but not required.

Git and GitHub

• Cloning a repository.
• Committing and pushing changes.

Pre-Session Requirement (Important)

Participants should download data before attending. CDS downloads can take significant time depending on queue/server load; doing this during the class can consume most of the hands-on session. If you do not have a CDS API key, you can download the pre-procured data using gdown from Google Drive (instructions in the notebook).

Data

The tutorial expects these files in data/:

File	Description
data/gph_se_asia.nc	Surface geopotential over SE Asia (static field)
data/t2m_2000.nc	ERA5 2m temperature (2000, 6-hourly, 0.25°)
data/t2m_2001.nc	ERA5 2m temperature (2001, 6-hourly, 0.25°)

CDS API Key Handling

(Ignore if downloading from Google Drive) The notebooks are set up to read a local API key file when DOWNLOAD_DATA=True.

• Create a file named CDS_API_KEY in the repository root.
• Put your CDS API key on a single line.
• The code writes ~/.cdsapirc automatically when download is enabled.
• Notebook source keeps REPLACE_WITH_YOUR_OWN_CDS_API_KEY as a safe placeholder.

CDS_API_KEY is ignored by git via .gitignore.

Installation and Setup

There are two options for participating in this practical:

• via a Local Install
• on Google Colab

We recommend the local install approach if you want to keep a persistent working copy and run reliably without session limits.

Local Install

1. Clone or fork the repository

Clone via HTTPS:

git clone https://github.com/Cambridge-ICCS/iccs-ss-2026-downscaling-tutorial.git
cd iccs-ss-2026-downscaling-tutorial

If you want to preserve your changes in your own GitHub account, fork first and then clone your fork.

2. Create and activate a virtual environment

The guide below assumes a local unix installation (Linux, macOS). Windows users are instructed to use Windows Subsystem for Linux (WSL).

python3 -m venv .venv
source .venv/bin/activate

3. Install dependencies

pip install --upgrade pip
pip install -r requirements.txt

4. Run Jupyter Notebook

From repository root:

jupyter notebook

Then open one of:

• notebooks/unet_temperature_downscaling_exercise.ipynb
• notebooks/unet_temperature_downscaling_solution.ipynb

5. (Optional) Register the environment as a notebook kernel

If your notebook server does not detect the environment, run:

python -m ipykernel install --user --name=downscaling-venv

Google Colab

Use the badge at the top of this README to open the repository in Colab.

Notes:

• Running in Colab requires a Google account.
• Colab sessions are ephemeral, so save your notebook outputs and changes externally.

License

See LICENSE.

Contribution Guidelines and Support

If you spot an issue with these materials, please open an issue in the repository.

Contributions are welcome. A typical workflow is:

1. Fork the repository.
2. Create a branch with your changes.
3. Open a pull request describing the updates.

For support adapting or delivering this practical, please contact ICCS via GitHub.