koomited/AIMS-PROJECT

Thesis project for AI For Science Master project : Fine tune Aurora model for weather prediction in South Africa # AMy Paper Title --- This repository is the official implementation of Towards Affordable Weather Forecasting in South Africa: Parameter-Efficient Fine-Tuning of an AI Weather Foundation Model. --- # Overview Accurate weather forecasts are critical for climate resilience, but traditional numerical weather prediction (NWP) systems are too costly and resource-intensive for many African regions. AI weather prediction (AIWP) models offer a faster and more affordable alternative, but they usually struggle in Africa due to limited meteorological infrastructure and scarce local data. This project explores fine-tuning a pretrained AI foundation model (Aurora) on South African weather data. By adapting state-of-the-art models to regional conditions, we aim to deliver affordable, locally adapted forecasts that can support communities, researchers, and policymakers across the continent. Key goals: - Make weather forecasting more accessible in low-resource settings - Improve predictive accuracy in Africa through regional fine-tuning - Contribute to climate resilience with AI-driven solutions # How to run? ## Clone the repository ```bash git clone github.com ``` ## Create a virtual environment and activate it ```bash python3 -m venv venv source venv/bin/activate ``` ## Install the requirements ```bash pip install -r requirements.txt ``` ## Pretrained models Make sure you download the checkpoint of the pretrained models you need, not the fine tuned one if you want to train, first from Hugging Face and save them in a folder named `model` in your root directory. You can explore how to do that with this notebook `notebooks/model_surgery.ipynb` Note that we download the ones we need, the small and large 0.25° resolution models. You can donwload other version of the model from Huggin Face To run the code in this project a linux environment particulary ubuntu is required. Also notice that we build the whole project on Google …