For the department Systems Analysis (SYS) at the Stuttgart site, we are seeking as soon as possible a student (m/f/d) to work on a

Master Thesis (m/f/d) in the field of Machine Learning Methods in Numerical Weather Prediction

In 2022 a group of researchers affiliated to Google published GraphCast ([Lam et al., arXiv:2212.12794]), a Graph Neural Network model for global weather forecasting. This model is a data-driven surrogate for one of the most advanced global numerical weather prediction (NWP) models , the Integrated Forecasting System High Resolution Model (IFS HRES) developed and operated by the European Centre for Medium Range Weather Forecasting (ECMWF). The IFS HRES forecasts the global state of the atmosphere on a grid with 9 km horizontal resolution on 137 vertical levels in 6 hour increments.

The goal of this master thesis will be to dive into machine learning advancements in weather forecasting by examining the GraphCast model and its ongoing research applications and implementing a forecasts pipeline using the GraphCast model on ZSW-local GPU architecture.

Your responsibilities:

• Description of the GraphCast model architecture and data basis
• Review of alternative approaches to data-driven weather forecasting (FourCastNet, FuxiWeather, PanguWeather)
• Review of the current state of 'derivative' research building on the GraphCast results with a focus on GNN modells
• Implementation of a framework to run GraphCast predictions on architecture provided by ZSW
• Evaluation of GraphCast predictions against ICON-D2 data
• Translating the Local Area Modeling approach [Oskarsson et al., arXiv:2309.17370] to ICON-D2 domain
• Training a GraphCast-based Local Area Model on historical ICON-D2 forecasts
• Evaluating model configurations