Autoregressive Deep Learning Earth System Models
15.05.2025Interested in AI and Climate Science? Simulating the Earth’s climate with Earth System Models (ESMs) is essential for understanding and predicting climate change — past, present, and future. These physics-based simulations are computationally expensive, but DL based approaches offer a more efficient alternative for climate projections.
While this is a fast-moving and fascinating research area, currently a major challenge is making sure DL models stay stable over long simulations so that the predictions remain physically consistent. We are actively exploring innovative model designs and training techniques to address this.
Possible thesis topics include (based on your interests and experience):
- Experimenting with different rollout strategies (e.g. latent rollouts, ensembling, different lead times
- Applying diffusion models to long-term climate simulations
- Adding memory states to existing DL models
If this sounds exciting to you, feel free to reach out — we’re happy to discuss project ideas and help you find the right fit.
Your profile
- Student of computer science, maths or physics
- Solid theoretical foundations in Deep Learning
- Practical experience programming with Python and PyTorch/Tensorflow
- Ideally experience with git
- Ideally experience in working with spatio-temporal/image/video data
Supervisor: Florian Gallusser
