Overview
Traditional numerical weather prediction solves complex physical equations on supercomputers, which is accurate but computationally expensive and often struggles to represent uncertainty. Recent machine-learning models can forecast weather quickly but may produce deterministic, overconfident predictions or unstable long-term rollouts.
Led by Martin Butz at the University of Tübingen, the project “Neural Generative Weather Forecasting” (2021–2025) investigated how generative neural models can improve probabilistic and long-range weather forecasting. The project focused on architectures that explicitly model uncertainty and generate multiple plausible future weather trajectories, rather than a single deterministic forecast.
Key results
- Developed DISTANA, a distributed spatiotemporal graph neural network that learns to predict weather dynamics from data while inferring hidden causal factors such as land–sea masks and other spatially varying influences.
- Combined DISTANA with active tuning, a retrospective latent-state inference method that adjusts hidden states based on past observations, improving stability and accuracy over long autoregressive rollouts.
- Designed generative weather-prediction models that produce ensembles of future states, enabling better-calibrated probabilistic forecasts and more realistic representation of uncertainty compared to non-generative deep-learning baselines.
- Advanced parsimonious deep-learning weather models (e.g., DLWP-HPX) that use coarse global meshes and reduced variable sets but still achieve competitive medium-range forecast skill, demonstrating that carefully designed inductive biases can yield efficient, stable models.
- Explored hybrid approaches that integrate neural components with physical constraints, aiming to combine the speed of data-driven models with the stability and interpretability of physics-based forecasting.
Why it matters
Accurate, uncertainty-aware weather forecasts are critical for agriculture, energy, disaster preparedness, and many other domains. Generative neural models can complement traditional numerical weather prediction by providing fast, probabilistic forecasts and by learning representations that capture complex spatiotemporal dependencies. The project’s methods contribute to more reliable ensemble forecasting, better-calibrated uncertainty estimates, and more efficient long-range predictions. As a research project, it primarily advances methodology and understanding; it does not replace operational forecasting systems but provides tools and insights that could inform future weather-prediction architectures.
This project produced the following peer-reviewed publications:
- Karlbauer, M., Menge, T., Otte, S., Lensch, H. P. A., Scholten, T., Wulfmeyer, V., & Butz, M. V. (2021). Latent State Inference in a Spatiotemporal Generative Model. arXiv:2009.09823 [cs.LG]. (Revised version 2021; related work on DISTANA and active tuning.)
- Karlbauer, M., Cresswell-Clay, N., Durran, D. R., Moreno, R. A., Kurth, T., Bonev, B., Brenowitz, N., & Butz, M. V. (2023). Advancing Parsimonious Deep Learning Weather Prediction using the HEALPix Mesh. arXiv:2311.06253 [cs.LG].
- Thuemmel, J., & Butz, M. V. (2023). Inductive Biases in Deep Learning Models for Weather. arXiv:2304.04664 [cs.LG].
- Additional related work by Martin Butz and collaborators on spatiotemporal generative models, causal wave dynamics, and neural processes for short-term weather forecasting, published in venues such as arXiv preprints and machine-learning and climate-AI workshops.
