Title: Generating Artificial Weather Data with Machine Learning

Project duration: March 2023 – present

Research Area: Compound Events, Climate Science, Meteorology

Studying rare weather events requires large observational datasets. A “once-in-a-century” flood by definition occurs only once every 100 years on expectation. Researchers interested in the impacts of extreme events often use models to simulate which weather situations lead to particularly strong consequences in a system of interest (hydrological catchments, forests, crops, …). Large datasets of “artificial weather data” (i.e. data that mimic the statistics of real weather) can help in finding patterns in the behaviour of these impact models – and thereby improve understanding of both model and real system. This PhD project employs techniques from probabilistic modelling and Machine Learning to generate such artificial data.

Aims

The goal of the project “Generating Artificial Weather Data with Machine Learning” is to develop a methodology that can be applied flexibly to generate artificial weather data for a range of impact models. The simulated data should capture statistical properties of real weather – including spatio-temporal dependencies, extremes, and multivariate dependencies.

Problem

The weather system is chaotic, with many components interacting non-linearly on multiple timescales. Developing weather generators that can capture the resulting dependence structures is a very challenging task. Additionally, the amount of observations that can used as training data is finite. And climate change introduces changes the statistics of weather.

Technology

• Statistical methods: resampling, similarity metrics, Markov Chains
• Deep Learning, especially techniques used for modelling videos and images

Team

Lead

• Jonathan Wider

Team Members

• Prof. Dr. Jakob Zscheischler