Join us
Internship
Frequencies of extreme precipitation events in the context of climate change at the European scale
A thesis work started in 2022 made it possible to develop a processing chain combining extreme value theory and bias correction method, in order to quantify the evolution of the frequencies (or return times) of certain bivariate compound climatic events during the XNUMXst century.
The objective of this internship is to deploy this approach on a European scale, and possibly for new bivariate composite events inspired by recent events, such as the Boris storm in September 2024. This internship involves several disciplines/skills. At the IT level, it will involve automating the current processing chain and setting up visualization tools. At the statistical level, questions of spatial dependencies can be addressed. At the climate science level, it will involve comparing the results obtained with current climate knowledge.
More information in the detailed offer below:
Generation of heavy tailed distributions with applications to rain generation
State-of-the-art generative models demonstrate remarkable performance on challenging high-dimensional data modalities. Current implementations predominantly utilize Variational AutoEncoders, generative adversarial networks or denoising diffusion generative models. However, the standard theoretical guarantees for these approaches usually rely on assumptions that do not hold for heavy-tailed distributions.
One such heavy-tailed distribution is the rainfall distribution derived from radar imaging of specific regions. While one might expect methods effective for typical image distributions to be well-suited for the task at hand, the unbounded and potentially heavy-tailed nature of the rainfall distribution poses significant challenges. Moreover, current research on climate change suggests that the likelihood of extreme rainfall events is expected to increase in the future. Therefore, generating data from the tail of this distribution is a critical task for urban planning and insurance.
The internship goal is to explore a subset of the following research lines.
• Understand the different theoretical guarantees holding for GAN, VAE and Diffusion models and identify their limitations when handling heavy-tailed data.
• Identify the main limitations of the existing algorithms in the specific benchmark datasets (homogeneous vs non-homogeneous, tail vs bulk generation, etc.) and explore new architectures adapted to such data.
• Explore adaptations of current proof strategies in the existing literature to adapt the frameworks to handling heavy tailed data. A natural first step will be to investigate VAE for heavy-tailed distributions.
More information in the detailed offer below:
Inference of spatio-temporal models by graph neural networks
The objective of this M2 internship is to train and validate an architecture based on graph convolutional networks (GNN) proposed during a previous internship and to develop an architecture adapted to the spatio-temporal context for the estimation of parameters of complex geostatistical models. A challenge of this generalization is the definition of a new spatio-temporal neighborhood structure in the GNN and the estimation of a larger number of parameters, which has never been implemented so far. We will then compare the proposed strategy with state-of-the-art spatio-temporal inference methods on simulated datasets. Finally, we will seek to adapt the approach for models derived from stochastic partial differential equations (SPDE). An application to a real dataset of solar radiation measurements will conclude the internship.
More information in the detailed offer below:
