Stochastic generation of precipitation by generative AI

A Stochastic Rainfall Generator (SRG) is a probabilistic model aimed at simulating synthetic rainfall data that has the same statistical signature as observations. This involves a phase of statistical learning of the probability density (pdf) of rain from observations, then a phase of sampling this pdf to generate synthetic rain fields. Different statistical models have been used in the past for stochastic generation of precipitation in order to simulate different properties of rain: random fields to model spatio-temporal dependencies, fractals to preserve links between scales, or point processes to take takes into account aggregates within rain fields.

Recently, generative AI has been successfully applied to the simulation of rain through deep generative models of rainfall (DGMR) which have given very promising results in the areas of short-term forecasting. from radar images (nowcasting) and the disaggregation of satellite images (downscaling). The main difference between DGMRs and traditional SRGs is that DGMRs replace the parametric statistical model of SRGs with a deep neural network. This innovation opens the way to the simulation of more complex rain fields with potentially more realistic spatio-temporal patterns.

In this project we will explore the capacity of generative AI to simulate different types of precipitation observed in a temperate climate (stratiform rain during the passage of a warm front, sleet, thunderstorms, etc.). For this, we will develop a deep neural network dedicated to the stochastic generation of high-resolution spatio-temporal rain fields, as well as a strategy to train this network from weather radar images. This generator will have a fully convolutional architecture which will make it easy to extend the space-time window during simulation. The neural network will be trained by adversarial learning.

Once the DGMR is obtained, we will compare its results to those of a SRG based on a parametric model. We will evaluate the realism of the simulated rainfall, the capacity of each model to reproduce the statistical signature of the observations, as well as their ability to simulate probabilistic ensembles reflecting the natural variability of rainfall observed in temperate climates. After evaluation, the generative AI rain simulator can be used for stochastic downscaling of climate projections, and to conduct studies on the impact on the hydrosphere of the change in the structure of rain under climate change.