The objective of the project is to better understand what controls the size of intense storms, also known as deep convective systems. The larger the storm is the more it has consequences in terms of extreme events or radiative effect. A particular pattern about the deep convective systems life cycle is that they simply linearly grow then linearly decay, so that with only three parameters -the maximal size reached, the total duration and the time of maximal duration- we can capture the full life cycle. We assume that knowing the beginning of the storm development we can predict its mature size. We try to test this hypothesis on global high resolution simulations, part of the project DYAMOND-nextgems. We use the cloud resolving model SAM and the tracking algorithm to detect deep convective systems is TOOCAN. We endeavor to apply machine learning algorithm with these data. We design a dataset that map early growth scalar features of the systems with their mature size. We compute this dataset in two steps.The first steps consists in slacking physical 2d fields surrounding the systems during their first 5 hours of development and link it to their mature size, as well as the total duration and the duration of growth. Additionally, the dataset contains meta-information about the system, especially to relieve morphological features from the raw data.
The second one consists in reducing the dimension of this dataset by computing associated scalar features to 2d physical fields. The different machine learning algorithms are trained on scalar vector.