A. Kahira in collaboration with L. P. Caron, worked on the application of convolutional neuronal networks to Hurricanes and cyclones observational databases to predict the number of such extreme events for the subsequent years.

• Extreme Events (Hurricanes/TCs). Deep Neural Networks from the field of computer Vision can be used for pattern recognition in a supervised way, exploiting reanalysis data and historical event databases (IBTrACS for TC). The idea here is to obtain segmentation masks with the identification of specific events and with the possibility of tracking their future state. The problem addressed here is similar to the one tackled by A. Kahira, but re-framing the goal (which entails using different networks) and using data with a finer temporal sampling.

• Deep Neural Networks as surrogate models (learning to replace parameterizations). We explore the use of encoder-decoder networks, such as those used for image translation in Computer Vision, for the task of learning precipitation fields from other predictor variables (geopotential, SST, etc). In a way what we do is to learn a transfer function, capturing information from predictor variables to produce precipitation, a variable highly dependant on parameterizations schemes.

Inference of surface pollutant concentrations from satellite and global scale meteorological and chemical simulations (AQ-WATCH project). Application to NO2 (TROPOMI sensor) and aerosols.

Statistical correction of air quality forecasts with machine learning algorithms (ML4AQ grant). Application to Mexico City.