The program aims to develop a cross-cutting data-centered approach that will complement and reinforce, through specific projects, other programs with more vertical thematic aims. The need to structure a program on this theme stems from the enormous and continuing growth in the availability of data, even from nontraditional sources, which, together with AI techniques, can be used even if not directly concerning the primary variables of interest. In addition, many policies define a new paradigm on the use and sharing of data, which are no longer considered an aid to achieve certain ends but become a goal in their own right.
The program builds on both existing know-how, such as modeling capabilities and process knowledge at a level of spatial detail far beyond that typical of the earth system modeling communities, and know-how of more recent interest to CIMA Research Foundation. These include fundamental AI techniques, particularly machine learning.
The first new field of action to be explored for the characterization of the program is based on the use of techniques that combine AI with data assimilation methodologies (e.g., Deep Data Assimilation), with potential for application both on the topic of impact-based forecasting in real time and on that of multi-hazard profiles with a probabilistic approach. This action may make it possible to solve some critical issues that still prevent the operational use of advanced data assimilation techniques in some elements of the modeling chain or the use of data not directly linked to the main predictive variables. Specifically, the first objectives to be pursued (in part already included in recently activated projects) are the coupling of 4DVAR and Ensemble Kalman Filter techniques; the application of deep neural networks in statistical downscaling to improve the prediction of air quality and wind energy production; the exploration of the topic of selection of observations for data assimilation and feature engineering with deep learning techniques; and, finally, the ‘execution of bias corrections for climate scenarios with machine/deep learning techniques.