Computational algorithms and synergistic models in the Arctic

The applied methodology is based on the optimization of a Transformer-type neural network architecture, historically developed for large linguistic models and here adapted for the processing of environmental time series. The computational system simultaneously processes a plurality of input vectors, including historical particulate matter measurements, dynamic projections provided by the Copernicus CAMS system, local meteorological variables and geographical parameters of the monitoring stations. This multifactorial system allows us to generate projections relating to PM10 concentrations at a time interval of 48 hours.

Alice Cuzzucoli, Cnr-Iia researcher and first author of the study, specified: “Comparing the model predictions with what actually happened, our results always proved to be significantly better than those of the classic models used so far, even in the evaluation of particularly extreme concentration peaks”.

The element of structural innovation lies in the non-exclusive integration of artificial intelligence within existing computing systems. The algorithm does not replace traditional dynamic models, but uses their outputs to refine the final statistical calculation.

Antonello Pasini, co-author of the study and scientist at the Cnr-Iia, clarified: “We obtain the best results by using AI in a synergistic way with respect to classic dynamic models, and not in an alternative way, also using their results as input“. The need to have such predictive tools also appears linked to the geopolitical and climatic changes taking place in the Arctic area, where the opening of new maritime trade routes and the increase in the frequency of high-latitude forest fires determine a quantitative variation in atmospheric emissions. constant monitoring and timely information are essential prerequisites for the protection of vulnerable ecosystems and the planning of health risk mitigation strategies for resident populations.