Abstract
In the era of artificial intelligence and big data, deep understanding of the Earth system and, in particular, the prediction of extreme weather events, requires to conceive our planet as a brick of a more complex heliophysical environment. In this perspective, the protection of satellite monitoring systems together with the exploitation for forecasting purposes of the data provided by these systems, are based on computational tools with predictive capacity and able to extract from the data the most significant features in terms of information content.
This project intends to build a pipeline of artificial intelligence (AI) techniques able to calibrate 1) numerical models based on the physics of space weather, with the aim of protecting satellite infrastructures designed for environmental and meteorological monitoring; and 2) numerical models based on the physics of the atmosphere, with the aim of anticipating the occurrence of extreme weather events and supporting decisions of civil protection agencies.
On a technical level, this project intends to integrate the most innovative machine and deep learning techniques with numerical models to simulate highly complex dynamical systems. On a strategic level, the involvement of territorial partners with notable experience in applied research, technology transfer, and territorial protection, will ensure systematic validation and a level of dissemination of the project results aimed to enhancing AIxtreme technological and social impacts.
Relevance
Weather monitoring on Earth and, in particular, forecasting of extreme events, require the availability of huge amounts of heterogeneous data that can be used to either initialize deterministic numerical models or to feed and train data-driven AI networks. Most of these datasets come from satellite systems whose protection from adverse impacts of space radiation and materials is therefore becoming more and more crucial. Indeed, activity on the solar surface is responsible for a particular type of events grouped under the name space weather. The Sun continuously emits particles from its corona into space and this stream of particles is known as solar wind. Earth's magnetic field and atmosphere protect us from the majority of the solar wind blast but it may happens that magnetic activity within the Sun causes intense explosions, called solar flares. Solar flares send tons of energy through space at the speed of light. Sometimes flares come with huge solar eruptions, called Coronal Mass Ejections. The solar wind gets much stronger during these storms and can be very dangerous. All of that extra radiation and ejected matter can damage the satellites we use for communications, navigation and monitoring the weather and climate on the Earth. The radiation from solar storms can also be dangerous for astronauts in space. At Earth, disruptions of power grids that provide electricity are possible.
For these reasons monitoring and forecasting space weather is of crucial relevance to protect critical infrastructures such as meteorological satellite services.
On the other hand, on-Earth weather anomalies possibly related to current climate change are becoming more and more frequent and more and more localized.
This fact implies the need of forecasting and nowcasting services able to predict in advance very seldom events that, however, may have significant impacts on modern societies' assets, including urban floods and deterioration of agricultural lands.
In the recent past several weather and space weather forecasting methods have been proposed. On one hand, there are the deterministic models simulating the propagation of CMEs and solar wing across the heliosphere (in the case of space weather prediction) and mimicking meteorological events often triggered by the sea (in the case of weather prediction). On the other hand, advanced machine learning techniques together with advanced technological skills have been exploited for building weather and space weather prediction systems, just looking at the data. Both approaches are characterized by strengths and weaknesses, and for this reason neither of them currently provides optimal predictions.
In this scenario AIxtreme intends to explore a third way that combines the strengths of the two previous approaches. The novelty of this new approach lies in the inclusion of the physical knowledge, typical of the deterministic approaches, into the learning process typical of a learning machine.
Objectives
The protection of critical satellite infrastructures and the exploitation of the environmental information they can collect, requires sophisticated forecasting tools in terms of predictive capacity. The unifying aim of the project is to provide a step forward in the development of this capacity, with the definition of tools, methods, algorithms, and software, for the prediction of extreme phenomena at both weather/space weather level. The main objective of AIxtreme project then deals with the development and validation of Artificial Intelligence methods for
the forecast of the space weather, as a fundamental element for the efficiency and protection of meteorological satellite systems, and for
the use of satellite measurements for climate monitoring, for estimating and forecasting the probability of occurrence and the intensity of extreme atmospheric events.
Methods
As already pointed out the prediction problem can be faced in two different ways. One can follows a purely data-driven approach (typical of machine learning/deep learning techniques) that can benefit from a huge amount of data to learn from, or deterministic models including physical knowledge on the propagation processes in space and on-Earth. Since in both cases the physics that generates extreme events is not completely known, in the presence of extreme events data-driven approaches have too few examples to learn from, while deterministic approaches cannot reproduce physical phenomena not described in the model. As a consequence, both approaches provide predictions with a level of accuracy that still has to be enhanced.
We here propose a different methodological Physics-Informed approach to the prediction problem where the knowledge of physical laws drives the learning process of a neural network.
The progress of the project and the effectiveness of the implemented computational techniques will be monitored through the definition of counterfactual protocols. All the developed techniques will be tested on the same historical data and, in particular
the performances of data-driven, purely AI approaches, will be compared to purely deterministic approaches, and
the performance of integrated Physics-Informed approaches will be compared to purely deterministic approaches.
The knowledge of the true outcome will allow us to quantitatively evaluate the potential of the developed techniques.
Partners
Partners of this project are characterized by different required expertise to combine Physics-based models and AI methods for Space Weather and Weather forecasting purposes. The Leading Unit of the whole project is the Dipartimento di Matematica, Università degli Studi di Genova (DIMA-UNIGE) where the Methods for Image and Data Analysis (MIDA) group has a longstanding experience in the development of computational techniques and AI methods with application to solar Physics problems. The project is subject to evaluation by the Impact Evaluation Unit of the Collegio Carlo Alberto (CCA).
The main project (AIxtreme) is then divided into two sub-projects focused on Space Weather (AIxtreme-I) and Weather (AIxtreme-II) forecasting topics, respectively, for which two groups of experts complementary with each other have been identified.
Leader of Project 1 (AIxtreme-I) is the Dipartimento di Scienze Matematiche, Politecnico di Torino (DISMA-POLITO) assisted by two territorial partners: Istituto Nazionale di Astrofisica - Osservatorio Astrofisico di Torino (OATo-INAF) and Aerospace Logistics Technology Engineering Company S.pA. (ALTEC), Torino. DISMA-POLITO can make at disposal of the project expertise in the numerical treatment of mathematical models described by partial differential equations (such as the models tracking/modeling space weather phenomena through the heliosphere) while OATo-INAF and ALTEC will contribute to the validation of the project’s results and to the transfer of the knowledge produced by the research activity into solar community and industrial contexts.
Leader of Project 2 (AIxtreme-II) is the Dipartimento di Ingegneria Civile, Chimica e Ambientale, Università degli Studi di Genova (DICCA-UNIGE), assisted by ARPAL - Agenzia Regionale per la Protezione dell’Ambiente Ligure, Genova, as territorial partner. DICCA-UNIGE is a leading group in the study and application of Weather Prediction Models for environmental needs and for this reason its research activity is particularly focused to the topics of this project. The transfer of knowledge produced by the project is here guaranteed from the involvement of the agency for the protection of the Ligurian environment (ARPAL).