Leggi in italiano

Mount Etna eruption, Sicily, Italy. Credit: Antonio Zanghì/ Moment/ Getty Images.

The way magma is stored and transported in volcanoes is difficult to investigate. Current models assume that volcanoes have multiple reservoirs at different levels, interconnected by a network of conduits that channel magma to the surface. Knowing how pressure and temperature vary between levels is important to understand the structure of a volcano’s plumbing system, also giving clues for predicting the rise of magma within the crust.

Lorenzo Chicchi and his team from the University of Florence have developed a machine- learning tool called GAIA (Geo Artificial Intelligence thermobArometry), that can estimate pressure and temperature conditions of magma by considering the components of a specific mineral found at the surface1.

The system is based on feed-forward neural networks, one of the most common architectures in machine learning, where artificial neurons are organized into different layers that are connected to each other. The initial layer receives the input signal, which is then processed and passed on to the next layer, until the final layer where the output is obtained. “The aim is to use the network to reproduce an unknown function starting from a dataset of examples,” says Chicchi. “The parameters are adjusted to make the network responses as consistent as possible with the initial dataset.”

Specifically, GAIA’s neural networks take as input the composition of the clinopyroxene mineral in erupted magmas, and return an estimate of the temperature and pressure at which the mineral crystallized. Clinopyroxenes are prevalent minerals in certain type of magmas, and their chemical composition is sensitive to crystallization conditions. As a training dataset, the team used the database of clinopyroxene compositions of the five most active Italian volcanoes (Somma-Vesuvius, Campi Flegrei, Etna, Stromboli and Vulcano).

The team trained multiple neural networks on distinct subsets of the dataset. Having multiple models enabled them to generate a range of estimates rather than just one, providing both an average value and an error margin for each clinopyroxene composition. In the end, they compared the results of GAIA with those from geophysical and geodetic surveys of the five Italian volcanoes, and found they were in good agreement.

“GAIA is a tool that can improve our understanding of the dynamic of volcano plumbing systems,” adds lead author, Simone Tommasini, from the University of Florence. “It can be applied to each single eruptive period of a given volcano to assess variations of the depth of magmatic reservoirs and also provide further basis for one of the key research areas in modern volcanology, that is the prevention and mitigation of volcanic hazard”.