A new machine learning model for predicting magma viscosity
A team of researchers led by Charles Le Losq (IPGP, IUF) has developed an innovative machine learning model capable of predicting the viscosity of magmas over a wide range of compositions, temperatures and pressures. This work represents a major advance in our understanding of magmatic processes, both on Earth and on exoplanets.
Molten magma on an exoplanet (generated by AI)
Publication date: 04/03/2025
Press, Research
Related teams :
Cosmochemistry, Astrophysics and Experimental Geophysics (CAGE)
An unprecedented database and a precise model
To train this model, the researchers compiled an exhaustive database containing almost 29,000 measurements of the viscosity of molten silicates, including high-pressure data up to 30 GPa. The model combines artificial neural networks with Gaussian processes, enabling accurate prediction of magma viscosity, even under extreme conditions such as those encountered on exoplanets.
Application to the exoplanet K2-141 b
By applying this model to the exoplanet K2-141 b, the team demonstrated the dominant role of temperature in controlling the viscosity of a magma ocean located on its diurnal surface. In addition, they determined that this planet probably has a thin atmosphere of rock vapour with a pressure of around 0.1 bar within a radius of 40° of the substellar point.
A notable result of the study is the interpretation of the night-time temperatures of K2-141 b. With values above 400 K, the researchers suggest the presence of a partially molten mantle beneath the surface, providing new insights into the internal structure and dynamics of lava planets.
An open-source library for the scientific community
The model developed is available in the form of an open-source Python library, gpvisc, enabling researchers around the world to explore the viscosity of magmas in various planetary and industrial contexts. This breakthrough opens up new perspectives for modelling magmatic processes and characterising rocky exoplanets.
This work was funded by the Labex UnivEarthS, ANR-10-LABX-0023 and ANR-18-IDEX-0001.
