Skip to content

Combining thermodynamic- and machine learning-based thermobarometry

Luca Caricchi1, Martin Miranda Muruzabal1, Clothilde Jost1, Mónica Ágreda-López2, Silvio Mollo3, Teresa Ubide Garralda4, Maurizio Petrelli2

  • Affiliations:  1Department of Earth Sciences, University of Geneva, Switzerland ; 2Department of Physics and Geology, University of Perugia, Italy; 3Department of Earth Sciences, Sapienza -- University of Rome, Italy; 4School of the Environment, The University of Queensland, St Lucia, QLD 4072, Australia 

  • Presentation type: Poster

  • Presentation time: Monday 16:30 - 18:30, Room Poster Hall

  • Poster Board Number: 146

  • Programme No: 3.1.38

  • Theme 3 > Session 1

Abstract

Over the past few decades, numerous thermodynamic-based and empirical models have been developed to describe mineral compositional changes and mineral-melt exchange reactions. More recently, machine-learning (ML) models have emerged, using only minerals (or melts) and mineral-melt pairs. However, current ML-thermobarometers do not incorporate thermodynamic constraints, relying solely on comparisons between the chemistry of the target phase and experimentally derived compositions. Importantly, these two approaches are not mutually exclusive and, ideally, should be integrated for more robust results. We present the results of preliminary tests designed to evaluate the efficacy of ML-thermobarometers from a thermodynamic perspective. These models were used to compute the melt composition (Xmelt) in equilibrium with natural clinopyroxene crystals at specific pressures (P) and temperatures (T). Initially, the recovered P-T-Xmelt data were tested for equilibrium with clinopyroxene by applying a series of thermodynamic equilibrium approaches from the literature. Subsequently, P-T estimates derived from thermodynamic-based and ML-thermobarometers were comparatively evaluated. Xmelt reconstructed by ML methods is effectively found to be in thermodynamic equilibrium with the natural crystals. However, while P-T estimates from thermodynamic-based thermobarometers are positively correlated according to the thermodynamic properties of melt and crystals, this behavior is less evident using ML- thermobarometers.