Tracking garnet dissolution kinetics in 3D using deep learning grain shape classification

Abstract

The kinetics of fluid-driven metamorphic reactions are challenging to study in nature because of the tendency of metamorphic systems to converge towards chemical equilibrium. However, in cases where mineral textures that reflect incomplete reactions are preserved, kinetic processes may be investigated. Atoll garnet, a texture formed by the dissolution of a garnet’s core, has been described in 2D from thin sections of rocks worldwide. Quantifying the extent of this dissolution reaction requires sample-wide examination of hundreds of individual grains in 3D. In this study, we quantified the distribution of atoll garnet using micro-computed tomography and grain shape analysis. A convolutional neural network was trained on human-labelled garnet grains for automated garnet classification. This approach was applied to a retrogressed mafic eclogite from the Zermatt-Saas Zone (Western Alps). Pervasive atoll-like resorption preferentially affected the larger porphyroblasts, suggesting that compositional zoning patterns exert a first-order control on dissolution rates. A kinetic model shows that the reactivity of metastable garnet to form atolls is favored at pressure-temperature conditions of 560±30 °C and 1.6±0.2 GPa. These conditions coincide with the release of water when lawsonite breaks down during exhumation of mafic eclogites. The model predicts dissolution rates that are 3–5 times faster for the garnet core than for the rim. This study shows that deep learning algorithms can perform automated textural analysis of crystal shapes in 3D and that these datasets have the potential to elucidate petrological processes, such as the kinetics of fluid-driven metamorphic reactions.