A Machine Learning Approach to Single Garnet Geothermometry and Application to Tracing the Fingerprint of Superdeep Diamonds

Abstract

Estimating the equilibration temperatures of mantle-derived garnets is crucial for assessing the diamond potential of kimberlites. Traditional garnet geothermometers require co-existing mineral data or costly trace element analysis, limiting their practical use. As an alternative approach, based on the major and minor element composition of garnet alone, we first re-calibrated an Mn-in-garnet thermometer using a newly compiled data set of garnets from well-equilibrated peridotitic xenoliths with well-constrained pressure-temperature (P-T) conditions. The re-calibrated Mn-in-garnet thermometer, however, is only of intermediate accuracy, with a relatively large discrepancy relative to the most reliable multi-phase thermometry, indicated by a high root mean square error value (RMSE = 79°C) across a temperature range from 900 to 1,400°C. In a second improve approach, we developed a new machine learning (ML)-based garnet thermometer that demonstrated superior performance, achieving significantly better accuracy and reduced discrepancies (average RMSE = 61°C). The ML-based garnet thermometer outperforms the Mn-in-garnet thermometer because it considers not only MnO but also other major and minor elements, particularly TiO₂, revealed by the ML model to be critical for accurate prediction of garnet temperatures. Applying the ML-based thermometer to garnet xenocrysts from kimberlites on the Slave and Kaapvaal cratons reveals that high numbers of sublithospheric (superdeep) diamonds are associated with significantly higher proportions of high-T (>1,200°C) high-Ti garnets, compared to kimberlites in which superdeep diamonds are either few or absent. This finding indicates that a number of kimberlites, not currently identified as containing superdeep diamond populations, are promising hosts of such diamonds.