Apatite (U-Th-Sm)/He date dispersion: First insights from machine learning algorithms

Abstract

Numerous parameters impact apatite (U-Th-Sm)/He (AHe) thermochronological dates, such as radiation damage, chemical content, crystal size and geometry, and their knowledge is essential for better geological interpretations. The present study investigates a new method based on advanced data mining techniques, to unravel the parameters that could play a role in He retention and thus on AHe date. The purpose is to decipher which factors influence the AHe date dispersion, and to exclude the impact of other parameters on helium retention. As an example, we use a dataset previously collected on apatite from basements rocks, sampled in French Brittany, where all samples underwent the same thermal history, and for which were reported a set of physical and chemical parameters. The dataset includes dimension and geometry, He, U, Th, Sm and major and trace element content for ∼35 crystals. The algorithm ranks the parameters according to their influence on helium retention, using predictive trees, which are commonly used in computing sciences. After looking at 100 simultaneous predictions, we compared the predicted and measured He content for each analyzed apatite crystal. For this particular case, the predictions confirmed the prominent role of the parent nuclides in the He production, as AHe dates can be predicted accurately with these parameters (especially U and Th). Additionally, the predictions without knowledge of the apatite chemical composition and dimension provided better results than using all available parameters (median error of 14% instead of 18%). Therefore, for this specific study, the apatite chemistry and crystal dimensions do not influence significantly He retention nor AHe date dispersion. Nevertheless, detailed inspection of analysis results suggests which parameters have the most discrimination ability, which in this study include crystal length, height, and Mn content. The latter may reveal an eventual influence on alpha damage annealing kinetics. Finally, this approach shows that some grains could never achieve good predictions, indicating that for these crystals the input parameters are not enough to predict the He content. We propose that such crystals are statistically different from the remaining dataset, and this suggests that machine learning has a strong potential to correct errors, or to detect anomalies.