Machine Learning-based AFT Annealing Parameter r mr0 from c-axis- projected Reduced Mean Length of Partially Annealed 252 Cf-derived FTs and LA-ICP-MS-derived Chemical Composition Data
AbstractAFT annealing parameter r mr0 relative to apatite standard B2 (Carlson et al., 1999; Ketcham et al., 1999) is re-calibrated here using a combination of natural apatite mixtures from sandstones and selected standards. The standards include well-studied DR, FC, TI, and RN-like (Ca-F-apatite end-member). A machine learning approach is used that predicts r mr0 based on independent chemical composition data from LA-ICP-MS. The c-axis-projected (e.g., Donelick et al., 1999), reduced mean length of partially annealed 252 Cf-derived FTs is measured and converted to r mr0 for each apatite grain studied. Absolute concentrations of Na, Mg, P, S, Cl, Ca, Mn, Fe, As, Sr, Y, 14 REEs, Th, U and relative concentrations of Al, Si, Sc, Br are determined for each grain by LA-ICP-MS using DR and other apatite species as matrix-matched standards (Donelick and Donelick, 2013). Of primary interest to r mr0 is: absolute Cl, Mn, Fe, Sr, ΣREEs, and annealing state of natural FTs (pre-annealed or not; t-T path dependence?); of secondary interest is relative Br, absolute Y, Th, U, and Pb-corrected UPb age (t-T path dependence?); of tertiary interest is everything else including the concentrations of individual REEs. Machine learning regression techniques applied to the data include linear regression, random forests, dense neural networks, and support vector models. We compared the validation results between these regression techniques and examine the importance of each chemical composition feature as determined by the model training. The best performing model identified so far is a random forest regressor, with a 5-fold cross validation mean absolute error of 0.057 +/-0.004 (1σ) on r mr0 .