Classifying detrital zircon U-Pb age distributions using automated machine learning

The prodigious use of detrital zircon U-Pb geochronology for provenance studies in recent decades has led many researchers to amass extensive datasets (>100,000 dates). When displayed as age distributions, individual samples are traditionally compared using visual inspection and statistical methods, which can become time-consuming and challenging when using large datasets. We propose that machine learning (ML) can more efficiently classify a sample by its source using detrital zircon U-Pb age distributions. Specifically, we hypothesize that automated machine learning (AutoML), which optimizes algorithm selection and hyperparameters, will outperform an unoptimized Random Forest (RF) classifier and the cross-correlation coefficient (R²), a commonly used metric for comparing age distributions. We test this approach using a well-constrained synthetic dataset and a natural dataset from the Jurassic-Eocene North American Cordillera. In synthetic experiments, AutoML models effectively classify samples by their sources when inter-source similarity across few sources is low to moderate and samples have more than ∼50 analyses. However, the effectiveness of AutoML is highly dependent on sample size and the variability of age modes within the data. Applied to the North American Cordillera dataset, AutoML achieves an ∼0.91 F₁ score when predicting between foreland and forearc basin tectonic settings and an ∼0.71 F₁ score when predicting subbasins within these settings, outperforming both RF and R². Moreover, AutoML identifies discriminating age populations between groups, with the average feature importance of 100 models highlighting the 145–125 Ma age range, corresponding to a magmatic lull of the Cordilleran magmatic arc. These results demonstrate AutoML's potential as a powerful predictive and interpretive tool in detrital zircon studies.