Applying Statistical and Machine Learning Methods to the Zircon Trace Elements Dataset: Implications for Detrital Provenance Studies

Abstract:

Zircons are common in terrestrial melts (as an accessory phase), robust against weathering, and have several useful geochemical properties. Here we investigate in-context zircon trace element (TE) chemistry (i.e. rare earth elements [REEs], Th, U, U/Yb, Eu anomalies) from the GeoRoc database and other published contributions in search of patterns that link zircon to specific host rock types (e.g., granite, rhyolite, granodiorite, gabbro, tonalite). We are also in the process of expanding the dataset we use for machine learning by gathering other published zircon datasets. We applied statistical, data science, and deep learning methods to study the zircon TE datasets and built a machine learning model to predict the zircon parent rock.

To start, we generated probability density functions for many TE and other variables, concentration histograms, and plots of rare earth element (REEs) vs. other REEs which were inspected for patterns. Certain zircon TE patterns became apparent when plotted using pooled data (n = ~500-~4000 depending on trace element and rock type). One major observation is that the TE concentration distributions for zircons from granitic or rhyolitic parent rocks are different with more zircons from granites generally having higher REE concentrations than zircons from rhyolites. For any given REE, between 5% to 24% of the granitic zircon dataset is greater than the concentration at the 90th quantile for the rhyolitic dataset.

Our machine learning algorithms were developed using the TensorFlow and Keras packages to build a neural network that could analyze TE and TE ratios and provide a prediction for the original host rock. After training on approximately 10³ zircons for zircons from each rock type (granites, rhyolites, and granodiorites), we applied the network to a test dataset and analyzed the prediction accuracy. Then, weights were retrieved from the network and studied to see what inputs the network had identified as important in predicting host rock type.

The observations developed for zircons of known provenance prime us for further study of detrital zircons, where the original host rock and magmatic formation environment is no longer clearly known. We applied the trained network to samples of interest (e.g. detrital Hadean zircons) to see what preliminary predictions the model would make.