Predictive Models for Detrital Titanite Provenance With Application to the Nanga Parbat—Haramosh Syntaxial Massif, Western Himalaya

Titanite is a versatile recorder of crystallization age, temperature, and host lithology, via the U–Pb system, the Zr‐in‐Ttn thermometer, and elemental composition, respectively. The paragenesis of titanite renders it especially useful for tracing detritus derived from lithologies that are infertile for the commonly used detrital zircon U‐Pb chronometer, such as sub‐anatectic metamorphism of calc‐silicates. Despite these advantages, detrital titanite analysis is underemployed, in part because the U–Pb system in titanite is often complicated by the incorporation of both inherited radiogenic Pb from precursor minerals during metamorphic reactions, and also bulk crustal common‐Pb. Recent systematic analyses of large titanite compositional data sets from diverse source rocks have revealed that the elemental composition of titanite is provenance‐specific. Here, we apply a workflow that incorporates a machine‐learning classifier to a large and representative compositional database for titanite, encompassing >11,000 analyses, with c. 6,700 points passed to our model. Only medians of the subcompositions for 205 rocks are used for our model. We reliably discriminate (>90%) between metamorphic and igneous titanite. Application of this classifier to a detrital case study from the Nanga Parbat‐Haramosh syntaxial massif of the western Himalaya reveals that titanite of different compositions formed during different orogenic events. Furthermore, titanite with significant common Pb solely derives from medium/low grade metasedimentary rocks. The method described here offers a pathway to increase the specificity of the provenance information derived from titanite; however, the published corpus of titanite data will have to be much larger before multi‐class source‐rock discrimination can be achieved reliably. The mineral titanite contains information about its age, the type of rock in which it formed, and the temperature at which it grew. By analyzing titanite in detritus, we can learn about the source of that detritus. Despite its potential, the use of titanite is still in its infancy. Much work remains to characterize titanite from all the potential rock types that erode to form sand in order to create predictive models. Here, we build on earlier research that aims to characterize titanite from different rocks. We assemble a large titanite database, which includes all common titanite‐bearing rock types and then analyze it to produce a model that can predict which rock type an individual titanite grain comes from. To do this, we use an approach that is becoming more common in similar studies: machine learning, specifically a method called random forest. Using our approach, we are able to discriminate titanite from igneous and metamorphic common rocks with high accuracy. This work thus increases the usefulness of titanite to understand changes on and below the Earth's surface in deep time. But much more data will be required for schemes that can determine the rock type of titanite in finer detail. Titanite compositions predict sediment‐source rock type using a random forest algorithmIgneous and metamorphic titanite are distinguished with >90% accuracyMore data will be required for accurate multi‐class predictive schemes for detrital titanite Titanite compositions predict sediment‐source rock type using a random forest algorithm Igneous and metamorphic titanite are distinguished with >90% accuracy More data will be required for accurate multi‐class predictive schemes for detrital titanite