This work investigates the feasibility of using pre-trained machine learning models to predict density (and other rock properties) during numerical geodynamic simulations of mangle convection. We built a large dataset of rock properties with the Gibbs Free Energy minimization program Perple_X (Connolly, 2009). Different regression algorithms were used to train Machine Learning models (RocMLMs), which we then evaluated in terms of accuracy and performance. We found that RocMLMs are able to predict rock properties up to 10$^1$–10$^3$ times faster than conventional approaches with equivalent accuracy. The speed of RocMLM predictions allows dynamic rock properties to be implemented in high-resolution numerical simulations of mantle convection for the first time.

This work investigates methods for interpolating surface heat flow data near subduction zones. We applied two different interpolations methods, Similarity and Kriging to the ThermoGlobe dataset (Jennings et al., 2021), and compared the results. Interpolations show that surface heat flow is complex and often discontinuous along strike near subduction zones, thus implying that the deep thermal structure and/or near-surface modifications are also discontinuous.

This work investigates where, and how many, rocks get detached from subducting oceanic plates beneath convergent margins. Over one-million numerical markers (representing rock fragments) from 64 numerical experiments were traced and classified as recovered, or not recovered, using a bespoke classification algorithm we wrote. Our results indicate that rocks are recovered from discrete depths, rather than continuously along the subduction interface.

This work involved input from many individuals during a pre-conference workshop at the Goldschmidt meeting in Honolulu, HI, during 8-9 July 2022. Our goal was to provide the geochemical community with recommendations to improve the existing infrastructure for storing and sharing geochemical data. Together we conceived OneGeochemistry: a community-driven initiative to coordinate the building of a global, online network of machine-readable data that is persistent, interoperable and reusable.

Online and PDF versions of my dissertation.

The notions that mechanical coupling in subduction zones regulate important seismic, volcanic, and geodynamic phenomena, and that depths of mechanical coupling may be invariant among diverse subduction zone segments (Wada & Wang, 2009), begs the following questions: where, how, and why does mechanical coupling occur along the interface between converging tectonic plates? This study investigates these questions by constructing 64 numerical geodynamic models of oceanic-continental convergent margins. Such a comprehensive suite of numerical models allows us to correlate mechanical coupling depths with thermo-kinematic boundary conditons—ultimately producing an expression for predicting coupling depths in real systems.

A lab exercise focusing on exploring the nuances of retrieving and analyzing data from online open-source geochemical repositories. I use this handout in my petrology courses. To be submitted to the “The On the Cutting Edge: Exemplary Reviewed Activity Collection” maintained by the NAGT.

A website showcasing random assortments of thin sections from Matt Kohn’s library. Useful for petrology courses—especially for remote teaching and learning.

Details about rock-forming processes like melting, (re)crystallization, fluid-rock interactions, and deformation are naturally encoded as sub-micrometer-scale features within mineral grains. Resolving these fine textural and compositional details is key to quantifying important aspects of our dynamic planet including: magmagenesis & volcanic hazards risks, lithospheric deformation & seismic hazards risks, geo-bio-hyrdo-atmosphere interactions & environmental hazards risks, carbon cycling & climate change, and concentration of critical elements required for transitioning to a sustainable energy future. However, the current state-of-the art for acquiring high-quality micrometer-scale images needed for quantitative petrology is slow and expensive because image quality (spatial resolution) scales with acquisition time.

Making reliable on-site decisions with portable low-cost instrumentation is key for exploring critical minerals deposits on Earth (Cai et al., 2022) and compositions of other rocky planets in our solar system (Berlanga et al., 2022). Raman spectroscopy shows promising opportunities as a field instrument that can quickly gather rich mineralogical information at low-cost (Cai et al., 2022). For example, many studies have built various ML classifiers to automatically identify minerals from Raman spectra that can be gathered in less than 30 seconds under laboratory conditions and achieve 80–90%+ accuracy in common mineral identification tasks (Berlanga et al., 2022; Cai et al., 2022; Carey et al., 2015; Ishikawa & Gulick, 2013; Jahoda et al., 2021; Liu et al., 2017; Sang et al., 2022; Sevetlidis & Pavlidis, 2019). However, all of the current systems are trained on a database (RRUFF, Lafuente et al., 2015) that is severely imbalanced in terms of the number of spectra across all mineral classes (Figure 2). Any classifier trained in this manner will necessarily converge on the same performance accuracy over time for a small subset of rock types. The only exception to this approach is the remarkable dataset gathered by Berlanga et al. (2022)—correctly claiming that ML classifiers can only achieve flexibility and high-performance accuracy on real-world mineral identification tasks by training on much larger and more balanced datasets than RRUFF.

Tectonic plate motions carry megatons of carbon stored near Earth’s surface into Earth’s deep interior each year as altered oceanic lithosphere and seafloor sediments subduct at convergent margins (Figure 3). Estimating unobservable carbon fluxes between Earth’s surface and interior is critical for projecting current climate trends. In principle, carbon and helium isotope compositions of diamonds and volcanic gasses can proxy for carbon recycling efficiency—i.e. how much subducted carbon is returned to Earth’s surface—yet interpretations of isotopic data remain unclear and highly-uncertain (Plank & Manning, 2019).