RocMLMs: Predicting Rock Properties through Machine Learning Models

Predicting the physical properties of rock is critical for modeling mantle convection. Density changes are the main driving force for mantle convection, for example, while elasticity/seismic properties allow probing of mantle dynamics and structure. Pressure-temperature (PT) conditions predicted by numerical thermo-mechanical models are often mapped to pre-computed pseudosections calculated by Gibbs Free Energy Minimization (GFEM) programs (e.g., Perple_X; Connolly, 2009) to predict phase changes and the associated evolution of rock properties (e.g., density, seismic wave velocities, and fluid contents). In principle, GFEM programs could be coupled to numerical geodynamic simulations (at each point in space and for each timestep) to establish models where phase assemblages and rock properties evolve self-consistently. In practice, this is currently intractable because GFEM programs remain too slow (10²–10⁴ ms per node) to be coupled to high-resolution numerical geodynamic models. While parallelization of GFEM calculations can increase efficiency dramatically (e.g., MAGEMin; Riel et al., 2022), predicting rock properties recursively during a geodynamic simulation requires GFEM efficiency on the order of ≤ 10⁰–10^-1 ms to be feasible. As an alternative to the GFEM approach, this study demonstrates the efficiency of predicting target rock properties through pre-trained machine learning models (referred to as RocMLMs). In our initial test case, RocMLMs are trained to predict density, seismic wave velocity, and melt fraction for dry upper mantle rocks based on an array of 128 Perple_X pseudosections (with 128x128 PT resolution) derived from 3111 harzburgite and lherzolite samples retrieved from the Earthchem.org repository. The training dataset size was reduced from 128¹³ to 128³ PTX examples by transforming the 11 oxide components of X (bulk rock composition) into a single Fertility Index value representing the estimated degree of melt extraction from a primitive mantle source. We show that Decision Tree, K-Neighbors, and single-layer Neural Network algorithms can predict rock properties up to 10³ times faster than commonly-used GFEM programs (with some performance trade-offs), attaining the required efficiency of 10⁰–10^-1 ms. Our results imply that implementing dynamic evolution of rock properties in geodynamic simulations is now possible with RocMLMs. Future generations of RocMLMs will include hydrated systems and a larger array of mantle compositions (e.g., dunites and pyroxenites).