RocMLMs: Predicting Rock Properties Through Machine Learning Models

• Published in: Journal of geophysical research. Machine learning and computation Vol. 1; no. 4
• Main Authors: Kerswell, Buchanan, Cerpa, Nestor G., Tommasi, Andréa, Godard, Marguerite, Padrón‐Navarta, José Alberto
• Format: Journal Article
• Language: English
• Published: American Geophysical Union/Wiley 01.12.2024; Wiley
• Subjects: Gibbs free energy minimization; machine learning; mantle transition zone; numerical geodynamic models; olivine phase transitions; Sciences of the Universe
• ISSN: 2993-5210; 2993-5210
• DOI: 10.1029/2024JH000264

Mineral phase transformations significantly alter the bulk density and elastic properties of mantle rocks and consequently have profound effects on mantle dynamics and seismic wave propagation. These changes in the physical properties of mantle rocks result from evolution in the equilibrium mineralogical composition, which can be predicted by the minimization of the Gibbs Free Energy with respect to pressure (P), temperature (T), and chemical composition (X). Thus, numerical models that simulate mantle convection and/or probe the elastic structure of the Earth's mantle must account for varying mineralogical compositions to be self‐consistent. Yet coupling Gibbs Free Energy minimization (GFEM) approaches with numerical geodynamic models is currently intractable for high‐resolution simulations because prediction speeds of widely‐used GFEM programs (100–102 ms) are impractical in many cases. As an alternative, this study introduces machine learning models (RocMLMs) that have been trained to predict thermodynamically self‐consistent rock properties at arbitrary PTX conditions between 1–28 GPa and 773–2,273 K, and dry mantle compositions ranging from fertile (lherzolitic) to refractory (harzburgitic) end‐members define9d with a large data set of published mantle compositions. RocMLMs are 101–103 times faster than GFEM calculations or GFEM‐based look‐up table approaches with equivalent accuracy. Depth profiles of RocMLMs predictions are nearly indistinguishable from reference models PREM and STW105, demonstrating good agreement between thermodynamic‐based predictions of density, Vp, and Vs and geophysical observations. RocMLMs are therefore capable, for the first time, of emulating dynamic evolution of density, Vp, and Vs due to partial melting and refertilization of dry mantle rocks in high‐resolution numerical geodynamic models. Plain Language Summary The mineralogical makeup of rocks within Earth's mantle largely determines how the mantle flows over geologic time, and how it responds to seismic waves triggered by earthquakes, because mineral assemblages control important rock properties such as density and stiffness (elasticity). The mineralogy of mantle rocks is not constant, however. It changes depending on three factors: pressure, temperature, and the chemical composition of the rock. Thus, it is important for computer simulations of mantle convection to account for the evolution of rock mineralogy. Computer programs that can predict rock properties based on thermodynamic calculations are available, but are generally too slow to be used in high‐resolution simulations. As an alternative approach, this study introduces machine learning models (RocMLMs) that have “learned” how to predict rock properties (density and elasticity) by “training” on a large data set of thermodynamic calculations. We demonstrate that RocMLMs can then predict rock properties up to 101–103 times faster than state‐of‐the‐art methods. We tested RocMLM predictions against reference mantle models based on observations of seismic waves and found good agreement. RocMLMs are therefore capable of fast and highly‐accurate predictions of changes in rock properties and can be implemented in high‐resolution computer simulations of mantle convection. Key Points RocMLMs predict rock properties up to 101–103 faster than commonly used methods RocMLMs trained with Neural Networks are more efficient compared to other regression algorithms RocMLM training data show good agreement with PREM and STW105 for an average mantle geotherm