A Machine‐Learning Surrogate for Differentiable Phase Equilibrium Modeling of Mantle Rocks on a Planetary Scale

Modeling processes within Earth's and other planetary bodies' mantle requires constraints from phase equilibrium to infer how changes in phase abundance and composition affect physical properties and geochemical processes. When applying such thermodynamic models on a planetary scale, performance becomes especially crucial. We present a fast and scalable machine‐learning surrogate that can approximate phase equilibrium predictions for mantle rocks. The model was calibrated on a large synthetic data set ( n = 2.0⋅10 6 ) generated using a Gibbs free energy minimizer. This data set covers upper‐ to transition‐zone mantle conditions in terms of pressure, temperature, and typical bulk rock composition. Evaluation of the model indicates uncertainties of less than ±0.01 molmol −1 for the prediction of phase fractions and less than ±0.005 molmol −1 for compositional variables within solid solutions. Gradient‐based analysis identifies phase transitions that are sensitive to bulk chemistry changes. Peridotite fertility exerts a secondary, yet locally decisive, control on the seismic discontinuities and buoyancy contrasts that influence mantle convection. The assessment of computational performance indicates a speed up of four orders of magnitude when emulating phase equilibrium predictions with the GPU accelerated surrogate.