IGyTheM: Integrating Geodynamics With Thermodynamics Using Machine Learning

Abstract

Achieving a comprehensive understanding of mantle convection requires integrated models that couple mantle dynamics with thermodynamic properties. Although advances in computational geodynamics have substantially improved our ability to simulate the underlying physical processes, incorporating thermodynamics into self‐consistent geodynamic models remains a major computational bottleneck as existing approaches are often cumbersome and prohibitively expensive for high‐resolution simulations. Here we present IGyTheM ( I ntegrating G eodynamics with The rmodynamics using M achine learning), a fast, flexible, and general framework for coupling geodynamics with phase equilibria and mineral properties. Within IGyTheM, we provide a machine‐learning–based thermodynamic surrogate trained on Gibbs energy minimization calculations. The surrogate predicts thermodynamic properties for arbitrary basalt–harzburgite mantle mixtures across mantle pressure–temperature conditions (0.0001–140 GPa, 273–4,273 K). Benchmark tests demonstrate that the surrogate achieves comparable accuracy while being more than two orders of magnitude faster than conventional Gibbs minimization calculations and lookup table approaches. We also provide a Python interface that allows users to supply inputs via an Excel table, enabling straightforward calling of the surrogate and convenient integration with existing petrological and geodynamic modeling tools. To demonstrate its application, we couple the surrogate with the community finite‐element mantle convection code Citcom, which includes thermochemical convection and temperature‐dependent rheology. In this configuration, thermodynamic properties are predicted rapidly during runtime with minimal computational overhead. This capability enables efficient large‐scale thermochemical mantle convection simulations and opens the door to integrated 2D and 3D geodynamic–geochemical modeling across broader parameter spaces and spatiotemporal scales than previously feasible.