Understanding the composition and physical properties of Earth’s deep interior is fundamental to deciphering planetary evolution, yet modeling these systems under extreme pressure–temperature (P–T) conditions remains a significant challenge. While first-principles simulations based on density functional theory (DFT) offer high accuracy, their prohibitive computational costs typically restrict studies to small scales and short timescales. Recently, machine learning potentials (MLPs) have emerged as a transformative bridge, providing near-first-principles fidelity at a fraction of the cost and enabling large-scale simulations of complex materials under deep-Earth conditions. This review synthesizes recent progress in applying MLPs to high P–T mineral physics, focusing on breakthroughs in modeling thermodynamic, transport, and elastic properties. We discuss methodological hurdles in extending MLPs across broad P–T ranges and evaluate emerging strategies—such as active learning, uncertainty quantification and temperature-dependent training—to overcome them. By uncovering phenomena like superionic conduction and defect-controlled behavior, MLPs are shifting our perspective of the deep Earth from idealized models toward complex, dynamic systems, paving the way for a predictive, data-driven era in mineral physics.
Li et al. (Mon,) studied this question.