Advances in the Application of Machine Learning to the Calculation of Mineral States and Properties in the Earth's Deep Interior

The deep interior of the Earth is under extreme high-temperature and high-pressure conditions. Its material composition, phase transition behavior, and physical properties are crucial for understanding the Earth's internal structure, dynamic processes, and evolution. Traditional experimental methods face challenges in maintaining thermodynamic states and diagnosing physical quantities under such extreme conditions. While first-principles calculations offer quantum-level precision, their computational efficiency limits their direct application to simulating deep-Earth minerals across large spatiotemporal scales. Machine learning methods present new opportunities. By constructing high-precision, efficient machine learning potential functions based on first-principles datasets, machine learning methods significantly extend the spatiotemporal scale of first-principles simulations, which provide revolutionary tools for studying the physical states, phase transitions, elasticity, and transport properties of deep-Earth minerals. This paper systematically reviews the progress of machine learning applications in studying major deep-Earth minerals, including those in the upper mantle, transition zone, lower mantle, subduction zone components, and core materials, and summarizes the representative achievements of machine learning methods in revealing phase transitions, thermal conductivity, diffusion, melting, and elastic properties, while also discussing current limitations and future research directions.