5:15 PM - 7:15 PM
[SCG58-P03] Machine-learning potential for olivine-spinel transformation dynamics simulation
★Invited Papers
Keywords:Machine-learning potential, Molecular dynamics, olivine-spinel transformation
Deep earthquakes occur in the transition zone between the upper and lower mantles, at depths of approximately 410 to 660 km. Since the pressure and temperature conditions in this zone differ significantly from those at shallow or intermediate depths, where most earthquakes occur, the mechanisms or triggers of deep earthquakes are also thought to be distinct from those of shallow earthquakes. However, studying the mechanisms of deep earthquakes is more challenging than studying shallower ones due to the extreme pressure and temperature conditions.
The transformation of super-pressurized olivine has been proposed as a potential trigger for deep earthquakes. However, it remains unclear which structural changes result from this transformation and whether it can induce earthquakes on a much larger length scale. To investigate the transformation from olivine to spinel under high-pressure and high-temperature conditions, we aim to develop a machine-learning interatomic potential (MLIP) for the Mg-Si-O ternary system using state-of-the-art MLIP techniques.
We employ a fast and interpretable MLIP constructed using B-spline functions for two-body and three-body interactions. This potential has been shown to accurately reproduce the energies and forces of not only olivine and spinel structures but also those of the intermediate structure suggested in a previous ab initio study. However, our results indicate that the stresses predicted by the MLIP, which was trained using only energies and forces, do not align well with ab initio calculations. This suggests that stress matching is crucial for applying MLIP to materials under high-pressure conditions, such as those found in the mantle transition zone.
In this study, we will discuss the accuracy and stability of the MLIP as well as the molecular dynamics (MD) simulation results of the olivine-to-spinel transition.
The transformation of super-pressurized olivine has been proposed as a potential trigger for deep earthquakes. However, it remains unclear which structural changes result from this transformation and whether it can induce earthquakes on a much larger length scale. To investigate the transformation from olivine to spinel under high-pressure and high-temperature conditions, we aim to develop a machine-learning interatomic potential (MLIP) for the Mg-Si-O ternary system using state-of-the-art MLIP techniques.
We employ a fast and interpretable MLIP constructed using B-spline functions for two-body and three-body interactions. This potential has been shown to accurately reproduce the energies and forces of not only olivine and spinel structures but also those of the intermediate structure suggested in a previous ab initio study. However, our results indicate that the stresses predicted by the MLIP, which was trained using only energies and forces, do not align well with ab initio calculations. This suggests that stress matching is crucial for applying MLIP to materials under high-pressure conditions, such as those found in the mantle transition zone.
In this study, we will discuss the accuracy and stability of the MLIP as well as the molecular dynamics (MD) simulation results of the olivine-to-spinel transition.