Deep machine-learning potential for atomistic simulation of δ-AlOOH at high pressures and temperatures

δ-AlOOH (δ) is a critical high-pressure hydrous phase for understanding the Earth’s geological water cycle. Our earlier ab initio study [1] offered a practical multi-configuration model for understanding the pressure-dependent behavior of the H-bond, including H-bond disorder, tunneling, and symmetrization. Since H-bond behavior in δ is likely typical of H-bonds in other hydrous or nominally anhydrous phases, it would be helpful to develop a more complete understanding of its behavior beyond the symmetrization transition pressures. However, further simulations on δ require large-scale simulations that are prohibitive to ab initio calculations.

This study adopts the DP-GEN scheme [2] to develop a deep machine-learning potential (DP) for δ. Our DP potential predicts forces and energies accurately, and DP-based MD simulations reproduce detailed features found in ab initio pair-correlation functions and finite-T equation of states. The simulation with DP potential predicts the P-T region of the superionic state of AlOOH. It helps interpret the recently reported [3] instability of δ near the cold slab geotherm.

[1] C. Luo, K. Umemoto, and R. Wentzcovitch, Phys. Rev. Research 4, 023223 (2022).

[2] Y. Zhang et al., Computer Physics Communications 253, 107206 (2020).

[3] Y. Duan, et al., Earth and Planetary Science Letters 494, 92 (2018).