17:15 〜 18:45
[SIT16-P06] Deep Learning Interatomic Potentials Developed for Understanding Deep Planetary Interiors
キーワード：惑星内部、氷惑星・氷衛星、第一原理計算、機械学習、高圧物性
Understanding the composition and thermodynamic properties of the internal structures of icy giants and icy moons is an essential part of elucidating the origin and evolution of planetary systems. The icy materials of the interiors can exist in warm dense matter conditions under high pressures and temperatures. The major molecules in the reducing nebula are H2O, CH4, NH3, and H2S [1], and these would be supplied and condensed in the outer regions of the protoplanetary disks. H2S has been detected in the atmospheres of Neptune and Uranus [2], and sulfide compounds are important components of icy bodies considering the possible variation of planetary systems. Hydrogen sulfide has unique properties, such as superconductivity, even though sulfur belongs to the same family as oxygen in the periodic table. Unlike water, the phase diagram of hydrogen sulfide has not been well studied. Experiments, e.g.) with a gas gun, have difficulty covering a wide range of temperatures and pressures. On the other hand, as for a theoretical approach to estimate unknown thermodynamical conditions under high pressure expected in interiors of icy bodies, ab initio molecular dynamics (AIMD) simulation is a powerful tool but highly computationally expensive.
In this study, we have developed a machine-learning scheme combined with AIMD simulations to obtain accurate thermodynamic properties in the interior of icy bodies, allowing MD simulations for larger systems and for longer times. First, we performed density functional theory molecular dynamics using VASP and Quantum Espresso for hydrogen sulfide (≈100 atoms per unit cell) for more than 10 ps with the NVT ensemble in the temperature range of 500 − 10,000 K and the density range of 1 − 5 g/cm3. Moreover, from the AIMD results (atomic coordinate, energy, and forces for each trajectory), the interatomic potential was generated using a neural network approach (NequIP) [3], and then we performed larger classical molecular dynamics calculations with the deep-learned potential using the LAMMPS program. For example, the radial distribution function for each condition was reasonably reproduced. We demonstrate the effectiveness of our scheme to elucidate planetary interiors.
References
[1] Johnson, Torrence V. et al.: 2012, The Astrophysical Journal, 757.2, 192. [2] Irwin, P. G. J. et al.: 2018, Nature Astronomy 2, 420. [3] Batzner, Simon et al.: 2022, Nature communications, 13.1, 2453.
In this study, we have developed a machine-learning scheme combined with AIMD simulations to obtain accurate thermodynamic properties in the interior of icy bodies, allowing MD simulations for larger systems and for longer times. First, we performed density functional theory molecular dynamics using VASP and Quantum Espresso for hydrogen sulfide (≈100 atoms per unit cell) for more than 10 ps with the NVT ensemble in the temperature range of 500 − 10,000 K and the density range of 1 − 5 g/cm3. Moreover, from the AIMD results (atomic coordinate, energy, and forces for each trajectory), the interatomic potential was generated using a neural network approach (NequIP) [3], and then we performed larger classical molecular dynamics calculations with the deep-learned potential using the LAMMPS program. For example, the radial distribution function for each condition was reasonably reproduced. We demonstrate the effectiveness of our scheme to elucidate planetary interiors.
References
[1] Johnson, Torrence V. et al.: 2012, The Astrophysical Journal, 757.2, 192. [2] Irwin, P. G. J. et al.: 2018, Nature Astronomy 2, 420. [3] Batzner, Simon et al.: 2022, Nature communications, 13.1, 2453.