20 September 2021 in Physical Review B
We develop a neuroevolution-potential (NEP) framework for generating neural network-based machine-learning potentials. They are trained using an evolutionary strategy for performing large-scale molecular dynamics (MD) simulations. A descriptor of the atomic environment is constructed based on Chebyshev and Legendre polynomials. The method is implemented in graphic processing units within the open-source gpumd package, which can attain a computational speed over atom-step per second using one Nvidia Tesla V100. Furthermore, per-atom heat current is available in NEP, which paves the way for efficient and accurate MD simulations of heat transport in materials with strong phonon anharmonicity or spatial disorder, which usually cannot be accurately treated either with traditional empirical potentials or with perturbative methods.
26 March 2018 in The Journal of Physical Chemistry Letters
21 June 2021 in Physical Review B
We study the role of nonperturbative phonon scattering in strongly anharmonic materials having ultralow lattice thermal conductivity with unusual temperature dependence. We take as an example and investigate its lattice dynamics using perturbation theory (PT) up to the fourth order and molecular dynamics (MD) with a machine-learning potential. We find distinct differences of phonon linewidth between PT and MD in the whole Brillouin zone. The comparison between the theoretical phonon linewidths and experiments suggests that PT severely underestimates the phonon scatterings, even when the fourth-order anharmonicity is included. Moreover, we extend our calculations to higher temperatures and evaluate the two-channel thermal conductivity based on the unified theory developed by Simoncelli et al. [ Nat. Phys. 15, 809 (2019)]. We find a crucial coherence contribution to the total thermal conductivity at high temperatures. Our results pave the path for future studies of phonon properties and lattice thermal conductivities of strongly anharmonic crystals beyond the conventional PT realm.
3 October 2017 in Chemistry of Materials
16 June 2020 in Nature Communications
The investigation of salts in water at extreme conditions is crucial to understanding the properties of aqueous fluids in the Earth. We report first principles (FP) and classical molecular dynamics simulations of NaCl in the dilute limit, at temperatures and pressures relevant to the Earth’s upper mantle. Similar to ambient conditions, we observe two metastable states of the salt: the contact (CIP) and the solvent-shared ion-pair (SIP), which are entropically and enthalpically favored, respectively. We find that the free energy barrier between the CIP and SIP minima increases at extreme conditions, and that the stability of the CIP is enhanced in FP simulations, consistent with the decrease of the dielectric constant of water. The minimum free energy path between the CIP and SIP becomes smoother at high pressure, and the relative stability of the two configurations is affected by water self-dissociation, which can only be described properly by FP simulations.
9 September 2019 in Journal of Applied Physics
Due to the compatibility with the well-developed Si-based semiconductor technology, the properties of silicene and silicene-based materials have attracted tremendous attention. Among them, the thermal conductivity (TC) is of special importance for electronic devices. However, unlike graphene, the poor quality of empirical potentials hinders the reliable evaluation of TC for silicene using molecular dynamics (MD). Here, we present a Gaussian approximation potential (GAP) for silicene based on ab initio derived training data. The potential can precisely describe the geometries, mechanical properties, as well as phonon dispersion of free-standing sheet, outperforming any other empirical ones. Using sinusoidal approach-to-equilibrium MD simulations based on the GAP potential, the TC of silicene is found to be at room temperature. Importantly, our result achieves a good agreement with Boltzmann transport equation (BTE) based first-principles predictions ( ), such that the TC value of silicene is confirmed via both MD and BTE; thus, we prove that the accuracy of machine learning potentials, like GAP, can enable a faithful prediction of TC at a density functional theory (DFT) level.