Molecular Dynamics Simulations of Molten Magnesium Chloride Using Machine‐Learning‐Based Deep Potential

Abstract

In previous work, molten magnesium chloride has been investigated using first-principles molecular dynamics (FPMD) simulations based on density functional theory (DFT). However, such simulations are computationally intensive and therefore are restricted in terms of simulated size and time. In this work, a machine learning-based deep potential (DP) is trained to accelerate the molecular dynamics simulation of molten magnesium chloride. The trained DP can accurately describe the energies and forces with the prediction errors in energy and force being 1.76 × 10 −3 eV/atom and 4.76 × 10 −2 eV Å −1 , respectively. Applying the deep potential molecular dynamics (DPMD) approach, simulations can be performed with more than 1000 atoms, which is infeasible for FPMD simulations. Additionally, the partial radial distribution functions, angle distribution functions, densities, and self-diffusion coefficients predicted by DPMD simulations are also in reasonable agreement with FPMD or experimental results. This work shows that the DP enables higher efficiency and similar accuracy relative to DFT, exhibiting a bright application prospect in modeling molten salt systems.

Unraveling Local Structure of Molten Salts via X-ray Scattering, Raman Spectroscopy, and Ab Initio Molecular Dynamics

“…This is consistent with the value of n Cl = 4.9 obtained by integrating the radial distribution function g ( r ) for Mg–Cl up to the first minimum in g ( r ) using PBE-D3 (Figure S8). A smaller value of n Cl = 4.6 was recently reported for molten MgCl 2 with PBE-D2 using a much smaller simulation box, shorter equilibration time, and more importantly a large-core 1s 2 2s 2 2p 6 pseudopotential for Mg employed in many very recent publications. − This leaves no valence electrons on Mg 2+ , leading to a poor description of the interaction between the essentially nonpolarizable Mg 2+ cation and the chloride anion. As a result, such treatment gives rise not only to the underpredicted CN, but also to the underestimated Mg–Cl bond distance (2.35 Å) compared to the experimental (2.42 Å) or computed value in this work (2.41 Å) using a small-core 1s 2 pseudopotential for Mg (see below.)…”

“…Reported in their work the CN = 4.6 along with the Mg–Cl bond distance of 2.35 Å are much smaller than the values suggested in our work. The main source of inconsistency is likely due to a poor description of a standard Mg pseudopotential in these studies, − which included 1s 2 , 2s 2 2p 6 electrons in the core and only 3s 2 electrons in valence. With no valence electrons left for divalent Mg 2+ , this ion is essentially treated as a nonpolarizable ion, leading to the Mg–Cl distance being too short and the CN being too low.…”

“…There has been a burst of recent interest − from different groups in studying molten MgCl 2 and various molten salts mixtures with alkali chlorides using ab initio and machine learning potentials trained based on the AIMD data. Reported in their work the CN = 4.6 along with the Mg–Cl bond distance of 2.35 Å are much smaller than the values suggested in our work.…”

Deep potentials for materials science

“…Other DPs were developed to calculate transport properties of silicate in the mantle [144][145][146]. DPs were also employed in large-scale calculations of thermodynamic, transport, and structural properties in different molten salts [149][150][151][152][153][154][155][156][157].…”

Deep Potentials for Materials Science

“…Other DPs were developed to calculate transport properties of silicate in the mantle [151][152][153]. DPs were also employed in large-scale calculations of thermodynamic, transport, and structural properties in different molten salts [156][157][158][159][160][161][162][163][164].…”