Development of Deep Potentials of Molten MgCl₂–NaCl and MgCl₂–KCl Salts Driven by Machine Learning

Abstract

Molten MgCl 2 -based chlorides have emerged as potential thermal storage and heat transfer materials due to high thermal stabilities and lower costs. In this work, deep potential molecular dynamics (DPMD) simulations by a method combination of the first principle, classical molecular dynamics, and machine learning are performed to systemically study the relationships of structures and thermophysical properties of molten MgCl 2 − NaCl (MN) and MgCl 2 −KCl (MK) eutectic salts at the temperature range of 800−1000 K. The densities, radial distribution functions, coordination numbers, potential mean forces, specific heat capacities, viscosities, and thermal conductivities of these two chlorides are successfully reproduced under extended temperatures by DPMD with a larger size (5.2 nm) and longer timescale (5 ns). It is concluded that the higher specific heat capacity of molten MK is originated from the strong potential mean force of Mg−Cl bonds, whereas the molten MN performs better in heat transfer due to the larger thermal conductivity and lower viscosity, attributed to the weak interaction between Mg and Cl ions. Innovatively, the plausibility and reliability of microscopic structures and macroscopic properties for molten MN and MK verify the extensibilities of these two deep potentials in temperatures, and these DPMD results also provide detailed technical parameters to the simulations of other formulated MN and MK salts.

Data Efficient and Stability Indicated Sampling for Developing Reactive Machine Learning Potential to Achieve Ultralong Simulation in Lithium-Metal Batteries

“…Then, the small time-length of MLPMD further causes inadequate sampling, thus making the model evolution too slow in the first few rounds (Figure c). The issue could be addressed by using data augmentation to increase the diversity of initial training data, e.g., perturbing DFT-relaxed structures, collecting DFT-labeled data from MD simulations with different temperatures, pressures, unit cells, thermodynamic ensembles, etc., see refs , and . However, it needs extra computational costs.…”

Development of a Neural Network Potential for Modeling Molten LiCl/KCl Salts: Bridging Efficiency and Accuracy

“…However, previous studies on molten salt NNPs have shown that the NNP’s performance becomes less sensitive to changes in the hyper-parameters once a certain threshold is reached. ,,− , These include but are not limited to the size of the hidden layers for the embedding and hidden networks, the prefactor values for the average atomic energy and forces, the r c and r cs values, and the maximum number of neighbors for each atom. For instance, in the investigation of NNPs for molten MgCl 2 –NaCl and MgCl 2 –KCl salts by Xu et al., it was observed that the RMSE values for average atomic energies and forces did not experience significant changes with larger neural networks, greater cutoffs for atomic interactions, or a larger number of maximum neighboring atoms. Considering this, along with the fact that our training protocols in this study adequately reproduce various experimental properties (as discussed in Section ), further fine-tuning of the NN hyperparameters was not explored, and the selected hyperparameters were proved to be sufficient for achieving accurate results.…”

Development of NaCl–MgCl2–CaCl2 Ternary Salt for High-Temperature Thermal Energy Storage Using Machine Learning

“…Xu et al studied two eutectic chloride molten salts, MgCl 2 −NaCl (MN) and MgCl 2 −KCl (MK). 32 Compared to individual salts, the coordination numbers of Mg−Cl, Na−Cl, and K−Cl in molten MN and MK were found to be higher, indicating that cations have a stronger ability to trap Cl − in the molten mixtures. These researches showed that DPMD not only has the same precision as AIMD in structural prediction but also can simulate with faster calculation speed and larger calculation dimension.…”

“…The basic transport properties such as density, self-diffusion coefficient, shear viscosity, and conductivity were in good agreement with experimental data. Xu et al studied two eutectic chloride molten salts, MgCl 2 –NaCl (MN) and MgCl 2 –KCl (MK) . Compared to individual salts, the coordination numbers of Mg–Cl, Na–Cl, and K–Cl in molten MN and MK were found to be higher, indicating that cations have a stronger ability to trap Cl – in the molten mixtures.…”