Theoretical prediction on the local structure and transport properties of molten alkali chlorides by deep potentials

doi:10.1016/j.jmst.2020.09.040

Journal of Materials Science & Technology

Volume 75, 10 June 2021, Pages 78-85

https://doi.org/10.1016/j.jmst.2020.09.040 Get rights and content

Highlights

•
Machine learning-based deep potentials were developed for molten LiCl, NaCl, and KCl.
•
It was demonstrated that deep potentials can achieve DFT accuracy.
•
The local structure and transport properties of these salts were investigated using deep potentials.
•
The results predicted by deep potentials matched well with the AIMD and experimental data.

Abstract

In this work, the local structure and transport properties of three typical alkali chlorides (LiCl, NaCl, and KCl) were investigated by our newly trained deep potentials (DPs). We extracted datasets from ab initio molecular dynamics (AIMD) calculations and used these to train and validate the DPs. Large-scale and long-time molecular dynamics simulations were performed over a wider range of temperatures than AIMD to confirm the reliability and generality of the DPs. We demonstrated that the generated DPs can serve as a powerful tool for simulating alkali chlorides; the DPs also provide results with accuracy that is comparable to that of AIMD and efficiency that is similar to that of empirical potentials. The partial radial distribution functions and angle distribution functions predicted using the DPs are in close agreement with those derived from AIMD. The estimated densities, self-diffusion coefficients, shear viscosities, and electrical conductivities also matched well with the AIMD and experimental data. This work provides confidence that DPs can be used to explore other systems, including mixtures of chlorides or entirely different salts.

Graphical abstract

Introduction

Molecular dynamics (MD) simulations using predefined potentials have been employed in numerous studies in the past few decades. Despite the success of MD simulations, it has limitations; for example, the accuracy of the results depends heavily on the quality of the predefined potential. Ab initio molecular dynamics (AIMD) simulations, which use electronic structure calculations to compute the forces that act on nuclei, have shown a higher accuracy. Unfortunately, it is computationally intensive and therefore, it is generally used on extremely small systems with short simulation times. A feasible way to address the issue of accuracy versus efficiency is to make an accurate, efficient, and transferable interatomic potential, but this is frequently time-consuming and requires a lot of expertise [1].

Recently, machine learning (ML) techniques have become of great interest to researchers, partly because they can serve as powerful tools for generating interatomic potentials [2,3]. With training data extracted from AIMD simulations, ML interatomic potentials can easily achieve ab initio accuracy. This means that MD simulations can be performed with accuracy that is comparable to that of AIMD and with efficiency that is similar to that of empirical potentials. Thus far, several kinds of ML potentials have been reported in the literature; these ML potentials include the Behler-Parrinello neural network potential (BNNP) [2], Gaussian approximation potential (GAP) [3], spectral neighbor analysis potential (SNAP) [4], and deep potential (DP) [5,6]. During the training process, atomic coordinates cannot be used directly because their format does not preserve the symmetry of the system. Usually, descriptors are used to convert the atomic coordinates into vectors to preserve the symmetry of the system. In BNNP, radial and angular symmetry functions are used as descriptors to preserve the symmetry. Although it is workable, fixing the symmetry functions is a painstaking process for systems with many atomic species. In the deep potential molecular dynamics (DPMD) scheme proposed by Zhang et al. [5,6], a local coordinate frame is established for every atom and its neighbors to preserve the symmetry, and this is considered to be more flexible. Although DPMD has been proven to be effective in studies of several systems, such as (Zr_0.2Hf_0.2Ti_0.2Nb_0.2Ta_0.2)C high entropy materials [7], water [8,9], Si [10], solid-state electrolytes [11], and alloy materials [[12], [13], [14]], there seems to be no report on molten alkali chlorides using this approach.

In this paper, DPs were trained for molten LiCl, NaCl, and KCl using the DeePMD-kit package [15]. The motivations for choosing such systems are as follows: (i) They are involved in many application fields, such as concentrating solar thermal power systems and nuclear reactors. (ii) Because these systems are the most typical molten salts, there is an extensive amount of quality data on structures and basic salt properties available to validate generated DPs. (iii) The newly generated DPs can be used to calculate the local structure and transport properties. (iv) The existing Born-Mayer-Huggins- Fumi-Tosi (BMHFT) potential can be used to compare the trained DPs [16,17]. It will be shown that the generated DPs can be used to predict the local structure and thermodynamics properties of molten LiCl, NaCl, and KCl at the DFT level.

The paper is organized as follows: The theory and training process of the DPs as well as the details of molecular dynamics simulations with DPs is briefly introduced in Section 2. In Section 3, the accuracy of the DPs is discussed by comparing the energies and forces of the DPs with those of DFT. Then, the local structure and transport properties obtained via DPMD simulations are compared with those obtained via AIMD simulations and experimental measurements through the partial radial distribution functions (RDFs), angle distribution functions (ADFs), densities, self-diffusion coefficients, shear viscosities, and electrical conductivities. Finally, conclusions are presented in Section 4.

Section snippets

Preparation of datasets

Datasets contain atomic configurations and related forces and energies, which can be extracted from AIMD simulations. The initial configurations of LiCl, NaCl, and KCl were prepared using the Packmol code [18]. All configurations consisted of 54 alkali metal cations and 54 Cl⁻ anions, and their sizes were fixed according to the experimental densities. AIMD simulations were performed at 900, 1073, and 1200 K for LiCl, 1080 and 1200 K for NaCl, and 1053 and 1200 K for KCl.

Classical MD simulations

Accuracy of trained DPs

The correlations between the energy and forces predicted using our DPs and those predicted using DFT for molten LiCl, NaCl, and KCl at multiple temperatures are presented in Fig. 1. The root-mean-square errors (RMSEs) of the energies and forces evaluated using our DPs for training and testing sets are summarized in Table S1. For the training set, the RMSEs of LiCl, NaCl, and KCl are 0.825, 0.426, and 0.427 meV/atom, respectively, for energies and 23.560, 17.714, and 18.267 meV/Å, respectively,

Conclusions

In this work, a scheme that is based on machine learning and a deep neural network was employed to generate interatom potentials for molten LiCl, NaCl, and KCl. From comparing the energies and forces that were evaluated using DFT and DPs, we demonstrate that DPs can easily achieve DFT accuracy. From comparing the results from the DPMD simulations with those from AIMD simulations and experimental measurements, we show that DPMD can be used to reliably calculate the radial distribution functions,

Declaration of Competing Interest

None.

Acknowledgments

The authors acknowledge the financial support provided by the National Natural Science Foundation of China (Grant U1407202 and Grant U1407126). The authors thank Haiou Ni, who provided support on software installation, including the DeePMD-kit and the LAMMPs. The authors also thank Gechuanqi Pan for many inspired discussions on MD simulations.

References (39)

F.-Z. Dai et al.
J. Mater. Sci. Technol.
(2020)
H. Wang et al.
Comput. Phys. Commun.
(2018)
M.P. Tosi et al.
J. Phys. Chem. Solids
(1964)
F.G. Fumi et al.
J. Phys. Chem. Solids
(1964)
S. Plimpton
J. Comput. Phys.
(1995)
G. Kresse et al.
Comput. Mater. Sci.
(1996)
A. Bengtson et al.
Comput. Mater. Sci.
(2014)
S. Wang et al.
Electrochim. Acta
(2019)
G.C. Pan et al.
Int. J. Heat Mass Transf.
(2016)
D. Marx et al.
Ab Initio Molecular Dynamics:Basic Theory and Advanced Methods
(2009)

J. Behler et al.

Phys. Rev. Lett.

(2007)

A.P. Bartok et al.

Phys. Rev. Lett.

(2010)

C.R. Trott et al.

SNAP: Strong Scaling High Fidelity Molecular Dynamics Simulations on Leadership-Class Computing Platforms

(2014)

L. Zhang et al.

Phys. Rev. Lett.

(2018)

L. Zhang et al.

End-to-end Symmetry Preserving Inter-atomic Potential Energy Model for Finite and Extended Systems

H.-Y. Ko et al.

Mol. Phys.

(2019)

G.M. Sommers et al.

Phys. Chem. Chem. Phys.

(2020)

R. Li et al.

Mater. Today Phys.

(2020)

A. Marcolongo et al.

ChemSystemsChem.

(2019)

Cited by (38)

Elucidating the local structure and properties of molten Na<inf>2</inf>CO<inf>3</inf>-K<inf>2</inf>CO<inf>3</inf> salts using Machine Learning-Driven molecular dynamics
2024, Chemical Engineering Science
Molten Na₂CO₃-K₂CO₃ binary salts have shown promise as electrolytes for fuel cell technology. Extensive investigations of the local structure and transport properties of these melts have been conducted via experiments and molecular dynamics (MD) simulations. With advancing machine learning capabilities, deep learning strategies are being widely incorporated into MD to generate highly accurate machine learning (ML) potentials. In this work, first-principle-derived Deep Potential is applied to study Na₂CO₃-K₂CO₃ melts at specified temperatures. A deep neural network acts to fit the first-principle data and produce the model. This model enables large-scale MD simulations with accuracy comparable to density functional theory (DFT). The local structure is analyzed. Temperature and concentration impacts on ion arrangements are observed. The concentration effect has little influence on the melt structure, but the radial and angular distributions between the ions are strongly affected by temperature. Density, viscosity, and diffusion values predicted by the DP model agree closely with experimental measurements. In comparison to conventional empirical potentials, the DP model demonstrates DFT-level accuracy for structural analysis and experimental-level property predictions. This study provides detailed insights into the microstructure and transport properties of Na₂CO₃-K₂CO₃ melts via the power of the DP model, showcasing their ability to advance understanding of complex molten salt systems. Meanwhile, a molten salt database is initially constructed using machine learning strategies.
Hydration MgCl<inf>2</inf>-NaCl-KCl molten salt using a novel approach for training machine learning potential
2024, Journal of Molecular Liquids
The structural information and thermophysical properties of high-temperature magnesium electrolytes play critical roles in controlling the energy consumption of industrial production and the quality of magnesium alloys. Large-scale molecular dynamics simulations can achieve accuracy similar to that of DFT while overcoming the limitations of high-temperature experiments by integrating machine learning (ML) strategies. However, many AIMD simulation tasks, computationally expensive due to configuration and volume optimization, are required to construct an initial dataset for the current ML potential training scheme. To bypass this issue, we developed a novel scheme using Deep Potential to generate ML potential that can be applied to multiple temperatures, improving efficiency without compromising accuracy. A short-time AIMD simulation at a certain target temperature was involved, followed by an iterative task using the DP-GEN package to generate an initial dataset based on a varying upper bound for model force deviation. This dataset was then used in a new DP-GEN task to train ML potential. Our method has been applied for the first time to the molten MgCl₂-NaCl-KCl system with trace amounts of water (hydration MNK). The trained potential effectively predicted interatomic energies and forces, with errors of 0.5517 meV/atom and 3.063 meV/Å, respectively, denoting the trained potential is similar accuracy relative to DFT. DPMD simulation, the large-scale methodology, was performed to investigate the structural information and thermophysical properties of the hydration MNK system. Through analyzing the obtained structural information, Coulombic interactions are the dominant forces in the melt and a stronger affinity between Na⁺ and water molecules relative to other cations. H₂O is unable to cross the first coordination layer of Cl-Mg. There are also differences in how the hydrogen and oxygen atoms of the water molecules interact with the cations in the melt. This work discusses the temperature-dependent evolution of several thermophysical properties, including density, self-diffusion coefficient, viscosity, and ionic conductivity. Our results show that density, viscosity, and ionic conductivity all exhibit negative relationships with temperature. In addition, the order of diffusion is as follows: $D_{{Mg}^{2+}} < D_{{Cl}^{-}} < D_{{Na}^{+}} < D_{K^{+}} < D_{H_{2} O}$ .
Path-integral molecular dynamics predictions of equilibrium H and O isotope fractionations between brucite and water
2023, Geochimica et Cosmochimica Acta
Precise prediction of hydrogen (H) isotope fractionations among different substances is a long-standing challenge in isotope geochemistry, as it needs treatments beyond the harmonic approximation. Path-integral molecular dynamics (PIMD) simulations have recently been proved to be valid in predicting equilibrium isotope fractionations for light elements. However, the lack of reliable force fields hinders the application of PIMD to the condensed phases. In this work, the deep potential models trained on the first-principles molecular dynamics (FPMD) data are applied to PIMD simulations to determine the D/H and ¹⁸O/¹⁶O isotope fractionation factors between brucite and water. To quantitatively assess the influence of quantum effects, the D/H and ¹⁸O/¹⁶O isotope fractionations between brucite and water are also determined under the framework of harmonic approximation.
By comparing the results of the PIMD and those of harmonic calculations, we find that H and O atoms in brucite and water are strongly affected by the anharmonic effects. The accuracy of the harmonic isotope fractionations mainly depends on the cancellation percentage of the anharmonic effects on the reduced partition function ratios (RPFRs) of the two substances. The isotope fractionations predicted by PIMD simulations are close to the experimental results at high temperatures. However, at low temperatures, the predicted isotope fractionations are different from those of experiments. The discrepancies are attributed to the approximations in the density functional theory (DFT) functionals used in this work, i.e., the inaccuracy in predicting the proton positions in brucite and water. Moreover, we find that the H isotope fractionations are extremely sensitive to the change of pressure. The direction of H isotope fractionation between brucite and water (at 300 K) will be even inversed as the pressure increases to more than 3 GPa. This strong pressure sensitivity may be a common characteristic of hydrous minerals and water systems. Therefore, the extrapolated H isotope fractionations used in the studies of dehydration of subducting slabs or deep Earth H distribution based on the isotope fractionations under low pressures may need to be rechecked.
Atomwise force fields for molten alkali chlorides (LiCl and KCl) and their mixtures: efficient parameterization via genetic algorithms
2023, Journal of Nuclear Materials
Citation Excerpt :
As a result, we chose to analyze the performance of the GA-derived force fields using highly accurate FPMD simulations. The RDFs obtained from the FPMD simulations show similar results as previously reported studies [23,88,89]. The RDFs produced from CMD simulations utilizing the GA-derived parameters and FPMD are shown in Fig. 9, while Tables S4–S6 (supporting information) list the nearest-neighbor distances or first peak positions of the RDFs obtained from the GA-derived parameters, previously reported force fields, and FPMD.
Over the last few decades, molten salts have gained significant attention in the field of green energy production. One useful method that can accelerate the discovery and optimization of desirable molten salts from a vast range of materials is classical molecular dynamics (CMD). However, because of the absence of CMD force fields that can reasonably reproduce experimental physical properties, the use of CMD to study molten salts remains a significant challenge. Hence, in this study, we report non-polarizable rigid ion models (RIMs) for the widely used LiCl, KCl, and LiCl–KCl molten salts, obtained via genetic algorithms (GAs). Based on the results presented in this study, GAs proved to be very effective in simultaneously optimizing a large number of force field parameters. With GAs, the force field generation process is automated and no manual intervention is required. The RIM force fields developed herein can reasonably reproduce experimental physical properties at different temperatures and various salt compositions, provides microstructural characteristics similar to those obtained using highly accurate and reliable first-principles MD, and an improved overall performance in contrast to previously reported polarizable force fields, all at a lower computational cost. In contrast to previous studies, we developed atomwise RIM force fields that employ mixing rules for interactions between ions of different types. This reduced the total number of optimizable terms, and thus resulted in RIM parameters that are transferrable between different molten salt systems.
Thermophysical properties of FLiBe using moment tensor potentials
2022, Journal of Molecular Liquids
Citation Excerpt :
Most of the published works on MLIPs for molten salts have used NNIP or DPMD. The number of data points that are typically used to train such potentials is in the tens of thousands [26,29–32] and in some cases hundreds of thousands [28,33–35]. Here, one data point consists of a single energy and a set of force and stress vectors on atoms from one periodic unit cell configuration (the unit cell typically contains about 100 atoms).
Fluoride salts are prospective materials for applications in some next-generation nuclear reactors and their thermophysical properties at various conditions are of interest. Experimental measurement of the properties of these salts is often difficult and, in some cases, unfeasible due to challenges from high temperatures, impurity control, and corrosivity. Therefore, accurate theoretical methods are needed for fluoride salt property prediction. In this work, we used moment tensor potentials (MTP) to approximate the potential energy surface of eutectic FLiBe (66.6% LiF – 33.3% BeF₂) predicted by the ab initio (DFT-D3) method. We then used the developed potential and molecular dynamics to obtain several thermophysical properties of FLiBe, including radial distribution functions, density, self-diffusion coefficients, thermal expansion, specific heat capacity, bulk modulus, viscosity, and thermal conductivity. Our results show that the MTP potential approximates the potential energy surface accurately and the overall approach yields very good agreement with experimental values. The converged fitting can be obtained with less than 600 configurations generated from DFT calculations, which data can be generated in just 1200 core hours on today's typical processors. The MTP potential is faster than many machine learning potentials and about one order of magnitude slower than widely used empirical molten salt potentials such as Tosi/Fumi.
Computational modelling of the local structure and thermophysical properties of ternary MgCl2-NaCl-KCl salt for thermal energy storage applications
2022, International Journal of Heat and Mass Transfer
Citation Excerpt :
Indeed, these numerical methods must be validated by experimental testing, and are precursor of the development of future algorithmic models that could be implemented in new power plants, for automated machine learning and digital twins development. Molten chlorides have already been studied for CSP application and some of these results will be compared to our calculations [8–25]. Indeed, assessment of potassium and sodium alkali chloride were the most investigated ones [19, 21, 23, 24], but magnesium chloride [22] and the ternary MgNaK chloride [22, 25, 26] attract more attention as they became salts of interest, offering the support of experimental studies to the design of accurate phase diagrams.
Molten salts as heat transfer fluids (HTF) for concentrated solar power (CSP) plant application are considered as the best thermal storage medium, and more precisely molten chlorides, presenting a wide operating range and coupled with competitive cost. Furthermore, MgCl2-NaCl-KCl (MgNaK) mixture appeared as the most promising one but need further studies to better understand its thermophysical properties. Indeed, its hydrated form leads to the formation of corrosive compounds. In this research, two different methods are used to model the ternary mixture. The dehydration process is evaluated by thermodynamical calculations with Thermocalc software. Then, the local structure, thermal conductivity and viscosity are estimated by means of molecular dynamics simulation, with LAMMPS package. The results were close to past simulations studies and experimental references, but discrepancies need to be further minimized regarding some variable fluctuations.

View all citing articles on Scopus

View full text

Theoretical prediction on the local structure and transport properties of molten alkali chlorides by deep potentials

Highlights

Abstract

Graphical abstract

Introduction

Section snippets

Preparation of datasets

Accuracy of trained DPs

Conclusions

Declaration of Competing Interest

Acknowledgments

J. Mater. Sci. Technol.

Comput. Phys. Commun.

J. Phys. Chem. Solids

J. Phys. Chem. Solids

J. Comput. Phys.

Comput. Mater. Sci.

Comput. Mater. Sci.

Electrochim. Acta

Int. J. Heat Mass Transf.

Ab Initio Molecular Dynamics:Basic Theory and Advanced Methods

Phys. Rev. Lett.

Phys. Rev. Lett.

SNAP: Strong Scaling High Fidelity Molecular Dynamics Simulations on Leadership-Class Computing Platforms

Phys. Rev. Lett.

End-to-end Symmetry Preserving Inter-atomic Potential Energy Model for Finite and Extended Systems

Mol. Phys.

Phys. Chem. Chem. Phys.

Mater. Today Phys.

ChemSystemsChem.