betaFeedback

WorksPhysics and Astronomy

Thermophysical Properties of Molten FLiNaK: a Moment Tensor Potential Approach

Abstract

Fluoride salts demonstrate significant potential for applications in next-generation nuclear reactors, necessitating a comprehensive understanding of their thermophysical properties for technological advancements. Experimental measurement of these properties poses challenges, due to factors such as high temperatures, impurity control, and corrosion. Consequently, precise computational modeling methods become essential for predicting the thermophysical properties of molten salts. In this work, we performed molecular dynamics (MD) simulations of several thermophysical properties of the eutectic salt mixture LiF-NaF-KF (FLiNaK) melt, including density, self-diffusion coefficients, viscosity, and thermal conductivity. We demonstrated the successful application of moment tensor potentials (MTP) as an accurate model for interatomic interactions in FLiNaK. Our results on thermophysical properties calculations exhibit strong agreement with experimental data. An important aspect of our methodology is the incorporation of an active learning scheme, which enables the generation of a robust and accurate potential, while maintaining a moderate-sized training dataset.

Introduction

Molten salts play a pivotal role in various industrial applications, including molten salt batteries Ong et al. [39]; Bell et al. [3]; Cui et al. [5]; Parasotchenko et al. [41]; LeBlanc [25], molten salt reactor systems Beneš and Konings [4]; Magnusson et al. [32], as well as in pyroprocessing methods for recycling nuclear fuel Zhitkov et al. [67, 66]; Mullabaev et al. (2022). Most important characteristics of melts from an application standpoint are temperature-dependent values of density, diffusivity, viscosity and thermal conductivity, which are usually measured experimentally. Despite the availability of numerous experimental measurements for different molten salts, reported thermophysical properties exhibit significant scatter and uncertainties, reaching up to 20% Romatovski and Hu [49]. This variability is attributed to challenges associated with impurities, high-temperature measurements, and deviations in composition.

As an alternative to experimental methods, accurate theoretical approaches based on ab initio molecular dynamics (AIMD) simulations have emerged as valuable tools for evaluating thermophysical properties in well-controlled conditions Porter et al. [44]. While AIMD has significantly advanced our understanding of the structure of various molten salts, essential properties such as diffusion coefficients, viscosity, and thermal conductivity necessitate larger unit cell sizes and extended simulation times for relevant statistical analysis Gheribi et al. [12]; Jabbari et al. [19]. Thus, there is a critical need for an efficient computational method to reliably predict the thermophysical properties of molten salts.

An alternative computational approach involves MD simulations, where interatomic interaction potential is fitted to first-principles calculations or experimental data Salanne et al. [52]; Galashev et al. [9]; Maxwell [34]. This overcomes limitations related to simulation cell sizes and time scales, enabling simulations of larger structures for more extended periods compared to AIMD. Properties derived from classical MD simulations generally achieve sufficient numerical precision, with errors dominated by the accuracy of the underlying interatomic potentials Lu et al. (2021); Lee et al. [26].

In the last decade, significant progress has been made in developing the so-called machine learning interatomic potentials (MLIPs) Deringer et al. [6]. MLIPs have demonstrated promise for MD simulations of molten salts modelling with near ab initio accuracy, on time and length scales comparable to traditional interatomic potentials Liang et al. [28]; Feng et al. [7]; Lam et al. [24]; Li et al. (2021); Rodriguez et al. [46]; Attarian et al. [2]. The moment tensor potential Shapeev [53] (MTP) was shown to be among the most efficient (in terms of data utilization) and accurate machine-learned models for interatomic interactions Zuo et al. [68]. In the case of molten LiF-BeF2_2 mixture, it was recently shown that MTP can be used to compute thermophysical properties with both high precision and low data utilization Attarian et al. [2].

In this study, we employ MTPs to calculate thermophysical properties of molten FLiNaK at eutectic composition (46.5-11.5-42 mol%). FLiNaK is among the most promising and well-investigated salts Locatelli et al. [30]; Maslennikova et al. [33]; Lizin et al. [29], with a lot of experimental and theoretical data. We fitted an MTP to approximate the potential energy surface of FLiNaK on the data calculated from first principles in the dispersion-corrected Density-Functional Theory (DFT-D3) framework. As will be demonstrated, MTP model used in couple with MD simulations allows to obtain thermophysical properties of FLiNaK in a good agreement with reported literature data, both experimental and theoretical.

Methodology

Moment Tensor Potential

In this work, we used MTP approach implemented in the MLIP-2 package Novikov et al. [38] to investigate the thermophysical properties of molten FLiNaK at finite temperatures. The potential energy of an atomic system as described by the MTP interatomic potential is defined as a sum of the energies of atomic environments of the individual atoms:

EMTP=∑i=1NV(ni),E_{\mathrm{MTP}} = \sum_{i=1}^{N} V(n_i),

where the index ii label NN atoms of the system, and nin_i describes the local atomic neighborhood around atom ii within a certain cutoff radius RcutR_{\mathrm{cut}} and the function VV is the moment tensor potential:

V(ni)=∑αξαBα(ni),V(n_i) = \sum_{\alpha} \xi_{\alpha} B_{\alpha}(n_i),

where ξα\xi_{\alpha} are the fitting parameters and Bα(ni)B_{\alpha}(n_i) are the basis functions that will be defined below. Moment tensors descriptors are used as representations of atomic environments and defined as:

Mμ,ν(ni)=∑jfμ(∣rij∣,zi,zj)rij⊗⋯⊗rij⏟ν times,M_{\mu,\nu}(n_i) = \sum_j f_{\mu}\left(\lvert r_{ij}\rvert,z_i,z_j\right) \underbrace{r_{ij} \otimes\cdots\otimes r_{ij}}_{\nu\ \text{times}},

where the index jj goes through all the neighbors of atom ii. The symbol “⊗\otimes” stands for the outer product of vectors, thus rij⊗⋯⊗rijr_{ij} \otimes\cdots\otimes r_{ij} is the tensor of rank ν\nu encoding the angular part which itself resembles moments of inertia. The function fμf_{\mu} represents the radial component of the moment tensor:

fμ(∣rij∣,zi,zj)=∑kcμ,zi,zj(k)Q(k)(r),f_{\mu}\left(\lvert r_{ij}\rvert,z_i,z_j\right) = \sum_k c_{\mu,z_i,z_j}^{(k)} Q^{(k)}(r),

where ziz_i and zjz_j denote the atomic species of atoms ii and jj, respectively, rijr_{ij} describes the positioning of atom jj relative to atom ii, cμ,zi,zj(k)c_{\mu,z_i,z_j}^{(k)} are the fitting parameters and

Q(k)(r):=Tk(r)(Rcut−r)2Q^{(k)}(r) := T_k(r)(R_{\mathrm{cut}}-r)^2

are the radial functions consisting of the Chebyshev polynomials Tk(r)T_k(r) on the interval [Rmin⁡,Rcut][R_{\min},R_{\mathrm{cut}}] with the term (Rcut−r)2(R_{\mathrm{cut}}-r)^2 that is introduced to ensure a smooth cut-off to zero. The descriptors Mμ,νM_{\mu,\nu} taking ν\nu equal to 0,1,2,…0, 1, 2, \ldots are tensors of different ranks that allow to define basis functions as all possible contractions of these tensors to a scalar, for instance:

B0(ni)=M0,0(ni),B1(ni)=M0,1(ni)⋅M0,1(ni),B2(ni)=M0,0(ni)(M0,2(ni):M0,2(ni)).\begin{aligned} B_0(n_i) &= M_{0,0}(n_i),\\ B_1(n_i) &= M_{0,1}(n_i)\cdot M_{0,1}(n_i),\\ B_2(n_i) &= M_{0,0}(n_i)\left(M_{0,2}(n_i):M_{0,2}(n_i)\right). \end{aligned}

Therefore the level of Mμ,νM_{\mu,\nu} is defined by lev⁡Mμ,ν=2μ+ν\operatorname{lev}M_{\mu,\nu}=2\mu+\nu and if BαB_{\alpha} is obtained from Mμ1,ν1M_{\mu_1,\nu_1}, Mμ2,ν2M_{\mu_2,\nu_2}, …\ldots, then lev⁡Bα=(2μ1+ν1)+(2μ2+ν2)+…\operatorname{lev}B_{\alpha}=(2\mu_1+\nu_1)+(2\mu_2+\nu_2)+\ldots. By including all basis functions such that lev⁡Bα<d\operatorname{lev}B_{\alpha}<d we obtain the moment tensor potential of level dd, which we denote as MTPd\mathrm{MTP}_d.

Dataset Generation and Potential Fitting

The MTP utilized in this study is trained using data computed within the Density-Functional Theory (DFT) framework. All DFT calculations were carried out using VASP (Vienna ab initio simulations package) Kresse and Furthmüller [21] with the projector augmented wave method Kresse and Joubert [22]. The Perdew-Burke-Ernzerhof generalized gradient approximation (PBE-GGA) Perdew et al. [42] was employed for the exchange–correlation functional, and the DFT-D3 method Grimme et al. [15] was utilized to account for dispersion forces.

The initial dataset was generated from four independent AIMD trajectories (at temperatures TT = 800 K, 1000 K, 1200 K, and 1400 K), which were conducted in the isothermal-isobaric ensemble (NPT) using the Nosé-Hoover thermostat Nosé [37]. Each AIMD trajectory was simulated for 2 ps with a 1 fs time step and for supercells containing 56 atoms (28 F atoms, 13 Li atoms, 3 Na atoms, and 12 K atoms) to closely mimic eutectic composition. The plane-wave basis set had an energy cutoff of 550 eV, and a single gamma point was used to sample the Brillouin zone. The initial structures for AIMD were generated by randomly distributing atoms in the cubic cell, which had the volume equal to the total volume of spheres, each with a radius matching the van der Waals radii of the respective atoms. The first picosecond of each AIMD simulation was discarded, and the second picosecond was subsampled with the 5 fs time intervals, resulting in a set of 800 samples. These samples were employed to train an initial MTP of level 14.

The root-mean-squared-errors (RMSEs) on this dataset are 2.1 meV/atom for energies and 48 meV/Å for forces. Previous studies Feng et al. [7]; Lam et al. [24] have indicated that RMSE of energy smaller than 5 meV/atom and RMSEs on forces smaller than 100 meV/Å are generally sufficient for predicting properties such as density, radial distribution functions, diffusion coefficient, and viscosity of molten salts. However, achieving this accuracy does not by itself ensure the robustness (ability to run long MD) of the potential. Generally, two strategies are employed to train a robust MLIP capable of capturing the entire range of local atomic environments: (i) including a variety of different systems which requires lengthy AIMD simulations, and (ii) utilizing an active learning (active sampling, learning-on-the-fly) strategy to selectively add data points on which the MLIP extrapolates significantly, i.e., the prediction of energies and forces of these structures is done with high uncertainty.

For the purpose of methodological validation, we have explored both strategies. We extended AIMD simulations by an additional 3 ps at each temperature, yielding (after subsampling) a dataset of 3200 samples. We shuffled this dataset and divided it into the training (20% of the total dataset) and test sets (remaining 80% of the dataset). Figure 1 presents a comparative analysis of DFT and MTP calculated energies and forces for both training and test sets. As presented in Table 1, the RMSE of MTP on energies is just 1.93 meV/atom, and on forces, it is 39 meV/Å. The displayed RMSE values in Figure 1 demonstrate that the MTP, trained on a 20% subset of configurations derived from AIMD simulations, accurately predicts the energies and forces of the remaining 80% (test set) with comparable errors to those of the training set. Low errors on the test set, suggests that the MTP is well-fitted and likely capable of extrapolating to diverse structures and chemical environments.

TrainTest
Energy (meV/atom)1.831.93
Forces (meV/Å)3939

Table 1. Potential energy and force RMSEs on training and test sets for the fitted Moment Tensor Potential.

Parity plots of the energies and forces for training and test sets

Figure 1. Parity plots of the energies (a) and forces (b) for training (blue) and test (orange) sets. The diagonal line in each figure shows the perfect fit.

However, even with these promising results, an ability to run long MD simulations without failure remains to be demonstrated. To assess stability of the fitted potential, we conducted MD simulations using MTP as the model for interatomic interactions in the micro-canonical (NVE) ensemble within LAMMPS (Large-scale Atomic/Molecular Massively Parallel Simulator) Thompson et al. [59]. Unfortunately, these simulations proved to be unstable, exhibiting instability after several tens of picoseconds. Notably, extending the potential training to the entire dataset did not alleviate this issue. This outcome underscores that while low energy and force errors are achieved, they do not guarantee the robust applicability of a MLIP for practical MD simulations, even under the same thermodynamic conditions as those employed in the training data.

We now revisit the active learning strategy, which means interactively selecting a diverse, but minimal, set of training data in the feature space to effectively fit the potential. Various active learning schemes are employed in the development of machine learning (ML) potentials [65, 54]. In our study, we utilized the D-optimality-based active learning procedure developed in Podryabinkin and Shapeev [43] and available in the MLIP-2 package.

Within the active learning algorithm, we initiate MD calculations in LAMMPS using a pretrained MTP. At each step of the MD simulations, the algorithm assesses the extrapolation grade γ\gamma of the atomic configuration based solely on atomic coordinates. Configurations with γ>2\gamma> 2 are added to the preselected set. When γ\gamma exceeds 10, the MD simulation halts, and all sufficiently different configurations from the preselected set are incorporated into the training, followed by the refitting of the potential. This procedure repeats until MD simulations can run without failure for 200 ps. In our case, MTP achieved robustness with a training set comprising 880 samples, indicating that only a small portion of the actively selected configurations required single-point DFT calculations. The potential trained in this manner demonstrates the ability to robustly conduct MD simulations for at least hundreds of picoseconds, showcasing the effectiveness of the active learning procedure in terms of reducing the necessary DFT data for training and enhancing potential robustness. The training set generated during the active learning procedure can later be used to train a MTP with enhanced accuracy, as will be demonstrated in this study.

Thus, the active learning scheme plays a crucial role in enhancing the robustness of any MLIP, and MTP, in particular. Its employment not only ensures the applicability of MTP at large length scales and extended time scales in the MTP-MD simulations, but also significantly reduces the number of required DFT calculations. In the subsequent sections of the paper, we present the results of our calculations on the thermophysical properties of FLiNaK. Specifically, we assess the MTP’s capability to accurately calculate the temperature dependencies in density, diffusivity, viscosity, and thermal conductivity.

Results

Radial Distribution Functions and Self-Diffusion Coefficients

To assess how well the developed MTP can predict local structural features of molten FLiNaK, we first calculated radial distribution functions (RDF) and diffusion coefficients of ions. The smallest unit cell of molten FLiNaK in an eutectic composition contains 400 atoms: 93 Li atoms, 200 F atoms, 23 Na atoms, and 84 K atoms, respectively. For MTP-MD simulations, the simulation cell was replicated in a 2×2×2 arrangement, which resulted in a supercell with 3200 atoms. The usage of larger supercells yields better statistics of calculations and MTP-MD allows to perform long simulations, which is inaccessible by AIMD. The first RDF peaks for F-Li, F-Na, and F-K are shown in Figure 2. RDF for Li-F shows a sharp peak at 1.842 Å and then rapidly decays. This suggests a strong bonding between Li and F within FLiNaK. On the other hand, the RDF for K-F has a wider first peak at 2.595 Å that does not decay to zero. This result suggests that the first nearest neighbor shell for K and F is more diffusive and not as well defined as for Li and F, which further indicated that the bonding between K and F in FLiNaK is not as strong as between Li and F. The values of the calculated first peak distances obtained using MTP-MD for F-Li, F-Na, and F-K are close to the experimental values Igarashi et al. [17] as shown in an inset of Figure 2. This indicates that MTP can accurately capture the local environment of ions in molten FLiNaK.

Radial distribution functions for F-Li, F-Na, and F-K ion pairs

Figure 2. Radial distribution functions, extracted from MTP-MD at temperature T=793T=793 K for F-Li, F-Na, and F-K ion pairs. Experimental nearest-neighbor peak distance for aforementioned ion pairs was determined from X-ray scattering experiment Igarashi et al. [17].

Self-diffusion of atoms in a melt is another fundamental property of liquid dynamics that provides important structural information Rollet et al. [47]. Figure 3(a, Figure 3b, Figure 3c, Figure 3d) show the temperature-dependence of self-diffusion coefficients of F, Li, Na, and K on a logarithmic scale, obtained from MTP-MD and experiment Umesaki et al. [61]. Although the experimental data is limited to a relatively short temperature intervals, for all four atomic species the calculated values and the slope of temperature dependencies are both close to the experimental results. Such results further validates our approach and demonstrated benefits of using atomistic simulations to accompany the experiment.

Figure 3. Temperature dependence of the self-diffusion coefficients for (a) Li, (b) K, (c) Na, (d) F. Experimental data is taken from Umesaki et al. [61]. Error bars are calculated from the standard deviation among 5 separately trained MTPs.

Density

We next study the temperature dependence of the density as one of the most technologically important thermophysical characteristics of the molten salt. While experimental measurements are feasible, computational modeling enables fast and inexpensive exploration of various salt mixtures with diverse components and compositions. The density of molten FLiNaK has been experimentally measured at different temperatures [49, 63] (); An et al. 2017), and recent efforts have employed ML potentials based on neural networks to calculate it Lee et al. [26].

We perform density calculations of molten FLiNaK, using the same simulation cell as for the diffusion calculations (comprising 3200 atoms at eutectic composition). First, we investigated the influence of different levels of MTP (i.e., varying numbers of parameters) in predicting the density of melt at T=1000T=1000 K. Figure 4 illustrates that beyond the level 16, increasing the complexity of MTP does not substantially enhance density values towards experimental results, while computational costs grow. For all subsequent calculations in this study, we employed the potential with level 16. Notably, such a clear convergence of the potential complexity and the property of interest (density in this case) as in Figure 4 is, however, not always observed. We also note that fitting of the MTPs is based on the Broyden-Fletcher-Goldfarb-Shanno method, and the optimized parameters depended on the initialization of MTP parameters, resulting in varying values and errors for the targeted property. To address this variability, we conducted 5 optimization sessions for each potential level, with error bars depicted in Figure 4. Our findings indicate that the level of as low as 10 can yield reasonably accurate results for molten FLiNaK density. The errors on energies and forces exhibit nearly identical patterns across different potential levels (from 10 to 20), aligning with observations in a recent study of molten LiF-BeF2_2, where MTP was also used Attarian et al. [2].

Figure 4. Density of FLiNaK calculated at T=1000K using MTP with different levels of complexity (number of parameters). The experimental data is taken from An et al. (2017); W. D. Powers and Greene [63]; Romatoski and Hu (2017). Error bars represent one standard deviation calculated over 5 separately trained MTPs.

Next, we proceed toward the comparison of our results with previous theoretical calculations. Recently, DeepMD potential was used to study molten FLiNaK Lee et al. [26]. We performed density calculations for temperatures T=600T=600 K, 800 K, 1000 K, 1200 K. Figure 5 reveals an overall agreement between our results and those reported in Lee et al. [26]. Notably, the configurations used for potential training in Lee et al. [26] were sampled at a fixed experimental density in the NVT ensemble, whereas we performed AIMD simulations with the NPT ensemble.

Figure 5. Temperature dependence of the density of FLiNaK calculated using MLIP. Olive disks represent results of MTP calculated on a dataset in which D3 dispersion corrections was not taken into account, while red disks represent MTP, which was trained on dataset with the D3 correction Grimme et al. [15]. Density calculated using DeepMD (orange triangles) is taken from Lee et al. [26]. Error bars represent one standard deviation calculated over 5 separately trained MTPs.

Given that FLiNaK is an ionic compound, accounting for long-range interactions becomes crucial in atomic dynamics calculations. Although the functional form of MTP generated by the MLIP-2 package currently lacks explicit terms for dispersion corrections, our tests of the convergence of energies and forces with respect to the cutoff radius of interactions showed no substantial improvements beyond a cutoff radius of 5.5 Å. This distance is deemed sufficient to include multiple neighbor shells in the liquid, with negligible electrostatic interactions expected beyond this distance for near-equilibrium FLiNaK configurations. While directly incorporating long-range interactions treatment into potential construction could potentially enhance simulation accuracy, such functionality has not been developed yet.

To the best of our knowledge, the literature lacks an in-depth exploration of the impact of long-range interactions on thermophysical properties calculations of molten salts. A study Frandsen et al. [8] demonstrated that including vdW-corrections increased the error in density determination by 4%. In our case, Figure 5 illustrates how a potential trained on data with the DFT-D3 correction leads to slightly different results (about a 10% difference) compared to a potential trained without the D3 correction. In our case, including long-range interactions during the dataset generation step improved agreement with experimental results. This test was performed by generating a dataset using the active learning methodology described earlier.

Finally, Figure 6 presents a comparison of temperature-dependent density of molten FliNaK obtained using MTP-MD with several experimental results [63, 49] (An et al. 2017). Our findings align well with the experimental results, despite not fixing experimentally-known density during the preparation of the training set. Given that experimental density may not be available for other salts, we emphasize this as an appropriate approach for conducting every MLIP benchmarks.

Figure 6. Temperature dependence of the density of FLiNaK. The experimental data is taken from An et al. (2017); W. D. Powers and Greene [63]; Romatoski and Hu (2017). Error bars represent one standard deviation calculated over 5 separately trained MTPs.

In the following subsections we will demonstrate how MTP can be used in conjunction with the Green-Kubo method [23, 13] to calculate viscosity of molten salt and with the Müller-Plathe non-equilibrium method [36] for thermal conductivity calculations. It should be noted that due to the high computational cost of conducting viscosity and thermal conductivity calculations the reported values are presented without error bars.

Viscosity

We calculated viscosity using the Green-Kubo (GK) approach Kubo et al. [23]; Green [13], employing the implementation available in the LAMMPS package. The GK approach involves the calculation of viscosity through the integral of the auto-correlation function of the off-diagonal elements of the stress tensor, expressed by the following relation:

η=VkBT∫0∞⟨Pαβ(t)Pαβ(0)⟩ dt,\eta= \frac{V}{k_{\mathrm{B}}T}\int_{0}^{\infty}\langle P_{\alpha\beta}(t)P_{\alpha\beta}(0)\rangle\,\mathrm{d}t,

where η\eta is viscosity, kBk_{\mathrm{B}} is the Boltzmann constant, and PαβP_{\alpha\beta} are the off-diagonal elements of the stress tensor.

In our viscosity calculations we again utilized a simulation cell containing 3200 atoms, with a simulation time step of 1 fs. For all temperatures, an auto-correlation time of 10 ps was selected ensuring the convergence of the auto-correlation function of the diagonal stress components to zero. After the initial equilibration at each temperature, the simulation was extended for 5 ns, employing the NVE ensemble.

Figure 7 illustrates the temperature-dependent viscosity of FLiNaK calculated here and experimentally J. Ambrosek and Allen [18]; Vriesema [62]; Rudenko et al. [50]. Our results exhibit excellent agreement with the experimental data. The relationship between viscosity and temperature established in our work is described by the following equation:

Figure 7. Temperature dependence of the viscosity of molten FLiNaK (eutectic composition). Experimental data is taken from J. Ambrosek and Allen [18]; Vriesema [62]; Rudenko et al. [50].

η=0.00543⋅exp⁡(5422T) [mPa⋅s].\eta= 0.00543 \cdot\exp\left(\frac{5422}{T}\right)\ [\mathrm{mPa}\cdot\mathrm{s}].

Thermal Conductivity

Before delving into the discussion of our heat transport calculations, it is essential to provide an overview of the available experimental data on the temperature dependence of a thermal conductivity (labeled as κ\kappa from now on). The earliest works provide κ\kappa values ranging from 1.3 W/mK to 4.5 W/mK, Vriesema [62]; Grele and Gedeon [14]; Hoffman and Lones [16]. Some experimental works Smirnov et al. [55]; Sohal et al. [56] indicate a slight increase in thermal conductivity with temperature, as depicted in Figure 8. Work Williams [64] indicated that the thermal conductivity at 973 K will be in the range 0.6–1.0 W/mK. Most recent experiments Gallagher et al. [10]; Robertson et al. [45]; Rudenko et al. [51] agree in suggesting a slight decrease of κ\kappa with temperature, indicating that the thermal conductivity value should be below 1 W/mK.

Figure 8. Temperature dependence of the thermal conductivity of molten FLiNaK (eutectic composition). The experimental results are taken from Smirnov et al. [55]; Sohal et al. [56]; Gallagher et al. [10]; Robertson et al. [45]; Rudenko et al. [51]. In Smirnov et al. [55] κ\kappa was measured with a correction of ±0.012\pm0.012 W/mK, introducing an uncertainty of 1-5%.

We determined thermal conductivity of molten FLiNaK using the Müller-Plathe non-equilibrium method Müller-Plathe [36]. It was demonstrated in Pan et al. [40] that this approach allows computing heat transport in molten salts in a better agreement with experimental values, comparably to the Green-Kubo method. We used a time step of 0.5 fs and the supercell with 4000 atoms (2.6 nm × 2.9 nm × 23 nm dimensionality) and a kinetic energy swap rate of 1 in every 1000 steps. For each temperature, after the initial equilibration in the NPT ensemble for 10 ps, the simulations were done for 4 ns in the NVE ensemble. Figure 8 shows results of our calculations and mentioned above experimental data. The linear fit of our results lead to the following equation for the thermal conductivity:

κ=0.880−0.116⋅10−3T [W/m⋅K]\kappa= 0.880 - 0.116 \cdot10^{-3}T\ [\mathrm{W/m \cdot K}]

As demonstrated in Figure 8, although the experimental data is slightly scattered, our calculated values agree well with the most recent experimental works Gallagher et al. [10]; Robertson et al. [45]; Rudenko et al. [51]. Taken into account the magnitude of thermal conductivity changes with temperature, we are of the opinion that it is reasonable to state that the thermal conductivity in FLiNaK stays nearly constant with temperature grows and its value is below 1 W/mK.

Summary and conclusions

In this work, we utilized a MTP and assessed its performance for the calculations of thermophysical properties of molten FLiNaK in a range of temperatures. Employing an active learning approach enabled the rapid training of a robust potential capable of predicting density, diffusion, viscosity, and thermal conductivity with near ab initio accuracy. While previous computational works demonstrated good agreement with experimental data, the MTP, coupled with active learning, stands out for its ability to deliver results much faster, thanks to its efficient utilization of data. This significantly reduces the computational cost compared to previous studies employing alternative MLIPs. This breakthrough paves the way for swift yet precise exploration of thermophysical properties across various salt systems.

Our analysis, encompassing radial distribution function and diffusivity calculations, demonstrates the MTP’s accuracy in predicting local structures of molten FLiNaK. The extrapolation of the potential to larger systems, achieved using a smaller simulation cell during the ML potential training, underscores its versatility in handling diverse local configurations. While the MTP has a medium-range cutoff distance (up to 5.5 Å in this study) and does not explicitly consider long-range interactions, our results indicate that it performs well in ionic systems. The absence of explicit charge treatment in the MTP, justified by local ionic solvation shells limiting long-range interactions, makes it a useful tool for fast and accurate computations of molten salt’s thermophysical properties. However, certain conditions must be met for optimal results. Our findings suggest that accounting for van der Waals dispersion during the generation of training data improves predictions of the temperature dependence of density. While the potential currently lacks an implicit treatment of long-range interactions, recent research Ko et al. (2021); Gao and Remsing [11] highlights the possible benefits of their explicit inclusion in charge systems. Such features should be included in the next generation of MTPs to improve modeling accuracy.

We also note that utilization of different DFT approximations might improve the results for molten salts calculations with MLIPs. In particular, it was recently shown in Tisi et al. [60] that training of a MLIP on the dataset generated at the DFT-PBE level of accuracy leads to the predicted values of the thermal conductivity of water 60% larger than in experiments. Retraining on the dataset generated with the strongly constrained and appropriately normed semilocal density functional (SCAN) Sun et al. [58, 57] decreases errors by two times. Proper benchmarking of the exchange-correlation density functional influence on the evaluated properties of molten salts is, to the best of our knowledge, currently missing in literature and have to be explored in future works.

Acknowledgements

N.R. and A.S. acknowledge funding from the Russian Science Foundation (Project No. 23-13-00332).

Our analysis, encompassing radial distribution function and diffusivity calculations, demonstrates the MTP’s accuracy in

References

  1. [2]Attarian, S., Morgan, D., Szlufarska, I., 2022. Thermophysical properties of fibre using moment tensor potentials. Journal of Molecular Liquids 368, 120803. URL: https://www.sciencedirect.com/science/article/pii/S0167732220234232X, doi:https://doi.org/10.1016/j.molliq.2022.120803.
  2. [3]Bell, S., Steinberg, T., Will, G., 2019. Corrosion mechanisms in molten salt thermal energy storage for concentrating solar power. Renewable and Sustainable Energy Reviews 114, 109328. URL: https://www.sciencedirect.com/science/article/pii/S1364032119305362, doi:https://doi.org/10.1016/j.rser.2019.109328.
  3. [4]Beneš, O., Konings, R., 2009. Thermodynamic properties and phase diagrams of fluoride salts for nuclear applications. Journal of Fluorine Chemistry 130, 22–29. URL: https://www.sciencedirect.com/science/article/pii/S0022113908002030, doi:https://doi.org/10.1016/j.jfluchem.2008.07.014. fluorine and Nuclear Energy.
  4. [5]Cui, K., Zhao, W., Li, S., Zhou, D., Liu, C., Qu, X., Li, P., 2022. Low-temperature and high-energy-density li-based liquid metal batteries based on licl–kcl molten salt electrolyte. ACS Sustainable Chemistry & Engineering 10, 1871–1879. URL: https://doi.org/10.1021/acssuschemeng.1c07560, doi:10.1021/acssuschemeng.1c07560.
  5. [6]Deringer, V.L., Caro, M.A., Csányi, G., 2019. Machine learning interatomic potentials as emerging tools for materials science. Advanced Materials 31, 1902765. URL: https://onlinelibrary.wiley.com/doi/abs/10.1002/adma.201902765, doi:https://doi.org/10.1002/adma.201902765.
  6. [7]Feng, T., Zhao, J., Liang, W., Lu, G., 2022. Molecular dynamics simulations of lanthanum chloride by deep learning potential. Computational Materials Science 210, 111014. URL: https://www.sciencedirect.com/science/article/pii/S0927025621007011, doi:https://doi.org/10.1016/j.commatsci.2021.111014.
  7. [8]Frandsen, B.A., Nickerson, S.D., Clark, A.D., Solano, A., Baral, R., Williams, J., Neuefeind, J., Memmott, M., 2020. The structure of molten flinak. Journal of Nuclear Materials 537, 152219. URL: https://www.sciencedirect.com/science/article/pii/S0022311520304736, doi:https://doi.org/10.1016/j.jnucmat.2020.152219.
  8. [9]Galashev, A.Y., Rakhmanova, O.R., Abramova, K.A., Katin, K.P., Maslov, M.M., Tkacheva, O.Y., Rudenko, A.V., Kataev, A.A., Zaikov, Y.P., 2023. Molecular dynamics and experimental study of the effect of cef3 and ndf3 additives on the physical properties of flinak. The Journal of Physical Chemistry B 127, 1197–1208. URL: https://doi.org/10.1021/acs.jpcb.2c06915, doi:10.1021/acs.jpcb.2c06915. PMID: 36696698.
  9. [10]Gallagher, R.C., Birri, A., Russell, N.G., Phan, A.T., Gheribi, A.E., 2022. Investigation of the thermal conductivity of molten lif-naf with experiments, theory, and equilibrium molecular dynamics. Journal of Molecular Liquids 361, 119151. URL: https://www.sciencedirect.com/science/article/pii/S0167732222006894, doi:https://doi.org/10.1016/j.molliq.2022.119151.
  10. [11]Gao, A., Remsing, R.C., 2022. Self-consistent determination of long-range electrostatics in neural network potentials. Nature Communications 13, 1572. URL: https://doi.org/10.1038/s41467-022-29243-2, doi:10.1038/s41467-022-29243-2.
  11. [12]Gheribi, A.E., Corradini, D., Dewan, L., Chartrand, P., Simon, C., Madden, P.A., Salanne, M., 2014. Prediction of the thermophysical properties of molten salt fast reactor fuel from first-principles. Molecular Physics 112, 1305–1312. URL: https://doi.org/10.1080/00268976.2014.897396, doi:10.1080/00268976.2014.897396.
  12. [13]Green, M.S., 2004. Markoff Random Processes and the Statistical Mechanics of Time-Dependent Phenomena. II. Irreversible Processes in Fluids. The Journal of Chemical Physics 22, 398–413. URL: https://doi.org/10.1063/1.1740082, doi:10.1063/1.1740082.
  13. [14]Grele, M.D., Gedeon, L., 1953. Forced-convection heat-transfer characteristics of molten sodium hydroxide.
  14. [15]Grimme, S., Antony, J., Ehrlich, S., Krieg, H., 2010. A consistent and accurate ab initio parametrization of density functional dispersion correction (DFT-D) for the 94 elements H-Pu. The Journal of Chemical Physics 132 154104. URL: https://doi.org/10.1063/1.3382344, doi:10.1063/1.3382344.
  15. [16]Hoffman, H., Lones, J., 1955. Fused salt heat transfer. part ii. forced convection heat transfer in circular tubes containing naf-kf-lif eutectic.
  16. [17]Igarashi, K., Okamoto, Y., Mochinaga, J., Ohno, H., 1988. X-ray diffraction study of molten eutectic lif–naf–kf mixture. J. Chem. Soc., Faraday Trans. 1 84, 4407–4415. URL: http://dx.doi.org/10.1039/F19888404407, doi:10.1039/F19888404407.
  17. [18]J. Ambrosek, M. Anderson, K.S. Allen, T., 2009. Current status of knowledge of the fluoride salt (flinak) heat transfer. Nuclear Technology 165, 166–173. URL: https://doi.org/10.13182/NT165–166, doi:10.13182/NT165–166.
  18. [19]Jabbari, F., Rajabpour, A., Saedodin, S., 2017. Thermal conductivity and viscosity of nanofluids: A review of recent molecular dynamics studies. Chemical Engineering Science 174, 67–81. URL: https://www.sciencedirect.com/science/article/pii/S0009250917305432, doi:https://doi.org/10.1016/j.ces.2017.08.034.
  19. [21]Kresse, G., Furthmüller, J., 1996. Efficient iterative schemes for ab initio total-energy calculations using a plane-wave basis set. Phys. Rev. B 54, 11169–11186. URL: https://link.aps.org/doi/10.1103/PhysRevB.54.11169, doi:10.1103/PhysRevB.54.11169.
  20. [22]Kresse, G., Joubert, D., 1999. From ultrasoft pseudopotentials to the projector augmented-wave method. Phys. Rev. B 59, 1758–1775. URL: https://link.aps.org/doi/10.1103/PhysRevB.59.1758, doi:10.1103/PhysRevB.59.1758.
  21. [23]Kubo, R., Yokota, M., Nakajima, S., 1957. Statistical-mechanical theory of irreversible processes. ii. response to thermal disturbance. Journal of the Physical Society of Japan 12, 1203–1211.
  22. [24]Lam, S.T., Li, Q.J., Ballinger, R., Forsberg, C., Li, J., 2021. Modeling lif and flibe molten salts with robust neural network interatomic potential. ACS Applied Materials & Interfaces 13, 24582–24592. URL: https://doi.org/10.1021/acsami.1c00604, doi:10.1021/acsami.1c00604. PMID: 34019760.
  23. [25]LeBlanc, D., 2010. Molten salt reactors: A new beginning for an old idea. Nuclear Engineering and Design 240, 1644–1656. URL: https://www.sciencedirect.com/science/article/pii/S0029549310000191, doi:https://doi.org/10.1016/j.nucengdes.2009.12.033.
  24. [26]Lee, S.C., Zhai, J., Z. Walter, N.P., Rose, M., Heuser, B.J., Z. Y., 2021. Comparative studies of the structural and transport properties of molten salt flinak using the machine-learned neural network and reparametrized classical forcefields. The Journal of Physical Chemistry B 125, 10562–10570. URL: https://doi.org/10.1021/acs.jpcb.1c05608, doi:10.1021/acs.jpcb.1c05608. PMID: 346946525.
  25. [28]Liang, W., Lu, G., 2020. Molecular dynamics simulations of molten magnesium chloride using machine-learning-based deep potential. Advanced Theory and Simulations 3, 2000180. URL: https://onlinelibrary.wiley.com/doi/abs/10.1002/adts.202000180, doi:https://doi.org/10.1002/adts.202000180.
  26. [29]Lizin, A., Tomilin, S., Ponomarev, L.I., Fedorov, Y.S., Hirose, Y., 2017. 12 - fast-spectrum, liquid-fueled reactors, in: Dolan, T.J. (Ed.), Molten Salt Reactors and Thorium Energy. Woodhead Publishing, pp. 375–433. URL: https://www.sciencedirect.com/science/article/pii/B9780081011263000129, doi:https://doi.org/10.1016/B978–0–08–101126–3.00012–9.
  27. [30]Locatelli, G., Mancini, M., Todeschini, N., 2013. Generation iv nuclear reactors: Current status and future prospects. Energy Policy 61, 1503–1520. URL: https://www.sciencedirect.com/science/article/pii/S0301421513006083, doi:https://doi.org/10.1016/j.enpol.2013.06.101.
  28. [32]Magnusson, J., Memmott, M., Munro, T., 2020. Review of thermophysical property methods applied to fueled and un-fueled molten salts. Annals of Nuclear Energy 146, 107608. URL: https://www.sciencedirect.com/science/article/pii/S0306454920303066, doi:https://doi.org/10.1016/j.anucene.2020.107608.
  29. [33]Maslennikova, A.A., Mushnikov, P.N., Dub, A.V., Tkacheva, O.Y., Zaikov, Y.P., Liu, Y.L., Shi, W.Q., 2023. Determination of the oxygen content in the lif-naf-kf melt. Materials 16. URL: https://www.mdpi.com/1996-1944/16/11/4197, doi:10.3390/ma16114197.DOI
  30. [34]Maxwell, C.I., 2022. Molecular dynamics study of fission gas behaviour and solubility in molten flinak salt. Journal of Nuclear Materials 563, 153633. URL: https://www.sciencedirect.com/science/article/pii/S0022311522001295, doi:https://doi.org/10.1016/j.jnucmat.2022.153633.
  31. [36]Müller-Plathe, F., 1997. A simple nonequilibrium molecular dynamics method for calculating the thermal conductivity. The Journal of Chemical Physics 106, 6082–6085. URL: https://doi.org/10.1063/1.473271, doi:10.1063/1.473271.
  32. [37]Nosé, S., 1984. A unified formulation of the constant temperature molecular dynamics methods. The Journal of chemical physics 81, 511–519.DOI
  33. [38]Novikov, I.S., Gubaev, K., Podryabinkin, E.V., Shapeev, A.V., 2020. The mlip package: moment tensor potentials with mpi and active learning. Machine Learning: Science and Technology 2, 025002. URL: https://dx.doi.org/10.1088/2632-2153/abc9fe, doi:10.1088/2632-2153/abc9fe.
  34. [39]Ong, T.C., Sarvghad, M., Lippiatt, K., Griggs, L., Ryan, H., Will, G., Steinberg, T.A., 2020. Review of the solubility, monitoring, and purification of impurities in molten salts for energy storage in concentrated solar power plants. Renewable and Sustainable Energy Reviews 131, 110006. URL: https://www.sciencedirect.com/science/article/pii/S1364032120302975, doi:https://doi.org/10.1016/j.rser.2020.110006.
  35. [40]Pan, G., Ding, J., Chen, P., Yan, H., Du, Y., Lee, D.J., Lu, Y., 2021. Finite-size effects on thermal property predictions of molten salts. Solar Energy Materials and Solar Cells 211, 110884. URL: https://www.sciencedirect.com/science/article/pii/S0927024820304827, doi:https://doi.org/10.1016/j.solmat.2020.110884.
  36. [41]Parasotchenko, Y., Gevel, T., Pavlenko, O.B., Gorshkov, L.V., Leonova, N.M., Suzdaltsev, A.V., Zaikov, Y.P., 2023. Choice of the composition of the chloride melts for the electrochemical synthesis of silicon. Silicon URL: https://api.semanticscholar.org/CorpusID:264982487.DOI
  37. [42]Perdew, J.P., Burke, K., Ernzerhof, M., 1996. Generalized gradient approximation made simple. Phys. Rev. Lett. 77, 3865–3868. URL: https://link.aps.org/doi/10.1103/PhysRevLett.77.3865, doi:10.1103/PhysRevLett.77.3865.DOI
  38. [43]Podryabinkin, E.V., Shapeev, A.V., 2017. Active learning of linearly parametrized interatomic potentials. Computational Materials Science 140, 171–180. URL: https://www.sciencedirect.com/science/article/pii/S0927025617304536, doi:https://doi.org/10.1016/j.commatsci.2017.08.031.
  39. [44]Porter, T., Vaka, M.M., Steenblik, P., Della Corte, D., 2022. Computational methods to simulate molten salt thermophysical properties. Communications Chemistry 5, 69. URL: https://doi.org/10.1038/s42004-022-00684-6, doi:10.1038/s42004-022-00684-6.
  40. [45]Robertson, S.G.I., Wiser, R., Yang, W., Kang, D., Choi, S., Baglietto, E., Short, M.P., 2022. The curious temperature dependence of fluoride molten salt thermal conductivity. Journal of Applied Physics 131.DOI
  41. [46]Rodriguez, A., Lam, S., Hu, M., 2021. Thermodynamic and transport properties of lif and flibe molten salts with deep learning potentials. ACS Applied Materials & Interfaces 13, 55367–55379. URL: https://doi.org/10.1021/acsami.1c17942, doi:10.1021/acsami.1c17942. pMID: 34767334.
  42. [47]Rollet, A.L., Sarou-Kanian, V., Bessada, C., 2009. Measuring self-diffusion coefficients up to 1500 k: A powerful tool to investigate the dynamics and the local structure of inorganic melts. Inorganic Chemistry 48, 10972–DOI
  43. [49]Romatovski, R., Hu, L., 2017. Fluoride salt coolant properties for nuclear reactor applications: A review. Annals of Nuclear Energy 109, 635–647. URL: https://www.sciencedirect.com/science/article/pii/S0306454917301391, doi:https://doi.org/10.1016/j.anucene.2017.05.036.
  44. [50]Rudenko, A., Kataev, A., Tkacheva, O., 2022a. Dynamic viscosity of the naf-kf-ndf3 molten system. Materials 15. URL: https://www.mdpi.com/1996-1944/15/14/4884, doi:10.3390/ma15144884.DOI
  45. [51]Rudenko, A., Redkin, A., Il’ina, E., Pershina, S., Mushnikov, P., Zaikov, Y., Kumkov, S., Liu, Y., Shi, W., 2022b. Thermal conductivity of flinak in a molten state. Materials 15. URL: https://www.mdpi.com/1996-1944/15/16/5603, doi:10.3390/ma15165603.DOI
  46. [52]Salanne, M., Simon, C., Turq, P., Madden, P.A., 2009. Heat-transport properties of molten fluorides: Determination from first-principles. Journal of Fluorine Chemistry 130, 38–44. URL: https://www.sciencedirect.com/science/article/pii/S0022113908002078, doi:https://doi.org/10.1016/j.jfluchem.2008.07.013. fluorine and Nuclear Energy.
  47. [53]Shapeev, A.V., 2016. Moment tensor potentials: A class of systematically improvable interatomic potentials. Multiscale Modeling & Simulation 14, 1153–1173. URL: https://doi.org/10.1137/15M1054183, doi:10.1137/15M1054183.
  48. [54]Sivaramam, G., Krishnamoorthy, A.N., Baur, M., Holm, C., Stan, M., Csányi, G., Benmore, C., Vázquez-Mayagoitia, Á., 2020. Machine-learned interatomic potentials by active learning: amorphous and liquid hafnium dioxide. npj Computational Materials 6, 104. URL: https://doi.org/10.1038/s41524-020-00367-7, doi:10.1038/s41524-020-00367-7.
  49. [55]Smirnov, M., Khokhlov, V., Filatov, E., 1987. Thermal conductivity of molten alkali halides and their mixtures. Electrochimica Acta 32, 1019–1026. URL: https://www.sciencedirect.com/science/article/pii/001346687900272, doi:https://doi.org/10.1016/0013-4686(87)90027-2.
  50. [56]Sohal, M., Ebner, M., Sabharwall, P., Sharpe, P., 2010. Engineering database of liquid salt thermophysical and thermochemical properties. Idaho Natl. Lab. Idaho Falls CrossRef doi:10.2172/980801.DOI
  51. [57]Sun, J., Remsing, R.C., Zhang, Y., Sun, Z., Ruzsinszky, A., Peng, H., Yang, Z., Paul, A., Waghmare, U., Wu, X., Klein, M.L., Perdew, J.P., 2016. Accurate first-principles structures and energies of diversely bonded systems from an efficient density functional. Nature Chemistry 8, 831–836. URL: https://doi.org/10.1038/nchem.2535, doi:10.1038/nchem.2535.
  52. [58]Sun, J., Ruzsinszky, A., Perdew, J.P., 2015. Strongly constrained and appropriately normed semilocal density functional. Phys. Rev. Lett. 115, 036402. URL: https://link.aps.org/doi/10.1103/PhysRevLett.115.036402, doi:10.1103/PhysRevLett.115.036402.DOI
  53. [59]Thompson, A.P., Aktulga, H.M., Berger, R., Bolintineanu, D.S., Brown, W.M., Crozier, P.S., in ’t Veld, P.J., Kohlmeyer, A., Moore, S.G., Nguyen, T.D., Shan, R., Stevens, M.J., Tranchida, J., Trott, C., Plimpton, S.J., 2022. LAMMPS - a flexible simulation tool for particle-based materials modeling at the atomic, meso, and continuum scales. Computer Physics Communications 271, 108171. URL: https://www.sciencedirect.com/science/article/pii/S0010465521002836, doi:https://doi.org/10.1016/j.cpc.2021.108171.
  54. [60]Tisi, D., Zhang, L., Bertossa, R., Wang, H., Car, R., Baroni, S., 2021. Heat transport in liquid water from first-principles and deep neural network simulations. Physical Review B 104, 242402.DOI
  55. [61]Umesaki, N., Iwamoto, N., Tsunawaki, Y., Ohno, H., Furukawa, K., 1981. Self-diffusion of lithium, sodium, potassium and fluorine in a molten lif + naf + kf eutectic mixture. J. Chem. Soc., Faraday Trans. 1 77, 169–175. URL: http://dx.doi.org/10.1039/F19817700169, doi:10.1039/F19817700169.
  56. [62]Vriesema, B., 1979. Aspects of molten fluorides as heat transfer agents for power generation.
  57. [63]W. D. Powers, S.I.C., Greene, N.D., 1963. Physical properties of molten reactor fuels and coolants. Nuclear Science and Engineering 17, 200–211. URL: https://doi.org/10.13182/NSE63-5, doi:10.13182/NSE63-5.
  58. [64]Williams, D., 2006. Assessment of candidate molten salt coolants for the ngnp/nhi heat-transfer loop.DOI
  59. [65]Zhang, H., Lin, D.Y., Wang, H., Car, R. E. W., 2019. Active learning of uniformly accurate interatomic potentials for materials simulation. Phys. Rev. Mater. 3, 023804. URL: https://link.aps.org/doi/10.1103/ PhysRevMaterials.3.023804, doi:10.1103/PhysRevMaterials.3.023804.DOI
  60. [66]Zhitkov, A., Potapov, A., Karimov, K., Kholkina, A., Shishkin, V., Dedyukhin, A., Zaykov, Y., 2022. Interaction between un and cdcl2 in molten licl–kcl eutectic. ii. experiment at 1023 k. Nuclear Engineering and Technology 54, 653–660. URL: https://www.sciencedirect.com/science/article/pii/S1738573321005192, doi:https://doi.org/10.1016/j.net.2021.08.030.
  61. [67]Zhitkov, A., Potapov, A., Karimov, K., Shishkin, V., Dedyukhin, A., Zaykov, Y., 2020. Interaction between un and cdcl2 in molten licl-kcl eutectic. i. experiment at 773 k. Nuclear Engineering and Technology 52, 123–134. URL: https://www.sciencedirect.com/science/article/pii/S1738573319301780, doi:https://doi.org/10.1016/j.net.2019.07.006.
  62. [68]Zuo, Y., Chen, C., Li, X., Deng, Z., Chen, Y., Behler, J., Csányi, G., Shapeev, A.V., Thompson, A.P., Wood, M.A., Ong, S.P., 2020. Performance and cost assessment of machine learning interatomic potentials. The Journal of Physical Chemistry A 124, 731–745. URL: https://doi.org/10.1021/acs.jpca.9b08723, doi:10.1021/acs.jpca.9b08723. pMID: 31916773.

Paper details

vlnc:W_n4d6fecr
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents