Machine-Learned Interatomic Potentials for Predicting Physicochemical Properties of Molten Metal–Salt Systems for Calcium Electrolysis. Journal: Electrochimica Acta
M. Polovinkin, N. Rybin, D. Maksimov, F. Valiev, A. Khudorozhkova, M. Laptev, A. Rudenko, A. Shapeev · 2026
Read Machine-Learned Interatomic Potentials for Predicting Physicochemical Properties of Molten Metal–Salt Systems for Calcium Electrolysis. Journal: Electrochimica Acta by M. Polovinkin, N. Rybin, D. Maksimov, F. Valiev, A. Khudorozhkova, M. Laptev, A. Rudenko, A. Shapeev (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-25330.txt.
Introduction
Polovinkin and co-authors predict the traits of molten metal-salt mixes for calcium making, using learned atom models. Hot experiments on such melts are slow and costly. The team trains Moment Tensor Potentials on quantum-grade data, then runs molecular dynamics with them. Two systems matter for calcium electrolysis: a calcium-copper melt and a chloride salt mix. The runs reproduce density, heat capacity, conductivity, and flow traits across heat and mix ranges. Errors stay within twenty percent of lab values, which validates the approach.
Catalog record
| CREATOR | M. Polovinkin, N. Rybin, D. Maksimov, F. Valiev, A. Khudorozhkova, M. Laptev, A. Rudenko, A. Shapeev |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 41261 |
| LICENSE | http://creativecommons.org/licenses/by-sa/4.0/ |
| RIGHTS | CC BY SA |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.25330
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Machine-Learned Interatomic Potentials for Predicting Physicochemical Properties of Molten Metal–Salt Systems for Calcium Electrolysis. Journal: Electrochimica Acta M. Polovinkin, N. Rybin, D. Maksimov, F. Valiev, A. Khudorozhkova, M. Laptev, A. Rudenko, A. Shapeev Abstract The design of efficient electrolysis devices for pure metal production requires accurate data on the properties of the melts used in the process. This work focuses on two key systems for calcium production: the molten Ca-Cu alloy and the CaCl 2 -KCl electrolyte. High-temperature experiments are often expensive and time-consuming; however, we demonstrate that molecular dynamics (MD) simulations driven by machine-learned Moment Tensor Potentials (MTPs), trained on highly accurate density functional theory data, offer an effective and accurate alternative. Our MTP-driven MD simulations accurately reproduce the structural, thermodynamic, and transport properties across a range of temperatures and compositions relevant to electrolysis systems. We report calculated densities, radial distribution functions, heat capacities, thermal conductivities, ionic conductivities (for the electrolyte), viscosities, and diffusion coefficients, with deviations from experimental data within 20%. The strong agreement between calculations and experiments validates the proposed approach, establishing a robust framework for the computational exploration and optimization of liquid systems in metallurgical applications. PACS 0000 , 1111 MSC 0000 , 1111 † † highlights: Moment Tensor Potentials for the molten Ca-Cu alloy and the CaCl 2 -KCl molten salt are developed for the first time The Moment Tensor Potential for the Ca-Cu alloy is compositionally transferable The same potentials are used to calculate structural, thermodynamic, and transport properties of the systems The ionic conductivity of the CaCl 2 -KCl molten salt predicted by Moment Tensor Potential molecular dynamics agrees with experiment 1 Introduction Metallic calcium is a critical element with a wide range of industrial applications, including steelmaking, battery production, and the synthesis of advanced alloys and magnetic materials [ 33 , 18 , 55 , 15 ] . As demand grows, improving calcium production methods is becoming increasingly important. The primary industrial route for producing high-purity calcium is molten salt electrolysis. This process involves two high-temperature liquid phases: a molten Ca-Cu alloy acting as a liquid cathode and a CaCl 2 -based molten salt electrolyte [ 54 ] . Optimizing the efficiency of production processes is a key industrial objective. Multiphysics modeling using digital twins can greatly assist in these efforts [ 50 ] . However, digital twins require precise knowledge of the structural, thermodynamic, and transport properties of both the electrolyte and the liquid cathode under operating conditions. This includes the temperature dependence of key properties such as density, viscosity, diffusion coefficients, thermal conductivity, heat capacity, and ionic conductivity. To the best of our knowledge, a gap exists in the available data for several physicochemical properties of the liquid Ca-Cu alloy and CaCl 2 -KCl molten salt. The temperature dependence of density has been measured for several compositions of molten Ca-Cu alloys [ 17 , 55 , 3 ] . In contrast, structural investigations have been limited to pure Ca and Cu melts [ 51 , 52 ] , leaving a gap in our understanding of the alloy structure. The heat capacity of Ca-Cu alloys remains unknown, and reported diffusion coefficients and viscosities require refinement, as they are available from only a single source [ 55 ] and contain inconsistencies. While the properties of the CaCl 2 -KCl molten salts have been studied for other compositions [ 22 , 45 , 6 ] , data for the specific 80:20 mass% ratio used in production are lacking. Acquiring these data through high-temperature experiments is often resource-intensive, expensive, and challenging, making such approaches impractical for the large-scale parameter screening—across composition, temperature, and pressure—that industrial optimization demands. Computational approaches offer a viable alternative to experiments [ 40 ] . While ab initio molecular dynamics (AIMD) can predict physicochemical properties, its computational cost is prohibitively high for sufficient configuration sampling and for the calculation of transport properties. Molecular dynamics (MD) modeling with machine-learning interatomic potentials (MLIPs), including Moment Tensor Potentials (MTPs), offers a favorable trade-off, providing accurate modeling at a fraction of the computational cost [ 56 , 20 ] . In this article, we develop and benchmark a computational methodology employing MTPs to fill data gaps for these systems. First, we construct diverse training sets from density functional theory (DFT) calculations to fit reliable potentials. We then employ MTP-MD to compute a comprehensive set of physicochemical properties for both systems involved in the electrolytic production of calcium. For the Ca-Cu alloy, we address notable gaps in the literature by reporting the composition dependence of heat capacity and by refining inconsistencies in the available viscosity and diffusion data. For the CaCl 2 -KCl electrolyte, we provide a complete set of properties, including ionic conductivity. To support our computational results, we conducted a series of experiments measuring the density, viscosity, and ionic conductivity of the molten salt. Combined with literature data, these results enable a proper benchmark. The computational framework established here is both general and transferable. It can be reliably applied to calculate the physicochemical properties of other molten salts and liquid alloys, providing a powerful tool for the design of new materials and the optimization of production processes. 2 Methods 2.1 MTP fitting To calculate the physicochemical properties, we conducted the MD simulation in LAMMPS [ 37 , 48 ] . The MTPs, implemented in MLIP-2 [ 35 ] , were used as the interaction potentials. In the MTP, the energy of an atomic configuration is represented as a sum of energies of local atomic environments. These energies are expanded through a set of basis functions, with linear coefficients determined by fitting energies, forces, and stresses to reference ones. For a comprehensive theoretical description of the MTPs, we refer the reader to the original publications [ 43 , 35 , 38 ] . For the systems of the Ca-Cu melt (for all compositions) and the CaCl 2 -KCl (80:20 mass%, 28:10 in molar fractions) electrolyte, we developed potentials containing 127 and 343 parameters, respectively. The potential for Ca-Cu liquid alloy was designed to be compositionally transferable. In the MTPs we used the Chebyshev radial basis and a cutoff distance of 5 \mathrm{\text{Å}} . Potentials were trained on atomic configurations sampled from the liquid bulk phase, with the simulation cell size ensuring a minimum length of twice the cutoff distance. The MTPs were trained in a multi-stage procedure developed to ensure both accuracy and stability in MD modeling. First, for each system, an initial training set was generated from a short 2 ps AIMD simulation. We sampled 150 evenly spaced atomic configurations from the last 1.5 ps of the trajectory. This set of configurations was used to train an initial MTP of level 10. To enhance the stability of the potential in MD, we employed an active learning procedure [ 39 ] . It consists of performing 100 ps isothermal-isobaric ensemble (NPT) MD modeling, during which atomic configurations with the highest extrapolation grade are automatically added to the training set. The initial AIMD produced atomic configurations that were structurally similar. To construct diverse and representative training sets, we performed a "resampling" step using a stable potential. For the Ca-Cu alloy, to ensure transferability across compositions, six separate 60 ps trajectories were run, with the calcium