[Paper Review] Machine-Learned Interatomic Potentials for Predicting Physicochemical Properties of Molten Metal-Salt Systems for Calcium Electrolysis

M. Polovinkin, N. Rybin+6 more|arXiv (Cornell University)|Mar 26, 2026

Molten salt chemistry and electrochemical processes0 citations

TL;DR

The paper develops and benchmarks Moment Tensor Potentials (MTPs) for Ca-Cu molten alloy and CaCl2-KCl molten salt, enabling MD predictions of structural, thermodynamic, and transport properties with good agreement to experiments. It demonstrates compositional transferability and provides a framework for electrolysis-relevant property predictions.

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.

Motivation & Objective

• Address data gaps for molten Ca-Cu alloy and CaCl2-KCl electrolyte relevant to calcium electrolysis
• Develop compositionally transferable MTPs trained on high-accuracy DFT data
• Validate MTP-MD predictions against experimental measurements across temperatures and compositions
• Provide a computational framework to predict densities, RDFs, heat capacities, viscosities, diffusion coefficients, and ionic conductivities

Proposed method

• Construct diverse training sets from DFT (VASP, PBE, DFT-D3 for Ca-Cu and dDsC for CaCl2-KCl) to fit Moment Tensor Potentials (MTPs)
• Train compositionally transferable MTP for Ca-Cu across all Ca molar fractions (0–1) and a separate MTP for CaCl2-KCl (80:20 mass%, 28:10 molar)
• Use active learning with NPT MD to enrich training data and ensure stability, then fit higher-level potentials
• Run MD with LAMMPS using MTPs to compute densities, RDFs, Cp, viscosities, diffusion, ionic conductivity; employ Green-Kubo and Nernst-Einstein formalisms for transport properties
• Benchmark MTP accuracy against DFT with RMSE ~5 meV/atom for energies and ~80–136 meV/Å for forces
• Validate transport and thermodynamic properties against experimental data and literature

Experimental results

Research questions

• RQ1 Can MTP-trained MD reproduce key structural and thermodynamic properties of molten Ca-Cu across compositions?
• RQ2 Can MTPs transfer across compositions for the Ca-Cu alloy and accurately predict CaCl2-KCl electrolyte properties?
• RQ3 Do MTP-driven MD predictions for density, RDFs, Cp, viscosity, diffusion, and ionic conductivity align with experimental data within acceptable errors?
• RQ4 Is the Green-Kubo approach for ionic conductivity more reliable than the Nernst-Einstein approximation for CaCl2-KCl in this system?
• RQ5 How do predicted transport properties vary with temperature and composition in these calcium electrolysis-relevant melts?

Key findings

• MTPs achieve ~5 meV/atom energy errors and 80–136 meV/Å force RMSE on validation sets
• Ca-Cu density matches literature within ~3–4% for 0.5–0.8 Ca molar fractions; pure Ca and Cu densities reproduce within ~3% and ~10% respectively at 1400 K
• Ca-Cu Cp increases nonlinearly with Ca content; literature Cp data for Ca-Cu alloys is unavailable, but MTP results lie between pure Ca and Cu values
• Viscosity across Ca-Cu compositions agrees with literature, resolving inconsistencies and supporting SES stability with a best-fit b ≈ 3.7±0.4 in the Stokes–Einstein relation
• CaCl2-KCl density matches experimental data within 2–3%; Cp ~980 J/(kg K) vs ~960 J/(kg K) experimental; thermal conductivity ~0.438 W/mK at 1000 K; viscosity and diffusion trends align with experiments
• Ionic conductivity from Green-Kubo agrees with experiments within 6–20% relative error, while Nernst-Einstein underpredicts conductivity and may misrepresent temperature dependence