Predicting Thermophysical Properties of Halide Salts Using On-the-Fly Machine-Learned Interatomic Potentials Across Multiple Exchange-Correlation Functionals
- 1Department of Chemical Engineering, Indian Institute of Technology Bombay, Powai, Mumbai, Maharashtra, India, 400076
Abstract
In silico modeling of molten salts is pivotal for developing emerging carbon-free energy technologies. Accurate prediction of their thermophysical properties using first-principles methods remains computationally intractable owing to the large length and timescales required. Further, the choice of exchange-correlation (XC) functional critically governs the predictive accuracy of thermophysical properties. Here, we train machine learning force fields to investigate the thermophysical properties of NaBr, KCl, and AgCl, representing alkali- and transition-metal halides, across their solid and liquid phases, systematically benchmarking ten XC functionals for NaBr (600-1400 K) and testing the transferability of the best-performing functional to KCl and AgCl. These MLIPs enable prediction of densities and ionic diffusion coefficients at length and timescales beyond ab initio simulations, while capturing the first-order solid-liquid transition. We find that BEEF-vdW accurately predicts the optimized lattice constants of NaBr and KCl, consequently reproducing experimental solid and liquid densities, with Na+ and Br- self-diffusion coefficients agreeing with experiments to within ~5%. However, we find that the BEEF-vdW functional inaccurately predicts AgCl density owing to its inaccurate lattice constant, whereas PBE-D3 yields excellent density agreement (~5% deviation) across both phases. This work establishes a foundation for high-throughput generation of molten-salt thermophysical databases using general machine learning potentials.0
Related
- Comparative Assessment of Exchange-Correlation Functionals for the Structural, Thermodynamic, and Transport Properties of Liquid Lithium
- Orbital-Free Surrogate Functionals Yield Transferable Interatomic Potentials and Electron Densities
- Thymol-based deep eutectic solvents: a comprehensive overview of classifications, preparation methods, stability, and thermophysical properties.
- Stationarity Constraints Improve Both Energies and Densities of Machine-Learned Exchange-Correlation Functionals
- Machine-learned exchange-correlation functionals for molecules, solids, and reactive surfaces
BibTeX
@misc{rudra2026predicting,
title = {{Predicting Thermophysical Properties of Halide Salts Using On-the-Fly Machine-Learned Interatomic Potentials Across Multiple Exchange-Correlation Functionals}},
author = {Pritam Rudra and Sudarshan Vijay and Ojus Mohan and PRITAM RUDRA},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008809/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008809/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Predicting Thermophysical Properties of Halide Salts Using On-the-Fly Machine-Learned Interatomic Potentials Across Multiple Exchange-Correlation Functionals
AU - Pritam Rudra
AU - Sudarshan Vijay
AU - Ojus Mohan
AU - PRITAM RUDRA
PY - 2026
DA - 2026/09/14
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008809/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008809/v1
AB - In silico modeling of molten salts is pivotal for developing emerging carbon-free energy technologies. Accurate prediction of their thermophysical properties using first-principles methods remains computationally intractable owing to the large length and timescales required. Further, the choice of exchange-correlation (XC) functional critically governs the predictive accuracy of thermophysical properties. Here, we train machine learning force fields to investigate the thermophysical properties of NaBr, KCl, and AgCl, representing alkali- and transition-metal halides, across their solid and liquid phases, systematically benchmarking ten XC functionals for NaBr (600-1400 K) and testing the transferability of the best-performing functional to KCl and AgCl. These MLIPs enable prediction of densities and ionic diffusion coefficients at length and timescales beyond ab initio simulations, while capturing the first-order solid-liquid transition. We find that BEEF-vdW accurately predicts the optimized lattice constants of NaBr and KCl, consequently reproducing experimental solid and liquid densities, with Na+ and Br- self-diffusion coefficients agreeing with experiments to within ~5%. However, we find that the BEEF-vdW functional inaccurately predicts AgCl density owing to its inaccurate lattice constant, whereas PBE-D3 yields excellent density agreement (~5% deviation) across both phases. This work establishes a foundation for high-throughput generation of molten-salt thermophysical databases using general machine learning potentials.0
ER -