Microstructure and Properties of Carbonate Salt Based on Machine Learning

Molten alkali carbonate are widely concerned as potential thermal storage and transfer materials in solar power utilization. As an effective method in molten salt research, computer simulations have been widely used. Local structure and some physical properties of K₂CO₃ and Na₂CO₃ at different temperatures were calculated by a complex simulation method including first-principle molecular dynamics, machine learning and classical molecular dynamics. In this method, first-principle molecular dynamics offered accuracy structure information, machine learning was used to create deep potential from structure information to describe potential energy of system, and classical molecular dynamics was used to perform large scale simulation. This complex method could reduce calculation errors caused by potential functions and empirical parameters. The calculation results showed that energy and force test errors of K₂CO₃ were 8.62×10⁻⁴ eV and 4.67×10⁸ eV/m, respectively; energy and force test errors of Na₂CO₃ were 1.19×10⁻³ eV and 5.31×10⁸ eV/m, respectively. In all simulation processes, the carbonate ion was a standard equilateral triangle structure in the system, and carbonate clusters were slightly loosened with the increase of temperature; the distance between anions and cations gradually increased with the increase of temperature. The result showed that simulated data of the property was in good agreement with the literature value. The deviations of density, specific heat capacity and thermal conductivity of K₂CO₃ were about 5.0%, 3.3% and 8.0%. The deviations of density, specific heat capacity and thermal conductivity of Na₂CO₃ were about 5.6%, 6.0% and 3.5%.