Abstract Molten eutectic salts consisting of ZnCl2 and other alkali chlorides are promising thermal storage and heat transfer fluid materials in the next generation concentrated solar thermal power. To go deep into the thermal and transport properties for a high order mixture, the microstructure information, as well as thermodynamics properties of individual components, have to be identified first. This work develops interatomic potentials of molten ZnCl2 based on neural-network machine learning approach for the first time. The machine learning potential is trained by fitting to the energies and forces of liquid structures ab initio molecular dynamics calculations. The developed machine learning potential is validated by comparing partial radial distribution functions, coordination numbers, and partial structure factors with AIMD and PIM potential. The machine learning potential yields a more precise description of the microstructures than the PIM potential which suffers from the analytical form. Furthermore, structural and thermophysical evolution with temperature are studied and the results are in good agreement with experimental values. The efficient machine learning potential with DFT accuracy from our study will provide a promising scheme for accurate molecular simulations of structures and dynamics of molten ZnCl2 mixtures.

中文翻译：

---

熔融 ZnCl2 及其混合物的 DFT 精确机器学习描述：1. 熔融 ZnCl2 的潜在发展和性质预测

摘要 由ZnCl2和其他碱金属氯化物组成的熔融共晶盐是下一代聚光太阳能热发电中很有前景的蓄热和传热流体材料。为了深入了解高阶混合物的热和传输特性，必须首先确定各个组分的微观结构信息以及热力学特性。这项工作首次基于神经网络机器学习方法开发了熔融 ZnCl2 的原子间势。通过从头算分子动力学计算拟合液体结构的能量和力来训练机器学习潜力。通过将部分径向分布函数、配位数和部分结构因子与 AIMD 和 PIM 潜力进行比较，验证了开发的机器学习潜力。与受分析形式影响的 PIM 势能相比，机器学习势能对微观结构产生更精确的描述。此外，研究了结构和热物理随温度的演变，结果与实验值非常吻合。我们研究中具有 DFT 精度的高效机器学习潜力将为熔融 ZnCl2 混合物的结构和动力学的精确分子模拟提供一个有前途的方案。