Assessing the Performance of a Deep Neural Network Potential for Salt Hydrates

Salt hydrates make up an important class of thermochemical materials (TCMs) with advantageous attributes for heat storage applications. Their properties vary significantly across different hydration states and undergo changes during hydration/dehydration cycles. Therefore, accurate models are essential for predicting key thermodynamic properties such as enthalpies and heat capacities as well as understanding the interplay between ionic crystals and water that governs these properties. Traditional force fields, including classical and reactive models, often fail to accurately describe TCMs due to cumbersome parametrization and limited transferability. To overcome these limitations, machine-learning-based potentials such as DeepMD offer a powerful alternative, enabling the study of complex salt hydrate systems with a high predictive accuracy. In this study, we investigated CaCl 2 · n H 2 O ( n = 0, 1/3, 1, 2, 4 (β, γ), and 6), MgCl 2 · n H 2 O ( n = 0, 1, 2, 4, and 6), and K 2 CO 3 · n H 2 O ( n = 0 and 1.5) using the DeepMD method to evaluate their fundamental properties. Our findings show that DeepMD achieves an average accuracy of over 80% when compared with DFT calculations and experimental data. The model accurately predicts hydration enthalpies and heat capacities for CaCl 2 hydrates, with deviations within 10% of experimental values. Additionally, it effectively reproduces lattice constants, heat capacities, and radial distribution functions across various hydration states. For MgCl 2 and K 2 CO 3 hydrates, hydration enthalpy predictions maintain a high accuracy of ∼80% (error ∼ 20%) and 76% (error ∼ 24%), respectively. A limited benchmark against a universal foundation model (MACE) indicates that system-specific training is necessary to achieve stable and accurate descriptions of salt-hydrate thermodynamics. These results demonstrate that DeepMD provides a reliable and efficient framework for studying salt hydrates, and its ability to capture complex hydration behavior makes it a valuable tool for designing and optimizing thermochemical materials for energy storage applications.