Compositional transferability of deep learning potentials: a case study for LiCl-KCl melt.

Context: One of the crucial issues related to machine learning potentials is the formation of representative dataset. For multicomponent systems, it is a general methodology to scan the composition range with a certain step. However, there is a lack of information on the compositional transferability of machine learning potentials. In this paper, we extend the knowledge in this area by studying the transferability of deep learning potential over the range of compositions of LiCl-KCl molten mixtures. The training dataset was formed using only the near-eutectic composition of 60% LiCl-40% KCl. Then, we tested the ability of the model to predict physicochemical properties of the melts far from the reference composition. It was found that for the composition range from 0 to 100% of LiCl, the calculated properties concur closely with those of other studies and ab initio calculations. Therefore, the model shows prominent non-intuitive compositional transferability. Moreover, the solid states and solid-liquid coexistence were reproduced. The calculated melting temperatures of LiCl and KCl show the errors of 6.6% and 0.4%, respectively. We argue that such good transferability stems from the local structure configurations that are typical both for pure LiCl and for pure KCl which are implicitly presented in the training dataset because of local fluctuations in composition.

Methods: To collect the data for the initial dataset, density functional theory was employed. Then, the DeePMD package was used to train a neural network potential. To calculate the properties of the melts, standard equilibrium and non-equilibrium molecular dynamic approaches were utilized.