Development of molten salt–based processes through thermodynamic evaluation assisted by machine learning

Abstract

Molten salt–based processes and hydrofluxes are highly sensitive to mixture composition and require knowledge of the combined melting point for successful materials syntheses. In particular processes using hydroxide–based fluxes (pure salt melts) and hydrofluxes (salt melts containing 15–50% H${}_{\text{2}}$O) have been shown to be interesting environments to synthesize inorganic materials in high oxidation states. The development of tools to predict these properties is desirable to inform the implementation of processes using these mixtures. In this work, we use an artificial neural network model to estimate the melting points of fluxes and hydrofluxes comprising of quaternary mixtures of NaOH, KOH, LiOH, and H₂O. A database of 1644 data points collected from 47 different sources was used in the training of the model. Melting points were predicted from the molar fractions of each component (4 independent variables). After training, the ANN model was able to approximate the melting points of the mixture with an $R^2$ of 0.996 for most conditions. Except for a region defined by the range 0.08 $\lesssim {\Phi }_{\text{LiOH}}\lesssim$ 0.14 and ${\Phi }_{\text{H}{}_2\text{O}}\lesssim$ 0.85, where the liquidus surface was multi–valued, preventing accurate representation by the ANN. The model was able to qualitatively recreate the binary curves and ternary liquidus surfaces of these mixtures with a root mean squared error of 6.1 °C (Full range −65 – 477 °C). In the future, this model could be used to aid the synthesis of materials in the quaternary mixtures investigated in this work.