Abstract
Understanding the chemical reactions and thermodynamic properties of molten salts and their aqueous solutions is crucial for designing their technological and environmental applications in the energy industry. While quantum chemical methods can describe these systems accurately, their high computational cost limits the size of systems and time scales. Machine learning models trained on ab initio data have been applied in large-scale molecular dynamics simulations, achieving similar accuracy with greater efficiency.

In this dissertation, machine learning models have been developed to model the chemical reactions in lithium carbonates (Li2CO3) and hydroxides (LiOH) at elevated temperatures. Initially, a machine learning model was generated for Li2CO3 using ab initio data for both vapor and liquid phases. Reactive vapor-liquid equilibria of Li2CO3 were studied using direct coexistence simulations, and the observed partial pressure of CO2 closely matched experiments. After demonstrating that machine learning-based simulations can efficiently describe reactive multiphase behavior, the same methodology was applied to LiOH. The new model revealed that decomposition produces H2O dissolved in the liquid phase, with some vaporized in the presence of the vapor phase. Comparison with the underlying ab initio method showed good agreement in terms of reaction kinetics and equilibrium liquid properties. Additionally, a general model for both Li2CO3 and LiOH was trained using data from the previous models as well as new data on mixtures.  This model was used to study decomposition reactions in pure and mixed systems and showed that LiOH can react with CO2 producing Li2CO3 and H2O. After demonstrating the ability of machine learning models to accurately predict thermodynamic and kinetic properties of fluids, a new model was generated for aqueous NaCl solutions to perform more challenging calculations — chemical potentials. Different methods were evaluated and modified to calculate chemical potentials of alkali halides in water with machine
learning potentials.

The studies in this dissertation demonstrate how a machine learning model can be developed to study reactive multiphase equilibria of molten salts, combining the accuracy of ab initio methods with the efficiency of classical molecular dynamics. The methodology described can be extended to other fluids and materials, particularly those with limited experimental data