(310b) Development of Ab-Initio-Based Machine Learning Models to Study Chemical Reactions in Molten Alkali Carbonates and Hydroxides

Alkali-metal carbonates and hydroxides are widely used in energy and environmental applications due to their appealing properties. At high temperatures, pure carbonates and hydroxides decompose into vapor carbon dioxide and water, respectively. Moreover, experiments show that hydroxides can capture carbon dioxide by producing carbonates and water. Modeling these chemical reactions in the pure and mixture systems of molten carbonates and hydroxides can be crucial in better understanding their reaction mechanisms and designing their applications. Chemical reactions can be accurately predicted using ab initio molecular dynamics (AIMD) simulations, but high computational costs demanding small system sizes and time scales make this approach not feasible for our study. Classical molecular dynamics (MD) simulations, on the other hand, can efficiently predict system properties at different conditions using large system sizes and long time scales. However, they are not able to model chemical reactions in the systems except via detailed, predefined reaction mechanisms making the observation of new reactions and unforeseen products challenging. Recent advances in machine learning technologies can help overcome these challenges by training deep neural networks on quantum chemical data. The machine-learning models have shown good results in retaining the accuracy of the underlying ab initio methods while being comparable in efficiency to classical MD simulations. In the current work, we generate machine-learning models based on the AIMD simulations for pure lithium carbonates and hydroxides as well as their mixtures to study chemical reactions in liquid phase and vapor-liquid equilibria. We perform direct coexistence simulations to analyze the reactive multiphase behavior of systems at different conditions. We evaluate our results in terms of system composition, equilibrium constants, and partial pressures of vapor components. Additionally, we study system and volume size effects on the obtained results and analyze how the addition of new components affects the equilibrium constants. In general, our machine learning model predictions are in good agreement with available experimental results.