Mining the intrinsic trends of CO2 solubility in blended solutions

CO2 solubility in trisodium phosphate (TSP) and its mixed solutions is a crucial information for CO2 absorption and utilization. However, with limited experimental data and large variations of experimental conditions, intrinsic trends of CO2 solubility under a specific set of conditions are difficult to be determined without comprehensive experiments. To address this, here, a machine learning based data- mining is proven a powerful method to explore the intrinsic trends of CO2 solubility trained from 299 data groups extracted from previous experimental literatures. A generalized machine learning input representation method was applied, for the first time, by quantifying the types and concentrations of the blended solutions. With a general regression neural network (GRNN) as the algorithm, we found that the intrinsic trends of CO2 solubility could be well- fitted with a limited amount of experimental data, having the average root mean square error (RMSE) lower than 0.038 mol CO2/mol solution. More importantly, it is shown that with a generalized input representation, machine learning can mine the relationships between CO2 solubility and various experimental conditions, which could help to better understand the intrinsic trends of CO2 solubility in blended solutions.