Prediction of Temperature-Dependent Henry’s Law Constants by Matrix Completion

Methods for predicting Henry's law constants <i>H</i>_<i>ij</i> describing the solubility of solutes <i>i</i> in solvents <i>j</i> as a function of temperature are essential in chemical engineering. While isothermal properties of binary mixtures can conveniently be predicted with matrix completion methods (MCMs) from machine learning, we advance their application to the temperature-dependent prediction of <i>H</i>_<i>ij</i> in the present work by combining them with physical equations describing the temperature dependence. For training the methods, experimental <i>H</i>_<i>ij</i> data for 122 solutes and 399 solvents ranging from 173.15 to 573.15 K were taken from the Dortmund Data Bank. Two MCMs are proposed: a data-driven MCM that relies solely on experimental data and a hybrid MCM that incorporates predictions from the established Predictive Soave-Redlich-Kwong (PSRK) equation of state (EoS), effectively combining physical knowledge and machine learning. The performance of these MCMs is assessed via leave-one-out analysis and compared to that of the PSRK-EoS, demonstrating superior prediction accuracy.