Prediction of Henry's law constants by matrix completion

Methods for predicting Henry's law constants H_ij are important as experimental data are scarce. We introduce a new machine learning approach for such predictions: matrix completion methods (MCMs) and demonstrate its applicability using a data base that contains experimental H_ij values for 101 solutes i and 247 solvents j at 298 K. Data on H_ij are only available for 2661 systems i + j. These H_ij are stored in a 101 × 247 matrix; the task of the MCM is to predict the missing entries. First, an entirely data-driven MCM is presented. Its predictive performance, evaluated using leave-one-out analysis, is similar to that of the Predictive Soave-Redlich-Kwong equation-of-state (PSRK-EoS), which, however, cannot be applied to all studied systems. Furthermore, a hybrid of MCM and PSRK-EoS is developed in a Bayesian framework, which yields an unprecedented performance for the prediction of H_ij of the studied data set.