Prediction of excess enthalpy in binary mixtures through probabilistic matrix completion, with a GP enforced smoothness constraint

Matrix completion (MC) methods have found success in application to the prediction of scalar valued thermophysical properties, and more recently for mixture data in our own group. In this work we employ MC using the probabilistic matrix factorization (PMF) framework to generate pseudo excess enthalpy data, H^E, for binary mixtures using data from established thermophysical property databases. We employ Gaussian processes (GP) to enforce smoothness of the pseudo excess enthalpy data across both composition and temperature. The equivalent kernel for the GP was derived from a modified version of the Redlich-Kister polynomial. We incorporate several thermodynamic considerations to improve the accuracy, robustness and interpretability of the estimates obtained, which we propose as surrogates in the absence of experimental data for fundamental thermodynamic model improvement.