Matrix Completion Methods for the Prediction of Thermodynamic Properties of Mixtures

The accurate prediction of thermodynamic properties is pivotal for chemical engineering as experimental data are scarce. While established physics-based methods face limitations in prediction accuracy and scope, emerging machine learning approaches, such as matrix completion methods (MCMs), offer promising alternatives. MCMs exploit the fact that experimental data for binary mixtures can be represented as elements of a sparse matrix, with rows and columns corresponding to the components that make up the mixture. Hence, MCMs can be used for closing the gaps in these matrices. In the present thesis, new methods for predicting thermodynamic properties of mixtures are developed that combine probabilistic MCMs with established physical methods. The resulting hybrid methods yield significantly improved predictions for key properties of binary mixtures, such as Henry’s law constants, activity coefficients at infinite dilution, and diffusion coefficients at infinite dilution, even when only limited experimental training data are available. In addition to predicting mixture properties directly, this thesis demonstrates that MCMs can also be applied to the pair-interaction parameters of physical group-contribution (GC) methods, which suffer from incomplete and improvable parameter sets limiting their applicability and accuracy. By using MCMs to infer the pair-interaction parameters, a comprehensive and consistent parameter set can be generated. This approach extends the applicability of widely used GC methods such as UNIFAC and modified UNIFAC (Dortmund), ultimately increasing their scope, predictive power, and robustness.