Molecular Descriptors in Matrix Completion Methods

点击翻译按钮获取中文摘要

Knowledge on properties of mixtures is of paramount interest in chemistry and chemical engineering. We have recently introduced matrix completion methods (MCMs) from machine learning as a novel technique for predicting properties of mixtures. These MCMs also open up new ways for characterizing the molecular components based on their behavior in mixtures. The idea behind an MCM is to store the set of available data on a certain property of binary mixtures in a matrix, where the rows and columns correspond, e.g., to solutes and solvents. The lack of experimental data usually results in a sparsely occupied matrix; the MCM then predicts the remaining entries, which is of high practical interest. Hereby, the MCM often yields better results than established physical benchmark methods. For solving the prediction task, the MCM determines descriptors of the pure components (called latent component features) based only on the mixture data. In the present work, we will explore these latent component features obtained from the MCM in detail. In particular, we are interested in finding out, how they are related to established molecular component descriptors. On the one side, this will enable us to substantially further enhance the predictions of the mixture properties by developing hybrid MCMs. On the other side, we aim at establishing the latent component features as a new class of pure component descriptors, which can be determined flexibly from mixture properties and can be used in many ways, also beyond the MCM. Our research will contribute to the research area “design and evaluation of molecular representations for machine learning” of SPP 2363, to which we will add expertise from engineering, machine learning, and visual data analytics, while profiting from the inspiring input from chemistry.