Hybridizing Physical and Data-driven Prediction Methods for Physicochemical Properties

Abstract

We present a generic way to hybridize physical and data-driven methods for predicting physicochemical properties. The approach 'distills' the physical method's predictions into a prior model and combines it with sparse experimental data using Bayesian inference. We apply the new approach to predict activity coefficients at infinite dilution and obtain significant improvements compared to the data-driven and physical baselines and established ensemble methods from the machine learning literature.Prediction methods for physicochemical properties are indispensable for process design and optimization in chemical engineering since experimental studies are expensive and tedious. The most widely used approaches are groupcontribution methods (GCMs) that model the properties of pure components or mixtures based on the structural groups that build up the components. [1][2][3][4] GCMs can also be used for predicting properties of mixtures of which the composition is unknown. [5][6][7] The most successful GCMs for mixtures are the different versions of UNIFAC 8-10 that model the excess Gibbs energy based on binary group-interaction parameters. The group-contribution concept greatly reduces the number of model parameters and the amount of data needed for fitting GCMs. However, the practical applicability of UNIFAC is still restricted, mainly due to necessary group-interaction parameters that have not been fitted yet.