Hybrid graph neural networks for the prediction of activity coefficients in separation processes

Abstract

The task of predicting properties of mixtures from the molecular structure of their components has been studied for decades. In the past, a broad spectrum of mechanis- tic models has been developed for predicting thermophysical properties of mixtures. These models have been the basis of many successful applications across various chemical and process engineering domains. Noticeable examples are the large chemical plants supporting today’s world economy through oil refinement. However, when the properties of novel complex mixtures are the focus, e.g., oc- curring in biorefineries or in chemical recycling of plastic waste streams, or when sustainable processes for the future circular economy need to be developed, the existing property models are often limited regarding their accuracy and predictive power and are not suited for the effective exploration of the vast chemical space. Of special importance is the modeling and calculation of phase equilibria, which is a cornerstone in the design of separation processes for molecular mixtures. For describing the non-ideal mixing behaviour of components in liquid mixtures, activity coefficients are often used. The exploration of alternative routes to separation requires the development of accurate and efficient predictive models that estimate the activity coefficients across large chemical spaces. In this dissertation several hybrid models are presented that combine the flexibility of graph neural networks (GNNs) with phenomenological/mechanistic modeling approaches of the thermodynamic behavior of mixtures. Two main arrangements for the construction of such hybrid models are explored: (i) a parallel arrangement in which the graph neural network serves as a corrector of a phenomenological model prediction. (ii) A serial arrangement in which the graph neural network is embedded in some form of mechanistic expression to preserve the physical constraints of the latter. The proposed hybrid graph neural network models are then presented from the simplest scenario to increasing levels of generality in predicting activity coefficients. The proposed models are extensively tested to evaluate their advantages and limita- tions compared to conventional methods (e.g., UNIFAC). The dissertation concludes with a series of case studies that demonstrate the utility of the proposed models in the context of supporting the early stages of separation process design. Overall, the results suggest that hybrid graph neural networks offer more efficient and accurate solutions for predicting activity coefficients compared to the standalone submodels. Such advantages can be exploited in practical scenarios in the context of separation process design. The implementation of hybrid graph neural networks could be of high relevance in the development of more sustainable separation processes.