(362n) Predicting Activity Coefficients at Infinite Dilution Using Hybrid Residual Graph Neural Networks

In the context of mathematical modeling of chemical processes, the availability of physicochemical data of the compounds involved is paramount. Besides this, data for mixtures of compounds must also be available. However, given the vastness of the chemical space of interest, collecting such data experimentally turns into an impracticable task [1]. This problem has historically motivated the use of predictive methods, such as UNIFAC or COSMO-RS. These methods have been largely used for predicting activity coefficients which, when measured at infinite dilution, provide useful insights into the performance of separator units [2] and play an important role in environmental studies [3].

The available models for predicting infinite dilution activity coefficients can be divided into two broader categories. Those which are derived from mechanistic or phenomenological knowledge (e.g., COMSO-RS and UNIFAC) and those which are based purely on machine learning techniques (e.g., Matrix Completion methods [4]). Despite the availability of many types of models, many deviations from reality are still present in systems of interest [5]. This calls for the development of new predictive models and methodologies that allow us to improve the existing ones.

In this work, we propose a collection of Graph Neural Networks (GNNs) for predicting activity coefficients at infinite dilution [6]. We use one GNN for constructing a solvent fingerprint containing relevant information for the prediction, and another GNN for doing so for the solute. These fingerprints are generated after performing convolutional operations on the molecular graphs followed by a global pooling operation. Both molecular fingerprints are then combined to generate a vectorial representation of the mixture which is later used to regress the corresponding infinite dilution activity coefficient. This framework is trained in an end-to-end fashion using backpropagation, which allows the method to go from SMILES strings to prediction of the infinite dilution activity coefficient directly.

Moreover, a series of hybrid residual GNNs were developed which combine the most popular mechanistic/phenomenological models for predicting infinite dilution activity coefficients with GNNs trained on the residuals of these same models. Overall, the predictions of the hybrid residual GNNs outperform those of the mechanistic or GNN models alone, while providing a larger application domain at the same time. This highlights the benefit of combining both approaches and opens the door to novel thermodynamic modeling paradigms that can potentially be used to increase our current understanding on activity coefficient prediction.

References:

[1] Gmehling, J., Kleiber, M., Kolbe, B. and Rarey, J., 2019. Chemical Thermodynamics for process simulation. John Wiley & Sons.

[2] Brouwer, T., Kersten, S.R., Bargeman, G. and Schuur, B., 2021. Trends in solvent impact on infinite dilution activity coefficients of solutes reviewed and visualized using an algorithm to support selection of solvents for greener fluid separations. Separation and Purification Technology, 272, p.118727.

[3] Harten, P., Martin, T., Gonzalez, M. and Young, D., 2020. The software tool to find greener solvent replacements, PARIS III. Environmental progress & sustainable energy, 39(1), p.13331.

[4] Damay, J., Jirasek, F., Kloft, M., Bortz, M. and Hasse, H., 2021. Predicting Activity Coefficients at Infinite Dilution for Varying Temperatures by Matrix Completion. Industrial & Engineering Chemistry Research, 60(40), pp.14564-14578.

[5] Brouwer, T. and Schuur, B., 2019. Model performances evaluated for infinite dilution activity coefficients prediction at 298.15 K. Industrial & Engineering Chemistry Research, 58(20), pp.8903-8914.

[6] Medina, E.I.S., Linke, S., Stoll, M. and Sundmacher, K., 2022. Graph neural networks for the prediction of infinite dilution activity coefficients. Digital Discovery.