(152h) Predicting Temperature-Dependent Activity Coefficients of Ionic Liquid-Solute Systems through Graph-Based Machine Learning

Ionic liquids (IL) have attractive physicochemical properties and are highly interesting solvents in the design of sustainable separation processes [1,2]. Properties of ILs can be adapted to the separation task at hand by combining anions, cations, and further structural groups so that the number of potential IL candidates is large including at least one million possible ILs [2]. To assess the suitability of an IL as a solvent candidate, the activity coefficient is of paramount importance. Compared to the large number of potential ILs, the experimental data on activity coefficients for ILs is, however, limited which calls for prediction models. Classical activity coefficient prediction models such as UNIFAC and COSMO-RS are well established in chemical engineering but are limited in terms of applicability and accuracy for IL-solute systems [3].

Recently, machine learning models have shown promising results for activity coefficient prediction, with two methods being actively investigated: Matrix completion methods (MCMs) and graph neural networks (GNNs) [3-6]. MCMs build on the idea that the task of activity coefficient prediction can be represented as filling entries of a matrix with rows as solutes and columns as solvents (or vice versa) and with each entry corresponding to the activity coefficient of a pair of solute and solvent [3,4]. MCMs have been applied to predict activity coefficients of IL-solute systems, showing superior accuracy compared to classical thermodynamic models specifically adapted to ILs like UNIFAC-IL and calibrated COSMO-RS [3]. A GNN, on the other hand, is a graph-based machine learning method that has shown great results in predicting physicochemical properties of molecules in recent years [5,6]. GNNs operate on the molecular graph with atoms as nodes and bonds as edges and thereby enable learning a direct mapping from the molecular graph to a molecular property of interest. Very recently, GNNs have been utilized for predicting activity coefficients of solvent-solute systems at a constant reference temperature with promising results [7,8].

We propose a GNN for predicting the temperature-dependent activity coefficient at infinite dilution of IL-solute systems. The proposed model extends GNNs for activity coefficient prediction to ILs and temperature-dependent activity coefficient values. To compare our GNN to state-of-the-art MCM models for IL-solute systems, we also implement an MCM approach and train both models on a large database including more than 40,000 data points [3,9]. The results show that GNNs and MCMs achieve comparable high prediction quality for predicting the activity coefficient of IL-solute systems. Whereas the applicability of MCMs is limited to ILs and solute molecules that have been included in model training, GNNs can be applied to ILs and solutes not seen during training. Our investigations show that the GNN allows for generalization with high accuracy, making it a promising tool for computer-aided design of ILs for chemical engineering applications. An open-source publication of the models and code is underway.

[1] Han, X., & Armstrong, D. W. (2007). Ionic liquids in separations. Accounts of Chemical Research, 40(11), 1079-1086.

[2] Rogers, R. D., & Seddon, K. R. (2003). Ionic liquids – Solvents of the future?. Science, 302(5646), 792-793.

[3] Chen, G., Song, Z., Qi, Z., & Sundmacher, K. (2021). Neural recommender system for the activity coefficient prediction and UNIFAC model extension of ionic liquid‐solute systems. AIChE Journal, 67(4), e17171.

[4] Jirasek, F., Alves, R. A., Damay, J., Vandermeulen, R. A., Bamler, R., Bortz, M., Mandt, S., Kloft, M. & Hasse, H. (2020). Machine learning in thermodynamics: Prediction of activity coefficients by matrix completion. The Journal of Physical Chemistry Letters, 11(3), 981-985.

[5] Medina, E. I. S., Linke, S., Stoll, M., & Sundmacher, K. (2022). Graph neural networks for the prediction of infinite dilution activity coefficients. Digital Discovery. Advance online publication.

[6] Felton, K. C., Ben-Safar, H., Alexei, A. A. (2022). DeepGamma: A deep learning model for activity coefficient prediction. 1st Annual AAAI Workshop on AI to Accelerate Science and Engineering (AI2ASE), 02-28-2022, virtual.

[7] Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O., & Dahl, G. E. (2017). Neural message passing for quantum chemistry. Proceedings of the 34th International Conference on Machine Learning, in Proceedings of Machine Learning Research 70:1263-1272.

[8] Wieder, O., Kohlbacher, S., Kuenemann, M., Garon, A., Ducrot, P., Seidel, T., & Langer, T. (2020). A compact review of molecular property prediction with graph neural networks. Drug Discovery Today: Technologies, 37, 1-12.

[9] Kazakov, A., Magee, J.W., Chirico, R.D., Paulechka, E., Diky, V.; Muzny, C.D., Kroenlein, K., Frenkel, M. (2013). NIST Standard Reference Database 147: NIST Ionic Liquids Database - ILThermo v2.0, National Institute of Standards and Technology, Gaithersburg, USA,
https://ilthermo.boulder.nist.gov (accessed on 4-4-2022).