(129b) Determining Ion Activity Coefficient Using Graph Convolutional Neural Networks in Ion-Exchange Membranes

Electrically driven ionic separations have held historical significance in water desalination and ultra-pure water production for applications in semiconductors, food, and pharmaceutical manufacturing. These electrically driven ionic separations encompass relying on ion-exchange membranes (IEMs). The chemistry and molecular structure of these IEMs significantly influence their ionic conductivity, permselectivity, and other transport properties, such as osmotic drag. Consequently, the chemistry and microstructure of IEMs exert a profound influence on these properties.

The partitioning of ions between polymeric ion-exchange membranes (IEMs) and the surrounding liquid is governed by the activity coefficients of the ions, which, in turn, have a significant impact on various ion transport processes within these membranes, notably conductivity. These coefficients are essential for comprehending phase equilibria, solubility, and chemical reactions, making them indispensable for a wide spectrum of industrial and scientific applications. In the past, estimating activity coefficients typically involved complex mathematical models and the collection of extensive experimental data.

This study introduces a computational framework designed to predict the activity coefficients of ions in charged Ion Exchange Membranes (IEMs). In recent years, machine learning (ML) has been integrated into materials design workflows as a complement to experiments and simulations to accelerate the discovery of a wide range of materials.[2] Graph Neural Networks (GNNs) have emerged as a powerful tool in the field of chemistry, offering an innovative approach to modeling and predicting diverse molecular properties, including the crucial domain of activity coefficients. Specifically, the framework utilizes Graph Convolutional Networks (GCN) to establish connections between the chemical structure of the polymer and the molecular-level attributes. This ultimately leads to the determination of macroscopic attributes, such as the activity coefficient, across a range of IEM materials, including random copolymer and block copolymer systems.

Data from our previous work based on the electrolyte molecular structure and composition alone with Molecular Dynamics simulations was used to train our model. Our dataset comprises a comprehensive combination of experimental and molecular dynamics (MD) data, which served as the foundational elements of the study conducted by Gallage Dona et al.[1] This method utilizes 137 different input variable categories derived from polymer characteristics, salt ions, the quantity of water molecules attached to ions, salt concentration, and twelve solvation descriptors obtained from MD simulations as the target variables.

The architecture encompasses a series of meticulously designed stages, each contributing to the model's exceptional performance. The initial phase of the architecture uses different data structures such as polymer structure and salt structure as an input to graph convolutional neural network. Following the graph-based transformation, the architecture leverages Convolutional Attention Networks (CAT) to seamlessly fuse the graph-based molecular representations with critical features such as salt concentration and Molecular Dynamics (MD) descriptors. The culmination of the architecture involves a meticulously crafted Feedforward Neural Network (FNN). This FNN component, driven by the consolidated data from the previous stages, serves as the predictive pinnacle, generating precise and insightful predictions for activity coefficients. Impressively, the ML-MD modeling strategy yielded highly accurate predictions of ion activity coefficients within IEMs.

References

[1] Hishara Keshani Gallage Dona, Teslim Olayiwola, Luis A. Briceno-Mena, Christopher G. Arges, Revati Kumar, and Jose A. Romagnoli, "Title of the Paper," Industrial & Engineering Chemistry Research, vol. 62, no. 24, pp. 9533-9548, 2023, doi: 10.1021/acs.iecr.3c00636.

[2] G. Bradford, J. Lopez, J. Ruza, M. A. Stolberg, R. Osterude, J. A. Johnson, R. Gomez-Bombarelli, and Y. Shao-Horn, "Chemistry-Informed Machine Learning for Polymer Electrolyte Discovery," ACS Cent. Sci., vol. 9, no. 2, pp. 206-216, Jan. 23, 2023, doi: 10.1021/acscentsci.2c01123.