Understanding confinement effects on ion permeability with computation: from first-principles to data-driven models

Developing structure-property relationships is essential for the improved design of membranes for water purification. By definition, macroscopic properties of water and ionic solutions are distinct in the confined pores of membranes. As a result, computational modeling can play an essential role in elucidating structure-property relationships. I will describe our efforts to build computational models to understand anomalous behavior of confined liquids in state-of-the-art nanoporous polyamide membranes. I will discuss how with a combination of first-principles and classical physics-based modeling, we have shown that shifts in local water dielectric constant alter local pKa of carboxyl sidechains in these materials and influence their ion selectivity[1]. I will then discuss how to identify how to improve design and understand relationships for ion-specific trends in selectivity we have built quantitative structure-property relationships using a data driven approach. We have combined a range of experimental and computational descriptors of ions and membrane materials to use as inputs to one-shot, direct feature selection machine learning (ML) models (i.e., LASSO and random forest) in order to build predictive ML models of the thermodynamic quantities observed experimentally to indicate preferential ion selectivity. This data-driven approach overcomes challenges of conventional physics-based modeling when the experimental system is sufficiently heterogeneous, making it difficult to model. Importantly, our ML models provide essential insight into the most important factors for advancing membrane materials design.

[1] Cody L. Ritt, Jay R. Werber, Mengyi Wang, Zhongyue Yang, Yumeng Zhao, Heather J. Kulik, and Menachem Elimelech "Ionization behavior of nanoporous polyamide membranes" Proc. Natl. Acad. Sci. in press.