(557f) Modeling Polymer Aging in Solvent Environments with Symbolic Regression

Polymeric membrane separations present an energy-efficient alternative to established, thermally-intensive separation processes. The majority of membrane polymers are glassy materials, which enable high selectivities, but their performance also changes over time due to conformational changes in the polymer microstructure towards a lower-energy state. These changes present a barrier to widespread membrane adoption due to uncertainty about long-term performance stability. Despite its practical importance, glassy polymer aging behavior is still a poorly understood phenomenon, especially in chemically harsh solvent environments. Here, we attempt to understand the relationship between operating environment and polymer aging behavior in solvents using symbolic regression. Symbolic regression is a machine learning paradigm wherein data is fit to a function, as in normal linear regression, but the algorithm additionally finds the most appropriate functional form for the provided descriptors. This strictly empirical approach has the benefit of uncovering non-obvious relationships that might otherwise be overlooked in physics-focused derivations. The inputs to the algorithm are solvent physical properties like kinetic diameter, solubility parameters, etc. We use high-resolution permeance vs. time data to develop models that predict the aging rate, the decline in flux due to aging, and the steady state permeance for Matrimid thin film composite membranes across a range of aromatic, non-polar, and polar solvents. We find that descriptors related to solvent condensability heavily influence the aging behavior of the polymer and that the models can predict Matrimid solvent aging behavior with high accuracy. This case study of a common polymer furthers our understanding of glassy polymer aging behavior in solvent-based membrane separations and highlights the ability of machine learning to provide quantitative structure-property relationships.