Explainable machine learning for unraveling solvent effects in polyimide organic solvent nanofiltration membranes

Understanding the effects of solvents on organic solvent nanofiltration currently depends on results obtained from small datasets, which slows down the industrial implementation of this technology. We present an in-depth study to identify and unify the effects of solvent parameters on solute rejection. For this purpose, we measured the rejection of 407 solutes in 11 common and green solvents using a polyimide membrane in a medium-throughput cross-flow nanofiltration system. Based on the large dataset, we experimentally verify that permeance and electronic effects of the solvent structure (Hildebrand parameters, electrotopological descriptors, and LogP) have strong impact on the average solute rejection. We furthermore identify the most important solvent parameters affecting solute rejection. Our dataset was used to build and test a graph neural network to predict the rejection of solutes. The results were rigorously tested against both internal and literature data, and demonstrated good generalization and robustness. Our model showed 0.124 (86.4% R²) and 0.123 (71.4 R²) root mean squared error for the internal and literature test sets, respectively. Explainable artificial intelligence helps understand and visualize the underlying effects of atoms and functional groups altering the rejection.