(380ai) Machine Learning Enabled Discovery of Novel Polymer Structures with Experimental Validation for Gas Separation Applications

Membrane separation technology has emerged as a compelling low energy alterative to traditional industrial separation methods that require significant energy input such as distillation or absorption. In addition to energy efficiency, polymeric membranes have smaller footprints, improved occupational safety, modularity, no toxic solvents present in operation, and fewer moving parts. One of the primary challenges in deploying membrane technology has been the limitations in membrane separation performance. Polymer membranes experience a well-known trade-off between gas permeability and selectivity. For much of the historical development of gas separation membranes, researchers have synthesized and screened novel polymer architectures to understand the structure-property relationships that govern the transport properties of polymers. While innovative macromolecular designs and polymers with finely tuned microporosity have shown promise, achieving the right balance of a rigid, ladder-like backbone (which hinders efficient polymer chain packing for high permeability), and restricted rotational freedom (for selective micropores) is complex. To help screen new polymer architectures and accelerate the discovery of novel structures with enhanced separation efficiency, our group has utilized machine learning models to generate novel polymer structures with high fractional free volume by combining different structural motifs seen in gas separation membranes. This model generates an interpretable grammar where the step-by-step derivation procedure which led to the final structure can be explained, allowing subject matter experts to critique on whether the mechanism that led to the structure makes practical sense. After generating novel structures by combining functional groups and structural motifs from a data set of reported polymers in new ways, the model is able to predict separation properties using a calculation of the fractional free volume that utilizes a revised and more accurate group contribution method. Importantly, experimental validation of the model was conducted on the novel structures to determine the accuracy of the predictions. Good agreement between computational and experimental permeability and selectivity for a variety of gases relevant to industrial separations demonstrates the utility of this machine learning process for predicting novel structures that may not be intuitive. Ongoing work has focused on generating and validating new structures that exceed the permeability/selectivity trade-off that defines the state of the art for polymer membranes.