(643f) Prospect of Using Machine Learning for Evaluating Gas Separation Membranes’ Transport Properties and Assisted Fabrication

Polymeric membranes are widely used to separate various mixed species due to their high removal ability and economic cost; however, membrane fabrication requires significant research to determine appropriate polymers and solvents and test characterization. Machine Learning methods are computational approaches that can be applied to membrane fabrication and facilitate this process. ML techniques used in membrane fabrication are categorized into two types. The first type breaks the polymer structures and defines each monomer using numerical measurements. Each polymer has a fingerprint, and this feature can be input into the ML algorithm as legible data by using molecular descriptors or molecular fingerprints. Molecular descriptors are experimentally measured or theoretically derived properties of a molecule compared to the molecular fingerprint, which is usually in the form of bits or uses vector elements. Molecular fingerprints specify the existence or frequencies of specific fragments, while molecular descriptors are used to illustrate a molecule’s physical, chemical, or topological characteristics. These methods represent molecules as a sequence of bits that computers can recognize: longer bit strings are more reliable through the similarity search since each significant bond in a molecule is defined separately as a bit and has more information about the molecular structure.

The second type is mapping a way to link the fingerprinted input and the target values. Classification, regression, dimensionality reduction, and clustering are the most common types of algorithms for chemical separation that can be applied to process data and predict the relationships between the input and output parameters using ML techniques. The ML technique is chosen and used based on the prediction model's target demand, the current dataset’s size, and various ML method features. Artificial neural networks (ANNs), deep learning (DL), support vector machine (SVM), Gaussian process regression (GPR), and random forest (RF) are the ML methods that can be used to analyze non-linear relationships and can be employed in prediction models when the interactions between components are unclear, such as with membrane fabrication.

This research aims to explore the application of an ML algorithm to polymeric membrane creation, review recent research that used ML to fabricate polymeric membranes for gas component separation and address the critical needs. Training Heteropolymers instead of Homopolymers, producing novel polymers by an inverse design approach, and using reliable datasets that are created under the same conditions, are the most crucial necessities that should be investigated. We have summarized dataset acquisition and training algorithms, ML methods, and examined methods to verify the results. We also report on future development prospects for the ML‐driven polymer‐based membrane design methods.