Polymer genome–based prediction of gas permeabilities in polymers

Abstract

Predicting gas permeabilities of polymers a priori is a long-standing challenge within the membrane research community that has important applications for membrane process design and ultimately widespread adoption of membrane technology. From early attempts based on free volume and cohesive energy to more recent group contribution methods, the ability to predict membrane permeability has improved in terms of accuracy. However, these models usually stay “within the paper”, i.e. limited model details are provided to the wider community such that adoption of these predictive platforms is limited. In this work, we combined an advanced polymer chemical structure fingerprinting method with a large experimental database of gas permeabilities to provide unprecedented prediction precision over a large range of polymer classes. No prior knowledge of the polymer is needed for the prediction other than the repeating unit chemical formula. In addition, we have incorporated this model into the existing Polymer Genome project to make it open to the membrane research community.

Polymer informatics with multi-task learning

“…Data for 36 different properties of over 13,000 polymers (corresponding to over 23,000 data points) were obtained from a variety of sources. 4 , 12 , 13 , 14 , 15 , 16 , 17 , 18 Table 1 shows a synopsis of the data. All polymers are “fingerprinted”, i.e., converted to a machine-readable numerical form, using methods described elsewhere 4 , 19 , 20 (and briefly in the experimental procedures section).…”

“… 112 [12.3,29.2] 4 , 14 Gas permeability c Barrer Exp. 2168 [0, 4.7] d 13 The total number of single data points is 23,616, and the total number of merged data points in the joint database is 13,766. a Experiments (Exp.…”

Imputation of Missing Gas Permeability Data for Polymer Membranes using Machine Learning

“…One of the main models for predicting polymer membrane performance is group contribution theory, where the chemical structure of a polymer is divided into smaller fragments and the fragments used in various ML models as input features [26][27][28]. Recently, hierarchical methods for fingerprinting polymers for property prediction have also been reported [29]. Such models were built upon chemical structures of polymers and are of great value for identifying structureproperty relationships.…”

“…Many polymers were studied for carbon capture from flue gas, where CO2/N2 separation is relevant, but they may be equally interesting for the strongly emerging new application field of biogas upgrading, where CO2/CH4 separation is important. Although it does not have the full predictive power of other methods [24,29], the advantage of the models presented in this work is that they do not require any knowledge about the polymer structure and they work for polymers with different measurement conditions (such as aging and solvent treatment), which makes it a fast and versatile approach. For the rapid screening of polymers, especially those produced via high-throughput techniques, the prediction of the full range of gas permeability from a single rapid measurement could be highly beneficial to researchers, especially as the chosen gas may be selected based on avoiding stringent local safety regulations (e.g.…”

Database Creation, Visualization, and Statistical Learning for Polymer Li+-Electrolyte Design

“…[19][20][21] Polymers have also been the recent subject of recent data-driven discovery, including the prediction of polymer glass transition temperature, T g , [22][23][24] exploration of new polymer electrolytes, [25][26][27] and prediction of gas diffusion in membranes. 28 Data visualization on its own can be immensely powerful, both to extract trends in ag-gregate data that may have been missed in smaller studies, as well as to highlight important gaps in published literature that should be corrected. [29][30][31] The addition of machine or statistical learning techniques can aid in situations where more complicated relationships exist between descriptors and the quantity of interest; in such cases simpler visualizations do not readily show trends.…”