AI-Driven Automated Discovery of Polymer Membranes for Carbon Capture

By Samudrapom DamAug 1 2023

In an article published in the journal Npj Computational Materials, researchers demonstrated artificial intelligence (AI)-driven automated discovery of polymer membranes for efficient carbon capture.

Background

New material discovery has traditionally been a resource-intensive and time-consuming process due to the application of a trial and error approach that involved the identification of known materials with properties similar to the target properties of new material and then combining/modifying them to achieve the desired outcome.

The development of robust simulation methods, such as the density functional theory (DFT) method, has improved the new material discovery process. Specifically, high-throughput computational materials screening and design (HCMSD) methods have significantly accelerated the discovery process.

However, HCMSD primarily depends on time-consuming ab initio calculations to model the chemical and physical processes. Thus, several computations are required to perform material screening or probe the phase space, which is a major limitation that makes HCMSD impractical for new material discovery. 

The advent of repositories with large simulation and experimental datasets has enabled the use of machine learning (ML) methods for materials discovery. ML-based materials design is more advantageous compared to HCMSD as it is not entirely dependent on ab initio simulations of quantum mechanical/classical molecular dynamics that occur in a chemical system.

Recently, the inverse materials design (IMD) method has received significant attention for materials discovery. In this method, an algorithm creates/develops optimized molecular structures depending on a pre-defined feature vector comprising material target properties. The IMD output undergoes physical validation to complete the materials discovery process.

The gas filtration performance of a polymer is determined by the complex morphology and heterogeneous internal structure of the amorphous polymer and the chemical and physical properties of monomer constituents. Thus, the validation and prediction of the gas permeability of a polymer membrane remain extremely challenging.

The physical validation for polymer membranes is required at the mesoscale to identify the amorphous material process-relevant properties. However, the automated ab initio simulation methods to validate complex materials at the mesoscale have not been developed until now.

Although ML can predict the gas separation performance of previously untested polymers when applied to known polymer repeat units, the method cannot generate optimized monomer units to create new polymer candidates and lacks automated physical performance outcome validation.

Thus, the absence of generative design, mesoscale physical performance validation, and automated training data creation in existing computational polymer membrane discovery frameworks has necessitated the identification of new approaches.

An AI-driven approach for automated polymer membrane discovery

In this study, researchers proposed an in silico, fully automatized materials discovery workflow to overcome the limitations of existing computational frameworks. The study aimed to realize automated complex material discovery through inverse molecular design based on mesoscale target features and process figures-of-merit.

The proposed workflow was applied to the physical validation and generative design of polymers optimized for carbon dioxide filtration under realistic pressure and temperature conditions to demonstrate the methodological advancements in small molecule discovery.

The study was limited to homo-polymers as reliable error margins and experimental data were available. Researchers investigated the multi-scale discovery regime by computationally validating and generating hundreds of polymer candidates designed for post-combustion carbon dioxide filtration. They also devised a representative elementary volume (REV)-enabling permeability simulations at 1000× the volume of an ML-generated, individual monomer to obtain a quantitative agreement.

The ML generative design sequence included extraction and selection of features, training of regression models, feature optimization, and structure generation based on graphs. Six regression models, including support vector regression, kernel ridge regression, random forest regression, elastic net regression, ridge regression, and lasso regression, were trained and cross-validated for molecular property prediction.

Significance of the Study

Researchers successfully demonstrated an end-to-end, fully automated computational discovery of polymer membranes to separate carbon dioxide. Every discovery step, including the creation of an automated training dataset and feature vector based on the graph-based generative inverse design of optimized monomer units and non-equilibrium molecular dynamics gas permeation/filtration simulation through the polymer membranes, was validated.

Molecular dynamics simulations performed using a minimum representative volume of the complex material successfully predicted the permeability and filtration dynamics of a polymer. Additionally, a quantitative agreement was obtained for computationally designed polymers between the carbon dioxide permeability predictions through the physical process simulations based on molecular dynamics and ML models. The discovery-to-validation time required per polymer candidate was on the order of 100 h using one graphics processing unit (GPU) and central processing unit (CPU), which offered a computational screening alternative before lab validation.

To summarize, the findings of this study demonstrated the effectiveness of the proposed automated computational discovery and physical validation of polymer membranes for carbon filtration. Moreover, the study can be crucial in advancing ML-generative design beyond small-molecule applications and substantially accelerating complex material discovery for scaled applications.