Grand canonical Monte Carlo (GCMC) simulations are frequently used to determine CO2 uptake by metal organic frameworks. However, GCMC can be quite computationally intensive. Machine learning (ML) algorithms can be much faster, but just as robust, if appropriate feature predictors can be identified. To evaluate the efficacy of an ML approach, approximately 10,000 hypothetical metal organic frameworks (hMOFs) were randomly selected from the hMOF Database (http://hmofs.northwestern.edu); GCMC simulations for CO2 adsorption were then run for 1,000 Monte Carlo cycles at 298K and 1 bar on those hMOFs. The GCMC simulation effort took over two weeks of "wall time" on the University of Connecticut BECAT Supercomputing Cluster. The results of the GCMC simulations provided the training data for use with a random forest ML algorithm. A novel combination of structural and chemical feature predictors, including void fraction, surface area, density, electronegativity, degree of saturation and metallic percentage among others, were used as part of the training process. The entire ML training took less than two hours on a laptop computer. The ML was then applied to the test dataset consisting of the remaining ~120,000 hMOFs in the database to predict CO2 uptake. Results were validated by carrying out follow-up GCMC on the best performing MOFs, defined as adsorbing more than 330 mg/g CO2, as predicted by the random forest. Calculation of CO2 adsorption for a single hMOF was virtually instantaneous. The mean absolute percentage error (MAPE) of the ML predictions with respect to GCMC was 10%. The MAPE results indicated that the quality of the ML algorithm in distinguishing high-performing candidates was high. The ML predictions were next evaluated on two well studied experimentally synthesized MOFs, IRMOF-1 and MOF-177. The ML predicted CO2 adsorption of 27 mg/g for IRMOF-1 as compared to the experimental result of 30 mg/g. Similarly, for MOF-177, the ML predicted CO2 adsorption of 49 mg/g compared to the actual 50 mg/g. All studies were done at 1 bar and 298 K. The results of these efforts indicate that ML is well suited to carry out fast, efficient, and robust material property prediction provided high quality feature predictors are used.

Combining structural and chemical feature predictors for fast and robust prediction of CO2 adsorption by metal organic frameworks using a machine learning approach

256th National Meeting (2018)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.