Active subsets as a tool for structural characterisation and selection of metal-organic frameworks

• Published in: Chemical engineering research & design 2022-03, Vol.179, p.424-434
• Main Authors: Yeardley, Aaron S., Milton, Robert A., Moghadam, Peyman Z., Cordiner, Joan, Brown, Solomon F.
• Format: Article
• Language: English
• Subjects: Active subspaces; Adsorption; Gas adsorption; Gaussian process; Metal-organic frameworks; Methane; Normal distribution; Oxygen; Reduced order modelling; Reduced order models; Regression models; Sensitivity analysis; Sobol’ indices; Storage; Subspaces
• ISSN: 0263-8762; 1744-3563
• DOI: 10.1016/j.cherd.2022.01.045

•Predicting a metal-organic frameworks deliverable capacity using Gaussian processes.•Locating active subspaces enables optimal dimension reduction of pore properties.•Applied to a practical exploration example comparing methane and oxygen storage.•Found top-performing oxygen structures irrelevant to the gas in the training data.•Presented an efficient method to predict and search for promising frameworks. To date over 80,000 metal-organic framework (MOF) structures have been synthesised and only ca.3% of these have had their adsorption capabilities measured for storing oxygen alone. As such, in order to aid the process of producing top-performing MOFs for storing various gases, accurate methods to predict the deliverable capacity of MOFs that have their synthesis method already known is increasingly important. For this purpose, this paper develops a reduced order model (ROM) that can predict the deliverable capacity of synthesised MOFs irrespective to the storage gas across similar gases. The ROM is constructed by identifying the active subspaces through a Sobol’ index-based global sensitivity analysis (GSA). The resulting Gaussian process (GP) regression model efficiently predicts the deliverable capacity given a MOFs pore properties with this reduced dimensional space. This approach was applied to a practical MOF exploration example by training a ROM with 2745 MOFs storing methane at 30bar. The ROM was robustly tested and analysed before using it to predict the deliverable capacity of 82,221 synthesised MOFs storing oxygen at 30bar. To ensure validity in the exploration example, the predictions produced from the methane trained ROM were compared to a separate ROM trained using the same MOFs but storing oxygen gas. The methane trained ROM was found to be in agreement with the oxygen trained ROM, and was shown to be a viable tool to identify the top-performing MOF structures for oxygen storage.