(572h) Machine Learning Approach for Construction of Fingerprint Kernels for Pore Structure Characterization of Metal-Organic Frameworks

Practical Metal-Organic Frameworks (MOFs) materials are not the ideal crystals. They usually contain secondary phases, residual chemicals, and various types of defects. It is of paramount importance to have an ability to evaluate the extent of sample crystallinity and accessibility of pore compartments for hosting guest molecules. To this end, we recently suggested the method of fingerprint isotherms based on the comparison of the experimentally measured isotherms and Monte Carlo generated isotherms of molecular models of ideal crystals. The proposed method relies on the decomposition, or compartmentalization, of the total isotherm into the fingerprint isotherms corresponding to adsorption in individual pore compartment of the 3d crystal pore network. This method was demonstrated using the examples of PCN-224, ZIF-412 and Cu-BTC structures using Ar, N₂ and CO₂ at their normal boiling temperatures.^1,2

One of the problems in the fingerprint method is how to divide the MOF pore network into individual pore compartment with invoking an arbitrary ad-hoc geometrical boundaries. In this work, we develop an unsupervised machine learning algorithm to calculate the fingerprint isotherms based on in-silico generated database of the spatial distribution of adsorbate molecules at different pressures. The algorithm learns from the distributions of adsorbate molecules at saturation conditions and then predicts which compartment a given adsorbate molecule belongs at any other pressure. To differentiate between the adsorbate molecules within different pore compartments, we use a combination of centroid and density-based clustering algorithms with cluster size constraints. Based on this procedure, we calculate the kernel of fingerprint isotherms and match it against the experimental isotherm to predict the accessibilities of pore compartments and overall non-ideality of the sample. The algorithm is applicable for calculating the kernels of fingerprint isotherms on any MOF. The proposed method is expected to be instrumental for the selection and design of novel MOF-based materials with improved properties for gas separations, energy storage, and catalysis.

1. Parashar, S.; Zhu, Q.; Dantas, S.; Neimark, A.V. Monte Carlo Simulations of Nanopore Compartmentalization Yield Fingerprint Adsorption Isotherms as a Rationale for Advanced Structure Characterization of Metal–Organic Frameworks. ACS Applied Nano Materials 2021 4 (5), 5531-5540. DOI: 10.1021/acsanm.1c00937

2. Dantas, S.; Neimark, A.V. Coupling Structural and Adsorption Properties of Metal–Organic Frameworks: From Pore Size Distribution to Pore Type Distribution. ACS Applied Materials & Interfaces 2020 12 (13), 15595-1560. DOI: 10.1021/acsami.0c01682