Efficient Xylene Isomer Separation: Accelerated Screening with Active Learning and Molecular Simulation

Abstract

Separating xylene isomers is vital in the petrochemical industry, yet it poses a considerable challenge due to their proximate boiling points, mandating selective adsorbents. This work utilizes active learning (AL) coupled with molecular simulations to rapidly screen 324,426 hypothetical metal–organic frameworks (hMOFs) to identify optimal materials for preferential para-xylene (pX) adsorption. To begin, a diverse subset, representative of the entire hMOF set, was curated using structural and chemical descriptors and evaluated through multiple screening methodologies. This comparative analysis highlighted the superior efficiency of AL in targeted screening processes, requiring on an average only 500 multicomponent Grand Canonical Monte Carlo simulations to identify the most pX-selective framework, encompassing 50.5% of the top 100 candidates. With an equivalent evaluation budget, both machine learning (ML) and evolutionary algorithms demonstrate an inadequate performance. While the former consistently fails to identify top performers, the latter continuously identifies significantly inferior materials. AL, on the other hand, surpasses rival approaches by effectively balancing exploration and exploitation, guiding simulations toward regions associated with high performance. Furthermore, we report the impact of different surrogate models, acquisition functions, and batch acquisition strategies on the convergence of our AL model. We found that the Gaussian process surrogate model coupled with expected improvement (EI) acquisition function and the Kriging-Believer upper bound (KBUB) acquisition strategy acquires the highest pX-selective MOF in just 86 acquisitions. Examining the top hMOF candidates revealed a complex correlation between the pX selectivity and structural features of hMOFs. In particular, the pcu topology, along with a pore size ranging from 5 to 6 Å, emerged as the dominant characteristic of top hMOFs. Furthermore, pressure-dependent simulations revealed optimal pressure maximizing pX uptake and selectivity. This computational workflow, integrating AL and molecular simulations, shows promise in accelerating data-driven material innovation for separation applications.