Screening of large covalent organic frameworks databases for urea removal from water using molecular simulations and machine learning

Efficient removal of urea from spent dialysate is a pivotal step toward the realization of wearable artificial kidneys. Covalent organic frameworks (COFs), with their tunable porosity and chemical versatility, represent a promising class of adsorbents for this purpose. In this work, high-throughput molecular Widom insertions were used to rank 993 COFs from the CORE-COF database for their infinite dilution urea binding strength and ideal urea/water selectivity, indicating COF–C6, EB-COF : Cl, COF–F6, and OH-TPB-BPTA-COF as the best candidates. The analysis of CORE-COF descriptors revealed that oxygen-to-carbon ratio (O/C), carbon fraction (Cfrac), and pore limiting diameter (PLD) are the dominant predictors of urea adsorption. The simulated urea binding strength was used to train and benchmark several ML models, using a nested cross validation. The Random Forest (RF) model was the one with the best performance (Rtrain2 = 0.92, Rtest2 = 0.62), although its accuracy is less satisfactory for the frameworks with high urea adsorption, that are underrepresented in the dataset. Improved performance in this regime will likely require additional training data targeting these rare structures and/or more informative interaction-sensitive features, which are still a challenge for COF structures. Using the best-performing RF model, we rapidly screened a larger dataset of 69 840 hypothetical COFs (hCOFs), which has limited feature-space overlap with CORE-COFs, thus representing a stringent test. Although the model systematically overestimates the hCOFS urea binding strength calculated from simulations, it allows to identify clear chemical patterns that most top hCOF candidates have, such as amide linkers and dia/hcb topologies. The combined simulation–ML strategy here proposed offers a scalable workflow to accelerate the discovery of optimal COFs for urea separation and dialysate regeneration in wearable artificial kidneys.