Discovering robust metal-organic frameworks with open copper sites for precombustion CO2 capture: Data-efficient exploration and exploitation by active learning

XY Wu and Q Liu and JW Jiang, CHEMICAL ENGINEERING JOURNAL, 521, 167021 (2025).

DOI: 10.1016/j.cej.2025.167021

As a carbon-free energy source, hydrogen (H-2) plays an essential role in transitioning toward environmentally sustainable and renewable energy systems. There has been tremendous interest in the development of nano- porous materials as adsorbents for H-2 purification. In this work, we aim to discover open copper site-containing metal-organic frameworks (OCS-MOFs) for precombustion CO2/H-2 separation by active learning (AL). Starting with only 3522 labelled OCS-MOFs, we couple a physics-rich descriptor set with AL to strategically interrogate a large unlabelled pool comprising 24,057 candidates. Encouragingly, with four similar to five rounds of AL acquisitions (800-1000 new calculations), top OCS-MOFs can be exploited that surpass the originally labelled benchmark, while exploratory accuracy on unseen structures rises from 0.39 to 0.88 in the coefficient of determination (R-2). To translate separation performance into deployable adsorbents, we cascade machine-learning models to assess thermal and activation stabilities, along with molecular dynamics simulations to predict mechanical stability, culminating in the identification of robust OCS-MOFs. Notably, we further propose H-M, a harmonic-mean metric which effectively encodes the trade-off between separation performance and Hill's bulk modulus. Based on HM ranking, we elucidate nbo topology combined with alkyne (C equivalent to C) moiety consistently promotes CO2/H-2 separation. This study demonstrates that AL with data-efficient exploration and exploitation can overcome cold- start bias, reduce computational cost by an order of magnitude, and accelerate the discovery of robust MOFs for H-2 purification.

Return to Publications page