Machine-Learning-Aided Computational Study of Covalent Organic Frameworks for Reversed C₂H₆/C₂H₄ Separation

Abstract

The efficient separation of ethane/ethene (C 2 H 6 / C 2 H 4 ) is imperative yet challenging in industrial processes. We herein combine machine learning (ML) and molecular simulation to predict optimal covalent organic frameworks (COFs) for reversed C 2 H 6 /C 2 H 4 separation before experimental efforts. Using molecular simulations, two out of 601 CoRE COFs were identified with excellent separation performance, and eight CoRE COFs exhibit high C 2 H 6 /C 2 H 4 selectivity surpassing all of the reported values, although these COFs have a relatively low working capacity. As for ML, we found that the random forest (RF) algorithm displays the highest accuracy (R 2 = 0.97) among the four different models, and the density (ρ) of COFs was identified as the key factor that influences the C 2 H 6 /C 2 H 4 selectivity. Moreover, the 10 best hypothetical COFs (hCOFs) with excellent selectivity were further predicted. Ultimately, the competitive adsorption behaviors of guests in COF-303 were disclosed, and the adsorption selectivity of COF-303 was enhanced by introducing the fluorine group. Results of this work could provide molecular-level insights for future design and synthesis of novel COFs that can directly remove low-concentration ethane from the C 2 H 4 /C 2 H 6 mixture.

Discovery of High-Performing Metal–Organic Frameworks for Efficient SF6/N2 Separation: A Combined Computational Screening, Machine Learning, and Experimental Study

“…Although TSN has been adopted as the primary assessment metric, there is another key parameter (Re%) to define the promising adsorbents for SF 6 /N 2 separation. Re% ≥80% could be considered as the promising MOFs for practical VSA application . The relationship between Re% and S ads,SF 6 /N 2 of MOFs was plotted in Figure a.…”

“…Zeo++ was used to calculate the structural properties of all of the MOFs, including PLD, global cavity diameter (GCD), largest cavity diameter (LCD), gravimetric accessible surface area (gASA), volumetric accessible surface area (vASA), Sa, density (ρ), void fraction (φ), accessible volume (AV), the total volume of unit cell ( V total ), and free volume ( V free ). More detailed information can be found in our previous work. , The range and distribution of MOF structural properties were visualized with violin plots in Figure . The PLD, GCD, and LCD are in the range of 5.13–19.65, 5.23–29.74, and 5.21–29.74 Å, respectively.…”

“…GCMC simulations were conducted using our in-house code (HT-CADSS) − to obtain the gas uptake in each MOF under vacuum-swing adsorption (VSA) conditions. The adsorption and desorption temperatures were set at 298 K and pressures at 1.0 and 0.1 bar, respectively.…”

“…To identify top-performance MOF adsorbents, we used simulated results to calculate various metrics which contain SF 6 /N 2 selectivity ( S ads , S F 6 / N 2 = N ads , S F 6 N ads , N 2 / y S F 6 y N 2 ) , working capacity (Δ N SF 6 = N ads,SF 6 – N des,SF 6 ), regenerability ( Re % = Δ N ads , S F 6 N ads , S F 6 × 100 ) , and trade-off between Δ N SF 6 and S ads,SF 6 /N 2 (TSN = Δ N SF 6 × ln( S ads,SF 6 /N 2 )). TSN is defined based on the adsorbent performance score metric . We take the logarithm for selectivity to avoid excessive selectivity affecting performance trade-offs.…”

“…More detailed information can be found in our previous work. 28,29 The range and distribution of MOF structural properties were visualized with violin plots in Figure 1. The PLD, GCD, and LCD are in the range of 5.13−19.65, 5.23−29.74, and 5.21− 29.74 Å, respectively.…”