Insights into CO₂/N₂ Selectivity in Porous Carbons from Deep Learning

Abstract

Porous carbons are an important class of porous material for carbon capture. The textural properties of porous carbons greatly influence their CO2 adsorption capacities. But it is still unclear what features are most conductive to achieving high CO2/N2 selectivity. Here, we trained deep neural networks from the experimental data of CO2 and N2 uptakes in porous carbons based on textural features of micropore volume, mesopore volume, and BET surface area. We then used the model to screen porous carbons and to predict CO2 and N2 uptakes, as well as CO2/N2 selectivity. We found that the highest CO2/N2 selectivity can be achieved not at the regions of highest CO2 uptake but at the regions of lowest N2 uptake where mesopores disrupt N2 adsorption. This insight will help guide experiments to synthesize better porous carbons for post-combustion CO2 capture.

Understanding the CO2/CH4/N2 Separation Performance of Nanoporous Amorphous N‐Doped Carbon Combined Hybrid Monte Carlo with Machine Learning

“…All training data were normalized within the range of 0–1. [ 29 ] The cross‐validation employed the leave‐one‐out method.…”

“…[ 24–26 ] It provides an opportunity to analyze big data, predict possible performance, accelerate material designing. [ 27–29 ] By using artificial neural networks, an important machine learning method, one can establish a relationship between key features as input and performance as output. According to CO 2 adsorption data of porous carbon, Zhang et al.…”

“…analyzed CO 2 /N 2 selectivity adsorbed in porous carbon by DNNs, and concluded that porous carbon with lowest N 2 uptake have more suitable micropore and mesopore volumes for better CO 2 /N 2 separation. [ 29 ] These aforementioned works mainly considered the influence of various structural features on adsorption and separation performance based on experimental data.…”

Machine Learning Enables Highly Accurate Predictions of Photophysical Properties of Organic Fluorescent Materials: Emission Wavelengths and Quantum Yields

“…Reported ML predictions are limited to Power Conversion Efficiency (PCE) [48][49][50][51][52] , gas absorption selectivity [53][54][55] and AIE effect 56 , most relying on expensive quantum mechanical calculations to generate input expressions. For solvated molecules, expression of solvent features is critical but rarely studied in detail 57 .…”

Machine Learning Enables Highly Accurate Predictions of Photophysical Properties of Organic Fluorescent Materials: Emission Wavelengths and Quantum Yields

“…As a successful example of ML-assisted material design, Aspuru-Guzik et al 32 achieved high-throughput pre-screening of thermally activated delayed fluorescence (TADF) organic light-emitting diodes (OLED) based on neural-network prediction of delayed fluorescence rate constant with data obtained from (TD-)DFT calculations. In contrast, only a few properties in the second category have been studied, examples being Power Conversion Efficiency (PCE) 35,36,37,38,39 , gas absorption selectivity 40,41,42 , and aggregation-induced emission (AIE) effect 43 . For these properties, the input expressions of existing ML models are usually obtained from expensive quantum mechanical calculations, limiting their application in large-scale fast virtual screening.…”