Machine learning of CO2 adsorption

“The literature suggests the separation conditions and the textural properties of the carbon-based adsorbents to be more impactful than their chemical composition (Zhu et al., 2020Yuan et al., 2021). Zhu et al. (2020) evaluated a total of 6422 data sets from 155 porous carbon materials (PCMs) for temperatures of 0 °C and 25 °C in a pressure range up to 1 bar with nitrogen contents between 0.2 – 12.2 wt% using the Random Forest model. This model was selected due to its high accuracy, insensitivity to noise and resistance to overfitting. Of the data, 80% was selected as the training group and the final fifth were left for results validation and evaluation. The findings of their work elucidated to a far stronger impact of pore volume than that of chemical composition., a stronger impact of pore. Quantitatively this meant that the chemical composition of the sorbent was 2.3 times less influential at 0°C and a pressure between 0.6 – 1 bar than pore volume. The Pearson correlation coefficients of some of the investigated parameters were: Micropore Volume = -0.034; Mesopore Volume = -0.142; Ultramicropore Volume = 0.144; Temperature = -0.38; Pressure = 0.796, with the negative values hindering the sorption process (positive – promoting) and the modulus depicting the strength of the interaction. Thus, a weakly positive correlation between mass percentage of N (0.038) and uptake has been described. Additionally, a well-developed ML model to correlate selectivity over N2 could not be made due to a lack of sufficient data. Nevertheless, when incorporating SFGs onto the adsorbent surface, considerations have to be made. Modelling general N-doping is not as transferrable to experimental studies as specific functionalities and does not provide sufficient insight into the sorption characteristics of the functionalised material (Figure 8).”

Leave a Comment