机器学习解析直接空气捕集用固体胺吸附剂的构效关系
周志斌 , 张智渊 , 邱雨晴 , 董越 , 赵国江 , 曾垌皓 , 吴晓宇 , 郭本帅 , 戴一阳 , 周利 , 刘冲 , 代忠德 , 吉旭
工程科学与技术 ›› 2025, Vol. 57 ›› Issue (05) : 79 -90.
机器学习解析直接空气捕集用固体胺吸附剂的构效关系
Machine Learning Analysis of the Structure‒Property Relationship of Amine-based Solid Adsorbents for Direct Air Capture
当前,直接空气捕集(DAC)技术蓬勃发展,已成为高效的负碳技术,也是实现二氧化碳净零排放的关键路径。高性能的吸附剂材料是保障DAC技术产业化部署的核心要素,其中固体胺吸附剂因性能优异备受关注。通过对全球固体胺吸附剂的CO2吸附实验数据进行系统性挖掘和整合,本文运用随机森林(RF)机器学习算法建立了吸附剂的24个特征描述符(涉及基底组分、孔隙性质、有机胺属性及实验条件等)与其CO2吸附容量之间的预测模型,获得了训练集决定系数R2=0.964、测试集R2=0.823的性能。单变量分析结果表明,固体胺吸附剂的CO2吸附容量受到多重因素影响,24个特征描述符与CO2吸附容量之间表现出明显的非线性关系。因此,本文进一步运用SHAP分析方法定量解析了机器学习模型的预测结果,揭示了决定固体胺吸附剂CO2吸附容量的关键参数。其中,有机胺负载量对吸附剂CO2吸附量的正向贡献度最高,可达0.46 mmol/g,其他重要的特征描述符还包括:测试所用CO2浓度、实验温度、吸附剂的BET表面积和孔隙率、基底的元素组成(如O、C、F)等。针对新型 DAC 固体胺吸附剂的开发,本文建议选择具有高孔隙率的杂化材料作为基底,并采用低聚合度的有机胺且实现其高负载量;同时,对吸附工艺也提出了一些建议,如采用更低的操作温度等。本研究聚焦新型DAC固体胺吸附剂的开发,有利于加快DAC负碳技术的开发与部署,助力“双碳”目标的顺利实现。
Objective Direct air capture (DAC) emerges as a promising negative emission technology to mitigate global warming. The performance of direct air capture adsorbents, particularly amine-based solid porous adsorbents, plays a critical role in the industrial deployment of DAC. Extensive experimental studies are conducted worldwide to investigate the CO2 capture capabilities of these adsorbents, generating a substantial volume of data. However, the CO2 capture performance is influenced by multiple parameters, requiring a systematic and comprehensive analysis to clarify the relationship between the structural and conditional parameters of amine-based adsorbents and their CO2 capture capabilities. Methods This study systematically compiled experimental data on amine-based solid DAC adsorbents from peer-reviewed scientific articles. Four different types of machine learning algorithms (random forest, artificial neural network, support vector machine, and ridge regression) were employed to construct a predictive model that correlated the features of amine-based solid adsorbents with their CO2 adsorption capacity values. In addition, the study utilized shapley additive explanations (SHAP) analysis to deconstruct the machine learning model’s predictive process, quantitatively revealing key parameters that determined the CO2 adsorption capacity of the adsorbents. Results and Discussions This study collected 629 valid data entries from 32 scientific publications, covering a wide range of CO2 capture capacities from 0 mmol/g to 5.0 mmol/g, to guide the design of new DAC adsorbents with enhanced CO2 capture performance. Each data entry was characterized by 24 descriptors, which encompassed information on the porous substrate components, textural properties, amine properties, and experimental conditions. An individual-variable analysis using the Pearson method revealed little linear correlation between the descriptors and CO2 capture capability, except for the amine loading in the adsorbents, with a Pearson correlation coefficient R = 0.543. Machine learning models were employed to uncover potential nonlinear and multivariable relationships. Four algorithms with good fitting capabilities and robustness against information noise were selected to build predictive models for the CO2 capacity of amine-based adsorbents, namely artificial neural network (ANN), support vector machine (SVM), ridge regression (Ridge), and random forest (RF). The dataset was split into training and test sets in a 4:1 ratio, and hyperparameters were optimized using grid search and validated with 5-fold cross-validation. After the optimization of hyperparameters, the RF model showed superior performance compared to the other selected models. The optimal RF model demonstrated the best performance in predicting CO2 adsorption capacity, with R2 = 0.823, MAE= 0.270, and RMSE= 0.372 in the test set. The performance of the RF model indicated that the 24 descriptors effectively covered the key factors that determined the CO2 capacity of amine-based adsorbents within the current experimental design space. A quantitative structure-property relationship (QSPR) analysis was conducted using the SHAP analysis method based on the mentioned reliable RF model. In this case, the SHAP method quantified the contribution of each descriptor of the DAC data entry to the output (predicted CO2 adsorption capacity) of the RF model, providing interpretability and insights into the model's decision-making process. The SHAP analysis results identified the most important descriptor influencing CO2 capacity, the amine loading, which exhibited a strong positive correlation. Other significant descriptors included the molecular weight of the incorporated amines (negative correlation), CO2 concentration (positive), porosity of the adsorbent (positive), and elemental contents in the substrate material, such as O (negative), C (positive), and F (positive). Accordingly, three experimental design strategies were proposed for further exploration of amine-based DAC adsorbents: 1) utilize substrates with high porosity, 2) avoid using amine-based polymers with excessively high molecular weight (> 10 000), and 3) select substrates containing C or F elements. These strategies contributed to the further enhancement of the CO2 adsorption performance of amine-based solid porous adsorbents and accelerated the development and deployment of DAC as a key negative emission technology. At the same time, the current limitations of chemical diversity and experimental space highlighted several areas and directions that required further research. For example, additional experimental studies were needed to investigate the CO2 capacity of amine-based adsorbents in humid and low-temperature environments, as well as the impacts of gas flow rate and gas mixture composition on the adsorbents’ CO2 capacities. In addition, an ideal DAC adsorbent should have met multiple criteria beyond CO2 capacity, such as efficient mass transfer and high stability in adsorption-desorption cycles. Currently available experimental data are still insufficient to address these aspects using machine learning strategies. Conclusions This research highlights the potential of machine learning in analyzing large-scale datasets to identify factors that influence the CO2 capture performance of DAC adsorbents. Critical factors were identified, including amine loading and adsorbent porosity. The findings provide insights that can guide the design of more effective DAC adsorbents and highlight areas requiring additional experimental research, particularly regarding the effects of environmental conditions and gas composition on adsorbent performance. This study contributes to the progress of DAC technologies as a feasible solution to achieving China's carbon dioxide peaking and carbon neutrality objectives by advancing the understanding of amine-based adsorbents.
| [1] |
United Nations Environment Programme.Emissions gap report 2023:Broken record-temperatures hit new highs,yet world fails to cut emissions (again)[R].Nairobi:United Nations Environment Programme,2023. doi:10.59117/20.500.11822/43922 |
| [2] |
|
| [3] |
|
| [4] |
|
| [5] |
|
| [6] |
|
| [7] |
|
| [8] |
侯正猛,罗佳顺,曹成, |
| [9] |
|
| [10] |
任世华,谢亚辰,焦小淼, |
| [11] |
|
| [12] |
陈彬,谢和平,刘涛, |
| [13] |
|
| [14] |
曹成,侯正猛,熊鹰, |
| [15] |
|
| [16] |
侯正猛,熊鹰,刘建华, |
| [17] |
|
| [18] |
|
| [19] |
|
| [20] |
|
| [21] |
|
| [22] |
|
| [23] |
|
| [24] |
|
| [25] |
|
| [26] |
|
| [27] |
|
| [28] |
|
| [29] |
张智渊,董越,邱雨晴, |
| [30] |
|
| [31] |
|
| [32] |
|
| [33] |
|
| [34] |
|
| [35] |
|
| [36] |
|
| [37] |
|
| [38] |
|
| [39] |
|
| [40] |
|
| [41] |
|
| [42] |
|
| [43] |
|
| [44] |
|
| [45] |
|
| [46] |
|
| [47] |
|
| [48] |
|
| [49] |
|
| [50] |
|
| [51] |
|
| [52] |
|
| [53] |
|
| [54] |
|
| [55] |
|
| [56] |
|
| [57] |
|
| [58] |
|
| [59] |
|
| [60] |
|
| [61] |
|
| [62] |
|
| [63] |
|
| [64] |
|
| [65] |
|
| [66] |
|
| [67] |
|
| [68] |
|
| [69] |
|
| [70] |
|
| [71] |
|
| [72] |
|
| [73] |
|
| [74] |
|
| [75] |
|
| [76] |
|
| [77] |
|
| [78] |
|
| [79] |
|
中国石油化工股份有限公司技术开发项目(323048)
/
| 〈 |
|
〉 |