Abstract:

Zeolite exhibits great potential for CO adsorption due to its well-organized pore structure and tunable chemical composition. However, numerous influencing factors hinder the design and optimization of adsorbents using traditional experimental methods and theoretical calculations. In this paper, a machine learning strategy of combining the structural characteristics of molecular sieve adsorbents with experimental data was proposed, and a comprehensive data set was constructed to express the structure-activity relationship between molecular sieve structure and CO adsorption performance by integrating literature data and zeolite structure information. Four integrated learning algorithms, gradient enhanced decision tree, random forest, extreme random tree and extreme gradient lifting (XGB), were used to predict CO adsorption, and nested cross validation was used to ensure the prediction accuracy of the evaluation model. The results show that XGB model performs best and shows excellent prediction accuracy. The fragmentation analysis of molecular sieve structure characteristics showed that triangular crystal system, Fd\overline{\mathrm{3}}m space group, near circular pores and interconnected cage structure were conducive to CO adsorption. The alkaline earth metals such as Ca²⁺ and Ba²⁺ in the framework of molecular sieve are conducive to the adsorption of CO. In the metal ion modified molecular sieve, the Cu(Ⅰ) loading has the most significant effect on the CO adsorption performance, and the optimal loading is 13%—15% by weight. When the specific surface area of the adsorbent is 230—400 m²/g and the total pore volume is 0.10—0.15 cm³/g, the adsorbent has the best performance. Its limited pore size is around 4.5 Å and the maximum pore size is in the range of 5.0—5.5 Å, showing relatively good CO adsorption performance. This study provides scientific guidance for the rational design and efficient screening of molecular sieve based co adsorbents.

Key words: molecular sieve, adsorbent, machine learning, carbon monoxide, adsorbent design

摘要：

分子筛因其规整的孔道结构和可调控的化学组成在CO吸附领域展现出巨大的应用潜力，但众多影响因素致使通过传统实验方法和理论计算进行吸附剂设计与优化的效率低下。本研究提出分子筛吸附剂的结构特征片段化与实验数据结合的机器学习策略，整合文献数据与沸石结构信息，构建了表达分子筛结构与CO吸附性能构效关系的综合数据集。采用梯度增强决策树、随机森林、极端随机树和极端梯度提升（XGB）四种集成学习算法预测CO吸附，并运用嵌套交叉验证确保评估模型的预测准确性。结果表明，XGB模型表现最优，显示出优异的预测精度。通过分子筛结构特征片段化分析发现，三角晶系、Fd\overline{\mathrm{3}}m空间群、近圆形孔道和相互连接的笼状结构有利于CO吸附。分子筛骨架中的Ca²⁺和Ba²⁺等碱土金属离子有利于CO吸附。在金属离子改性分子筛中，引入的Cu(Ⅰ)负载量对CO吸附性能的影响最为显著，最优负载量为吸附剂总质量的13%~15%。吸附剂的比表面积在230~400 m²/g、总孔体积在0.10~0.15 cm³/g时表现最佳，其限制性孔径在4.5 Å（1 Å=0.1 nm）附近、最大孔径在5.0~5.5 Å内表现出较好的CO吸附性能。本研究为分子筛基CO吸附剂的合理设计和高效筛选提供了科学指导。

关键词: 分子筛, 吸附, 机器学习, 一氧化碳, 吸附剂设计