金属铜具有优秀的催化性能，能将二氧化碳催化还原为各类碳氢化物和碳氧合物，因而被广泛用作二氧化碳还原反应的催化剂。然而，以金属铜催化的单一产物反应路径中往往存在较低的电流密度和法拉第效率，因此，新型二氧化碳还原反应催化剂的开发及调控吸引了广泛的关注。在众多新型催化剂中，一种将低浓度的掺杂金属原子分散于基底金属表面而形成的单原子合金催化剂，凭借其独特的物理特性和几何构型，打破了多相催化过程比例关系的限制，并有助于提高贵金属利用率。在二氧化碳还原反应过程中，一氧化碳吸附能是催化活性和选择性的重要指标。结合机器学习算法和根据第一性原理计算建立的一氧化碳吸附能数据集，本研究建立了多个机器学习模型，从不同的途径系统地完成了对铜以及铜基单原子合金表面一氧化碳吸附能的预测。在铜表面一氧化碳吸附能预测模型中，本研究建立了两个机器学习模型，以基础的体系结构和元素信息作为输入，对铜表面和团簇体系总能量以及一氧化碳吸附的铜表面和团簇体系总能量进行预测，并据此得到对应体系的一氧化碳吸附能。在模型构建之前的数据处理过程中，本研究设计并运用了多种采样手段，并证实了训练集的数据分布会对机器学习模型性能造成显著影响。结合合理的数据采样和机器学习算法，我们保证了模型在不同温度和构型的体系之间的迁移性。在铜基单原子合金表面一氧化碳吸附能预测模型中，本研究进行了详细的数据处理和特征工程。该模型以人工设计的特征集作为输入，直接以对应体系的一氧化碳吸附能为目标进行预测。通过系统的特征分析，我们发现了对体系一氧化碳吸附能影响最为突出的特征。基于迁移学习的原理，我们提出了一种跨组别学习策略，使得模型在各类单原子合金体系之间的迁移性大幅提高。
Cu has excellent catalytic performance and can reduce CO2 to various hydrocarbons and oxygenates, therefore it is widely used as a catalyst for CO2 reduction reaction (CO2RR). However, a specific product reaction pathway typically exhibits low current density and Faraday efficiency. Consequently, the design and modulation of novel catalysts for CO2RR have attracted extensive attention. Single atom alloy (SAA) catalyst is formed by dispersing low concentration of doping metal atoms on the surface of host metal. By virtue of the physical characteristics and geometric uniqueness of SAAs, it breaks the limits of scaling relationship among different intermediates and improves the utilization rate of noble metals. CO adsorption energy is an important indicator of catalytic activity and selectivity in CO2RR. Combining machine learning algorithms and CO adsorption energy dataset collected by ab initio calculation, several machine learning models were established in this study to predict the CO adsorption energy on the surface of Cu and Cu-based SAAs from different approaches.For the purpose of predicting CO absorption energies on Cu surfaces, two machine learning models were established on the structure and element information of the chemical systems. The models predict the total energies of Cu systems and CO absorbed Cu systems, and thus obtained CO absorption energies of the corresponding systems. So as to pre-process the dataset, a variety of sampling methods were designed and applied, and it was confirmed that data distribution of the training set has a significant impact on the performance of a machine learning model. With reasonable data sampling methods and machine learning algorithms utilized, the model considerably maintained both structure and temperature transferability. As for predicting CO adsorption energies on Cu-based SAAs, detailed data pre-processing and feature engineering were carried out. The predictive model takes artificially designed feature set as input and directly predicts CO adsorption energies of the corresponding systems. According to the feature analysis, we found the most prominent features that influenced CO adsorption energies on Cu-based SAA surfaces. Inspired by transfer learning, we proposed a cross-group learning strategy, which greatly improved the model transferability among Cu-based SAAs with different alloying elements.
