Machine Learning Based Electronic Structure Predictors in Single-Atom Alloys: A Model Study of CO Kink-Site Adsorption across Transition Metal Substrates

This work reports on a comprehensive analysis of the predictive capacity and underlying physicochemical trends provided by d-band based electronic structure features as applied to single-atom alloys (SAAs). Taking CO adsorption energies at kink sites as a model framework, SAA adsorption trends are examined across a range of substrates with vastly differing intrinsic CO adsorption trends. Through this approach, it is demonstrated that SAA adsorption properties can be highly transferable, often displaying atom-like behavior independent of the host substrate, particularly in groups 6 through 12 of the periodic table. The predictability of such SAA behavior is found, however, to be highly qualitative for single d-band based electronic structure features. Nevertheless, it is shown that predictive capacity can be greatly improved through the creation of a feature space comprised of as few as 8 electronic structure features. Intriguingly, following the framework of Hammer and Nørskov, the machine learning accuracy of d-band based electronic structure features is shown to be sensitive to the atomic configuration diversity present in the training ensemble with model accuracy systematically improving through restrictions in the configurational space. More directly, it is shown that elements to the far left of the transition metal block such as Zr and Hf may exhibit CO binding properties comparable to Cu in the CO2 reduction reaction. However, impurities from groups 6–10 are demonstrated to overbind in a highly transferable manner in line with established pure substrate trends and are likely to act as unwanted posing species concerning CO and the overall CO2 reduction reaction. The results of this work broadly lay out the predictive capabilities of d-band features as applied to SAAs, as well as their propensity for exhibiting transferable binding properties among d-band substrates.本工作针对基于d带（d-band）电子结构特征在单原子合金（single-atom alloys, SAAs）中的预测能力与内在理化趋势展开全面分析。以扭折位点处的一氧化碳（CO）吸附能作为模型框架，针对一系列本征一氧化碳吸附趋势差异显著的基底，系统考察了单原子合金的吸附演化规律。研究表明，单原子合金的吸附性能具备极强的可迁移性，通常展现出与载体基底无关的类原子行为，尤其在元素周期表第6至12族元素构成的基底中表现显著。不过，仅依靠单一d带电子结构特征时，此类单原子合金行为的可预测性往往仅为定性层面。但研究证实，通过构建仅包含8种电子结构特征的特征空间，模型的预测能力可得到大幅提升。值得注意的是，依托哈默与诺斯柯夫（Hammer and Nørskov）的理论框架，基于d带电子结构特征的机器学习精度对训练集内的原子构型多样性具有显著敏感性：模型精度随构型空间的约束筛选而系统性提升。更直接的结论是，过渡金属区左侧远端的元素（如锆（Zr）和铪（Hf））在二氧化碳还原反应（CO₂ reduction reaction）中，可能展现出与铜（Cu）相当的一氧化碳结合性能。然而，第6至10族的杂质元素会以高度可迁移的方式过度结合一氧化碳，这与已报道的纯基底吸附趋势一致，且此类杂质大概率会成为干扰一氧化碳及整体二氧化碳还原反应的有害物种。本研究成果系统阐明了d带特征应用于单原子合金的预测能力，以及其在d带基底间展现可迁移结合特性的内在倾向。