Atomic descriptors generated from coordination polyhedra in crystal structures

We developed atomic descriptors from local crystal structures, which will facilitate researchers’ use of machine learning to predict the properties of inorganic materials via materials informatics. We applied singular value decomposition to the occurrence matrix of local coordination polyhedra in crystal structures. We generated two atomic descriptors, each based on the coordination atoms and topology of the coordination polyhedra. As a result of atomic clustering using these descriptors, the composition descriptor proposed in previous research depends on the similarity between same-group atoms in the periodic table. In contrast, when using our original descriptors based on the coordination atoms and topology of the coordination polyhedra, the similarity between adjacent atoms in the periodic table as well as the similarity between same-group atoms was pertinent. When we used machine learning to predict the formation energy and band gap using these descriptors as inputs, the prediction accuracy and generalization ability increased compared with using a physical property descriptor.

本研究从局域晶体结构出发构建原子描述符，旨在助力研究者借助材料信息学方法，通过机器学习预测无机材料的各项性能。我们针对晶体结构中的局域配位多面体出现矩阵应用了奇异值分解（singular value decomposition）。据此生成了两类原子描述符，二者均基于配位多面体的配位原子与拓扑结构。借助此类描述符开展原子聚类实验后可发现，既往研究提出的成分描述符仅能反映元素周期表中同族原子间的相似性。与之形成对比的是，当使用我们基于配位原子与配位多面体拓扑结构提出的原创描述符时，元素周期表中相邻原子的相似性以及同族原子的相似性均具备关联性。以该类描述符作为输入开展机器学习预测形成能与带隙的实验中，相较于采用物理属性描述符的方案，模型的预测精度与泛化能力均得到了提升。