尖晶石氧化物能量和结构的第一性原理计算和机器学习

正尖晶石氧化物AB正尖晶石氧化物AB_2O4结构,可通过对A和B位点分别置换73种元素产生5 329种原子构型.利用高通量第一性原理方法计算了这5 329种正立方尖晶石氧化物结构的形成能和晶格常数,并提出了 "中心-环境"(center-environment,CE)特征模型,构建同时包含局部成分和结构信息的特征作为机器学习(machine learning,ML)方法的输入变量.结合第一性原理计算数据,使用随机森林算法开发了机器学习模型,准确有效地预测了尖晶石氧化物结构的形成能和晶格常数.通过比较机器学习预测的假想结构与实验结构的形成能,预测出了 361种更稳定的新尖晶石氧化物结构.讨论了与尖晶石氧化物稳定性相关的"好"与"坏"的构成元素,有助于指导实验合成新的、稳定的尖晶石氧化物.

First-principles computation and machine learning of the energies and structures of spinel oxides

The formal spinel oxides AB$_{2}$O$_{4}$ can have 5 329 configurations by substituting A and B with 73 elements. The first-principles method was applied to calculate the formation energies and lattice constants of 5 329 spinel oxides. To develop efficient machine learning (ML) methods, centre-environment (CE) feature models were proposed to construct the input variables of the ML methods containing local composition and structure information. Based on the first-principles computational data, random forest algorithm was used to develop an ML model to predict the formation energies and lattice constants of spinel oxides. By comparing the formation energies of hypothetical and experimental structures predicted by ML, 361 new and more stable spinel oxides were discovered. The “good” and “bad” stabilisation elements were disscussed, which helped in guiding theexperimental synthesis of novel stable spinel oxides.