Learning Materials Properties from Orbital Interactions
Learning Materials Properties from Orbital Interactions
复制标题
从轨道相互作用中学习材料特性
DOI：
10.1088/1742-6596/1290/1/012012
复制
发表时间：
2019
期刊：
影响因子：
--
通讯作者：
H. Dam
中科院分区：
文献类型：
--
作者：
T. Pham;Tran;Van;H. Kino;T. Miyake;H. Dam
发展了轨道场矩阵(OFM)描述符，重点是原子轨道，用于表示多元素化合物数据集中的物质结构。这些描述符基于原子价、壳层电子及其配位。除了原始的OFM和OFM1，在这项工作中，我们还提出了另一个版本OFM0，它是没有原子距离信息的OFM1，用于预测未优化结构的性质。我们关注晶系的形成能和相稳定性，而原子化能则是对分子的考察。为了强调识别具有相似性质的材料的能力，这里用决策树(DT)回归、随机森林(RF)回归和核岭回归(KRR)系统地检验了OFM、OFM1和OFM0的适用性。我们表明，OFM描述子家族能够很好地为固体和分子的性质建立预测模型。一个DT和一个树林(RF)模型的精度与KRR模型相当。由OFM_1估计的具有拉普拉斯核的KRR得到了最准确的预测，其形成能、相稳定性和原子化能的平均绝对误差分别为0.072 eV/原子、0.059 eV/原子和6.74千卡/摩尔。没有原子距离的OFM_0也给出了可以接受的预测，其MAE分别为0.090 eV/原子、0.069 eV/原子和7.77kcal/mo1。结果表明，我们的描述符对于寻找相似的材料非常有用。
Orbital field matrix (OFM) descriptors were developed with an emphasis on atomic orbitals for representing material structures in datasets of multi-element compounds. The descriptors were based on atomic valence shell electrons and their coordination. In addition to original OFM and OFM1 which is OFM with a column representing information on the center atom, in this work, we present another version, named OFM0, which is OFM1 without information on atomic distances, for predicting the properties of unoptimized structures. We focus on formation energy and phase stability of crystalline systems, while the atomization energy is examined for molecules. With the emphasis on the ability to identify materials with similar properties, here, the applicabilities of OFM, OFM1, and OFM0 are systematically examined with decision tree (DT) regression, random forest (RF) regression, and kernel ridge regression (KRR). We show that the family of OFM descriptors are highly capable to build predictive models for the properties of solids and molecules. The accuracy of a DT and a forest of trees (RF) is comparable to that of the KRR models. The KRR with a Laplacian kernel estimated by OFM1 yields the most accurate predictions, with the formation energy, phase stability, and atomization energy having mean absolute errors (MAEs) of 0.072 eV/atom, 0.059 eV/atom, and 6.74 kcal/mol, respectively. The OFM0 without atomic distances also yields acceptable predictions with respective MAEs of 0.090 eV/atom, 0.069 eV/atom, and 7.77 kcal/mol. The results imply that our descriptors are highly useful to find similar materials.
影响因子：
5.5
作者：
Lam Pham T;Kino H;Terakura K;Miyake T;Tsuda K;Takigawa I;Chi Dam H
通讯作者：
Chi Dam H