npj Computational Materials

A high-throughput data analysis and materials discovery tool for strongly correlated materials（用于强相关材料的高通量数据分析和材料发现工具）
Hasnain Hafiz, Adnan Ibne Khair, Hongchul Choi, Abdullah Mueen, Arun Bansil, Stephan Eidenbenz, John Wills, Jian-Xin Zhu, Alexander V. Balatsky & Towfiq Ahmed
npj Computational Materials 4:63 (2018)
doi:s41524-018-0120-9
Published online:22 November 2018
Abstract| Full Text | PDF OPEN

摘要:由于自旋轨道耦合效应、电子-电子相互作用和局域f电子与流动的导电电子之间的复杂相互作用，含f电子系统的建模具有挑战性。这种复杂性不仅丰富了材料的电子特性，也使其适用于各种技术应用。在此背景下，我们提出并实现了一种数据驱动的方法来帮助材料的发现过程。通过部署最先进的算法和查询工具，我们使用基于现有锕系和镧系化合物的大型模拟数据集来训练我们的学习模型。这样获得的机器学习模型可用于搜索和寻找具有所需电子和物理性能的新型稳定材料。本研究讨论了f电子数据库的基本结构，以及我们对结构数据文件进行清理和修正的方法。该数据库的应用实例包括成功预测双钙钛矿的稳定超结构，以及在f电子基材料的强相关特性中识别一些物理关联的趋势。　　

Abstract:Modeling of f-electron systems is challenging due to the complex interplay of the effects of spin–orbit coupling, electron–electron interactions, and the hybridization of the localized f-electrons with itinerant conduction electrons. This complexity drives not only the richness of electronic properties but also makes these materials suitable for diverse technological applications. In this context, we propose and implement a data-driven approach to aid the materials discovery process. By deploying state-of-the-art algorithms and query tools, we train our learning models using a large, simulated dataset based on existing actinide and lanthanide compounds. The machine-learned models so obtained can then be used to search for new classes of stable materials with desired electronic and physical properties. We discuss the basic structure of our f-electron database, and our approach towards cleaning and correcting the structure data files. Illustrative examples of the applications of our database include successful prediction of stable superstructures of double perovskites and identification of a number of physically-relevant trends in strongly correlated features of f-electron based materials.

Editorial Summary

Materials databases: Checked and analyzed （材料数据库：检查和分析）

f电子系统具有诸多有趣的特性，而这些特定化合物的数据库可以帮助新材料的发现。来自美国东北大学的Hasnain Hafiz教授和洛斯阿拉莫斯国家实验室的科研人员在本工作中提供了f电子结构数据库。与其它数据库不同，该计算数据是用所有电子生成的，从而更好地描述这些材料。实验信息有时会丢失一些重要的数据，但本研究使用人工神经网络来纠正这种不完整性，从而能够正确地确定晶体系统（精确度达99.1％）。为验证数据库的可靠性，该工作成功找出了8种已知的双钙钛矿（AA'BB'CC'）材料，并额外预测了4种未知的稳定双钙钛矿材料。此外，数据库中的电子结构分析工具还发现了元素周期表中f电子的局域化趋势。这种数据驱动的方法可以促进新型f电子材料的发现，并带来诸多新的应用。

F-electron systems can possess interesting properties, and a database on these specific compounds could aid materials discovery. Here, Hasnain Hafiz at Northeastern University, and colleagues at Los Alamos National Lab, present the f-electron structure database. In contrast to other databases, computational data is generated with all electrons, resulting in a better description of these materials. Experimental information can sometimes miss essential data, but here an artificial neural network is used to correct this incompleteness, enabling correct determination (with 99.1% accuracy) of a crystal system. To verify the database, eight known double perovskites (AA′BB′CC′) were successfully found, and four unknown stable double perovskites were predicted. Moreover, electronic structure analysis tools in the database identified f-electron localization trends across the periodic table. This data-driven approach could drive the discovery of new f-electron materials, and lead to new applications.