Liu Zeyu, Jiang Meng, Luo Tengfei
Department of Aerospace and Mechanical Engineering, University of Notre Dame, Notre Dame, IN 46556, USA.
Department of Computer Science and Engineering, University of Notre Dame, Notre Dame, IN 46556, USA.
Sci Adv. 2020 Nov 4;6(45). doi: 10.1126/sciadv.abd1356. Print 2020 Nov.
Electron properties are usually easier to obtain than phonon properties. The ability to leverage electron properties to help predict phonon properties can thus greatly benefit materials by design for applications like thermoelectrics and electronics. Here, we demonstrate the ability of using transfer learning (TL), where knowledge learned from training machine learning models on electronic bandgaps of 1245 semiconductors is transferred to improve the models, trained using only 124 data, for predicting various phonon properties (phonon bandgap, group velocity, and heat capacity). Compared to directly trained models, TL reduces the mean absolute errors of prediction by 65, 14, and 54% respectively, for the three phonon properties. The TL models are further validated using several semiconductors outside of the 1245 database. Results also indicate that TL can leverage not-so-accurate proxy properties, as long as they encode composition-property relation, to improve models for target properties, a notable feature to materials informatics in general.
电子性质通常比声子性质更容易获得。利用电子性质来帮助预测声子性质的能力，因此可以极大地有益于通过设计用于热电和电子等应用的材料。在这里，我们展示了使用迁移学习（TL）的能力，即从对1245种半导体的电子带隙训练机器学习模型中学到的知识被转移，以改进仅使用124个数据训练的模型，用于预测各种声子性质（声子带隙、群速度和热容量）。与直接训练的模型相比，对于三种声子性质，迁移学习分别将预测的平均绝对误差降低了65%、14%和54%。迁移学习模型使用1245个数据库之外的几种半导体进一步验证。结果还表明，迁移学习可以利用不太准确的代理性质，只要它们编码成分-性质关系，来改进目标性质的模型，这通常是材料信息学的一个显著特征。
