Big-data-accelerated aperiodic Si/Ge superlattice prediction for quenching thermal conduction via pattern analysis
Yida Liu, Run Hu, Yan Wang, Jinglong Ma, Zhangcan Yang, Xiaobing Luo
Abstract
Thermal conductivity of material is one of the basic physical properties and plays an important role in manipulating thermal energy. In order to accelerate the prediction of material structure with desired thermal property, machine learning algorithm has been widely adopted. However, in the optimization of multivariable material structure such as different lengths or proportions, the machine learning algorithm may be required to be reconducted again and again for each variable, which will consume a lot of computing resources. Recently, it has been found that the thermal conductivity of aperiodic superlattices is closely related to the degree of the structural randomness, which can also be reflected in their local pattern structures. Inspired by these, we propose a new pattern analysis method, in which machine learning only needs to be carried out for one time, and through which the optimal structure of different variables with low thermal conductivity can be obtained. To verify the method, we compare the thermal conductivities of the structure obtained by pattern analysis, conventional machine learning, and previous literature, respectively. The pattern analysis method is validated to greatly reduce the prediction time of multivariable structure with high enough accuracy and may promote further development of material informatics.