基于数据驱动的晶体直径模型辨识方法研究

直拉法硅单晶生长是一个多场多相耦合、物理变化复杂,且具有大滞后和非线性现象的过程,基于单晶硅生长系统内部机理所构建的机理模型由于存在诸多假设而无法应用于工程实际。因此,本文以现有CL120-97单晶炉拉晶车间的长期、海量晶体生长数据为基础,忽略炉内复杂的晶体生长环境,对影响晶体直径的拉晶参数进行关联性分析及特征量化,探寻拉晶数据中所蕴藏的规律信息,进而建立基于数据驱动的BP神经网络晶体直径预测模型,并针对现有BP神经网络易陷入局部极小值的问题,采用遗传算法对BP神经网络的阈值和权值进行优化,以提高晶体直径预测的准确性。通过实际拉晶数据对模型预测结果进行验证,结果表明,对任意选取的8组拉晶数据进行直径预测,预测的平均相对百分比误差为0.095 71%,证明该模型对于等径阶段晶体直径的预测是可行的。

Research on Identification Method of Crystal Diameter Model Based on Data Driven

Czochralski silicon single crystal growth is a dynamic time-varying process with multi-field and multi-phase coupling, complex physical changes, nonlinearity and large hysteresis. However, mechanism models based on a large number of assumptions are difficult to apply in practice. Therefore, this article is based on long-term and massive crystal growth data from the CL120-97 single crystal furnace of crystal pulling workshop, which ignores the complex crystal growth environment in the furnace, analyzes the correlation of the crystal pulling parameters that affect the crystal diameter, mines the regular information contained in the data, and further builds a crystal diameter prediction model based on the BP neural network. Aiming at the problem that the existing BP neural network model is easy to fall into the local minimum, the genetic algorithm is used to optimize the weight and threshold of the BP neural network to improve the accuracy of the crystal diameter prediction of the algorithm. The model prediction results are verified by actual crystal pulling data. The results show that the average relative percentage error of prediction is 0.095 71% for diameter prediction of 8 groups of randomly selected crystal pulling data. It is proved that the model is feasible for the prediction of crystal diameter in the equal-diameter stage.