基于优化插值与差值神经网络算法的硅片刻蚀深度预测模型

Prediction model of silicon wafer etching depth based on optimal insert-value and difference-value neural network algorithm

针对半导体加工工艺复杂、成本高、工艺数据量少,一般的人工神经网络(ANN)算法无法准确预测其加工工艺性能的问题,提出一种基于优化插值与差值神经网络(OIDNN)算法的适用于小样本的硅片刻蚀深度预测模型。首先,分别由实验得到刻蚀深度的实验数据,由计算机辅助设计(TCAD)技术仿真得到刻蚀深度的模拟数据,并划分为训练集、验证集和预测集;将TCAD模拟数据作为额外输入参数插入ANN1模型,同时,将实验数据与TCAD模拟数据的差值作为ANN2模型的输出参数,得到两份预测结果;最后将两份预测结果作为输入参数,经ANN3模型训练选择权重,得到最终预测结果。OIDNN算法在不同大小的样本数量下,所得预测刻蚀深度和实验刻蚀深度之间平均的均方误差(MSE)为0.009 5μm,相较于ANN减小80%以上,相较于自适应权值神经网络(AWNN)减小85%以上。实验结果表明,所提模型可以有效提高预测的准确度,提高算法的收敛速度,并且适用于小样本的工程应用场景。

Due to the complexity of semiconductor processing technology,high cost and small amount of process data,the general Artificial Neural Network(ANN)algorithm cannot accurately predict the processing performance. Here,a prediction model of silicon wafer etching depth based on Optimal Insert-value and Difference-value Neural Network(OIDNN)algorithm were proposed,which is suitable for small samples. Firstly,the experimental data of etching depth was obtained by experiment,and the simulation data of etching depth was obtained by Technology Computer Aided Design(TCAD)simulation,both data was divided into training set,test set and prediction set. TCAD simulation data was inserted into ANN1 model as additional input parameters,and the difference between experimental data and TCAD simulation data was taken as the output parameters of ANN2 model to get two prediction results. Finally,the two prediction results were taken as the input parameters,the weights were selected by ANN3 model training,and the final prediction results were obtained. The experimental results show that the Mean Square Error(MSE) between predicted etching depth and experimental etching depth is 0. 009 5 μm,which is 80% less than that of ANN and is 85% less than that of AWNN(Adaptive Weighting Neural Network). It is proved that this model can effectively improve the accuracy of prediction,accelerate the convergence speed of the algorithm,and is suitable for the small sample engineering application scenarios.