Research On Prediction Method Of Wafer Surface Material Removal Rate Based On Improved Deep Belief Network

In the Chemical Mechanical Polishing(CMP)process,the Material Removal Rate(MRR)of wafer surface is a key measure.MRR will profoundly affect the final quality and yield of the wafer product.However,currently,MRR can only be obtained after the CMP is completed,which limits the improvement of wafer quality and hinders the exploration and understanding of the CMP mechanism.If the MRR was predicted accurately in the CMP process,it would be helpful to deduce the CMP mechanism and improve the quality control level of wafer.The study is also helpful to promote the rapid development of China’s semiconductor industry.Therefore,this paper summarizes and analyzes the existing MRR prediction models,and proposes a method that combines Deep Belief Networks(DBN),Online Sequential Extreme Learning Machine(OS-ELM)and improved Particle Swarm Optimization(PSO)model.It aims to further improve the accuracy of MRR prediction in the CMP.First,based on the nonlinear and high-dimensional characteristics of CMP data,this paper analyses DBN and OS-ELM,and builds a DBN-ELM-based MRR prediction model.In order to improve the DBN’s lack of nonlinear ability and long training time,then the prediction performance of the model is improved.In order to verify the validity of the MRR prediction model based on DBN-OSELM,this paper uses the data set in the Fault Prediction and Health Management Data Competition for simulation.The simulation results show that the R~2,MAE and RMSE of the DBN-OSELM model are0.99248、1.78 and 2.56,respectively.Compared with the DBN model,the DBN-OSELM model has improved all predictors.Secondly,in view of the difficulty of parameter adjustment of the DBN-OSELM model,this paper uses the improved PSO algorithm to propose a greedy parameter optimization method,and builds a MRR prediction model based on PSO-DBN-OSELM.The model combines the powerful feature extraction ability of DBN,the strong nonlinear ability of OS-ELM and the characteristics of online update.It also avoids the uncertainty of manual parameter adjustment and achieves a better prediction effect.The simulation results show that the optimization time of the greedy optimization scheme is 27%of the usual scheme,an d the final RMSE is reduced by 5.5%compared with the usual scheme,which shows that it has higher optimization efficiency.And the R~2,MAE and RMSE of the PSO-DBN-OSELM model are 0.99354,1.68,and 2.41,respectively.Compared with the DBN-OSELM model,the prediction indicators are further improved.Moreover,compared with MRR prediction models in other literature,the lowest RMSE is 2.59,this model also has higher prediction accuracy.Finally,based on the PSO-DBN-OSLEM model,this paper builds a prototype system.It is useful to predict the removal rate of wafer surface material and promote the realization of MRR online prediction in the CMP proces s.Also it can assist researchers to adjust the parameters of CMP process in real-time,then achieve the purpose of increasing wafer production capacity.