Research On Data-based Prediction Approaches Of Material Removal Rate In Chemical Mechanical Polishing Process Of Semiconductor Chip

Semiconductor chip manufacturing technology is one of the most significant technologies in the information industry,which plays a vital role in accomplishing industrialization of the country.In the semiconductor chip manufacturing process,chemical mechanical polishing technology is currently the only surface precision processing technology that can achieve the global planarization.The material removal rate of semiconductor wafer refers to the thickness of the erased workpiece in a unit time,which is a key indicator in the chemical mechanical polishing process.Accurate prediction of the material removal rate is of great significance to the semiconductor wafer process modeling of the chemical mechanical polishing technology,the manufacturability design of the integrated circuit layout,and the improvement of the semiconductor chip quality.From the perspective of the data-based,this paper develops machine learning and deep learning approaches to analyze historical data measured by sensors in the chemical mechanical polishing process to predict the material removal rate.The specific implementation process is to firstly extract the data collected by the sensors during the chemical mechanical polishing process of the semiconductor wafer.In addition,the data preprocessing is performed,and the process parameters that have a significant impact on the material removal rate are obtained through the feature selection,which are used as the input of the learning algorithm model.Furthermore,the machine learning approaches,namely,random forest,extreme random tree,gradient boosting tree,extreme gradient boosting optimized by genetic algorithm and deep learning approaches,namely,convolutional neural network and residual convolutional neural network are developed to train the models,respectively.Finally,the trained models are employed to predict the material removal rate,and the model prediction performance is evaluated by corresponding indicators.According to the comparison of the experimental results,the best prediction approach is selected.The public dataset provided by the PHM2016 Data Challenge is used to verify the material removal rate prediction approaches based on the machine learning and the deep learning.Experimental results show that the deep learning approach based on the residual convolutional neural network proposed in this paper has the best prediction performance in predicting material removal rate outperforms all the existing literature.