Particle contamination detection in plasma etching process using Artificial Neural Network

Particle contamination on wafer during fabrication process has become major issue in semiconductor manufacturing. Particles on wafer can cause the circuit to malfunction affecting manufacturing cost and productivity. Currently, detection of particles is conducted after the etching process was completed. Therefore, developing a system where an early particles detection is crucial. In this study, a method for particle contamination detection in plasma etching process using Artificial Neural Network (ANN) is proposed. The proposed method comprised of four steps. First, data collection from Statistical Process Control (SPC) and Advanced Process Control (APC), Infineon Technology, Kulim. Three collected datasets are 45i01, 45i08 and 45il0. Secondly, collected datasets is pre-processed to remove missing values and unwanted attributes. Thirdly, feature selection is used to select the most relevant features that correlated with the number of particles contamination. Three feature selection techniques are used which are Minimum Redundancy Maximum Relevance (mRMR), Least-Square Feature Selection (LSFS) and Maximum Likelihood Feature Selection (MLFS). Finally, datasets with the selected features together with datasets without features selected are used as inputs for training and testing ANN in particles contamination detection. In this work, multilayer perceptron (MLP) network is used for ANN model. The datasets with and without feature selection is used to evaluate the performance of MLP network. Simulation results indicated that CNT WF V and ESC RF HOUR parameters are the most relevant features correlating to particle contamination as both features were voted by the three feature selection methods. Simulation results also showed that 45i01 and 45i 10 without feature selection method have lowest error and maximum R2 where 45i01 have 176.59 and 13.29 errors and R2 is 0.84. As for 45il0, errors are 165.77 and 12.88 and R2 is 0.82. 45i08 with feature selection method using MLFS have 291.51 and 17.07 errors and 0.75 R2.