The key performance indicators in Chemical Mechanical Planarization (CMP) processes are usually assessed by measuring the material removal rate (MRR) and Within-Wafer-Nonuniformity (WIWNU), which are vitally dependent on the processing variables including down pressure, wafer rotation, polishing pad rotation, polishing table rotation, slurry flow, and the condition of the polishing pad etc. MRR is critical to the WIWNU also since MRR can infer the end-point in the polishing process. In this study, empirical approaches were conducted to model the MRR with the production CMP settings. With the collected data from real semiconductor manufacturing processes, correlation and principle component analysis (PCA) were conducted to select the features mostly related to the CMP processes, then neural network (NN) and adaptive neuro fuzzy inference system (ANFIS) based models were proposed to understand processing variables in CMP process and estimate the MRR. The NN and ANFIS models were compared on the performance metrics of 1) mean square error (MSE), and determination coefficient (R2) based on bootstrap. The bootstrap based evaluation shows that NN achieved a MSE of 9.68e03 with the R2 value of 0.81 in the training stage and MSE of 9.59e3 with the R2 value of 0.81 in the validation stage; ANFIS achieved a MSE of 126.24 with the R2 value of 0.9102 in the training stage and MSE of 6.17e4 with the R2 value of 0.3133 in the validation stage. The empirical models are promising to be integrated with the data-driven based control of CMP processes.

This content is only available via PDF.
You do not currently have access to this content.