Application of Constrained Experiment Design Scheme and Design Method Evaluation in PECVD BKM Process Development and Establishment of Predictive Models_Christy PanSpeakers: Christy Pan , Process Engineer, Applied Materials
the speech topic: Application of Constrained Experiment Design Scheme and Design Method Evaluation in PECVD BKM Process Development and Establishment of Predictive Model
Summary:
The film properties of PECVD applied in metal oxide processes are critical to display device performance. Film thickness uniformity, stress and deposition rate are all key process attributes. Exploring the complex science of metal oxide thin film deposition requires a structured and systematic approach to DOE in order to get the job done efficiently and smoothly while saving time and cost. Before the process of fine-tuning the DOE begins, conduct a gage R&R test to ensure good data quality and minimize the size of the DOE experiment. In addition, the gage R&R noise sigma is compared to the regression noise RMSE to isolate the measurement noise from the regression residual noise. Due to DOE budget constraints, we compared (1) Taguchi L27 DOE, (2) classical partial factorial DOE, (3) custom DOE, and (4) deterministic screening (DSD) based on 21-27 similar-sized experimental scales ) various response surface DOE methods in DOE. From the results, DSD showed the best orthogonality (correlation colormap), prediction uniformity, and power at the same or smaller experimental scale. The folded DSD design can greatly improve the rotatability of DOEs at the smallest experimental scale. Design constraints are also discussed in this article. We found a design constraint due to the physical limitations of the plasma, the effect of which can be minimized by designing the structure of the DSD. To identify a safer approach to risk in terms of design constraints, we also investigated power analysis, colormaps, scatterplot matrices, prediction variance, and design space proportions to evaluate the DSD design structure under design constraints. The next challenge is how to avoid any model overfitting. Based on least squares regression modeling, the predictive model can be modified by removing or adding factors at the desired alpha risk level. Nevertheless, it is difficult to clearly determine whether the threshold for the significance level of the effect term should be set at 0.05 or 0.10. We observed overfitting behavior when using higher p-value thresholds. ANOVA results can indicate whether overfitting is significant by adding more insignificant terms to the model. Lower ANOVA p-values and reduced risk of overfitting (higher adjusted R-squared) can be obtained after removing effect terms with higher p-values in the model. After establishing the transfer function, perform a global sensitivity analysis using the prediction profiler to rank the sensitivities of the input variables. The results of the JMP objective sensitivity analysis were consistent with the subjective quality functional expansion (House of Quality). To simulate long-term production general variation through transfer functions, we perform Monte Carlo simulations, incorporating noise variation for all key factors, to calculate the non-conforming defect rate. Robust design using JMP profiler to find insensitive process windows (optimized with first derivative). Model validation was performed by checking that the optimal process parameter settings were within the region of high predictive power based on the uniformity scatterplot. According to the Monte Carlo model, the optimal process settings can achieve ~0% off-spec performance. JMP tools are powerful and can be of great help in process improvement and validation.
ANOVADOEDSDModelingMonte Carlo SimulationTaguchi DOEexperimental designvariance analysisTaguchi DesignDefinitive Screening DesignMonte Carlo Simulationpredictive modeling
