Predicting component performance during semiconductor manufacturing is challenging, and this unpredictability can lead to reduced yields and increased costs. This report aims to build a prediction model that uses online inspection data (including 3 response variables ( Y ) and more than 300 predictor variables ( X ), with a data set of approximately 500 data points) to predict component performance.
The project was evaluated using the JMP multivariate platform, which showed relatively weak YY competition, strong YX causality, and severe XX multicollinearity. Both the standard least squares and stepwise algorithms showed severe overfitting risks when all 300+ predictors were considered. Therefore, we applied four data mining techniques, from which a few key X variables were selected for RSM modeling, including: response screening, partitioning, neural network, and predictor screening. Since only regular JMP software was used, we used the contribution pattern and ranking of predictor screening as a reference for evaluating the other three data mining platforms. Among them, the predictor screening algorithm (random forest) can solve the recursive partitioning problem and make the contribution pattern more uniform.
Based on the contribution patterns of predictor screening, we ran adaptation models, neural networks, and partitioning models to predict the performance of the three devices. This JMP project successfully improved our predictive capabilities and gained valuable insights from augmented reality applications. At the same time, reliable prediction models can also significantly shorten the new product development cycle. The results of this study can provide an important reference for the semiconductor manufacturing industry to improve the prediction capabilities of component performance.
