Speakers :
Jiaping Shen, Process Support Engineer, Applied Materials
The speech topic :
Application of JMP DOE analysis and modeling optimization in predicting epitaxial silicon growth rate
Speech abstract :
In semiconductor manufacturing, epitaxial silicon is widely used in applications such as wafer substrates, pMOS SiGe, nMOS SiP, and trench filling. The growth rate of the epitaxial silicon layer is affected by multiple steps and multiple process parameters. In order to more efficiently establish such a strong interactive prediction model, we need a thorough and comprehensive DOE evaluation and design a highly orthogonal DOE. This project uses historical data of epitaxial silicon growth rate as a starting point for design evaluation. For the establishment of a predictive model, the historical data structure is weak, the D efficiency is low, the design uniformity is poor, and the effect correlation and prediction variance are high. During the modeling, due to the lack of model freedom, the stepwise algorithm of RSM is unstable, resulting in poor model repeatability and too wide confidence intervals for the optimal design point. In addition, two pairs of strong interactions were observed in the RSM model: one pair of interactions may be attributed to the high correlation between the effects; while the other pair of interactions is related to the competitive mechanism of the process. Robust design and Monte Carlo simulation were then used for tolerance design, and the design space profiler was used for tolerance allocation analysis to simulate future technology requirements. In order to improve the existing DOE structure and optimize the prediction model at the lowest cost, we did not use a completely new DOE. Instead, we used the enhanced design method to improve the existing DOE structure through three methods: (1) removing non-orthogonal data; (2) default enhancement algorithm; (3) center enhancement algorithm. Through a very comprehensive enhancement design, JMP recommended the best four enhancement data points, which ultimately improved the DOE structure, greatly improved the efficiency of using historical data, and significantly shortened the process development cycle and cost.
