A dynamic sampling strategy based on confidence level of virtual metrology predictions

Metrology is a costly and time consuming activity in semiconductor fabrication; for this reason, Dynamic Sampling strategies and Virtual Metrology approaches have proliferated in the past recent years. Both Dynamic Sampling strategies and Virtual Metrology techniques aim at minimizing the amount of performed measures while keeping acceptable levels of production quality. In this work we study a Dynamic Sampling scheme recently proposed in literature that takes into account the availability of a Virtual Metrology module in the advanced process control architecture. The idea supporting the investigated strategy is based on the availability of a confidence level in the Virtual Metrology predictions; in our implementation of this scheme, this is achieved by exploiting a popular Machine Learning approach for supervised learning tasks, called Random Forests. The aforementioned scheme is tested on a real industrial dataset related to Plasma Etching and it is compared with classical metrology strategies.