UC Berkeley

The semiconductor manufacturing industry currently faces many challenges in terms of metrology and process control. With the delay of EUV and the advent of high aspect-ratio 3D structures, there is an increase both the number of complex processing steps and systematic/random errors, and optical resolution for metrology has now reached its limit for sub-14nm devices. The industry now requires real-time wafer-to-wafer control and in-line metrology, such as scatterometry or virtual metrology, for effective process monitoring.

Data models provide a quick and flexible way for integrating different forms of information. For example, in metrology, often times it is useful to combine sensor data, previous measurements, and other types of signals to extract the best possible measurement. Moreover, as the number of process steps continues to increase, explicit physical modeling of each step becomes extremely time-consuming and empirical data models will quickly become an effective alternative. In this dissertation, we discuss the application and usefulness of empirical data inference models in the context of W2W advanced process control, especially

focusing on wafer fault detection.

We first use virtual metrology, a type of in-line metrology technique, to determine whether the introduction of such data inference models is actually useful for the fab. Moreover, results show that the effective cost is determined by not only the model type and accuracy, but also the resulting false and missed alarm patterns. In the next chapter, we demonstrate an application of data models to fault detection by constructing a support vector machine classifier (SVM) and using only the diffraction signatures from scatterometry measurements to detect alarms. In the last chapter, we develop an algorithm for the SVM that allows one to choose the optimal false and missed alarm combination based on an asymmetric cost function. Moreover, our algorithm can be generalized for optimal hyperparameter selection

for any SVM problem.