Structural Feature-Based Fault-Detection Approach for the Recipes of Similar Products

The sensor signals (i.e., data streams of process parameters) of semiconductor processes exhibit nonlinear, multimodal trajectories with some common structural features. In this paper, we propose a process fault-detection approach based on the structural features of the sensor signals, such as the geometric shape, length, and height. The approach aims at constructing a shared univariate model and a multivariate model. The shared univariate model is set up for individual process parameters and clusters the process recipes of similar products. The result is a tree where the leaf nodes and intermediate nodes correspond to individual recipes and feature-based fault-detection criteria, respectively. The recipes with the same parent nodes share the criteria specified in the nodes. On the other hand, the multivariate model is constructed for a process recipe. It builds a Hotelling's $T^{2}$ that considers the correlations between the signal structures of the process parameters. We demonstrated that the test results of the two models using the data collected from a work-site etch process were encouraging.