<title>Monitoring of a RTA process using multi-PCA</title>

Abstract

Process faults usually lead to changes in the normal relationship among process variables. These changes can be detected by a Principle Component Analysis (PCA) model based on the data from normal batches of operation. Therefore, monitoring process variables via a PCA model may lead to the earlier detection of process fault than traditional SPC method which depends on periodic information from test wafers. However, PCA is a linear method, and does not explain the relationship among process variables with time. Because the relationship among process variables for a wafer processing equipment is both nonlinear (relationship changes dramatically from one processing step to another) and time dependent (dependent on different processing steps), applying PCA method directly to the monitoring of such processes may prove to be difficult. A multi-PCA modeling technique is proposed in our study of process monitoring and fault detection for a semiconductor manufacturing tool. A series of local PCA models can be built with each local model only describing a local relationship among process variables at a particular time. During monitoring, if an observation from the k th sample of the current run is consistent with the local PCA model built on data around the k th sampling time of previous normal runs, this observation can be considered normal. A method is also proposed to eliminate redundant local PCA models. The proposed method has been implemented for the monitoring of a commercial Rapid Thermal Anneal (RTA) tool. The RTA process is a typical single wafer process. For the same product, all wafers should be processed according to the same recipe. First, the data from normal lots were collected and verified by wafer electrical test (WET) data to be normal data. A multi-PCA model was built based on all data from these normal production lots. In the modeling, only 2 or 3 principal components were necessary for each local PCA model to explain 99% variance of each sub-matrix of data. This paper will discuss using Multi-PCA as a modeling method for detecting real-time process variations based on equipment signals, with abnormal process signals being indicated by a single parameter - Squared Prediction Error (SPE) - from the PCA model. Issues related to the use of this technique in a state-of-the-art semiconductor fab will also be discussed.