| The fabrication of modern semiconductor products requires thousands of processing steps. A key element in achieving high yields and throughput with short cycle-times is to monitor the equipment to ensure proper processing at each step. This thesis develops a monitoring method suitable for real-time fault detection, fault diagnosis, and wafer state prediction. Because not all wafer states can be directly measured while the wafers are being processed in each piece of equipment, we use real-time signals sensitive to the equipment state to infer the condition of the wafer. This set of real-time signals is monitored and analyzed by the system, which consists of three distinct modules.; The fault detection module employs time series modeling and multivariate statistics to detect run-time errors on a second-to-second basis. When a malfunction is detected, the fault diagnosis module assigns a cause to the problem. Two methods for diagnosis were investigated. The first uses discriminant analysis techniques, while the second uses a combination of clustering algorithms and neural network models. Examples of faults which have been detected and diagnosed on a plasma etcher include various levels of miscalibrations in mass flow controllers, pressure gauges, and radio frequency (RF) power generators.; In addition, the system predicts the wafer state after each process step. Generally, models for wafer states are built using the input settings of the equipment. Experimental results in this thesis, however, demonstrate that models built with the selected real-time signals, which we call chamber state based (CSB) models, are more accurate than models built with machine input settings for the prediction of key wafer states of plasma processes especially after the machine has aged significantly since the original model was created.; The system as a whole has the potential to reduce the overall cost of ownership of semiconductor equipment by increasing both the wafer yield and throughput of product wafers, and decreasing the down-time and mean-time-to-repair of the equipment. Furthermore, this system does not depend upon monitor wafers or expensive metrology; rather, it uses real-time signals collected automatically and non-invasively from the equipment. As such, it will enable inexpensive run-to-run and real-time control applications. The system has been developed and tested on the Lam Rainbow 4400 and Lam TCP 9600 plasma etch equipment. |
