| Manufacturing integrated circuits (IC's) with increased density and complexity on larger and larger substrates requires the stringent control of hundreds, or even thousands of process variables. Individual IC process steps are conducted by complex pieces of fabrication equipment. When unreliable performance causes this equipment to vary beyond specified limits, overall product quality is jeopardized. Since process shifts resulting from faulty equipment can degrade semiconductor products to an unacceptable level, it is essential that root causes for the malfunctions be diagnosed and corrected quickly to prevent the continued occurrence of expensive misprocessing.;Although in-line measurements and electrical test data have historically been used to detect process fluctuations, these methods alone have become inadequate for rapidly identifying problems in processes with a narrow range of acceptable performance. However, with the advent of highly proficient sensors designed to monitor process conditions in-situ, it has become feasible to perform such malfunction diagnosis on a real-time basis. Therefore, methods of equipment diagnosis which utilizes these capabilities are critical to the overall success of the semiconductor production process.;This dissertation presents a general methodology for the automated diagnosis of IC fabrication equipment. The techniques presented combine the best aspects of quantitative algorithmic, qualitative experiential, and neural network approaches. By using neural network-based modeling techniques in conjunction with statistical inference methods, evidential belief is generated from equipment maintenance history, real-time sensor data and in-line measurements. The cause of failures is inferred by integrating this belief using Dempster-Shafer evidential reasoning techniques. This methodology is applied to the identification of faults in the Plasma Therm 720/740 Dual Chamber reactive ion etching (RIE) system. |
