Automated Defect Sizing Gives Near-SEM Accuracy

Alexander E. Braun, Associate Editor -- Semiconductor International, 6/1/1999

I n-line defect inspection and classification's goals are defect reduction and yield prediction. Some fabs consider all defects a problem and aim to reduce overall defectivity levels. Others focus yield prediction to enable key decisions, such as increasing wafer starts, to compensate for an expected yield loss and scrapping lots early on in the process to prevent a costly investment in wafers with low expected yield. Still others do both.

The industry has developed complex yield models to make accurate yield predictions. These vary from fab to fab, with many customized to suit a particular IC manufacturer's product mix and device types. Regardless of the model, however, three inputs are needed to accurately predict yield in all models: the number of defects detected in-line, their classifications and their size.

Today's patterned wafer inspection tools and ADC systems have made the first two inputs relatively accurate. However, sizing is generally inaccurate and can result in erroneous yield predictions (Fig. 1) because fabs relied on in-line sizing outputed by inspection tools. In-line size error can vary considerably, depending on the inspector used, with darkfield tools being less accurate than brightfield. And even brightfield systems' sizing capability is inherently limited by the inspection pixels used, which are typically large (0.62 to 0.39 µm) because of manufacturing throughput constraints associated with smaller, high-resolution pixels.

Fig. 1. By providing near-SEM-equivalent accuracy, the HRDC sizing module enables users to detect small sizing errors that would otherwise result in a large yield prediction errors, leading to better processing decisions. (Source: KLA-Tencor)

An analogy to the situation is the "image stabilization" feature in video cameras. Before it, regardless of how good the optics, zoom, electronics, tape or other camera components were, if the operator moved while shooting, the image was blurred. Yield models are the same way: no matter how sophisticated, if the sizing is inaccurate or fudged as a default value, the model produces a "blurry" yield estimate.

Traditionally, the most accurate method to get accurate sizing data has been SEM measurement. This requires moving the wafer out of the inspector to an off-line SEM. This is costly in terms of throughput and possible added contamination through the extra handling.

KLA-Tencor (San Jose, Calif.) has come up with the industry's first defect sizing capability as an optional module for its Impact ADC platform. It allows users to get "common-denominator" sizing of defects detected by any inspection system. Because KLA-Tencor's on-line High-Resolution Defect Classification (HRDC) solution classifies defects at the highest optical magnification possible, detailed size information can be extracted to provide accurate and consistent defect sizing across all inspection platforms.

The HRDC can extract up to nine different size features at 0.13 µm resolution. The result is near-SEM-equivalent sizing accuracy, without requiring an additional step. The advantages of this new capability are consistent sizing across different inspection tools (brightfield, darkfield, etc.); accurate input for yield models, resulting in accurate predictions; and the ability to increase or decrease fab wafer starts based on accurate yield predictions. The system is easy to implement and will provide customizable or automatic reports.