Flexible hybrid defect classification for semiconductor manufacturing

Abstract

Hybrid methods for classifying defects in semiconductor manufacturing are provided. The methods include applying a flexible sequence of rules for defects to inspection data. The sequence of rules includes deterministic rules, statistical rules, hybrid rules, or some combination thereof. The rules included in the sequence may be selected by a user using a graphical interface The method also includes classifying the defects based on results of applying the sequence of rules to the inspection data.

Description

[0001] Examples of fully rule-based approaches include Run Time Classification (RTC) provided on the AIT II, AIT III, and AIT XP systems, which are commercially available from KLA-Tencor, San Jose, Calif., the early release of on-the-fly (OTF) classification methods on the Compass tools commercially available from Applied Materials Inc., Santa Clara, Calif., and gray level binning for voltage contrast defects, which is commercially available from Hermes MicroVision, Milpitas, Calif. The setup of such a classifier is relatively simple and easy for the user to understand. Many of these approaches provide some user assistance by showing how the defects have been separated through a variety of graphical means and by showing examples of defects in each bin. Deterministic rule-based classifiers generally have a high throughput. [0002] Examples of statistical (trained) classification are the current automatic defect classification (ADC) and inline ADC (iADC) products on the 23xx, AIT, eSxx...

AI Technical Summary

Benefits of technology

[0001] Examples of fully rule-based approaches include Run Time Classification (RTC) provided on the AIT II, AIT III, and AIT XP systems, which are commercially available from KLA-Tencor, San Jose, Calif., the early release of on-the-fly (OTF) classification methods on the Compass tools commercially available from Applied Materials Inc., Santa Clara, Calif., and gray level binning for voltage contrast defects, which is commercially available from Hermes MicroVision, Milpitas, Calif. The setup of such a classifier is relatively simple and easy for the user to understand. Many of these approaches provide some user assistance by showing how the defects have been separated through a variety of graphical means and by showing examples of defects in each bin. Deterministic rule-based classifiers generally have a high throughput.

[0007] Accordingly, it may be advantageous to develop computer-implemented methods for classifying defects that eliminate one or more of the disadvantages described above. SUMMARY OF THE INVENTION

[0018] In one embodiment, the sequence of rules may include only statistical rules. In one such embodiment, these rules may be organized into groups to aid in user understanding of the rules and to allow selectivity in the characteristics to be used. This selectivity has three major advantages: classification can be done with significantly fewer examples, which may include abstract examples; classification can be stable over more specimens; and classification can be executed more quickly. In another embodiment, the statistical rules may be weighted separately. In an additional embodiment, the statistical rules and the hybrid rules may be organized into groups for selection to aid in user understanding of these rules and to provide classifications that reflect the intent of the user.

Problems solved by technology

For example, many deterministic methods do not include all of the characteristics of the defects that are relevant to good classification.

In addition, fixed boundaries often do not work well over time on different specimens.

The deterministic rule based methods are also generally inflexible in the usage of rules and defect characteristics.

In addition, these methods generally include some restrictions on the number and kinds of characteristics and how they are combined.

Furthermore, these methods generally have user interface deficiencies in being able to create the classification recipe.

For example, the user interface can be complex to navigate, and the final results may not be clear.

One disadvantage of the fully trained approaches is that these methods generally rely on having a sufficient population of the defects for each bin available for training.

In addition, these methods work in a way that may not reflect the intentions of the user because these methods function as a black box (i.e., the user is unable to select the characteristics or characteristic groups to be used to do the classification).

Furthermore, these methods often neglect non-appearance characteristics that can be important in separating defects for purposes of analysis.

Lastly, fully trained classifiers are generally slower to execute than deterministic rules, particularly ones trained with a large number of characteristics.

The inflexible, hybrid methods have disadvantages such as that these methods often do not account for novel ways that the user might want to separate defects for a particular image or specimen.

In addition, these methods rigidly restrict the paths used to bin the defects.

Images