Browse Prior Art Database

A Unified Outlier Management Method for Trace Data Analysis in Semiconductor Manufacturing

IP.com Disclosure Number: IPCOM000244103D
Publication Date: 2015-Nov-09
Document File: 4 page(s) / 94K

Publishing Venue

The IP.com Prior Art Database

Abstract

Disclosed is a unified outlier management method for process trace data (PTD) in semiconductor manufacturing that takes into consideration the domain knowledge. The proposed method takes advantage of information theoretic and statistical techniques, and it consists of two steps – entropy based data complexity reduction and Cumulative Sum (CUSUM) based abrupt change detection.

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 38% of the total text.

Page 01 of 4

A Unified Outlier Management Method for Trace Data Analysis in Semiconductor Manufacturing

Given the complexity of wafer fabrication processing operations and large -scale data collected from the processes, data mining and statistical analysis play a key role in semiconductor manufacturing for product quality control and the detection of anomalies .

Process trace data (PTD) is an important data type with a very large aggregate volume. PTD consists of the time series that are recorded in each sensor during each step , and modern manufacturing tools are equipped with many different types of sensors that record a variety of chemical, physical, and mechanical process measurements, typically to provide feedback to the operators or the process control mechanisms on the tools .

Some applications that benefit from the analysis of the PTD are used to improve tool stability and matching. For a specified period, the trace data associated with a sensor and a process step consists of a vector of the summary statistics (e.g., median, standard deviation) for all wafers processed by this chamber via this process during this period, which is termed the Trace Data Vector (TDV). Based on this definition, TDV is
a type of univariate data.

The problem addressed herein is that the complexities in a manufacturing environment result in complex manufacturing process data, which poses challenges to addressing outlier detection problems for the analysis of PTD .

Current methods for outlier/anomaly detection are classified into the following categories:

• Classification based (require labeling of the normal instances)

• Nearest neighbor based (oriented to multivariate data)

• Clustering based (oriented to multivariate data)

• Spectral (for handling high dimensional data sets)

• Statistical (make distributional assumptions for the data )

• Information theoretic techniques (effective if there are a fairly large amount of

outliers present in the data)

The novel contribution is a unified outlier management method for PTD in semiconductor manufacturing that takes into consideration the domain knowledge . The proposed method takes advantage of information theoretic and statistical techniques , and it consists of two steps - entropy based data complexity reduction and Cumulative Sum (CUSUM) based abrupt change detection . Novel outlier detection algorithms for both steps take into account the practical requirements and engineering insights of the manufacturing engineers.

Entropy based data complexity reduction uses entropy to measure the data complexity and attempts to identify the minimum subset (outliers) that can maximize the difference in the complexity between the original data set and the data set after the outliers are removed. This method does not need to make assumptions about the data distribution ,

1


Page 02 of 4

but may not be effective when there are only a small number of outliers in the data .

To address this problem, in the second step,...