System and method for determining endpoint in etch processes using partial least squares discriminant analysis in the time domain of optical emission spectra

• US 6,830,939 B2
• Filed: 08/28/2002
• Issued: 12/14/2004
• Est. Priority Date: 08/28/2002
• Status: Active Grant

• Current Assignee: Verity Instruments Incorporated
• Original Assignee: Verity Instruments Incorporated
• Inventors: Harvey, Kenneth C., Hosch, Jimmy W., Wise, Barry M., Gallagher, Neal B.
• Primary Examiner(s): Goudreau, George A.
• Application Number:
  US10/232,987
• Publication Number:
  US 20040045934A1
• Time in Patent Office: 839 Days
• Field of Search: 438/8, 438/9, 438/16, 438/714, 216/60, 216/67, 134/1.1, 134/1.2, 134/22.11
• US Class Current: 438/8
• CPC Class Codes: G01N 21/73 using plasma burners or torches
• CPC Class Codes: H01J 37/32935 Monitoring and controlling tubes by information coming from the object and/or discharge
• CPC Class Codes: H01J 37/32963 End-point detection

Abstract

The present invention is directed to a system, method and software product for creating a predictive model of the endpoint of etch processes using Partial Least Squares Discriminant Analysis (PLS-DA). Calibration data is collected from a calibration wafer using optical emission spectroscopy (OES). The data may be non-periodic or periodic with time and periodic signals may be sampled synchronously or non-synchronously. The OES data is arranged in a spectra matrix X having one row for each data sample. The OES data is processed depending upon whether or not it is synchronous. Synchronous data is arranged in an unfolded spectra matrix X having one row for each period of data samples. A previewed endpoint signal is plotted using wavelengths known to exhibit good endpoint characteristics. Regions of stable intensity values in the endpoint plot that are associated with either the etch region or the post-etch region are identified by sample number. An X-block is created from the processed OES data samples associated with the two regions of stable intensity values. Non-periodic OES data and asynchronously sampled periodic OES data are arranged in a X-block by one sample per row. Synchronously sampled periodic OES data are arranged in the X-block by one period per row. A y-block is created by assigning a discriminate variable value of “1” to OES samples associated with the class, i.e. the etch, and assigning a discriminate value of “0” to all samples not in the class, i.e. the post-etch. A b-vector is regressed from the X- and y-blocks using PLS and is used with the appropriate algorithm for processing real-time OES data from a production etch process for detecting an endpoint.

Claims

1. A method creating a predictive model of an event in an etch process using partial least squares discriminant analysis (PLS-DA) comprising:

collecting calibration data for an etch process;

using at least a portion of the calibration data, identifying a feature associated with the event;

creating a predictor matrix, wherein the predictor matrix includes data for the feature associated with the event;

creating a response matrix, wherein the response matrix is comprised of a first discriminate variable value for the feature associated with the event; and

finding a predictive model for the event by regressing the response matrix and the predictor matrix using PLS regression.

2. The method recited in claim 1, wherein the event is an endpoint event and the calibration data is collected using optical emission spectroscopy (OES), each OES data sample having intensity values for a plurality of simultaneously sampled discrete wavelengths, wherein identifying a feature associated with the event further comprises:

plotting an endpoint signal of intensity values from at least a portion of the OES data samples.

3. The method recited in claim 2 further comprises:

identifying a first region of stable intensity values on the endpoint signal associated with the endpoint event in an etch process; and

selecting OES data samples associated with the first region of stable intensity values.

4. The method recited in claim 3, wherein the first region of stable intensity values on the endpoint signal representing an etch process occurring prior to the endpoint event.

5. The method recited in claim 3, wherein the first region of stable intensity values on the endpoint signal representing a post-etch process occurring subsequent to the endpoint event.

6. The method recited in claim 2 further comprises:

identifying a first region of stable intensity values on the endpoint signal;

identifying a second region of stable intensity values on the endpoint signal, wherein one of the first and second region of stable intensity values represents an etch process occurring prior to the endpoint event and the other of the first and second region of stable intensity values represents a post-etch process occurring subsequent to the endpoint event;

selecting OES data samples associated with the first region of stable intensity values and the second region of stable intensity values.

7. The method recited in claim 6, wherein the data for the feature associated with the event from the calibration data comprises OES data samples associated with one of the first and second region of stable intensity values.

8. The method recited in claim 6, wherein the first discriminate variable value represents the first region and a second discriminate variable value represents all other regions.

9. The method recited in claim 8, wherein the method further comprises:

calculating a resultant endpoint signal by applying the predictive model to the calibration data; and

comparing the resultant endpoint signal to the endpoint signal from the at least a portion of the OES data.

10. The method recited in claim 8, wherein the OES data samples contain non-periodic intensity values, subsequent to collecting calibration data for an etch process the method further comprises:

processing the calibration data; and

creating the predictor matrix further comprises;

arranging the OES data samples by having one OES data sample in each row of the predictor matrix.

11. The method recited in claim 10, wherein processing the calibration data further comprises:

removing OES data samples with startup or shutdown transients; and

removing saturated wavelengths of OES data samples.

12. The method recited in claim 8, wherein the OES data samples are asynchronously sampled periodic intensity values, subsequent to collecting calibration data for an etch process the method further comprises:

filtering the periodic intensity values; and

creating the predictor matrix further comprises;

arranging the OES data samples by having one OES data sample in each row of the predictor matrix.

13. The method recited in claim 8, wherein the OES data samples are synchronously sampled with a period of the intensity values, subsequent to collecting calibration data for an etch process the method further comprises:

processing the calibration data; and

creating the predictor matrix further comprises;

arranging the OES data samples by having one period of OES data samples in each row of the predictor matrix.

14. The method recited in claim 13, wherein processing the calibration data further comprises:

removing OES data samples with startup or shutdown transients;

removing OES data samples with noise associated with the period;

selecting integral periods of OES data samples by removing periods having OES data samples associated with the first region of stable intensity values and OES data samples associated with another region; and

removing saturated wavelengths of OES data samples.

15. The method recited in claim 1 further comprising:

detecting an event in a production etch process by filtering production data with the predictive model.

16. A method creating a predictive model of an endpoint of an etch process using partial least squares discriminant analysis (PLS-DA) comprising:

collecting data samples from a calibration wafer using optical emission spectroscopy (OES), each OES data sample having intensity values for a plurality of simultaneously sampled discrete wavelengths;

plotting an endpoint signal from the OES data samples using a plotting method for identifying a location of an endpoint transition;

identifying an etch region of stable intensity values and a post-etch region of stable intensity values on the endpoint signal;

creating a predictor matrix by arranging OES data samples associated with the etch region and the post-etch region by one OES sample in each row;

creating a response matrix with a first discriminate variable value for the OES data samples associated with the etch region and a second discriminate variable value for the OES data samples associated with the post-etch region;

finding a predictive model for the endpoint by regressing the response matrix and the predictor matrix using PLS regression; and

filtering real-time data from a production wafer using the predictive model.

17. A method creating a predictive model of an endpoint of an etch process using partial least squares discriminant analysis (PLS-DA) comprising:

collecting data samples from a calibration wafer using optical emission spectroscopy (OES), each OES data sample having intensity values for a plurality of simultaneously sampled discrete wavelengths;

finding a time derivative of the optical emission spectroscopy (OES) data samples;

plotting an endpoint signal from the OES data samples using a plotting method for identifying a location of an endpoint transition;

identifying an etch region of stable intensity values, a transition region of evolving intensity values and a post-etch region of stable intensity values on the endpoint signal;

creating a predictor matrix by arranging OES sample derivative values associated with the etch region, the transition region and the post-etch region by one OES sample derivative value in each row;

creating a response matrix with a first discriminate variable value for OES sample derivative values associated with the transition region and a second discriminate variable value for all other OES sample derivative values;

finding a predictive model for the endpoint by regressing the response matrix and the predictor matrix using PLS regression; and

filtering real-time data from a production wafer using the predictive model.

18. A method creating a predictive model of an endpoint of an etch process using partial least squares discriminant analysis (PLS-DA) comprising:

collecting data samples from a calibration wafer using optical emission spectroscopy (OES), each OES data sample having intensity values for a plurality of simultaneously sampled discrete wavelengths, wherein the OES data is synchronously sampled with a period of intensity data;

removing ambiguous and misleading OES data samples;

plotting an endpoint signal from the OES data samples using a plotting method for identifying a location of an endpoint transition;

identifying an etch region of stable intensity values and a post-etch region of stable intensity values on the endpoint signal;

creating a predictor matrix by arranging OES data samples associated with the etch region and the post-etch region by one period of OES samples in each row;

creating a response matrix with a first discriminate variable value for the OES data samples associated with the etch region and a second discriminate variable value for the OES data samples associated with the post-etch region;

finding a predictive model for the endpoint by regressing the response matrix and the predictor matrix using PLS regression; and

filtering synchronously sampled real-time periodic data from a production wafer using the predictive model.

19. A method creating a predictive model of an event in an etch process using partial least squares discriminant analysis (PLS-DA) comprising:

collecting data samples from a calibration wafer, each data sample having a plurality simultaneously recorded readings, said event occurring in the data samples from a calibration wafer;

previewing an event plot from the data samples using previewing method for identifying a location of an event transition;

identifying the event transition region on the event plot;

creating a predictor matrix by arranging simultaneously recorded readings associated with the event transition region and simultaneously recorded readings not associated with the event transition region, by one simultaneously recorded reading in each row;

creating a response matrix with a first discriminate variable value for the simultaneously recorded readings associated with the event transition region and a second discriminate variable value simultaneously recorded readings not associated with the event transition region;

finding a predictive model for the event by regressing the response matrix and the predictor matrix using PLS regression; and

filtering real-time data from a production wafer using the predictive model.

20. A method creating a predictive model of an event in a chamber cleaning process using partial least squares discriminant analysis (PLS-DA) comprising:

collecting calibration data for a chamber cleaning process;

using at least a portion of the calibration data, identifying a feature associated with the event;

creating a predictor matrix, wherein the predictor matrix includes data for the feature associated with the event;

creating a response matrix, wherein the response matrix is comprised of a first discriminate variable value for the feature associated with the event; and

finding a predictive model for the event by regressing the response matrix and the predictor matrix using PLS regression.

21. The method recited in claim 20, wherein the event is an endpoint event and the calibration data is collected using optical emission spectroscopy (OES), each OES data sample having intensity values for a plurality of simultaneously sampled discrete wavelengths, wherein identifying a feature associated with the event further comprises:

plotting an endpoint signal of intensity values from at least a portion of the OES data samples.

22. The method recited in claim 21 further comprises:

identifying a first region of stable intensity values on the endpoint signal associated with the endpoint event in a chamber cleaning process; and

selecting OES data samples associated with the first region of stable intensity values.

23. The method recited in claim 22, wherein the first region of stable intensity values on the endpoint signal represents a post-etch process after all interior surfaces of the chamber have cleared of film occurring subsequent to the endpoint event.

24. The method recited in claim 22, wherein the first region of stable intensity values on the endpoint signal represents an etch process occurring prior to the endpoint event.

25. The method recited in claim 22 further comprises:

identifying a first region of stable intensity values on the endpoint signal;

identifying a second region of stable intensity values on the endpoint signal, wherein one of the first and second region of stable intensity values represents an etch process occurring prior to the endpoint event and the other of the first and second region of stable intensity values represents a post-etch process after all interior surfaces of the chamber have cleared of film occurring subsequent to the endpoint event;

selecting OES data samples associated with the first region of stable intensity values and the second region of stable intensity values.

26. The method recited in claim 22 further comprises:

identifying a first region of stable intensity values on the endpoint signal representing a post-etch process after all interior surfaces of the chamber have cleared of film occurring subsequent to the endpoint event;

identifying a second region of non-stable intensity values on the endpoint signal representing an etch process for a plurality of interior surfaces of the chamber, wherein the second region is identified from a single spectra or a region of spectra having changing intensities.

27. A method creating a predictive model of a fault event in a semiconductor fabricating process using partial least squares discriminant analysis (PLS-DA) comprising:

collecting calibration data for a semiconductor fabricating process;

using at least a portion of the calibration data, identifying a feature associated with the fault event;

creating a predictor matrix, wherein the predictor matrix includes data for the feature associated with the fault event;

creating a response matrix, wherein the response matrix is comprised of a first discriminate variable value for the feature associated with the fault event; and

finding a predictive model for the event by regressing the response matrix and the predictor matrix using PLS regression.

28. The method recited in claim 27, wherein collecting calibration data further comprises:

generating data from a measurement taken for a process parameter associated with a failure mode in the semiconductor fabricating; and

changing the magnitude of the measurement.

29. The method recited in claim 28, wherein the measurement is one of gas pressure, gas flow rate, throttle valve position, RF tuning capacitor position, RF power, and temperature.