AVS 45th International Symposium
    Manufacturing Science and Technology Group Thursday Sessions
       Session MS-ThM

Paper MS-ThM10
Multivariate Spectral Analysis of Optical Emission Spectroscopy for use in Low-Open Area Endpoint Detection

Thursday, November 5, 1998, 11:20 am, Room 317

Session: Sensors and Support Technology
Presenter: B. Goodlin, Massachusetts Institute of Technology
Authors: D. White, Massachusetts Institute of Technology
B. Goodlin, Massachusetts Institute of Technology
A. Gower, Massachusetts Institute of Technology
D. Boning, Massachusetts Institute of Technology
H. Sawin, Massachusetts Institute of Technology
T. Dalton, Digital Equipment Corporation
Correspondent: Click to Email
As device dimensions continue to shrink, the need for tighter control of semiconductor processes is increasing. In particular, accurate determination of endpoint in plasma etching processes is essential to decrease defects due to both incomplete clearing of the etched material and excessive overetch of the underlying material, leading to a loss of dimension control. This is particularly challenging for low open area etches (<1%), where traditional sensors are at the limits of their sensitivities in determining endpoint. Many sensors have been utilized for the purposes of determining endpoint including optical emission spectroscopy(OES), laser interferometry, optical emission interferometry, mass spectrometry, and rf impedance monitoring, but OES is the most widely used sensor. Traditional endpoint algorithms using OES observe only a few selected wavelengths corresponding to major product and reactant species, thus utilizing only a small fraction of the data provided by OES. For instance in an oxide etch process, using C@sub 2@F@sub 6@, we might follow the emission lines corresponding to a reactant species C@sub 2@ (e.g. 516 nm) and a product species SiF (e.g. 440 nm) during an etch process. Endpoint would be indicated by an increase in the ratio of the C@sub 2@ line intensity to the SiF line intensity. Since both lines are changing in intensity at endpoint we say that these lines are correlated or covarying. The OES spectrum, however, consists of a number of other emission lines which also correspond to reactant and product species, including many more lines corresponding to the many different excitations of C@sub 2@ and SiF. All of these lines have correlated changes that occur at endpoint, so by throwing away all of the spectra except a few spectral lines, the traditional endpointing algorithms do not take full advantage of all of the information available, resulting in a lower signal to noise ratio than that resulting if all of the lines were kept. In this paper, we examine the use of a multivariate technique called principal component analysis (PCA) which utilizes the entire OES spectrum and thus demonstrates superior signal to noise over the traditional univariate methods. We then demonstrate the technique for real-time endpoint detection in an industrial oxide contact etching process with low open areas (~1%). Lastly, implementation issues such as adjusting for process drift due to window fogging and PCA model validation are discussed.