GB2454793A

GB2454793A - Assignment of particles to processes of origin

Info

Publication number: GB2454793A
Application number: GB0820835A
Authority: GB
Inventors: Maximilian Staudacher; Linus Goerlitz; Fred Hamprecht
Original assignee: Robert Bosch GmbH
Current assignee: Robert Bosch GmbH
Priority date: 2007-11-13
Filing date: 2008-11-13
Publication date: 2009-05-20
Anticipated expiration: 2028-11-13
Also published as: GB0820835D0; DE102007054099A1; GB2454793B

Abstract

A method for assigning particles to a process of origin, in which a magnified, highly-resolved and high-contrast image of a particle is recorded and the exact position of the particle is determined. Subsequent to this, the Hu invariants and a weighted gradient-phase histogram of the surface of the particle are calculated, to provide an indication of the shape and texture of the surface of the particle. Finally, an integrated classification of the Hu invariants and of the gradient-phase histogram is undertaken for the purpose of identifying the process in which the particle originated. The method may be used to identify the origin of particles in residual-dirt analysis in semiconductor processing.

Description

Ti tie Assigiunent of particles to processes of origin State of the Art The invention relates to a method for assigning particles to their process of origin.

With methods that require great purity in terms of the production process, in particular methods of such a type in which individual particles already result in damage to the product, or even in the product becoming unusable, it is desirable if the process of origin of particles of such a type, which result in such a contamination, is known.

Processes that require such purity are, for example, methods for producing wafers such as are employed in semiconductor technology.

Methods with which particles can be assigned to their process of origin are still not known at the present time.

It is merely known -for example, from DE-A 10 2005 004 599 -to survey an object of measurement on a semiconductor wafer with a measuring instrument. In this case, in a first evaluation step a first target object is sought on : the basis of the measured intensity profile on the semiconductor wafer, in order to determine the characteristic parameters of said object as a first result S.....

* * 30 of measurement. However, if this evaluation fails or shows 55..

large deviations from known process specifications, a *. further evaluation step with another target object is sought, whereby the characteristic parameter thereof is likewise determined. By this means, it is possible to search i.n succession for differing target objects with only one measuring rule. Structural elements that have been formed on a semiconductor wafer, for example, are regarded as target objects. The measuring instrument with which the target objects are sought is generally a scanning electron microscope.

Disclosure of the Invention

Advantages of the invention A method according to the invention for assigning particles to their process of origin includes the following steps: (a) recording a magnified, highly-resolved, high-contrast image of the particle, (b) determining the exact position of the particle in the image, (c) digitising the image of the particle, calculating the Hu invariants, and calculating a weighted gradient-phase histogram on the surface of the particle, (d) integrated classification of the Hu invariants and of the gradient-phase histogram for the purpose of identifying the process of origin.

By virtue of the method according to the invention, it is possible for particles to be assigned to the causative * process on the basis of a backscatter image produced by a :.. scanning electron microscope. In particular, particles can *.S.

also be assigned, the source of which cannot be identified via a material that has been used uniquely. As a result, sources of particles are identifiable, and the causes of process instabilities can be identified in targeted manner.

By virtue of the targeted identification of process instabilities, intervention can be undertaken in targeted manner in the event of cleanliness problems. A further advantage of the method according to the invention is that said method can be integrated without great effort into an existing method for residual-dirt analysis.

The position of the particle is ordinarily determined by an image-editing method. For this purpose, in general the digitised photograph produced by the scanning electron microscope is subjected to a standard image-editing method.

Since the particle and the background exhibit differing grey-scale values or textures in the case of a grey-scale image, or differing colours or textures in the case of a colour photograph, the particle can be detected on the basis of the differences in the grey tone or on the basis of the colour differences, and in this way the position of the particle can be determined.

If the method for assigning particles to their process of origin is part of a comprehensive analytical method -for example, a residual- dirt analysis -it is possible to determine the position of the particle by deduction from S. : the detection of the particle in a preceding analytical step. In this case, in the preceding step it is firstly detected whether particles are present at all. If e..

* * 30 particles are found, the location of the particle can also S...

be derived from this. * *1

A magnified image of the particle is then recorded. The maximum magnification in this connection is dependent on the method being employed. Ordinarily, recording is undertaken with the aid of a scanning electron microscope.

In conventional processes for residual-dirt analysis processes at most a 2000-fold magnification arises by this means.

The characteristic properties of a particle are the shape and the editing texture thereof. These fluctuate very considerably with the material that is present. For this reason, it is preferred to determine the material of the particle prior to the assignment to a process of origin.

Determination of the material is undertaken by means of an EDX spectrum which is recorded from the entire surface of the particle and evaluated with standard software.

In particularly preferred manner, the determination of the material of the particle is undertaken directly following the recording of the magnified, highly-resolved and high-contrast image.

After the determination of the material of the particle, the Hu invariants and the weighted gradient-phase histogram on the surface of the particle are calculated. The Hu invariants serve for characterising the shape of the particle. By way of Hu invariants, the invariant object moments introduced by Hu in M.K. Hu, "Visual Pattern Recognition by Moment Invariants", IRE Transactions on Information Theory, 1962, pp 179-187, are utilised for the *...

purpose of characterisirig the shape. These serve to obtain :: 30 an invariance under rotation and scaling in connection with

S

***,. the shape description. * S.

:.. The texture of the particle is described by a weighted S...

gradient-phase histogram. For this purpose, for each pixel -.5-of the surface of the particle in the digitised image of the particle the discrete gradient is calculated, and the latter is resolved into amplitude and phase. In this connection the calculation of the gradient is generally undertaken using methods that are conventional in image processing, by convolution with a suitable mask.

After the resolving of the gradients into amplitude and phase, the phases of the gradients of all the pixels of the surface of the particle are summarised in a histogram. The contribution to the histogram is weighted by the amplitude.

The summarising of the gradients of all the pixels of the surface of the particle is preferentially undertaken by 2-degree steps. In the course of surnmarising the gradients, antiparallel gradients are identified. By this means, it.

is possible to use only directions between 00 and 180°. In this manner, important editing edges in the image are given a greater influence on the histogram, since large magnitudes in the case of the gradient have to be reckoned with therein. The histogram that has arisen is rotated so far that the maximum value of the histogram is situated at position 1. By this means, it is made possible for the histogram to be invariant under rotations. An invariance in relation to differing particle sizes and a robustness in relation to fluctuations in illumination are obtained by virtue of the fact that the histogram is finally sum-S...

: normalised. In this manner, a set of shape invariants and a phase histogram with 90 entries are obtained for each particle.

*.S.S. * 30

In order to be able to assign the particle that has been :.. described in this way to a process of origin, an integrated classification of the Hu invariants and of the gradient-phase histogram is finally undertaken. This is preferentially undertaken with the aid of a k-nearest neighbour classifier and a mixed Euclidean-symmetrised Kullback-Lejbler divergence semi-metric. The mixed Euclidian-symrnetrjsed Kuliback-Leibler divergence semi-metric is used as a separation measure for the k-nearest neighbour classifier. For this purpose, use is made of a Euclidian metric for 2 sets of shape parameters x, y, defined as L2(x,y) = -y)2 X,yE and, in the case of the histograms, use is made of a syrnmetrjsed Kullback-Lejbler divergence which is given by KL3 (pjq) = .((x) log +q(x,) log with histograms p, q, and both separation values are mixed in material-dependent manner. Mixing is undertaken in accordance with the following equation: FM(P,Q) = (1 2). L2 (Pob Qob1) + 2* KL (PHu,ogrQH,s,ogr) In this case the optimal value of the mixing parameter A can be inferred by cross-validation via a test set. * a a...

The parameter k of the k-nearest neighbour classifier is also preferentially determined by cross-validation on a *...

test set. The test set can be easily created by separate :.. preliminary experiments. In these preliminary experiments * 30 the provenance of the process is known. *... * a

In order to enable an assignment of the particle to the process of origin, reference values are determined from test data, preferentially for differing processes of origin. Assignment to a process of origin is undertaken by comparison of the Hu invariants and of the gradient-phase histogram with the reference values. In this connection the particle is assigned in each instance to that process of origin to which a reference particle appertains, the features of which are closest to the features of the particle to be classified.

The method according to the invention is preferentially employed for the purpose of identifying sources of particles in residual-dirt analysis. This comes in useful, in particular, in manufacturing processes that have to be carried out under high-purity conditions. Manufacturing processes of such a type are, for example, the production of semiconductor wafers.

Brief Description of the Drawings

An embodiment of the invention is represented in the drawing and elucidated in greater detail in the following

description.

The single Figure shows a flow chart of the method *: according to the invention. S... * .

Embodiments of the Invention S..... * *

In the single Figure a flow chart of the method according * to the invention is represented. S. * * S. *SS. * . ***S

In a first step 1, a magnified, highly-resolved and high-contrast image of a particle is recorded. A scanning electron microscope is preferentially employed for this purpose. However, any other apparatus can be employed with which a magnified, highly-resolved and high-contrast image of the particle can be recorded. Ordinarily, however, only scanning electron microscopes provide sufficient magnifications. The particle is preferably recorded with a magnification with which the particle is imaged in image-filling manner.

From the image that has been recorded in such a manner, the exact position of the particle is ascertained in a second step 3. This is preferentially undertaken using conventional methods of image processing. For this purpose it is, for example, possible to ascertain the position of the particle on the basis of grey-scale differences or colour differences in the image. In order to be able to employ the methods of image processing, however, it is important to represent the image firstly in digital form, in order to enable an appropriate evaluation using methods of image processing. This evaluation is generally undertaken by using a computer. As an alternative to the use of the standard methods of image processing for the purpose of detecting the exact position of the particle, it is also possible to derive the position of the particle by deduction by means of results of earlier steps of a process chain. The earlier steps of the process chain arise, for S...

example, by reason of a particle-detection process in :: 30 cleanliness analysis, such as is employed, for example, in * *.* processes that are carried out in a high-purity atmosphere.

* Also in the case of the particle-detection process in f:::' cleanliness analysis, a magnified image of an object is firstly recorded, and said image is examined for particles using methods of image processing. The images obtained in this way can then be magnified further in a renewed recording. In this case it is preferred if the first step 1 and the second step 3 are interchanged. That is to say, firstly the position of the particle is detected and, subsequent to this, the recording is undertaken of the magnified, highly-resolved and high-contrast image of the particle.

In a third step 5, after the recording of the image in the first step 1 and after the detection of the exact position of the particle in step 3, the Hu invariants for the digitised image of the particle are calculated. With the aid of the Hu invaria.rits a characterisatjon of the shape of the particle is undertaken. With the aid of the Hu invariants it is possible for the shape of the particle to be described in such a manner that the latter is invariant also under rotation and scaling of the image. In this manner it is possible for a characterisation of the shape of the particle to be obtained that is independent of the magnification and of the location of the particle.

After the calculation of the Hu invarjants for the digitised image of the particle in the third step 5, a weighted gradient-phase histogram on the surface of the particle is calculated. This calculation is undertaken in a fourth Step 7. The calculation of the weighted gradient- s.., phase histogram is undertaken with the aid of the high- ***5 contrast grey-scale image that was recorded in the first :": 30 step 1. By virtue of the weighted gradient-phase histogram, it is possible for the texture of the surface of the particle to be described. * 5* *1 * 5* * **s * I 5',

The texture and the shape of a particle are generally characteristic of the respective process of origin. For this reason, on the basis of the shape of the particle and on the basis of the surface texture it is possible to infer the process of origin of the particle that has been described in such a manner. For the weighted gradient-phase histogram, for each pixel of the surface of the particle in the high-contrast grey-scale image the discrete gradient is calculated, and the latter is resolved into amplitude and phase. The phases of the gradients of all the pixels of the surface of the particle are summarised in a histogram, and the contribution to the histogram is weighted by the amplitude. In this connection the summarising of the phases in a histogram is undertaken in steps which, on the one hand, are chosen to be not too large, in order to be sufficiently robust in relation to rotation of the object. On the other hand, however, too many bins -that is to say, a value that is too small -result in unnecessary variations. A summarising in 2° steps has proved suitable. In the course of summarising the gradients, antiparallel gradients are identified.

Consequently only directions between 0° and 1800 are used.

In this manner, important editing edges in the image are given a greater influence on the histogram, since large magnitudes in the case of the gradient have to be reckoned with therein. The histogram that has arisen is rotated so far that the maximum value of the histogram is situated at *::::* position 1. By this means, it is ensured that the histogram is invariant under rotations. Finally the histogram is also sum-normalised, in order to obtain an invariance in relation to differing particle sizes and a * ** robustness in relation to fluctuations in illumination. By reason of the 2-degree steps, the weighted gradient- phase histogram that has been created in this way has 90 entries.

In addition, 7 shape invariants, which result from the calculation of the Hu invariants, are obtained per particle.

In a final step 9, for the purpose of identifying the generating process an integrated classification of the Ru invariants arid of the gradient-phase histogram is carried out. This is preferentially undertaken with the aid of a k-nearest neighbour classifier and a mixed Euclidean-symmetrised Kuliback-Leibler divergence semi-metric. The parameter k of the k-nearest neighbour classifier and the mixing parameter A of the Euclidean-symmetrised Kuliback-Leibler divergence semi-metric are preferentially determined by cross-validation on a test set. In this connection the parameter k represents the number of nearest neighbours of the k-nearest neighbour classifier.

For the test set, the method is carried out for a particle that has arisen by virtue of a known process of origin. By this means, reference values arise for the Ru invariants and the gradient-phase histogram. In this manner, particles that were produced by known differing methods of origin can be determined, and the data thereof can serve as reference values.

The identification of the generating process then results by comparison of the ascertained values for the recorded particle with the values for particles, the process of origin of which is known. Since identical processes of :m: 30 origin for each particle yield similar values in each **..

instance, the process of origin can consequently be * ** inferred by comparison with these values. S. * * *. *SSI * S *.SS

However, since the Hu invariants and the gradient-phase histogram depend greatly on the material of the particle, in order to be able to assign the process of origin correctly it is necessary firstly to ascertain the material of the particle. The differing values for various materials result, in particular, from differing grey-scale values for various particles that are recorded.

In order to identify the material of the particle, an EDX spectrum is recorded. To this end, the scanning range of the electron beam is restricted to the surface of the particle, and the excited X-ray radiation is registered.

This EDX spectrum is examined for the materials contained therein, and the proportions thereof are quantified.

The method according to the invention is preferentially employed, as previously described, in connection with a residual-dirt analysis such as is employed, for example, in the inspection of semiconductor wafers. ***S * * * ** S S... * S

S

S..... * . S.. S * * * *S

S * S. S.., * I *.*.

Claims

Claims 1. Method for assigning particles to a process of origin, including the following steps: (a) recording a magnified, highly-resolved and high-contrast image of the particle, (b) determining the exact position of the particle in the image, (c) digitising the image of the particle, calculating the Hu invariants, and calculating a weighted gradient-phase histogram on the surface of the particle, (d) integrated classification of the Hu invariants and of the gradient-phase histogram for the purpose of identifying the process of origin.
2. Method according to Claim 1, characterised in that the position of the particle is determined by an image-editing method.
3. Method according to Claim 1, characterised in that the position of the particle is determined by deduction from the detection of the particle in a preceding analytical step.

:m: 30
4. Method according to one of Claims 1 to 3, characterised S...

in that prior to the assignment to a process of origin the material of the particle is determined. * *. * S * S. * .S. * S S...
5. Method according to one of Claims 1 to 4, characterised in that the integrated classification of the Hu invariants and of the gradient-phase histogram is undertaken with the aid of a k-nearest neighbour classifier and a mixed Euclidean-symrnetrised Kuilback-Leibler divergence semi-metric.
6. Method according to Claim 5, characterised in that the parameter k of the k-nearest neighbour classifier is determined by cross-validation on a test set.
7. Method according to Claim 5 or 6, characterised in that the mixing parameter of the Euclidean-syrnrnetrised Kuliback-Leibler divergence semi-metric is determined by cross-validation on a test set.
8. Method according to one of Claims 1 to 7, characterised in that reference values are determined from test data for differing processes of origin, and the assignment to a process of origin is undertaken by comparison of the Hu invariants and of the gradient-phase histogram with the reference values.
9. Use of the method according to one of Claims 1 to 8 for the purpose of identifying sources of particles in residual-dirt analysis. S. * * S * S S... * 5 S...

S

**SS.S * . * S S... * S. *. . * S. * SSS * . S...