WO2021223956A1

WO2021223956A1 - Method and device for inspecting the surface of a mirror

Info

Publication number: WO2021223956A1
Application number: PCT/EP2021/059218
Authority: WO
Inventors: Daniel Kraehmer; Steffen STIRNER; Manuel JOSEPH-HAENLE; Alexander Freytag; Daniel SCHMIED; Frank Eisert
Original assignee: Carl Zeiss Smt Gmbh
Priority date: 2020-05-08
Filing date: 2021-04-08
Publication date: 2021-11-11
Also published as: DE102020112496A1

Abstract

The invention relates to a method and a device for inspecting the surface of a mirror, in particular for microlithography. A method according to the invention comprises the following steps: Carrying out at least one roughness measurement on the surface of the mirror, and automatically evaluating measurement data obtained during this roughness measurement on the basis of a model, this model being generated using a method involving artificial intelligence, in which method, in a learning phase, the model is trained on the basis of a multiplicity of training data, which training data comprise both measurement data from previously performed roughness measurements and evaluations assigned to these measurement data.

Description

Method and device for inspection of the

Surface of a mirror

The present application claims the priority of the German patent application DE 10 2020 112 496.1, filed on May 8, 2020. The content of this DE application is incorporated by reference into the present application text.

BACKGROUND OF THE INVENTION

Field of invention

The invention relates to a method and a device for inspecting the surface of a mirror, in particular for microlithography.

State of the art

Microlithography is used to manufacture microstructured components such as integrated circuits or LCDs. The microlithography process is carried out in what is known as a projection exposure system, which has an illumination device and a projection lens. The image of a mask (= reticle) illuminated by the lighting device is projected by means of the projection lens onto a substrate (e.g. a silicon wafer) coated with a light-sensitive layer (photoresist) and arranged in the image plane of the projection lens, in order to create the mask structure on the light-sensitive Transfer coating of the substrate. In projection exposure systems designed for the EUV range, ie at wavelengths below 15 nm (eg about 13.5 nm or about 7 nm), mirrors are used as optical components for the imaging process due to the lack of availability of suitable light-permeable refractive materials.

Approaches to increasing the numerical aperture on the image side to increase the resolving power (NA) go hand in hand with an enlargement of the mirror surfaces. At the same time, in practice there is a need to adjust the effect of the optical element in question on the system wavefront in the respective optical system as precisely as possible or taking into account a given specification. In addition to locally long-wave surface defects of the respective optical element that are relevant for the effect on the system wavefront, short-wave surface defects with regard to scattered light losses can also be of importance in local terms.

In particular, the framework conditions that exist in microlithography applications for surface processing in the area of final correction typically require high-precision processing with accuracies in the nanometer or even picometer range.

During mirror production, a roughness evaluation is typically carried out, e.g. using interferometric or microscopic measurements (e.g. carried out with a scanning force microscope) in order to decide on the basis of this evaluation whether further surface processing is required or not. With the increasing dimensions of EUV mirrors (e.g. with diameters in the order of 1 meter or more), this roughness assessment represents a demanding challenge in view of the high accuracy requirements that exist.

In this context, it is particularly common to first evaluate the measurement data with regard to measurement artifacts (which indicate an incorrect implementation of the roughness measurement) as well as those that occur in a strongly localized manner (not by to carry out defects on the surface due to the statistical roughness properties.

In practice, the problem arises in particular that such a visual evaluation of the measurement data obtained during a roughness measurement leads to a high workload for the personnel performing the control during the production process and is also subject to a high degree of subjective error. In this context, it should be noted that not every "visible" error during the visual inspection has a significant impact on the final appearance, which is why the inspection is usually different depending on the experience of the staff. It should also be noted that no person performing the visual inspection is able to achieve a reproducibility of even approximately 100%.

Further disadvantages can result from the fact that the measurement data obtained during a roughness measurement are usually evaluated with a time delay (possibly several days), which on the one hand has a negative impact on the throughput achieved in production and on the other hand - at necessary rejection of a large number of incorrect measurements carried out in the meantime - is associated with a significant additional effort due to the necessary re-measurement of the mirror in question.

SUMMARY OF THE INVENTION

It is an object of the present invention to provide a method and a device for inspecting the surface of a mirror, in particular for microlithography, which are also designed for mirrors that have comparatively large dimensions or for optical systems with a high numerical aperture , enable a comparatively fast and reliable characterization. This object is achieved by the method and the device according to the features of the independent patent claims.

A method for inspecting the surface of a mirror, in particular for microlithography, has the following steps:

- Carrying out at least one roughness measurement on the surface of the mirror; and

automatic evaluation of measurement data obtained during this roughness measurement on the basis of a model;

This model is generated using an artificial intelligence method, the model being trained on the basis of a large number of training data in a learning phase, the training data each comprising measurement data from previously performed roughness measurements and evaluations assigned to these measurement data.

In the context of the present application, the term "roughness measurement" is intended to include various optical and tactile measurement methods or technologies, e.g. using an interferometer, a microinterferometer, a scanning force microscope (AFM), a stylus device, as well as other measurement methods.

The invention is based, in particular, on the concept of performing an automatic evaluation of the measurement data obtained during this roughness measurement with the aid of a model and using an artificial intelligence method when inspecting the surface of a mirror after a roughness measurement has been carried out. In particular, the invention includes the concept of training the model on which the automatic evaluation of the measurement data obtained in the roughness measurement is based in advance by using a large number of measurements that have already been carried out and the evaluation data assigned to the measurement data obtained as training data. As a result, in the invention, an evaluation - conventionally carried out on the basis of visual assessment by experienced personnel - is replaced by an automatic system Manufacture of the mirror can be avoided.

In particular, the invention also makes use of the fact that suitable training data for training the model - typically from roughness measurements carried out in the past and evaluations already assigned to these measurement data - are usually available in very large numbers (e.g. several thousand).

According to one embodiment, the automatic evaluation of the measurement data obtained during the roughness measurement comprises the following steps:

- In a first step, assessing whether the measurement data are free of measurement artifacts indicating an incorrect implementation of the roughness measurement; and

- Assess, in a second step, whether the measurement data indicate one or more local defects on the surface that are not caused by the statistical roughness properties.

According to one embodiment, the assessment in the second step only takes place if, according to the assessment from the first step, the measurement data are free of measurement artifacts.

In embodiments, if such measurement artifacts are present, the relevant measurement artifacts can also be classified automatically (ie the assignment to a specific class of measurement artifacts, such as “focusing error”, “too much tilting of the sample”, “drift during measurement”, “vibration during the Measurement "). Furthermore, if there are local defects on the surface that are not caused by the statistical roughness properties, an automated classification of the defects in question can also take place (whereby different types of defects can also occur at the same time).

According to one embodiment, the artificial intelligence method comprises machine learning, in particular “deep learning”, further in particular “convolutional neural networks” (CNNs).

In contrast to other methods of automated image analysis, such as an extraction of information-bearing features (e.g. SIFT or HOG) with subsequent transformation of these features (e.g. PCA or Fisher-En coding) and subsequent application of classic machine learning models for classification (e.g. SupportVectorMachine , RandomForest, nearest neighbor classifier or the like), the use of CNNs has the advantage that a step of feature extraction and the subsequent transformation does not have to be carried out before the actual one, at great expense and using justified assumptions Classifier can be used. Rather, CNNs allow the simultaneous learning of information-carrying features, transformation and classifier by optimizing a single, common optimization criterion (e.g. classification accuracy). This represents a substantial advantage especially for the context of the present application, since it is not known in advance which feature extractions are suitable for the respective artifact types or defect types.

A concrete implementation can include “convolutional neural networks” (CNNs), possibly pre-learned on other data sets, optimized on the basis of variants of the gradient descent (eg Adam). In such an implementation, a CNN can be designed to output a single value (eg the probability of the presence of a defect). Alternatively, a CNN can be designed to output two (“binary”) or more (“multi-class”) values that indicate class affiliations for which the classes are mutually exclusive (eg the probabilities for the mutually exclusive classes “defect” and “defect-free”). Alternatively, a CNN can also be designed to output two or more values that indicate class memberships in which the classes can also occur independently of each other (“multi label”, e.g. the probabilities of independently occurring defect types “scratches” and "Trough"). Such training would, as an optimization criterion in the learning step, for example, minimize the quadratic error, the Softmax cross entropy, or the sigmoid cross entropy.

Suitable CNN structures consist of sequences of convolution layers (“Convolutions”), non-linear activation functions (e.g. tangent hyperbolicus or piecewise linear functions “ReLU”) and normalization layers (e.g. batch normalization) and provide outputs from one layer to one or more subsequent layers Disposal. Well-known examples of such CNNs are AlexNet, VGG, Resnet, InceptionResnet, MobileNet, Xception and DenseNet.

For robust learning, the data are repeatedly and slightly changed in the learning step (“augmentation”), for example by mirroring on the x and y axes (“data flipping”) or by cutting out randomly selected image sections (“data cropping”). In addition, suitable regularization procedures are used in the learning step in order to prevent the mere memorization of the data (e.g. "drop-out").

According to one embodiment, the artificial intelligence method comprises monitored learning (= “supervised learning”). However, the invention is not limited to this. Thus, in further embodiments, the artificial intelligence method can also include partially monitored learning (= “semi-supervised learning”) or unsupervised learning (= “unsupervised learning”).

According to one embodiment, the automatic evaluation takes place immediately after the respective roughness measurement has been carried out. According to one embodiment, the model, on the basis of which the automatic evaluation of the measurement data obtained during the roughness measurement takes place, is selected from a plurality of previously trained models depending on a measurement setting of the respective roughness measurement.

According to one embodiment, in the learning phase, the model is trained on the basis of training data which were obtained for different measurement settings of the respective roughness measurement. Previous manual test results can be used as target values. Furthermore, the assignments of several different people performing the control can be used to create the training data. If necessary, only those examples can be used that have been assigned sufficiently clearly (whereby the remaining examples can be automatically assigned to a class “unclear”, for example).

In particular, a single model can be used for evaluating measurements that have been generated with different measurement settings. The measurement setting (s) can in particular include the enlargement of the respective roughness measurement.

However, the invention is not restricted to this. In other embodiments, several models can also be used and trained for different measurement settings.

According to one embodiment, depending on the evaluation of the measurement data obtained during the roughness measurement (e.g. if the roughness measurement in question contained a measurement artifact), a roughness measurement is carried out again on the surface of the mirror.

According to one embodiment, the mirror has an optically effective surface in the form of a free-form surface. According to a further embodiment, the mirror is designed for a working wavelength of less than 30 nm, in particular less than 15 nm. However, the invention is not limited to this either, and in other applications the mirror can also be designed for a different working wavelength.

The invention further relates to a device for inspecting the surface of a mirror, in particular for microlithography, which is configured to carry out a method having the features described above.

Further embodiments of the invention can be found in the description and the subclaims.

The invention is explained in more detail below using an exemplary embodiment shown in the accompanying figures.

BRIEF DESCRIPTION OF THE DRAWINGS

Show it:

FIG. 1 shows a flow chart to explain the possible sequence of a method according to the invention; and

Figure 2 is a schematic illustration to explain the possible

Construction of a microlithographic projection exposure system designed for operation in the EUV. DETAILED DESCRIPTION OF PREFERRED EMBODIMENTS

In the following, an exemplary sequence of a method according to the invention for inspecting the surface of a mirror (such as an EUV mirror) is described with reference to the flowchart shown in FIG.

The method according to the invention contains in particular the concept, as explained below, based on a model after the roughness measurement has been carried out on the surface of the mirror to be inspected and using an artificial intelligence method, an automatic evaluation on the one hand with regard to any measurement artifacts and on the other hand, possibly also with regard to local surface defects (not caused by the statistical roughness properties). The said application of an artificial intelligence method in turn includes the training of a model (or several models) in a learning phase on the basis of training data, this training data being measurement data from previously performed roughness measurements and evaluations assigned to these measurement data (i.e. measurement artifact (s)) yes or no, if yes, which measurement artifact type (s) and / or local (s) surface defect (s) include yes or no, if so, which (s) surface defect type (s)). In principle, according to the invention, different models for the two above-mentioned evaluations (with regard to measurement artifacts and with regard to local surface defects) or a common model for both evaluations can be used or trained.

According to the flowchart of FIG. 1, after the start of a measurement job in step S10, an evaluation tool (step S11) which implements the above evaluation and the start of the actual measurement sequence of the roughness measurement (step S12) are started in parallel Measuring point (step S13), the implementation and evaluation of the actual roughness measurement (step S14) and the storage of the obtained measurement result (step S15). After the measurement result has been received and saved in step S15, a trigger signal is used to initially trigger the evaluation according to the invention for the presence of any measurement artifacts (step S16). In this case, the evaluation can be based in particular on whether or not a characteristic value (“score”) obtained when the model is used exceeds a predetermined threshold value (“operating point”). By selecting this threshold value or working point, it can be influenced how high a “false alarm rate” of the evaluation is.

If, according to the following query in step S17, the measurement result obtained (ie the measurement data determined during the roughness measurement carried out) contains at least one measurement artifact indicating that the roughness measurement was carried out incorrectly, the message "repeat measurement" is output according to step S22 with the result that as a result of the query carried out on the storage of the measurement result in step S15 in step S23, the measurement sequence is carried out again, returning to step S13.

If, on the other hand, according to the above query from step S17, the measurement data obtained during the roughness measurement carried out are free of measurement artifacts, the subsequent evaluation takes place according to FIG. and the corresponding storage of the evaluation result (step S19).

Then, in step S20, a query or check is made as to whether or not a re-measurement should be carried out in view of the evaluation result stored in step S19. If so, the method goes correspondingly to step S22, ie the output of the message “repeat measurement”. If not, the message “next measuring point” is output in step S21. Furthermore, according to FIG. 1, in step S24 a query is made “All measuring points ready?” To the effect that all measuring points on the surface to be inspected have been processed. This query takes place both after the “next measuring point” message output in step S21 and if, according to the query in step S23, no further repetition of the measurement sequence (in the absence of a “repeat measurement” message according to step S22) is to take place. If, according to the query in step S24, not all of the measuring points have been processed, a return to step S13 takes place, otherwise the measuring job ends (step S25).

In further embodiments, instead of the immediate evaluation of each measurement, a plurality of measurements can first be collected over a suitable period of time before these measurements are then evaluated in a model-based manner according to the invention.

Furthermore, according to the invention, several models or, alternatively, only a single model can be used and trained for different measurement settings (e.g. magnifications).

FIG. 2 shows a schematic illustration of an exemplary projection exposure system designed for operation in the EUV.

According to FIG. 2, an illumination device in a projection exposure system 200 designed for EUV has a field facet mirror 203 and a pupil facet mirror 204. All of the mirrors in the projection exposure system 200 can be freeform surfaces. The light from a light source unit, which comprises a plasma light source 201 and a collector mirror 202, is directed onto the field facet mirror 203. In the light path after the pupil facet mirror 204, a first telescope mirror 205 and a second telescope mirror 206 are arranged. In the light path below a deflecting mirror 207 is arranged, which directs the radiation hitting it onto an object field in the object plane of a projection lens comprising six mirrors 251-256. A reflective structure-bearing mask 221 is placed at the location of the object field a mask table 220, which is imaged with the aid of the projection lens in an image plane in which a substrate 261 coated with a light-sensitive layer (photoresist) is located on a wafer table 260.

The invention can be used in the characterization or inspection of the surface during or after the production of any mirror of this projection exposure system. However, the invention is not limited to the implementation in the production of mirrors for operation in the EUV, but also in the inspection of mirrors for other working wavelengths (e.g. in the VUV range or at wavelengths less than 250 nm) and also for use in other optical systems (not intended for microlithography). If the invention has also been described on the basis of specific embodiments, numerous variations and alternative embodiments will be apparent to the person skilled in the art, e.g. by combining and / or exchanging features of individual embodiments. Accordingly, it is understood by a person skilled in the art that such variations and alternative embodiments are also encompassed by the present invention, and the scope of the invention is limited only within the meaning of the attached patent claims and their equivalents.

Claims

1. A method for inspecting the surface of a mirror, in particular for microlithography, the method comprising the following steps:

• Carrying out at least one roughness measurement on the surface of the mirror; and

• automatic evaluation of measurement data obtained during this roughness measurement on the basis of a model;

2. The method according to claim 1, characterized in that the automatic cal evaluation of the measurement data obtained during the roughness measurement comprises the following steps: i. In a first step, assessing whether the measurement data are free from measurement artifacts indicating an incorrect implementation of the roughness measurement; and ii. In a second step, assess whether the measurement data indicate one or more local defects on the surface that are not caused by the statistical roughness properties.

3. The method according to claim 2, characterized in that if measurement artifacts are present, the relevant measurement artifacts are automatically classified.

4. The method according to claim 2 or 3, characterized in that if there are local defects on the surface that are not caused by the statistical roughness properties, an automated classification of the relevant defects takes place.

5. The method according to any one of claims 2 to 4, characterized in that the assessment in the second step only takes place if, according to the assessment from the first step, the measurement data are free of measurement artifacts.

6. The method according to any one of the preceding claims, characterized in that the artificial intelligence method comprises machine learning, in particular “deep learning”, further in particular “convolutional neural networks”.

7. The method according to any one of the preceding claims, characterized in that the artificial intelligence method comprises monitored learning (= "supervised learning").

8. The method according to any one of the preceding claims, characterized in that the automatic evaluation takes place immediately after the respective roughness measurement has been carried out.

9. The method according to any one of claims 1 to 8, characterized in that the model, on the basis of which the automatic evaluation of the measurement data obtained in the roughness measurement takes place, is selected from a plurality of previously trained models depending on a measurement setting of the respective roughness measurement.

10. The method according to any one of claims 1 to 8, characterized in that in the learning phase, the model is trained on the basis of training data obtained for different measurement settings of the respective roughness measurement.

11. The method according to claim 9 or 10, characterized in that the measurement setting comprises an increase in the respective roughness measurement.

12. The method according to any one of the preceding claims, characterized in that depending on the evaluation of the measurement data obtained during the roughness measurement, a roughness measurement is carried out again on the surface of the mirror.

13. The method according to any one of the preceding claims, characterized in that the mirror has an optically effective surface in the form of a free-form surface.

14. The method according to any one of the preceding claims, characterized in that the mirror is designed for a working wavelength of less than 30 nm, in particular less than 15 nm.

15. Device for inspecting the surface of a mirror, in particular for microlithography, characterized in that the device is configured to carry out a method according to one of the preceding claims.