Method for providing a predictive model for crack detection, method for crack detection on a semiconductor structure and device for crack detection on a semiconductor structure

Abstract

Method for providing a predictive model for crack detection on a semiconductor structure, which semiconductor structure is a photovoltaic solar cell, a precursor of a photovoltaic solar cell in the manufacturing process, in particular semiconductor material for the manufacture of a solar cell, comprising the following process steps: A. Providing a reference semiconductor structure which has at least one crack; B. Providing crack data for the at least one crack, which crack data includes geometric location data regarding the location of the crack on the reference semiconductor structure; C. Spatially resolved measurement of the reference semiconductor structure by spatially resolved measurement for a plurality of local measurement points of photoluminescence radiation generated in the semiconductor structure and / or spatially resolved measurement of the IR absorption of the semiconductor structure and D. Creating a predictive model by training a learning algorithm using the spatially resolved measurement data determined in process step C and the crack data provided in process step B, wherein the training of the learning algorithm comprises the following process steps: D1 Create at least one descriptor for at least one local descriptor point by specifying or determining a test region for the descriptor point and creating the descriptor based on the measurement data within the test region, which descriptor is a feature vector and / or a feature distribution and / or a feature histogram and D2 Training the learning algorithm using the descriptor and the crack data.

Description

[0001] The invention relates to a method for providing a prediction model for crack detection on a semiconductor structure according to claim 1 and a method for crack detection on a semiconductor structure according to claim 2. The semiconductor structure is a photovoltaic solar cell, a precursor of a photovoltaic solar cell in the manufacturing process or, in particular, semiconductor material for manufacturing such a photovoltaic solar cell.

[0002] Photovoltaic solar cells, which are constructed from semiconductor materials, have long been used to convert electromagnetic radiation into electrical energy. In typical photovoltaic solar cells, the semiconductor material represents a significant portion of the total cost of manufacturing the solar cell. Therefore, more cost-effective materials, such as multicrystalline silicon, are increasingly being used. Furthermore, one goal of solar cell research is to reduce the thickness of the semiconductor wafers used in the production of photovoltaic solar cells, thus reducing material costs.

[0003] However, this increases the risk of cracks compromising the stability of the semiconductor material. Such cracks can arise, for example, due to material defects in a silicon ingot, due to stresses during the production of the wafer raw material, due to defects and mechanical stress during wafer sawing, and due to mechanical stress during transport and handling. In addition, the semiconductor material is exposed to mechanical and thermal stresses during the manufacturing process.

[0004] If the semiconductor material breaks during the manufacturing process, high costs arise because, among other things, production must be stopped. However, even such semiconductor material that survives the entire manufacturing process despite a crack can lead to a significant loss of performance in the photovoltaic solar cell when used later.

[0005] For this reason, there is a high demand for methods for detecting cracks in a semiconductor structure, which may be the aforementioned semiconductor raw material, such as a semiconductor wafer for manufacturing a solar cell, the precursor of a solar cell in the manufacturing process, or the finished solar cell. Since the cracks typically have dimensions in the micrometer range, they are also referred to as microcracks.

[0006] It is known to detect cracks using so-called IR transmitted light images. For this purpose, the semiconductor structure is subjected to a transmitted light measurement using radiation in the infrared range and then measured using an imaging method such as a CCD camera. Due to the different absorption in the IR range, a user can visually detect cracks in the image obtained in this way. Crack detection using photoluminescence is also known: For this purpose, photoluminescence radiation is generated in the semiconductor structure and measured with spatial resolution using an imaging method such as a CCD camera. The application of photoluminescence measurements is generally known and is described, for example, in Trupke, T. Progress with luminescence imaging for the characterisation of silicon wafers and solar cells. in Proceedings of the 22nd European Photovoltaic Solar Energy Conference. 2007. Milan, Italy.However, it is often not possible to distinguish cracks from other effects such as recombination-active defects.

[0007] Furthermore, crack detection using resonance ultrasonic vibration (RUV) is known and described, for example, in Monastyrskyi, A., et al. Resonance Ultrasonic Vibrations for In-line Crack Detection in Silicon Wafers and Solar Cells. in Proceedings of 33rd PVSC. 2008.

[0008] DE 199 14 115 A1 discloses a classification of wafers, solar cells, or solar modules with regard to defects and their quality. For this purpose, samples, e.g., in the form of solar cells, are prepared in a training system. These samples exhibit a characteristic surface texture as well as near-surface defects such as cracks or breakouts. document

[0009] DEMANT M ET AL, “Analysis of luminescence images applying pattern recognition techniques”, EU PVSEC PROCEEDINGS : 25TH EUROPEAN PHOTOVOLTAIC SOLAR ENERGY CONFERENCE AND EXHIBITION / 5TH WORLD CONFERENCE ON PHOTOVOLTAIC ENERGY CONVERSION, 6-10 SEPTEMBER 2010, VALENCIA, SPAIN, WIP-RENEWABLE ENERGIES, DE, (20100901), ISBN 978-3-936338-26-3, pages 1078 - 1082, XP009162280 [X] 2-6,11,14 discloses a pattern recognition method for evaluating the quality of wafers with surface defects such as deformations, deposits, and / or cracks (see sections 2.1-2.3). These defects lead to a deviation in the measurable open-circuit voltage of the wafer compared to a defect-free wafer.

[0010] The invention is based on the object of creating a reliable and, in particular, industrially applicable method for crack detection in a semiconductor structure which is a photovoltaic solar cell, a precursor in the manufacturing process of a photovoltaic solar cell or raw material for the manufacture of the photovoltaic solar cell.

[0011] This object is achieved by a method for providing a prediction model for crack detection on a semiconductor structure according to claim 1 and by a method for crack detection on a semiconductor structure according to claim 2. Advantageous embodiments of the methods according to the invention can be found in claims 3 to 13. The invention is further achieved by a device for crack detection on a semiconductor structure according to claim 14. The wording of all claims is hereby incorporated into the description by reference.

[0012] The invention is based on the applicant's finding that learning algorithms are suitable for creating a prediction model for crack detection in a semiconductor structure, so that crack detection is carried out based on the prediction model. The method and device according to the invention thus differ fundamentally from the crack detection methods previously used in the photovoltaic field, since machine learning is used for the first time using a learning algorithm for crack detection.

[0013] As previously described, the crack detection methods and apparatus according to the invention relate to the photovoltaic field. Accordingly, the term "semiconductor structure" is used here and below to refer to raw material such as silicon wafers, and in particular multicrystalline silicon wafers, for the production of a solar cell, as well as the precursor of a photovoltaic solar cell at any point in the manufacturing process and the finished solar cell.

[0014] The method according to the invention for providing a prediction model for crack detection on a semiconductor structure comprises the following method steps: In a method step A, a reference semiconductor structure is provided, which reference semiconductor structure has at least one crack.

[0015] In a method step B, crack data are provided for the at least one crack. The crack data include geometric location data regarding the position of the crack on the reference semiconductor structure.

[0016] The semiconductor structure typically represents a planar element, so that the crack can typically be described with respect to the position of the front or back of the semiconductor structure, i.e., typically using two-dimensional geometric data. It is within the scope of the invention that the crack data comprise a plurality of points and / or lines that describe the extent of the crack and / or that the crack data comprise a crack center at which at least two lines of a star-shaped crack intersect.

[0017] It is also within the scope of the invention for the reference semiconductor structure to be a semiconductor structure that has a crack, the geometric position of which has been measured using another measurement method. Alternatively or additionally, it is also possible to induce the crack at a predetermined geometric position in the reference semiconductor structure by scratching or similar methods.

[0018] In a method step C, the spatially resolved measurement of the reference semiconductor structure is carried out. For this purpose, a spatially resolved measurement of the photoluminescence radiation generated in the semiconductor structure and / or a spatially resolved measurement of the infrared absorption (IR absorption) of the semiconductor structure is carried out for a plurality of local measurement points. To carry out these spatially resolved imaging methods, known measuring equipment and, in particular, CCD cameras for spatially resolved measurement can be used. It is essential that a spatially resolved measurement is carried out for a plurality of local measurement points. The measurement points cover at least one area on the surface of the semiconductor structure that includes the crack.

[0019] In process step D, a prediction model is created. For this purpose, a learning algorithm is trained using the spatially resolved measurement data obtained in process step C and the crack data provided in process step B.

[0020] In contrast to previously known crack detection methods, a physical model is not used to distinguish those measurement points associated with a crack from the other measurement points based on the measurement signals. Instead, a learning algorithm is trained based on the measurement data and the crack data, i.e., the description of the existing crack known with regard to the reference semiconductor structure. The specifications and criteria (i.e., the characteristics of these descriptions) that characterize the crack and distinguish it from other elements, such as recombination-active geometric structures, with regard to the measurement signals are thus not predetermined, but rather developed by training the learning algorithm.

[0021] Training the learning algorithm involves the following steps: In a method step D1, at least one descriptor is created for at least one local descriptor point. A test region is specified or determined for the descriptor point, and the descriptor is created based on the measurement data within the test region. This descriptor is a feature vector and / or a feature distribution and / or a feature histogram.

[0022] The test region thus comprises at least a subset of the measuring points from method step C. It is also within the scope of the invention that the test region comprises all measuring points from method step C.

[0023] The descriptor point is a local point for which the said descriptor is formed, so that a description for the local descriptor point is formed based on the feature vector and / or the feature distribution and / or the feature histogram, taking the test region into account. The features of the feature vector and / or the feature distribution and / or the feature histogram are based on the measurement data determined in method step C. It is within the scope of the invention that the features arise directly from the measurement data. However, it is particularly advantageous for the features to be a further processing and / or correlation, in particular a structural description, of multiple measurement data.

[0024] In a process step D2, the learning algorithm is trained using the descriptor and the crack data.

[0025] Training can be carried out in a manner known per se using a learning algorithm known per se. It is also within the scope of the invention to train the learning algorithm using only one descriptor point and associated descriptor and / or using only one crack and the associated geometric data. However, training is advantageously carried out using a plurality of descriptor points and correspondingly created descriptors. Alternatively, and in particular additionally, it is advantageous to carry out the aforementioned method using a plurality of cracks and their associated crack data.

[0026] As already described, the learning algorithm can be trained in a conventional manner. The key requirement here is that, based on the given crack data, it is known whether a crack is present at the descriptor point. This information is crucial for the classification performed by the learning algorithm during training.

[0027] It is within the scope of the invention to use multiple reference semiconductor structures to train the learning algorithm. It is not absolutely necessary for each reference semiconductor structure to have a crack, since descriptor points and descriptors without a crack are also useful for training. However, at least one reference semiconductor structure must have at least one crack, as described above.

[0028] The method according to the invention for crack detection in a semiconductor structure comprises the following method steps: In a process step A, the semiconductor structure is provided.

[0029] In contrast to the reference semiconductor structure, in the semiconductor structure used for crack detection, the geometric position of the crack or the information as to whether the semiconductor structure has a crack at all is of course not known in advance.

[0030] In a method step B, a prediction model is provided. The prediction model was created by training a learning algorithm, preferably by means of a method according to claim 1 or an advantageous embodiment thereof.

[0031] In a method step C, a spatially resolved measurement of the semiconductor structure is carried out by spatially resolved measurement for a plurality of local measurement points of photoluminescence radiation generated in the semiconductor structure and / or spatially resolved measurement of the IR absorption of the semiconductor structure. Method step C can thus be identical to method step C according to claim 1 or an advantageous embodiment thereof. In particular, it is advantageous to perform the same type of measurement when creating the prediction model and during crack detection, i.e., a spatially resolved measurement of photoluminescence radiation or a spatially resolved measurement of the IR absorption.

[0032] In a method step D, a determination is made for at least one local test point as to whether a crack is present at this test point. The local test point represents a point on a surface of the semiconductor structure and preferably corresponds to one of the measurement points according to method step C.

[0033] The determination includes the following procedural steps: In a method step D1, at least one descriptor is created for the test point, in which a test region is specified and determined for the test point and the descriptor is created on the basis of the measurement data within the test region, which descriptor is a feature vector and / or a feature distribution and / or a feature histogram.

[0034] Method step D1 in crack detection is thus analogous to method step D1 in the creation of the prediction model according to claim 1, wherein in claim 1 the descriptor is created for a corresponding descriptor point and in claim 2 the descriptor is created for a corresponding test point. All preceding and subsequent explanations regarding descriptor creation can thus be applied - at least analogously - both with regard to method step D1 according to claim 1 and with regard to method step D1 according to claim 2.

[0035] Therefore, it is also essential for crack detection that a descriptor is created for the test point, which thus represents features in the form of a feature vector and / or a feature distribution and / or a feature histogram, so that properties for the test point are mapped in the descriptor, taking into account the measurement data obtained in process step C within the test region.

[0036] In a process step D2, the descriptor and the prediction model are used to determine whether there is a crack at the test point.

[0037] The method for crack detection according to the invention is thus characterized in that only a prediction model has to be provided, for example by carrying out a method according to claim 1, and then a measurement according to method step C is carried out in a manner known per se as a photoluminescence measurement or measurement of the IR absorption and by descriptor formation it is possible to determine by means of the prediction model whether there is a crack at the test point.

[0038] It is within the scope of the invention to perform the aforementioned determination according to method step D2 at only one test point. Preferably, the determination is performed at a plurality of test points. Thus, it is within the scope of the invention that each of the local measuring points at which a measurement is performed in method step C is successively selected as a test point.

[0039] As a result, both the determination of whether the semiconductor structure has a crack and the determination of the geometric position and / or extent of the crack are ensured by means of the crack detection method according to the invention, without having to specify a relationship between the measurement data according to method step C and abstract crack properties by means of hypotheses, specifications or derivations from a physical model.

[0040] Investigations by the applicant have shown that the development of a prediction model according to claim 1 and the application according to claim 2 are outstandingly suitable for detecting cracks in a semiconductor structure. The method for crack detection according to the invention thus represents a significant improvement over previously known methods for crack detection. In particular, by training the learning algorithm when carrying out the method according to claim 1, even when using a photoluminescence measurement method in the respective method steps C, a distinction can be made with high accuracy between a crack on the one hand and other geometric structures such as recombination-active structures, wherein the criteria for distinguishing a crack structure from other structures do not have to be predetermined, but are simply developed by training the learning algorithm.

[0041] The methods according to the invention further have the advantage that a division between creating the prediction model on the one hand and applying the prediction model for crack detection on the other hand is possible: For example, a manufacturer can train a learning algorithm with high accuracy by precisely preparing a reference semiconductor structure and / or training it on the basis of a large number of cracks and crack data, and the prediction model thus developed can be made available to the user.

[0042] Therefore, it is not necessary for the user to train the learning algorithm. The user can also apply the manufacturer-trained prediction model directly for crack detection.

[0043] In the method for crack detection according to the invention, a crack reconstruction is preferably carried out in a method step E in which the geometric data of the crack shape are determined. Thus, starting from a test point that has been assigned to a crack, a reconstruction is carried out in order to determine at least the partial, preferably the complete, geometric shape of a crack. In particular, it is advantageous in method step E to determine an orientation for each measurement point within a local reconstruction region around the test point and to determine the measurement points associated with the crack by means of a similarity comparison with a star-shaped and / or linear pattern through pattern recognition.The above-mentioned preferred embodiments are particularly advantageous when using photoluminescence measuring methods and in particular when using multicrystalline silicon semiconductor structures.

[0044] This preferred embodiment of the method according to the invention is based on the applicant's finding that cracks in semiconductor structures typically exhibit a star-shaped pattern. This allows crack reconstruction to be performed using a combination of known methods for detecting and reconstructing a star-shaped and / or linear pattern.

[0045] As previously described, it is within the scope of the invention to successively use a plurality of location points as descriptor points or test points in method step D when creating the prediction model and / or during crack detection. For example, all local measurement points measured in the respective method steps C can be used.

[0046] However, it is advantageous to limit the number of descriptor points and / or checkpoints using methods for keypoint determination. Such methods are also known as methods for determining a point of interest (POI). In particular, it is advantageous to use a method for finding keypoints to determine the descriptor point and / or checkpoint, preferably using keypoint determination according to at least one of the SIFT, GLOH, and / or SURF methods or known methods for feature detection such as the Canny or Harris detector, or similar, derived, or analogous methods. The aforementioned methods are known per se and are described, for example, in: - GLOH: Krystian Mikolajczyk and Cordelia Schmid “A performance evaluation of local descriptors”, IEEE Transactions on Pattern Analysis and Machine Intelligence, 10, 27, pp 1615--1630, 2005. - SIFT: Lowe, David G. (1999). „Object recognition from local scale-invariant features“. Proceedings of the International Conference on Computer Vision. 2. pp. 1150-1157. doi:10.1109 / ICCV.1999.790410. http: / / doi.ieeecs.org / 10.1109 / ICCV.1999.790410. - SURF: Bay, H., Tuytelaars, T., Gool, L.V., „SURF: Speeded Up Robust Features“, Proceedings of the ninth European Conference on Computer Vision, May 2006. - Harris: C. Harris and M. Stephens: A combined corner and edge detector. In: Proceedings of the 4th Alvey Vision Conference. 1988, S. 147-151.

[0047] Investigations by the applicant have shown that, for key point determination, it is particularly advantageous to determine the at least one key point using an orientation-controllable filter. Advantageously, the determination is carried out using a filter, in particular for calculating a derivative of the measured data, preferably the second derivative. Investigations by the applicant have shown that, in particular, a bar filter formed from the second derivative of a Gaussian filter is suitable for key point determination.

[0048] To select the key points, it is within the scope of the invention to determine the key points within the entire area, in particular starting from all measurement points. It is also within the scope of the invention to specify restricted geometric areas on the surface of the semiconductor structure and / or a subset of the measurement points measured in method step C within which key points are determined.

[0049] When determining key points, various properties and characteristics of a local region or measurement point can be used to identify relevant image sections, such as the intensity characteristics of the measurement signal or the distribution of the intensity values ​​of the measurement signal. These characteristics can be mapped, for example, by determining moments such as variance, mean, symmetry, or other known moments and used to determine the key point.

[0050] Likewise, the use of the characteristics of features such as corners, edges, lines, or round structures with regard to their geometric characteristics is possible. Such features can be determined, for example, from filter responses when using filters and / or from the combination of several filters – possibly with pre- and further processing of the results. Examples of applicable filters are Sobel, Prewitt, Gaussian filters, Difference-of-Gaussian, and Laplacian-of-Gaussian. Likewise, the use of tensors such as the Harris detector and Determinant of Hessian is within the scope of the invention, as is the use of controllable filters and the use of wavelets such as Gabor and Haar wavelets. Furthermore, the use of the SUSAN corner detector or the Canny edge detector is within the scope of the invention.

[0051] Investigations by the applicant have shown that filtering with anisotropic bar filters, which represent the second derivative of the Gaussian filter and are described, for example, in J. Malik, Contour and Texture Analysis for Image Segmentation. International Journal of Computer Vision 43(1), pp. 7-27, 2001, is particularly advantageous.

[0052] It is particularly advantageous to use these filters with different spatial orientations. The positive filter responses are correlated, preferably summed, for different spatial orientations of the filter. This results in a form of Haar integration over the rotation group SO(2), as described, for example, in H. Schulz-Mirbach. Application of Invariance Principles for Feature Extraction in Pattern Recognition. Dissertation, Technical University of Hamburg-Harburg, February 1995. Series 10, No. 372, VDI-Verlag.

[0053] Investigations by the applicant have shown that the resulting value is particularly elevated in the center of a crack and is also rotationally invariant. This makes it easy to determine a crack center regardless of the rotation, i.e., the current course of the crack lines on the respective measurement image. This has the advantageous effect that, regardless of the rotation of the cracks or the semiconductor structure during measurement, equally high values ​​are obtained during analysis and thus equally detected.

[0054] Furthermore, the use of filters with different scales is within the scope of the invention. This provides a robust result against changes in crack size.

[0055] It is particularly advantageous to use the evaluation of the results at different scale levels to estimate the size of the area to be considered: Correspondingly high results at a high scale level indicate a larger crack, so that a larger image section is advantageously used to determine the key points, for crack detection and / or crack reconstruction, compared to results for which results are available at lower scale levels.

[0056] In general, the determination of key points has the advantage that the number of potential crack regions can be limited to a significantly smaller number of key points, thus accelerating the process.

[0057] It is also within the scope of the invention to first determine key points using other measurement methods. For example, the detection result of an IR measurement can be used in a conventional manner to determine key points. Different measurement methods can also be combined: For example, the determination of key points based on an IR measurement is within the scope of the invention, after which a photoluminescence measurement for the key points previously obtained using the IR measurement is evaluated for crack detection.

[0058] Preferably, the descriptor in method step D1 according to claim 1 and / or claim 2 is determined according to at least one of the methods SIFT, GLOH, HOG, LESH, and / or SURF and / or a variant of these methods, in which a descriptor is generated based on the features of a test region. Alternatively or additionally, it is advantageous to use the maximum filter response of a calibratable and rotatable filter, preferably a bar filter as described above, to generate the descriptor. The LESH method is also known per se and is described, for example, in Sarfraz, S., Hellwich, O.: "Head Pose Estimation in Face Recognition across Pose Scenarios", Proceedings of VISAPP 2008, Int. conference on Computer Vision Theory and Applications, Madeira, Portugal, pp. 235-242, January 2008 (Best Student Paper Award).

[0059] In particular, it is advantageous that the descriptor is created comprehensively by the following process steps: In a method step D1.1, an amplitude and an orientation are determined for each location point of the test region, for example for each measuring point within the test region, preferably by determining a characteristic value for each measuring point of the test region, which characteristic value is more preferably the gradient of the measured value for the respective measuring point or the maximum filter response and orientation of a controllable filter for the respective measuring point.

[0060] In a process step D1.2, a feature histogram is created for at least selected key points of the test region.

[0061] It is particularly advantageous that a main crack orientation is additionally determined and that the feature histogram is corrected based on the main crack orientation in order to obtain a rotation-invariant feature histogram.

[0062] When creating the descriptor, the test region around the descriptor point or test point is described based on features. This feature description can consist, for example, of one or more features, a feature distribution or their parameters, or histograms of specific characteristics of these features. Known feature descriptions can be used according to the descriptors using the known algorithms SIFT, GLOH, HOG, LESH, or SURF. The basis for determining the descriptors is the measurement data acquired in process step C.

[0063] Advantageously, in the aforementioned preferred embodiment, the main crack orientation is determined between method steps D1.1 and D1.2, and in method step D1.2, the feature histogram is created by sampling the measurement points within the test region using a sampling algorithm. The use of known sampling algorithms according to GLOH and / or SIFT or variants thereof is particularly advantageous here.

[0064] To create the descriptors, a specific sampling and / or division of the image section can be used. In particular, known sampling methods according to the aforementioned algorithms can be used.

[0065] To form the descriptors, the selection of features can be done as is known in the algorithms SIFT, GLOH, HOG, LESH or SURF.

[0066] Investigations by the applicant have shown that, in particular, the use of the maximum filter response of the anisotropic bar filter, as described above, is advantageous here as well. Instead of using the gradient orientation and gradient amplitude as known from SIFT and GLOH, the orientation and amplitude of the bar filter with the maximum filter response are thus advantageously used.

[0067] It is within the scope of the invention to pre- or post-process the feature description when creating the descriptors. In particular, rotation invariance can be achieved as described above by correcting the descriptor according to the determined main orientation of the crack. Furthermore, robustness against contrast changes or noise can be achieved by scaling and / or smoothing the descriptor.

[0068] Preferably, in method step D1 according to claim 1 and / or claim 2, the test region is determined comprising the following method steps: First, at least one local key point is determined as described above, in particular with regard to claims 5 and 6.

[0069] Subsequently, the measurement points within the test region are determined by specifying a geometric extent and shape for the test region. Preferably, a rectangular, elliptical, or circular region is specified for the test region, with the key point as its center.

[0070] In an advantageous embodiment of the method according to the invention, the geometric position of grain boundaries is specified as grain boundary data within the semiconductor structure and / or determined by means of a spatially resolved measuring method and a weighting is carried out depending on the grain boundary data in that - when determining a key point as described above, those location points which lie on a grain boundary are not selected as a key point or are given a lower weighting than the other location points when determining the key point and / or - when creating a descriptor in process step D when creating the prediction model and / or during crack detection, those location points which lie on a grain boundary are not taken into account or are given a lower weighting than the other location points and / or - when reconstructing the crack, as described above, those location points which lie on a grain boundary are not taken into account or are given a lower weighting than the other location points.

[0071] This preferred embodiment is therefore particularly applicable when the inventive method is applied to multicrystalline silicon wafers. If the geometric position of the grain boundaries is known in advance or can be easily determined using the available measuring equipment, the susceptibility to errors in crack detection can be further reduced by using the information about the position of the grain boundaries in the inventive method as described above.

[0072] The learning algorithm used to create the prediction model or to perform crack detection can be a known learning algorithm. In particular, the use of a neural network, a Bayesian classifier, or a kernel machine is within the scope of the invention. Investigations by the applicant have shown that the use of a support vector machine algorithm is particularly advantageous. This is due to the fact that the support vector machine determines the optimal hyperplane for separating the feature descriptions of the "crack" and "no crack" classes to solve the classification task. This results in a significant advantage because the separation surface is optimally located between the classes. Feature descriptions that do not lie on the class boundary have no negative influence on the course of the boundary.To solve the classification problem using a separation plane, the data is mapped into a higher-dimensional space using the so-called "kernel trick." This procedure is well known and is described, for example, in Bernhard Schölkopf and Alex Smola: Learning with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond (Adaptive Computation and Machine Learning), MIT Press, Cambridge, MA, 2002, ISBN 0-262-19475-9.

[0073] As previously described, the learning algorithm uses the relationship between the feature description of the descriptors and the known information regarding whether a crack is present at a given location to develop the prediction model. Using the prediction model, unknown image sections or the descriptors generated for them can then be classified as cracks or non-cracks during crack detection.

[0074] In the method for crack detection according to the invention, as described above, it is advantageous to perform a crack construction in method step E. Here, the crack structures can be reconstructed based on structural properties of the region around the corresponding test point in the considered test region. The structural properties of the location points belonging to the crack are distinguished from structural properties that do not belong to the crack. Structural properties can be position, intensity values, gradients, and filter responses.

[0075] Investigations by the applicant have shown that pattern recognition is advantageously carried out assuming a star-shaped or linear structure of the crack.

[0076] In particular, in the case of linear or star-shaped cracks, it is advantageous to design the reconstruction in such a way that first a crack center is determined, preferably as previously explained in the description of the finding of key points, and defined as the center of a test region.

[0077] Subsequently, structural properties of the measurement points in the test region are extracted. The description of the measurement points surrounding the key point is compared with the description of a corresponding point in a star-shaped pattern using well-known pattern recognition. Structures that display a similar structure, such as orientation, to the regions of the star-shaped pattern are therefore highly likely to belong to the crack and can therefore be classified as belonging to the crack. To calculate the similarity of the structural property, additional weighting can advantageously be applied, for example, using the intensity values ​​of the respective location point, gradients, or a corresponding filter response. If the similarity is high, a decision is made as to whether the location point belongs to a crack. This decision can be made, for example, using the well-known hysteresis method starting from the crack center.

[0078] Investigations by the applicant have shown that the orientation and intensity of the structure are particularly suitable for describing the structural property.

[0079] Here, too, filtering with the previously described anisotropic bar filters in different orientations has proven advantageous for calculating the orientation. Advantageously, the maximum filter response and its orientation are used as information about the structure.

[0080] Further features and preferred embodiments of the invention are explained below with reference to figures and exemplary embodiments. In the following: Fig. 1 in partial image a) a photoluminescence image of a multicrystalline silicon wafer which has a crack approximately in the middle, whereby in partial image b) an enlarged detail of the region around the crack is shown; Fig. 2 a representation of a maximum filter response when applying a bar filter to the region containing the crack; Fig. 3 in part a) an unsmoothed orientation histogram of the crack environment and in part b) the orientation histogram after smoothing and Fig. 4 the result after crack reconstruction.

[0081] Embodiments of the method according to the invention for providing a prediction model for crack detection and of the method according to the invention for crack detection are explained with reference to the figures.

[0082] The reference semiconductor structure is a multicrystalline silicon wafer, approximately square in shape, with an edge length of approximately 15.6 cm and a thickness of approximately 180 µm. Such microcrystalline silicon wafers represent a typical starting material for the production of photovoltaic solar cells.

[0083] To form this semiconductor structure as a reference semiconductor structure, a metal tip, a so-called plunger, was dropped from a predetermined height onto the semiconductor structure using a plunger machine, so that a crack forms and, in addition, the geometric position, i.e. the X and Y coordinates of the crack center, which corresponds to the point of impact of the metal tip on the wafer, is known.

[0084] The reference semiconductor structure prepared in this way is provided in a process step A. Furthermore, as described above, the spatial coordinates of the crack center are known and are thus provided as crack data in a process step B.

[0085] In process step C, the reference semiconductor structure is measured with spatial resolution. A conventional photoluminescence measurement apparatus is used: Optical excitation of the semiconductor by irradiation with electromagnetic radiation generates electron-hole pairs. Due to the recombination processes, photoluminescence radiation is emitted from the reference semiconductor structure, which is measured with spatial resolution using a silicon CCD camera. As a result of process step C, a measurement value is thus available for a large number of measurement points distributed in a grid pattern across the surface of the reference semiconductor structure. This measurement value corresponds to the intensity of the emitted photoluminescence radiation for this measurement point or at least correlates with the intensity.

[0086] The measurement result is in Fig. 1a. The high spatial resolution of (1024x1024) points is clearly visible. Approximately in the center, a star-shaped crack structure can already be seen. The region surrounding the crack structure is marked by a black rectangle. Fig. Figure 1b shows an enlarged section of this rectangle, where the crack structure (within the black circle) is more clearly visible.

[0087] However, the Fig. 1a and Fig. 1b shows that a large number of other structures have similar measured values ​​and therefore it is a high demand on the method used to separate the measuring points belonging to a crack from the other measuring points based on the measuring signals.

[0088] Therefore, in the exemplary embodiment of the method according to the invention, in a method step D, a prediction model is created by training a learning algorithm based on the spatially resolved measurement data determined in method step C and the crack data provided in step B: It would be possible to create a descriptor for each measurement point as described above and then train the learning algorithm.

[0089] In the embodiment described here, however, potential crack centers are first detected to locate key points, so that the processing time and effort are significantly reduced, since the subsequent steps are only carried out for the key points.

[0090] In principle, the previously described methods for locating key points (so-called points of interest, POI) are applicable here. However, the applicant's investigations have shown that, due to the star-shaped structure of a crack, the use of anisotropic bar filters is particularly advantageous for locating key points. In the present embodiment, the second derivative of the anisotropic Gaussian filter is used. This filter is created in various orientations, in this case in the orientations 0,π6,2π6,3π6,4π6,5π6.

[0091] After filtering the measured values, the positive filter responses of the individual bar filters are summed. This results in a high filter response, particularly in the crack center, i.e., the point or area where crack lines intersect. To further narrow down the potential crack centers, measurement points are extracted whose summed filter responses exceed a specified threshold, preferably according to the following formula 1: Ibinary(x,y)={1,if A(x,y)≥c⋅max(A)0,otherwise

[0092] Here, A(x,y) represents the sum of the positive filter responses of the image with the bar filter at different orientations for the measuring point with coordinates (x, y). Each measuring point is thus assigned a binary value, which is selected as 1 if the filter response for this measuring point is greater than or equal to a threshold value multiplied by the maximum filter response of all measured values. In all other cases, the value 0 is assigned. The sensitivity is regulated by the choice of threshold value c; the larger c is selected, the fewer key points are obtained and vice versa. The threshold value c is determined empirically, preferably in such a way that all crack centers are included. It has been shown that crack centers usually have very high values.

[0093] In Fig. Figure 2 shows a processing of the measurement data, showing the summed filter responses for each pixel. It is clearly visible that a particularly high value is obtained at the crack center.

[0094] Subsequently, preferably connected structures are reduced to a single location point. This is done in a conventional manner using a morphological operation in which objects are shrunk to points. The result, i.e., the remaining points, represent key points, each of which is potentially a crack center.

[0095] For each key point, a descriptor is then created in a process step D1. Thus, each key point is successively selected as a descriptor point, a test region is specified for the descriptor point, and a descriptor is created.

[0096] Two variants of descriptor creation are described below: Firstly, descriptor creation using gradients and secondly, descriptor creation using bar filters.

[0097] When creating the descriptor in a first variant of the exemplary embodiment using gradients, the creation is analogous to the well-known GLOH method. First, gradient orientations are considered. However, since noise in the measured values ​​has a strong influence on the gradients, the measured values ​​are first smoothed by connecting a Gaussian filter. This procedure is well-known. By convolving the measured data with the derivative of the Gaussian filter in the x and y directions, the measured data are smoothed and, at the same time, the gradients in the x and y directions are determined. From this, the orientation and amplitude of the gradient vector for the desired position (x, y) are determined. The orientation in two-dimensional space can be represented with the values ​​[0, 2 π].

[0098] In method step D1 of the exemplary embodiment, a test region is first defined for each key point. In the present example, this is defined as a circle around the key point with a radius of 40 pixels (measurement points). Subsequently, the gradient orientation and gradient amplitude are determined for each measurement point within the test region, as previously described, in a method step B1.1.

[0099] To align the gradient orientations according to the structure around the key point, the orientation of the structure around the key point is first determined as follows: The previously determined gradient orientations within the test region are mapped in an orientation histogram, which has, for example, 36 bins, i.e. 36 individual categories, each covering a uniform orientation range. The result of such a classification is Fig. 3a. It may happen that between two bins with maximum values, a bin with a low value lies. Preferably, such a low-value orientation is not included in the calculation of the keypoint orientation. It is therefore advantageous to smooth the histogram using known methods. The result of such smoothing is shown in Fig. 3b.

[0100] To obtain a rotation-invariant result, i.e., a rotation-invariant descriptor for the keypoint, the test region is rotated around the previously determined orientation at the keypoint. Depending on the type of descriptor generation, this can be done indirectly by correcting the gradient orientation and the descriptor according to the keypoint orientation.

[0101] Based on this, the key point descriptor is determined for the key point, which was determined in particular depending on the measurement data of the measurement points within the test region.

[0102] Previously, the keypoints were determined, their orientation calculated, and the test region rotated according to the keypoint orientation. This procedure is well known and described, for example, in (Lowe, DG, Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 2004: pp. 91-110.).

[0103] A descriptor is then created for each key point: This can be done analogously to the well-known GLOH method.

[0104] A simple way to create descriptors is, for example, to create a feature vector based on the gradient orientations. Each orientation or orientation range represents a feature. The characteristic value of a specific orientation results from the sum of the gradient amplitudes with this orientation.

[0105] In addition to the aforementioned descriptor creation, conventional sampling methods can also be applied. Polar sampling is particularly advantageous here, as the cracks to be examined typically exhibit higher gradient amplitudes near the crack center compared to the more distant measurement points. Therefore, polar sampling, whose center coincides with the key point, is particularly advantageous. A further advantage is that the aforementioned rotation of the test region according to the key point orientation can be easily implemented in polar coordinates.

[0106] An additional improvement in descriptor generation is achieved, analogous to GLOH, by dividing the test region into different subregions. A feature vector is created for each subregion. The descriptor is then derived from a combination of the feature vectors of the various subregions. This provides additional information about the structure of the region, which is extremely beneficial for distinguishing cracks from other structures.

[0107] Investigations by the applicant have shown that, in a variant of the previously described use of gradients for generating descriptors, it is advantageous to use the filter responses of the previously described bar filter to calculate the descriptors. For this purpose, the orientation of the bar filter that generates the maximum filter response is selected instead of the gradient orientation.

[0108] Analogous to the previously described procedure, the corresponding intensity of the maximum filter response is selected instead of the gradient amplitude. The difference between this variant and the variant using gradients is particularly evident in the amplitude matrix: While in the variant using gradients, the edges of the cracks exhibit maximum values, in the variant using the filter response of the bar filter, the crack itself exhibits the maximum value.

[0109] Fundamentally, this variant of descriptor generation differs from the previously described only in the choice of input: Instead of the gradient values ​​of the test region surrounding the keypoint, the filter response of the anisotropic bar filter is now used. The rest of the process is analogous.

[0110] Regardless of the variant chosen, a learning algorithm is finally trained in a process step D2 using the created descriptors and the crack data provided in process step B.

[0111] Investigations by the applicant have shown that a support vector machine in particular can be used reliably.

[0112] The result of this embodiment of the method according to the invention is thus a prediction model created by training the support vector machine.

[0113] In a second embodiment, an embodiment of a method for crack detection according to the invention will now be described.

[0114] For this purpose, the prediction model described above is used.

[0115] In a process step A, a semiconductor structure is provided and in a process step B, the previously described prediction model is provided.

[0116] Subsequently, a spatially resolved measurement of the semiconductor structure is performed in process step C. This is preferably carried out analogously to the spatially resolved photoluminescence measurement described in the first embodiment according to process step C described therein.

[0117] As a result, measurement data are available for a large number of local measurement points of the semiconductor structure, which correlate with the respective local intensity of the generated photoluminescence radiation.

[0118] In a process step D, it is determined for at least one local test point whether a crack is present at this test point.

[0119] Here too, it is fundamentally possible to carry out such a determination for each of the measuring points, for example.

[0120] However, it is particularly advantageous, as described above, to first determine key points for crack detection and to perform the aforementioned test only for the identified key points. The key points can be determined as described for the first embodiment.

[0121] For each test point, a descriptor is created in a method step D1, in which a test region is specified or determined for the test point. This descriptor creation is carried out in the same way as previously described in the first embodiment.

[0122] In contrast to the first embodiment, during crack detection, in a method step D2, the prediction model is applied in such a way that, using the descriptor and the prediction model, it is determined whether a crack is present at the test point. Using the prediction model, classification is thus possible by classifying corresponding key points as points belonging to a crack or points not belonging to a crack.

[0123] In the second embodiment of the method according to the invention described here, a crack reconstruction is then carried out for the key points classified as belonging to a crack in method step D: Crack construction is performed using the well-known hysteresis method: starting at the key point, which is typically the crack center, the crack is reconstructed in its original shape, point by point. The reconstruction is based on the assumption that the crack propagates in a star-shaped pattern, and based on this, the similarity of the crack's surroundings to a star-shaped pattern is determined. A measurement point in the vicinity of the crack center therefore most likely belongs to the crack if this measurement point lies on a beam structure that propagates from the crack center. It is therefore advantageous to calculate the orientation and filter response of a beam filter for each location point for the reconstruction, as described above.

[0124] The orientations of the measurement points in the vicinity of the crack center are compared with the orientations of a star pattern. Preferably, the cosine between the angular differences can be calculated as a similarity measure and weighted by the amplitude of the respective filter response.

[0125] Preferably, the resulting image, which is created by the hysteresis, is additionally examined for its morphological properties. In particular, it is advantageous to use the area, orientation, and length of the crack to correlate with the stability of the wafer.

[0126] The result of such a crack reconstruction is Fig. 4 shown.

Claims

[1] Method for providing a prediction model for crack detection on a semiconductor structure, which semiconductor structure is a photovoltaic solar cell, a precursor of a photovoltaic solar cell in the manufacturing process, in particular semiconductor material for manufacturing a solar cell, comprising the following method steps: A. Providing a reference semiconductor structure, which reference semiconductor structure has at least one crack; B. Providing crack data relating to the at least one crack, which crack data comprises geometric location data regarding the position of the crack on the reference semiconductor structure; C. Spatially resolved measurement of the reference semiconductor structure by spatially resolved measurement for a plurality of local measuring points of photoluminescence radiation generated in the semiconductor structure and / or spatially resolved measurement of the IR absorption of the semiconductor structure and D. Creating a prediction model by training a learning algorithm based on the spatially resolved measurement data determined in process step C and the crack data provided in process step B, whereby the training of the learning algorithm comprises the following process steps: D1 Creating at least one descriptor for at least one local descriptor point by specifying or determining a test region for the descriptor point and creating the descriptor based on the measurement data within the test region, which descriptor is a feature vector and / or a feature distribution and / or a feature histogram and D2 Training the learning algorithm using the descriptor and the crack data. [2] Method for crack detection in a semiconductor structure, which semiconductor structure is a photovoltaic solar cell, a precursor of a photovoltaic solar cell in the manufacturing process, in particular semiconductor material for the production of a solar cell, comprising the following method steps: A. Providing the semiconductor structure; B. Providing a prediction model, which prediction model was created by training a learning algorithm, preferably according to claim 1; C. Spatially resolved measurement of the semiconductor structure by spatially resolved measurement for a plurality of local measuring points of photoluminescence radiation generated in the semiconductor structure and / or spatially resolved measurement of the IR absorption of the semiconductor structure; D. Determination for at least one local test point whether a crack is present at that test point, comprising the following procedural steps: D1 Creating at least one descriptor for the test point by specifying or determining a test region for the test point and creating the descriptor based on the measurement data within the test region, which descriptor is a feature vector and / or a feature distribution and / or a feature histogram and D2 Determine whether there is a crack at the test point using the descriptor and the prediction model. [3] Method according to claim 2, characterized by that when a crack is detected according to process step D, a crack reconstruction is carried out in a process step E by determining the geometric data of the crack shape. [4] Method according to claim 3, characterized bythat in method step E an orientation is determined for a local reconstruction region around the test point for each measuring point within the reconstruction region and the measuring points belonging to the crack are determined by means of a similarity comparison with a star-shaped pattern by pattern recognition. [5] Method according to one of the preceding claims, characterized by that in method step D1 according to claim 1 the at least one descriptor point and / or in method according to claim 2 the at least one check point is determined by means of a method for finding key points, preferably according to at least one of the methods SIFT, GLOH and / or SURF. [6] Method according to claim 5, characterized bythat the determination of the at least one key point is carried out by means of an orientation-controllable filter, preferably by means of a bar filter, in particular using a derivative, preferably the second derivative, of a Gaussian filter. [7] Method according to one of the preceding claims, characterized by , that the descriptor in method step D1 according to claim 1 and / or claim 2 is determined according to at least one of the methods SIFT, GLOH, HOG, LESH and / or SURF and / or that the maximum filter response and orientation of a scalable and rotatable filter, preferably a bar filter, is used to create the descriptor. [8] Method according to claim 7, characterized by that the descriptor is created comprehensively following procedural steps: D1.1 Determination of an amplitude and an orientation for each location point of the test region, preferably by determining a characteristic value for each measuring point of the test region, which characteristic value is more preferably the gradient of the measured value for the respective measuring point or the maximum filter response and orientation of a controllable filter for the respective measuring point; D1.2 Creation of a feature histogram for at least selected key points of the test region. [9] Method according to claim 8, characterized by , that a main crack orientation is additionally determined and that a correction of the feature histogram is carried out based on the main crack orientation in order to obtain a rotationally invariant feature histogram. [10] Method according to one of claims 8 to 9, characterized by , that the main crack orientation is determined between process steps D1.1 and D1.2 and that in method step D1.2 the feature histogram is created by sampling the measuring points within the test region using a sampling algorithm according to GLOH and / or SIFT and / or a variant of GLOH or SIFT. [11] Method according to one of the preceding claims, characterized by , that the test region is determined in method step D1 according to claim 1 and / or according to claim 2 comprising the following method steps: - Determination of at least one local key point according to one of claims 5 to 6, - Determination of the measuring points within the test region based on a given geometric extension of the test region around the key point, preferably a rectangular, elliptical or circular test region whose center is the key point. [12] Method according to one of the preceding claims, characterized by , that the geometric position of grain boundaries is specified as grain boundary data within the semiconductor structure and / or determined by means of a spatially resolved measuring method and that a weighting is carried out depending on the grain boundary data by - when determining a key point according to claim 5, those location points which lie on a grain boundary are not selected as a key point or are given a lower weighting than the other location points when determining the key point and / or - when creating a descriptor in process step D according to claim 1 and / or 2, those location points which lie on a grain boundary are not taken into account or are given a lower weighting than the other location points and / or - when reconstructing the crack according to claim 3, those location points which lie on a grain boundary are not taken into account or are given a lower weighting than the other location points. [13] Method according to one of the preceding claims, characterized by that the learning algorithm is a neural network or a Bayesian classifier or a kernel machine-based learning algorithm, particularly preferably a support vector machine algorithm. [14] Device for crack detection on a semiconductor structure, which semi- conductor structure is a photovoltaic solar cell or a precursor of a photovoltaic solar cell in the manufacturing process, including the semiconductor raw material, comprising a measuring unit for the spatially resolved measurement of luminescence radiation generated in the semiconductor structure and / or spatially resolved measurement of the IR absorption of the semiconductor structure and an evaluation unit for crack detection based on spatially resolved measurement data determined by means of the measuring unit, characterized by , that the evaluation unit is designed for crack detection according to one of claims 2 to 13.

Images