Computer-implemented method for evaluating rods made of semiconductor material

Abstract

The invention relates to a computer-implemented method for evaluating rods made of semiconductor material, wherein the rods are produced by means of a production process in a vapour deposition reactor, wherein the method comprises the following steps: controlling an imaging device having at least one camera to capture one or more images of a batch of rods in a reaction chamber of the vapor deposition reactor after completion of the production process for the batch of rods; and processing input data in at least one machine-learning model, in order to obtain output data, wherein the input data is based on the one or more images, wherein the output data comprises one or more output elements for evaluating the rods.

Description

[0001] Wa12439P / Be

[0002] 1

[0003] Computer-implemented method for evaluating rods made of semiconductor material

[0004] The invention relates to a computer-implemented method for evaluating rods made of semiconductor material, wherein the rods are produced by means of a manufacturing process in a gas phase deposition reactor (300), the method comprising the following steps.

[0005] - Controlling an imaging device with at least one camera to capture one or more images of a batch of rods in a reaction chamber of the gas-phase deposition reactor after completion of the manufacturing process for the batch of rods, and

[0006] - Processing input data in at least one machine-learned model to obtain output data,

[0007] - where the input data is based on one or more images,

[0008] - where the output data includes one or more output elements for evaluating the bars .

[0009] Several examples of the disclosure relate to the evaluation of rods made from a semiconductor material, where the rods are obtained in a manufacturing process in a vapor deposition reactor. Several examples of the disclosure specifically relate to the use of a machine-learned model to enable this evaluation.

[0010] Single-crystal silicon is used in the semiconductor industry to manufacture electronic components such as transistors and memory chips. Polycrystalline silicon (polysilicon) serves as a starting material in the production of single-crystal silicon, for example by crucible pulling (Czochralski process) or by zone melting. Wa12439P / Be

[0011] 2

[0012] (Float zone process). Furthermore, polysilicon is required for the production of multicrystalline silicon, for example by means of ingot casting.

[0013] Silicon carbide (SiC) is also used as a semiconductor material for various electronic components. SiC is characterized by high radiation hardness, a wide band gap, a high saturated electron drift velocity, and a high operating temperature. SiC possesses excellent electronic properties, including the absorption and emission of high-energy protons in the blue, violet, and ultraviolet regions of the spectrum. SiC-based electronic components can be used at temperatures up to 250 °C and exhibit high oxidation resistance. This provides advantages over pure silicon-based electronic components.

[0014] Polycrystalline SiC (Poly-SiC) serves as the starting material for the production of single-crystal SiC-based electronic components, for example using Physical Vapor Transport (PVT) processes.

[0015] Polysilicon and poly-SiC can be produced by a chemical vapor deposition process, which in the case of polysilicon is known as the Siemens process. In this process, support materials are heated by direct electric current in a vapor deposition reactor (hereinafter also referred to simply as the reactor), and a reaction gas containing a silicon-containing component and hydrogen is introduced. For the production of poly-SiC, a carbon-containing component is also required. The silicon-containing component is usually monosilane (SiHA or Wa12439P / Be).

[0016] 3

[0017] Chlorosilane / chlorosilane mixture. A typical example is trichlorosilane (SiHCl, TCS). For polysilicon production, SiFü or TCS in a mixture with hydrogen is predominantly used. For poly-SiC production, for example, a methylsilane or a TCS-methane mixture in a mixture with hydrogen can be used.

[0018] The design of a typical gas-phase deposition reactor for polysilicon production is described, for example, in US 2012 / 0100302 Al. The design of a typical gas-phase deposition reactor for poly-SiC production can be found in US 2023 / 0141427 Al. Generally, essentially identical reactors can be used. The base of such a reactor (base plate) is typically equipped with electrodes that hold the support materials. These support materials are usually filaments (also called filament rods or thin rods), which can consist of materials such as silicon, graphite, or SiC. Typically, two support materials are connected by a bridge to form a pair, which creates an electrical circuit via the electrodes.

[0019] The surface temperature of the support bodies is typically above 1000°C during deposition. At these temperatures, the silicon-containing, or the silicon-containing and carbon-containing, components of the reaction gas decompose, and elemental silicon (SiC) is deposited from the gas phase onto the support bodies. This increases their diameter. Once a predetermined diameter is reached, the deposition process is stopped, the reactor is opened by lifting the reactor hood (also called the reactor mantle), and the resulting rods of semiconductor material (polysilicon or poly-SiC) are removed. After removing the bridge, cylindrical rods are obtained. Wa12439P / Be

[0020] 4

[0021] The morphology of polysilicon and poly-SiC rods can influence their performance during further processing. Fundamentally, the morphology of a polysilicon or SiC rod is determined by the parameters of the deposition process (e.g., rod temperature, reaction gas composition, specific flow rate). Depending on these parameters, pronounced interfaces, including holes and grooves, can form. These are generally not distributed homogeneously within the rod. For example, varying the parameters can result in rods with different (mostly concentric) morphology zones, as described in EP 2 662 335 Al. The dependence of the morphology on the rod temperature is described, for example, in US 2012 / 0322175 Al. The morphology of the rods can vary from compact and smooth to porous and fractured.Compact polysilicon, or poly-SiC, is essentially free of cracks, pores, joints, and fissures. A porous and fractured morphology can, for example, negatively affect the crystallization behavior of polysilicon. This is particularly evident in the CZ process for the production of single-crystal silicon. Here, the use of fractured and porous polysilicon leads to economically unacceptable yields. In general, particularly compact polysilicon results in significantly higher yields in the CZ process.

[0022] Accordingly, polysilicon, in particular, is distinguished and classified not only by purity and fracture size but also by its morphology. Since the term morphology can encompass various parameters such as porosity (sum of closed and open porosity), specific surface area, roughness, luster, and color, a Wa12439P / Be

[0023] Reproducible determination of morphology presents a challenge.

[0024] US Patent 2023 / 0011307 describes a method in which 2D or 3D images are generated after the removal of silicon rods or fragments produced from them. From these images, at least two surface structure parameters are generated using various image processing techniques and combined into a single morphology parameter. This parameter is then used for classification. A disadvantage of this method is that removal of the rods is mandatory. The mechanical stress caused by removal can create additional cracks, which may lead to incorrect conclusions about the deposition process. Furthermore, only a small section of the rod surface or internal structure can usually be examined. Inferences about the position of the rods are also generally not possible.

[0025] US Patent 2022 / 0234900 describes a process in which thermographic images of the rods are generated through a sight glass during polysilicon deposition, and a morphology index is generated from these images using image processing. This index allows the quality of the deposited polysilicon to be determined and influenced at any given time. A disadvantage is that cracks or fractures in the rods may only appear during cooling, i.e., after the deposition process is complete. Furthermore, a pyrometer mounted in front of a sight glass typically only allows observation of a small section of the rod. Additionally, deposits on the sight glass or dust formation during deposition can affect image quality. Wa12439P / Be

[0026] 6

[0027] The invention is based on the objective of obtaining an automated method for evaluating rods made of a semiconductor material, which are produced by means of a manufacturing process in a gas phase deposition reactor.

[0028] This task is solved by the features of the independent patent claims. The features of the dependent patent claims define implementation forms.

[0029] The following describes techniques for the automated evaluation of rods made of a semiconductor material using one or more machine-learned models. Such a machine-learned model receives input data that is based, for example, at least partially on images of a batch of rods in the reaction chamber of the vapor deposition reactor. Alternatively or additionally, other measurement modalities can also be used. The one or more machine-learned models can then provide output data comprising one or more output elements related to this evaluation of the rods.

[0030] A computer-implemented method for evaluating rods made of a semiconductor material is disclosed. The rods are manufactured using a fabrication process in a vapor deposition reactor. The method includes controlling an imaging device. The imaging device comprises at least one camera. The imaging device is controlled to acquire one or more images of a batch of rods. The batch of rods is located in a reaction chamber of the vapor deposition reactor. The images are acquired after completion of the Wa12439P / Be

[0031] 7

[0032] The manufacturing process for the batch of rods is recorded. Furthermore, the process also includes processing input data in at least one machine-learned model. Output data is thus obtained. The input data is based on one or more images. The output data comprises one or more output elements for evaluating the rods.

[0033] Generally, after the manufacturing process is complete, the rods are connected to a base plate of the reactor via electrodes.

[0034] The use of one or more machine-learned models enables robust evaluation of measurement data, such as images of a batch of rods, particularly under varying environmental conditions and / or different reactors. For example, images acquired at different times may exhibit different brightness levels, contrast ratios, and other characteristics. Furthermore, the perspective of the captured images may vary. The same applies to images acquired using different imaging devices for different reactors. When using non-machine-learned models to evaluate the images, the results may be less robust in such situations.For example, variations in the contrast ratio or brightness, or even in the perspective on the bars, can lead to an incorrect evaluation by conventional, non-machine-learned models.

[0035] In contrast, machine-learned models can be trained on the basis of training data that exhibit a corresponding variance, such as the Wa12439P / Be used during inference.

[0036] 8

[0037] Images. Furthermore, machine-learned models can be used, which, compared to classical models, can generalize better and also handle new, previously unseen situations, or at least identify them as such. Moreover, with machine-learned models, it is possible, for example, to apply them to new reactor types or new types of output elements (that is, to further train output elements of images or other measurement data (fine-tuning)), so that they also deliver good results with new input data or changed scenes. In contrast, with conventional models, it may be necessary to fundamentally reparameterize them or to use an entirely new model.

[0038] In principle, different types of images can be used in the various examples shown here. This means that the images used can have different informational content, depending on the variant.

[0039] In particular, one or more lateral images can be captured, that is, images depicting the reaction space from one or more lateral perspectives. The longitudinal axis of the rods extends into the image plane. Alternatively or additionally, it would also be conceivable to capture one or more images from a zenith perspective, in which the rods are depicted from above.

[0040] Preferably, the one or more images therefore include a second image depicting the reaction space from the zenith perspective Wa12439P / Be

[0041] 9

[0042] An image taken from a zenith perspective contains complementary information to one or more lateral images. In particular, in such an image, which shows the rods from above, few or no details of the rods' lateral surfaces are visible. This is different in the lateral images. On the other hand, such an image taken from a zenith perspective may reveal the arrangement of the rods in the corresponding batch.

[0043] The rods can be arranged along two or more concentric circles (thus, from a position along the circumference of the reaction space, the rods appear arranged in a grid pattern). The rods can be arranged offset from one another along the concentric circles. To also depict rods located closer to the center of the reaction space in a side view, the perspective can be appropriately chosen. In particular, the azimuth angle of the perspective can be selected so that the degree of occlusion of rods located closer to the center of the reaction space by rods located further away from the center is reduced or even minimized. In this way, with a suitable choice of perspective, rods located inside the reaction space can also be visible from the outside.

[0044] In general, multiple images can be captured. For example, multiple lateral images can be captured. Such lateral images can then show several different lateral perspectives of the batch of rods. For instance, these lateral perspectives can be offset from each other along a longitudinal axis of the reaction chamber (the longitudinal axis of the rods is aligned with the longitudinal axis of the reaction chamber). Alternatively or additionally, the multiple lateral Wa12439P / Be

[0045] 10

[0046] Perspectives depict the reaction space from different azimuth angles.

[0047] Preferably, the one or more images can comprise several first images, wherein the several first images depict the reaction space from several lateral perspectives that are offset from each other at least along a longitudinal axis of the reaction space.

[0048] It may also be provided that the multiple first images depict the reaction space from the multiple perspectives that depict the reaction space from different azimuth angles.

[0049] Preferably, the azimuth angle of the multiple perspectives is selected to reduce the degree of obscuration of rods in the images.

[0050] By using multiple lateral perspectives, the batch of rods can be comprehensively measured. In particular, the different perspectives on the batch of rods allow for the detection of certain defects or misalignments of the rods that cannot be clearly identified using a single perspective.

[0051] It is conceivable that the various lateral images have an overlapping area. This means that certain areas of the batch of rods or the reaction chamber are depicted from different perspectives in different images. In this way, it is possible to identify certain properties of the rods—such as specific defects or cracks—particularly well, because the corresponding areas are shown multiple times and from different perspectives. Wa12439P / Be

[0052] Figure 11 is shown. In such a case, several images exhibiting a corresponding overlap area can be processed together as input data in the machine-learned algorithm. This allows the machine-learned algorithm to draw further conclusions from the redundant information present in the two images, or from variations in the appearance of the respective area between the images, within the framework of the evaluation of the bars.

[0053] As previously described, multiple images depicting the batch of rods from different perspectives were processed as input data in at least one machine-learned model. Furthermore, it is conceivable that multiple images depicting the batch of rods from different perspectives could be consolidated before processing in the at least one machine-learned model: If multiple images are captured (e.g., multiple side views), these can also be captured as tiled images, which are then assembled into a mosaic image. Thus, a panorama stitching of the multiple images can be performed to obtain a composite side view.

[0054] Preferably, the method according to the invention comprises performing a panorama stitching of the several first images to obtain a composite first image, wherein the input data is based on the composite first image.

[0055] For example, a 360° image of the batch of bars can be obtained. This can then serve as the basis for processing in at least one machine-learned model. The input data for the at least one machine-Wa12439P / Be

[0056] Twelve learned models can therefore be based on the composite side view.

[0057] Such panoramic stitching can be facilitated by overlapping the images. This enables image registration. Image registration involves a transformation between the image coordinate systems of both images. This is achieved by comparing the image information in the overlap area. Image registration can, in particular, be multimodal. Such multimodal image registration can register images that depict specific objects (in this case, the semiconductor rods) from different perspectives.

[0058] Preferably, the method includes performing an image registration of the first several images, wherein the panorama stitching is based on the image registration.

[0059] Furthermore, brightness and contrast ratios can vary significantly between images if they show different perspectives of the batch of rods. Corresponding flexibility in the image registration algorithm is then helpful. One example of an image registration algorithm is the "Mutual Information" algorithm. This algorithm can handle not only different perspectives but also varying brightness and contrast ratios.

[0060] Basically, there are different possibilities for a hardware implementation of a corresponding imaging device with at least one camera. For example, at least one of the cameras could be located in the area of ​​the lower edge of a reactor hood of the gas phase deposition device Wa12439P / Be

[0061] Thirteen reactors are arranged in the structure. Typically, the reactor is bell-shaped and comprises a base plate with a removable reactor hood. After completion of the manufacturing process, this reactor hood is lifted from the base plate and moved upwards (along the longitudinal axis of the rods) by a lifting device, usually motorized (crane). The reactor hood may also be rotatable around a longitudinal axis. If the camera is mounted at the bottom of the reactor hood, the camera moves relative to the rods as the reactor hood is moved.

[0062] When the reactor hood is raised, lateral perspectives are enabled for the images, offset from each other along the longitudinal axis of the reaction chamber. If the reactor hood is rotated around an axis parallel to the longitudinal axis of the reaction chamber, multiple perspectives are enabled, depicting the reaction chamber at different azimuth angles (which can, in particular, reduce the degree of obscuration of the internal rods, as discussed above). The camera can then be controlled in sync with the movement or raising of the reactor hood to capture the images. In this way, multiple perspectives are made possible by the simultaneous movement of the camera and the reactor hood.

[0063] Such an implementation makes it possible to capture a large number of images from different perspectives (both from different azimuth angles and offset vertically along the longitudinal axis of the rods) with just one camera. Furthermore, it may be possible to obtain pose information for the camera based on the control data of the lifting device. This pose information for the camera can then be combined with the various captured images. Wa12439P / Be

[0064] 14 can be associated, so that a corresponding registration can be made more accurate.

[0065] Preferably, at least one of the at least one camera used to capture the multiple first images is arranged in the region of a lower edge of a removable and optionally rotatable reactor hood of the gas phase deposition reactor, wherein the imaging device is controlled to capture the one or more images synchronously with the lifting of the reactor hood by a motorized lifting device, so that the multiple perspectives are formed by moving the at least one camera together with the reactor hood.

[0066] Multiple cameras can also be arranged, for example, on opposite sides along the circumference of the reactor hood. One or more cameras could also be positioned separately from the reactor hood, for example, on a separate tripod, such as on a rail system.

[0067] In principle, the camera could be mounted in a way that allows it to swivel. This would allow for different perspectives to be achieved by swiveling the camera, without having to move the camera itself.

[0068] Several examples have been described above in which images are captured from one or more perspectives, such as a zenith perspective and / or side perspectives. Alternatively or in addition to such a variation of perspective, other Wa12439P / Be

[0069] 15

[0070] Imaging properties can be varied. For example, the lighting conditions can be changed.

[0071] Preferably, the one or more images include a second image depicting the reaction space from the zenith perspective.

[0072] The images can be captured using ambient lighting. In other words, no separate lighting device is required to capture the images. For example, backlighting from a factory floor can be used.

[0073] Preferably, the imaging device comprises an illumination device that provides at least one planar illumination configuration.

[0074] Such an area lighting configuration can, for example, serve to achieve uniform illumination of the rods in a reaction chamber, enabling better visibility of details on the rods by the at least one machine-learned model. An area lighting configuration can also help to minimize shadowing on the surface of the rods. This configuration can be achieved by an arrangement of LEDs or other light sources on the ceiling or walls of the reactor hood. Alternatively or additionally, luminaires or reflective surfaces can be used to improve the illumination in the reaction chamber.

[0075] Structured lighting can also be used in general. With structured lighting, a Wa12439P / Be can be used.

[0076] 16. A specific light pattern, for example a grid pattern, is projected onto the rods so that certain deformations of the surface of the rods can be revealed by a corresponding deformation of the grid pattern.

[0077] Preferably, several images are captured, with different images being captured with different lighting configurations.

[0078] Preferably, the imaging device comprises a switchable lighting device that provides multiple lighting configurations, wherein the one or more images comprise one or more pairs of images that depict the arrangement of the rods in the reaction space from the same perspective and with different lighting configurations.

[0079] By varying the lighting configuration, it may be possible to highlight certain surface properties of the rods. For example, in a pair of images, the first image could depict the rods with diffuse illumination, while the second image of the same pair could show the rods from the same perspective but with diffuse illumination switched off (i.e., without any targeted lighting, using only ambient light). The second image could also depict the rods illuminated by a point light source. It would also be conceivable to vary the wavelength of the light within the different lighting configurations.

[0080] By differentiating or, more generally, comparing the two images of an image pair associated with different lighting configurations, certain Wa12439P / Be can be determined.

[0081] 17 properties of the surface of the rods that depend on the lighting configuration are made visible.

[0082] Besides varying the perspective and / or lighting configurations, other imaging properties can also be modified. For example, a specific digital or hardware filter could be used during imaging. The corresponding filter parameters could be varied. This means that a captured image can be digitally post-processed before being processed by at least one machine-learned model.

[0083] For example, difference images or mosaic images obtained through stitching can be considered. Another technique would be image preprocessing, for instance, to extract a three-dimensional geometry of the rods. Structured lighting could be used for this purpose. As a further example, a so-called "shape from shading" (SFS) technique could be employed. SFS is based on the assumption that the brightness of a point on an object depends on the object's orientation relative to the light source and the viewer. The SFS technique can be used to extract the three-dimensional geometry of the rods. This typically involves capturing multiple images of the rods from different perspectives. Each image is then analyzed to determine the brightness of each point on the object.By comparing the brightness levels in the different images, the three-dimensional geometry of the object can then be reconstructed. Wa12439P / Be.

[0084] 18

[0085] In another variant of image preprocessing, the images can be segmented to obtain segmentation information for the rods.

[0086] Preferably, the method according to the invention further comprises performing a segmentation of the rods in the one or more image data to obtain segmentation information, wherein the input data is further based on the segmentation information.

[0087] This segmentation information can then be passed to the at least one machine-learned model (for example, along with one or more images). An example of such segmentation information would be an instance segmentation mask that masks individual bars separately.

[0088] However, it is not necessarily required to define a segmentation mask that separates different rods from each other. It would also be conceivable for the segmentation mask to simply separate the foreground (i.e., the rods) from the background (i.e., the environment or reactor back wall, etc.). In particular, such a two-stage approach with upstream segmentation can enable a particularly robust determination of rod properties.

[0089] It is also conceivable that the segmentation model is also machine-trained, but not, or at least less, domain-specifically trained than the downstream machine-trained model for determining the output data, which provides one or more output elements for evaluating the bars. Wa12439P / Be

[0090] 19

[0091] Preferably, the segmentation is performed using a machine-learned segmentation model that is domain-unspecific.

[0092] This is based on the understanding that, initially, more robust rod detection is enabled as a comparatively less complex task using a broadly trained machine-learned model. This also improves the complexity of the downstream processing of the image data within corresponding mask areas by the at least one machine-learned model that provides the rod evaluation. The complexity of the downstream machine-learned model for rod evaluation can thus be reduced. For example, the downstream machine-learned model no longer needs to differentiate between foreground and background. For instance, the model used for segmentation might not be specifically trained for segmenting rods made of semiconductor material. It could, for example, be a so-called foundation model for segmentation; see, e.g., Kirillov, Alexander, et al."Segment anything." Proceedings of the IEEE / CVF International Conference on Computer Vision. 2023.

[0093] Another option would be for the segmentation model to be specifically trained on segmenting semiconductor rods in a reactor, but also trained for multiple reactor types, arrangements of rod batches, etc. If the downstream machine-learned model, which provides the one or more output elements for rod evaluation, receives the segmentation information that already identifies individual rods, then it may be unnecessary for this Wa12439P / Be

[0094] At least one downstream machine-learned model was trained for different arrangements of rod batches, different reactor types, etc. This is based on the understanding that the local properties of the rods are typically similar for different arrangements of rod batches and different reactor types.

[0095] Furthermore, the downstream machine-learned model can operate with a smaller receptive field because it is already known a priori whether a particular image area represents a rod or the background. Typically, this differentiation between foreground and background requires a relatively large receptive field, while the detection of local cracks or defects, or surface roughness, works well even with a correspondingly smaller receptive field. Reducing the size of the receptive field reduces the complexity of the downstream machine-learned network.

[0096] Various scenarios have been described above in which the input data for the at least one machine-learned model is image-based. Alternatively or additionally to such input data for the machine-learned model, which is based on one or more images of the rods, other forms of measurement modalities or prior knowledge for the input data are also conceivable.

[0097] For example, the input data can still be based on geometry reference data specific to a reference arrangement of rods defined by the respective vapor deposition reactor. Wa12439P / Be

[0098] 21

[0099] In other words, such a reference arrangement can, for example, specify the arrangement of the filaments and bridges between individual filaments. Such geometric reference data can, for example, have a 100% contrast, i.e., comprise a mask image. Preferably, the geometric reference data comprises a mask image that depicts the reference arrangement.

[0100] For example, the mask image can have the same resolution as an image captured by a camera, with a corresponding mask area marked at the various locations where a filament is situated. Using such geometric reference data, it is possible to detect deviations in the arrangement of the rods from a reactor-specific SO11 arrangement. For example, a tilting of individual rods or a torsion of certain rods can be detected. It can be detected, for instance, if two adjacent rods are tilted in such a way that they are touching each other.

[0101] According to another version, the input data is still based on measurement data with depth information that indicates a topography of an arrangement of the rods in the reaction space.

[0102] Depth information can be obtained, for example, via an ultrasonic sensor or a stereoscopic camera. Depth information can also be obtained using a light-based time-of-flight method. Using this depth information, it is particularly possible to distinguish between rods that are located closer to the camera (typically further out in the reactor) and those rods that are located further away.

[0103] 22 are located further away from the camera (then typically further inside the reactor).

[0104] It is possible that the rods may exhibit different properties depending on their position within the reactor. This can be the case, in particular, if the process parameter values ​​of the manufacturing process are not homogeneous within the reaction chamber. By considering the positioning of the rods relative to the camera (and thus implicitly also within the reaction chamber), the detection of position-specific properties of the rods by at least one machine-learned model can be enabled.

[0105] The input data can also be based on perspective information for perspectives of one or more images.

[0106] For example, the orientation and / or arrangement of a camera in a reference coordinate system can be specified. This can be done for each image. In this way, based on an analysis of the images, the positioning of the one or more rods depicted in the respective image can be carried out in the reference coordinate system. A camera model can be used for this purpose. If the arrangement of the reaction space in the reference coordinate system is known, this in turn allows conclusions to be drawn about the arrangement of the respective one or more rods depicted in the image within the reaction space.

[0107] Such a positioning of the rods in the reaction chamber is particularly helpful when a localization of certain characteristic areas on the surface of the Wa12439P / Be is required.

[0108] 23

[0109] The bars, which are identified based on at least one model, are desired in a reference coordinate system. Such localization can be helpful if such characteristic areas are to be drawn in a (for example, photogrammetrically generated) 3D model of the batch of bars.

[0110] The input data can still be based on one or more nominal or measured process parameter values ​​of the manufacturing process.

[0111] Nominal process parameter values ​​can be those values ​​specified by a control system. Measured process parameter values ​​can be those values ​​measured at a specific point, for example, within the reactor during the manufacturing process. Typical examples include reactor temperature and reactor pressure. The gas flow rate can also be specified. Furthermore, the gas composition can vary.

[0112] Different reaction gases, such as silane, hydrogen, or dopant gases like phosphine or diborane, typically influence the electrical and structural properties of the semiconductor material. A corresponding indicator for the process parameter values ​​of these or other process parameters can be passed to the at least one machine-learned model in the form of corresponding values. In particular, it is also possible to pass corresponding time series of process parameter values ​​during the manufacturing process to the at least one machine-learned model. For example, the temperature can be specified as a function of time. Wa12439P / Be

[0113] 24

[0114] Since the properties of the rods typically vary depending on the corresponding process parameter values, it may be possible to identify deviations from the process-specific norm by passing the corresponding process parameter values ​​to the at least one machine-learned model.

[0115] More generally, this means that in addition to input data relating to the result of the manufacturing process (for example, images of the batch of rods captured after completion of the manufacturing process), input data relating to the execution of the manufacturing process can also be used. This allows for the reliable identification of certain rod properties relevant to the evaluation of the rods, which correlate with the presence of specific process parameter values ​​or anomalies during the manufacturing process – even if such rod properties have only a weak signature in the image data, i.e., are poorly visible.

[0116] It was explained above that in various examples, in addition to image-based input data, non-image-based input data can also be considered in the at least one machine-learned model. In the various examples revealed here, it is also conceivable that the input data includes both image data and one or more feature vectors. The one or more feature vectors then relate to measurement modalities that do not involve image data. Wa12439P / Be

[0117] 25

[0118] The at least one machine-learned model can then encode the one or more feature vectors on the one hand and the image-based input data (image data) on the other in different coding branches. This means that the image data and the one or more feature vectors are mapped into different latent feature spaces. The corresponding encoded representations can be combined at an output of the coding branches. This means that a fusion of different measurement modalities takes place in a latent, machine-learned feature space. Different coding branches can be trained in separate training processes. This promotes modularity, for example, if only certain types of input data are available for specific reactors. However, a combined training process would also be possible.By using different coding branches, the number of parameters can be adapted to the specific coding task. Typically, image data is relatively high-dimensional, so the corresponding coding branch has many machine-learned weights across a large number of layers. On the other hand, time-series data for a process parameter value, for example, can be relatively low-dimensional.

[0119] According to another implementation, the input data includes image data, wherein the input data further includes one or more feature vectors, and wherein the image data and the one or more feature vectors are encoded in different coding branches of a deep neural network.

[0120] In principle, different architectures for the one or more machine-learned models or the coding branches Wa12439P / Be can be used in the various techniques described here.

[0121] 26 can be used. The core structure of such a system can be based, for example, on convolutional neural networks (CNNs), recurrent neural networks (RNNs), transformer models, or a hybrid architecture, depending on the requirements for accuracy and latency.

[0122] CNNs are particularly suitable for pattern recognition within visual data (here, image data after completion of the manufacturing process) and implement multiple convolution layers that extract features at different levels of abstraction.

[0123] RNNs, on the other hand, which are characterized by their ability to process data with sequential dependencies, are particularly advantageous in the processing of time series data (here, for example, feature vectors that specify a time sequence of measurement data for process parameter values ​​during the manufacturing process).

[0124] Transformer models use mechanisms of attention control to capture relationships between data points in large sequences, making them suitable for tasks where context sensitivity is critical.

[0125] Hybrid architectures combine elements of these approaches to leverage the advantages of different architectures while minimizing their disadvantages, resulting in improved generalizability. Such models are configured through layering and the integration of specific functional units, which increase coding efficiency and improve adaptability to changing data characteristics. Wa12439P / Be

[0126] 27

[0127] According to another embodiment, the at least one machine-learned model comprises a domain-specific trained machine-learned model, wherein the domain-specific trained machine-learned model is trained based on training data for several different arrangements of the rods in gas phase deposition reactors of different types.

[0128] The one or more machine-learned models can, in principle, be trained for multiple arrangements of rods in reactors of different types. This means that the training dataset used to train the machine-learned model includes images showing batches of rods in different arrangements. In particular, the training dataset can also include images showing the batch of rods in the same arrangement as the images acquired during inference.

[0129] It is also conceivable that the training dataset does not contain images showing the same arrangement of rods observed in the reactor during inference. In such a case, it can be particularly helpful to perform the task of locating the rods in the images and the detection of defects using separate models.

[0130] In the various examples mentioned, only a single machine-learned model can be used. This model includes weights trained together. However, it would also be conceivable to use multiple machine-learned models that include specific weights in separate training sessions. For example, as described above, different coding branches could be used for sub- Wa12439P / Be

[0131] 28 different types of input data can be trained separately. It would also be conceivable to use several machine-learned models sequentially to solve different subtasks. For example, a first machine-learned model could perform preprocessing of images; and a second machine-learned model could then process specific data based on this preprocessing.

[0132] The preceding sections described aspects related to the input data for the machine-learned network, as well as aspects related to the processing of this input data within the machine-learned network. The following sections focus more on aspects related to the output data provided by the machine-learned network.

[0133] In the various examples described, different types of output elements can be determined in the output data.

[0134] Preferably, the one or more output elements include a quality measure determined by regression.

[0135] It may still be preferable for at least one of the output elements to provide a classification, for example, with regard to one or more predefined quality criteria. For example, it could be checked whether the surface meets or fails to meet certain (minimum) requirements for roughness, defect density, or number of cracks, etc. Wa12439P / Be

[0136] 29

[0137] According to a preferred embodiment, at least one of the one or more output elements of the output data is indicative of a batch-global property of the rods.

[0138] This means that a general evaluation of the bars is carried out at the batch level, without differentiating between individual bars or even individual positions along the surface of the bars. For example, it can be stated whether the batch of bars obtained through the manufacturing process as a whole meets a certain requirement, such as that no adjacent bars are touching or that none of the bars has a crack longer than X centimeters.

[0139] Batch-global properties can be expressed not only in the form of a classification, as described above, but also as a regression. For example, a quality measure could be determined through a regression.

[0140] The batch-global property of the rods is therefore not characteristic of a specific rod of the batch, but is valid batch-wide.

[0141] Preferably, at least one batch-global property of the rods is selected from the following group:

[0142] - Defect density of defects in the rods ,

[0143] - Measure of the surface quality of the rods ,

[0144] - Measure of surface roughness of the bars ,

[0145] - Deviation of an arrangement of the rods in the reaction chamber from a reference arrangement of the rods ,

[0146] - Production success for the batch of rods. Wa12439P / Be

[0147] 30

[0148] Typically, the annotation effort required to generate training data for a batch-global property of the bars is comparatively low. This is because no localization information needs to be provided. Thus, it is relatively easy and quick to generate the corresponding baseline for the batch-global output element. Accordingly, the reliability can be...

[0149] Robustness in determining such a batch-global output element should be comparatively high.

[0150] Alternatively or in addition to such a batch-global output element, one or more output elements relating to individual bars can also be obtained. For example, one or more bar-specific output elements can be determined. For instance, it can be determined for each of several bars in a corresponding batch whether that bar exhibits properties that meet one or more specifications. Further localization beyond the individual bar is not necessarily required. Such a technique would be desirable if certain bars are to be removed as scrap.

[0151] If a specific rod were identified that exceeds a certain critical defect density, this rod could be removed from the batch as rejects. In such a scenario, further differentiation regarding the local origin of such a rod-specific characteristic is unnecessary. Alternatively, a rod indicator could be issued that uniquely identifies the respective rod within the corresponding batch (for example, according to a predefined counting scheme).

[0152] It may be desirable to achieve a finer differentiation beyond the identification of individual rods Wa12439P / Be

[0153] 31 to determine the location for certain properties of the bars that are relevant for the assessment.

[0154] Thus, at least one of the one or more output elements of the output data can be indicative of a local property of the rods and include associated localization information.

[0155] It is therefore possible to provide localization information that marks an area associated with a specific local property of a rod in one or more images.

[0156] Preferably, at least one of the one or more output elements, which is indicative of the local property of the bars, includes associated localization information that marks an associated area in the one or more images.

[0157] Such localization information can be provided, for example, in the form of a bonding box, a center point localization, or position coordinates. This allows for the localization of specific defects or characteristic points of interest on the rod surface.

[0158] Preferably, the local property of the bars is selected from the following group:

[0159] - Cracking on a rod surface,

[0160] - Contact between two adjacent rods,

[0161] - Rod diameter,

[0162] - Inclination angle of the respective rod ,

[0163] - Surface properties of a local surface of the respective rod. Wa12439P / Be

[0164] 32

[0165] If such localization information is available, the rod could then be broken into several individual pieces, some of which could be reused, and others removed from the batch as scrap. This can be done by taking the localization information into account, ensuring that those areas or parts of the rods that exhibit high quality are fed into further processing. Other parts that do not exhibit high quality can be removed as scrap.

[0166] If such rod-specific properties are determined, or if further localization information is available for certain properties of a rod, then a corresponding area can be marked in a model of the rods. Such a model can be a so-called digital twin, which is generated after completion of the manufacturing process. For example, such a digital twin could be generated photogrammetrically.

[0167] Preferably, the areas indicated by the localization information are marked in a model of the bars, which is optionally generated photogrammetrically.

[0168] The model can therefore specify a 3-D arrangement and shape of the rods.

[0169] Techniques have been described above in which bars are evaluated according to pre-known criteria. With appropriate training, a user can define bar-specific or batch-global labels and use them to train at least one machine-learned model. In this way, the machine-learned model learns specific Wa12439P / Be labels present in the training data.

[0170] 33

[0171] To recognize features in the data present during inference.

[0172] Another strength of machine-learned models is the detection of anomalies, that is, features that are present in the data available for inference but are not represented in the training data during training.

[0173] Preferably, the machine-learned model includes anomaly detection. Accordingly, the output data can include one or more output elements that are indicative of an anomaly.

[0174] Anomalies can be detected in the appearance of the rod surfaces. For example, certain surface topographies or textures of the rods that deviate from the norm seen in the training data can be identified.

[0175] Several examples have been described above of how specific properties of the rods can be identified using image data. For instance, certain irregularities on the rod surface or even cracks can be detected. Areas of the rods that are particularly rough can be identified. Rods that are tilted or exhibit tension or torsion could be detected. Such characteristics, which describe a property of the rods, make it possible to check whether a particular rod or part of a rod is suitable for further processing or must be rejected.

[0176] However, in addition to such symptomatic recognition of the properties of the rods, it can also be helpful to consider Wa12439P / Be

[0177] 34 to identify the underlying cause in the manufacturing process (English: “root cause”).

[0178] According to a preferred variant, at least one of the one or more output elements of the output data, which is indicative of a process anomaly of the manufacturing process, is assumed to be the root cause of a property of the rods.

[0179] This means that such a process anomaly is a hidden observable, which can be inferred from image data depicting the rods as a result of the manufacturing process. For example, it is possible to detect certain structural defects in the reactor, such as thermal bridges, leaks, etc. In this way, predictive maintenance can be enabled. Based on such output data, it is conceivable that a corresponding process parameter value from the manufacturing process could be adjusted for a subsequent batch of rods.

[0180] Preferably, therefore, based on the output data, an adjustment of a process parameter value of the manufacturing process is made for a subsequent batch of the rods.

[0181] This allows for better process results in subsequent manufacturing processes. The quality of the rods can be increased.

[0182] Another application of the output data involves controlling further processing of the rods. For example, the shredding of the rods could be set. Specific fragments or parts of rods can be marked to be removed as scrap. Wa12439P / Be

[0183] 35

[0184] In addition to using the output data from the at least one machine-learned model for process control or final inspection of the bars, the output data can alternatively or additionally be used for continuous learning of the at least one machine-learned model. For example, a user could acknowledge the correctness of individual or all output elements of the data, so that a corresponding label is received as a baseline for a subsequent training step of the at least one machine-learned model. Besides such labels based on manual annotations, pseudo-labels could also be used, which are associated with a certain degree of uncertainty but do not require manual annotations.To reduce uncertainty in determining appropriate labels, validation could be performed based on an alternative evaluation of the images or other input data. This process could, for example, include validating the output data by comparing it to one or more parameters of an alternatively created model of the bars. Depending on the outcome of such validation, the output data could then optionally be stored in a training dataset for retraining the at least one machine-learned model.

[0185] Preferably, the method according to the invention therefore comprises the following steps:

[0186] - Validating the output data based on a comparison with one or more parameters of an alternatively created model of the bars and optionally based on user input, and Wa12439P / Be

[0187] 36

[0188] - depending on the validation of the output data, optionally saving the output data in a training dataset for retraining the machine-learned model.

[0189] Another aspect of the invention relates to an electronic data processing device comprising at least one processor and at least one memory, wherein the at least one processor is configured to load and execute program code from the at least one memory, wherein when the at least one processor executes the program code, it performs the method according to the invention as described.

[0190] The electronic data processing device generally performs at least parts of the techniques described above.

[0191] The features outlined above and those described below can be used not only in the explicitly stated combinations, but also in other combinations or in isolation.

[0192] FIG. 1 is a schematic sectional view through a gas phase deposition reactor in which a batch of rods is arranged.

[0193] FIG. 2 schematically illustrates a machine-learned model according to various examples.

[0194] FIG. 3 illustrates a schematic mask image that can be used as input to the machine-learned model. FIG. 4 is a flowchart of an exemplary procedure. FIG. 5 schematically illustrates an electronic data processing device according to various examples. Wa12439P / Be

[0195] 37

[0196] The described properties, features and advantages of the invention, as well as the manner in which these are achieved, become clearer and more easily understood in connection with the following description of the exemplary embodiments, which are explained in more detail in connection with the drawings.

[0197] The present invention is explained in more detail below with reference to preferred embodiments and the drawings. In the figures, identical reference numerals denote identical or similar elements. The figures are schematic representations of various embodiments of the invention. Elements depicted in the figures are not necessarily shown to scale. Rather, the various elements depicted in the figures are represented in such a way that their function and general purpose are understandable to a person skilled in the art. Connections and couplings between functional units and elements shown in the figures can also be implemented as indirect connections or couplings. A connection or coupling can be implemented as a wired or wireless connection. Functional units can be implemented as hardware, software, or a combination of hardware and software.

[0198] The following describes techniques that enable the evaluation of rods made of semiconductor material, in particular polysilicon or poly-SiC. The rods are obtained during a manufacturing process in a vapor deposition reactor. The evaluation metric can also be a morphology metric, as described in US 2023 / 0011307 Al. Depending on the deposition parameters, semiconductor material (especially polysilicon) with different surface properties (morphology) can form, even within the same semiconductor rod (Wa12439P / Be).

[0199] 38

[0200] Surfaces with varying morphologies can occur. Here, surface texture refers specifically to the fragmentation of the semiconductor material, resulting from the frequency and arrangement of holes, pores, grooves, and cracks. Surface texture can also be understood as the porosity of the semiconductor material. Furthermore, dust deposits and dendrite growth are also subsumed under the term surface texture.

[0201] Early detection of cracks in semiconductor rods is crucial for safety, as these cracks can lead to rod breakage during removal. Therefore, a preferred design includes at least one metric as an evaluation parameter, allowing conclusions to be drawn about the number and severity of cracks (crack formation) in the rods. In particular, the length, width, and, if applicable, depth of a crack can be taken into account.

[0202] For example, the rods can be evaluated with regard to one or more of the following criteria: surface finish, crack formation, rod diameter, angle of inclination, rod position, color, and combinations thereof. Particularly preferably, at least the surface finish is recorded as an evaluation criterion. Varying rod diameters (within the same rod or between multiple rods) can indicate an inhomogeneous deposition process. Furthermore, different comminution or crushing tools may be required depending on the rod diameter.

[0203] To evaluate the bars, one or more machine-learned models are used, the input data being Wa12439P / Be

[0204] 39 characteristic features of the rods are mapped and processed. In particular, image-based input data is discussed, which is based on one or more images depicting the batch of rods when arranged in the reaction chamber of the reactor.

[0205] The evaluation can be performed as a classification task and / or a regression task. Batch-global characteristics can be determined for the evaluation. Alternatively or additionally, rod-specific characteristics can be determined for the evaluation, possibly together with localization information that places the corresponding characteristics on a surface of the rods.

[0206] FIG. 1 is a sectional view perpendicular to a longitudinal axis (in the z-direction; the z-direction is oriented perpendicular to the plane of the drawing) of a gas-phase deposition reactor 300. Shown are rods 309 of a corresponding batch, which are arranged in a reaction chamber 305 of the gas-phase deposition reactor 300. The rods 309 extend perpendicular to the plane of the drawing of FIG. 1, but may optionally have a tilt angle to the z-axis.

[0207] In the present case, rods 309 are arranged along two imaginary concentric circles, each having its center on a central axis 306 of the reaction space 305. The central axis 306 extends in the z-direction. These concentric circles extend in the plane of FIG. 1.

[0208] Reaction chamber 305 is bounded by a movable reactor hood to which a camera is attached. Typically, the camera is mounted on the lower edge of the reactor hood (Wa12439P / Be).

[0209] 40. It is stated that the camera is not located in the reaction chamber when the reaction chamber 205 is closed (i.e., when the reactor hood is lowered and touching a base plate of the reaction chamber 305). The camera can, for example, be attached to a flange or projection of the reactor hood.

[0210] Figure 1 shows perspectives 311, 312, and 313, for which images of the reaction chamber 305 and the rods 309, respectively, are captured by the camera. These perspectives 311, 312, and 313 can be generated by rotating the reactor hood about the central axis 306 as it is lifted. Thus, Figure 1 shows several lateral perspectives 311, 312, and 313, offset along the azimuth angle cp. Alternatively, several lateral perspectives could be provided that are offset from each other along the longitudinal axis of the reaction chamber 305 (i.e., along the z-axis). It is evident that the degree of occlusion for the internal rods 309 is comparatively high for perspective 312, particularly compared to perspectives 311 and 313. Therefore, lateral perspectives 311 and 313 may be preferable.

[0211] In FIG. 1, the arrows at the origin of each perspective 311, 312, 313 show that further perspectives can be generated by rotating the camera. Even more perspectives can be generated by raising the reactor hood along its longitudinal axis, so that the camera is positioned at several locations offset along the longitudinal axis when capturing corresponding images (this is not visible in FIG. 1 due to the sectional view). Wa12439P / Be

[0212] 41

[0213] FIG. 1 also shows a lighting device 320 which can provide an area lighting configuration with a corresponding extended light source.

[0214] Figure 1 shows a variant in which a camera is attached to a reactor hood to capture corresponding images of the rods 309. However, other mounting options for the camera are also conceivable. For example, the camera could be mounted on a separate tripod located in the vicinity of the reactor 300.

[0215] Figure 2 shows an exemplary machine-learned model 210, which is used to determine an evaluation of the rods 309. In the example shown in Figure 2, the machine-learned model 210 comprises two separate coding branches 211 and 212. Coding branch 211 receives as input three images 201, 202, and 203. In principle, more or fewer images could also be used as input to the machine-learned model 210. Images input to the machine-learned model 210 can be camera images captured of the rods 309 by a camera moved along the circumference of the reactor 300 (see Figure 1, where corresponding perspectives 311, 312, and 313 are shown). In other words, this means that images 201, 202, 203 show different perspectives of the reaction space 305 (e.g. a bird's-eye view from the zenith above the reaction space and one or more side views, see FIG. 1).

[0216] Alternatively or in addition to such images that have different perspectives, it would also be conceivable to process images from the machine-learned model 210 that have the same perspective, but with different angles. Wa12439P / Be

[0217] 42 different lighting configurations are captured. It would also be conceivable to process the difference images, e.g., the difference between two different lighting configurations (such as area lighting on / off or different lighting directions). Furthermore, it would be conceivable to process images with structured lighting. Mosaic images could also be processed, which are composed of a large number of corresponding tile images using panorama stitching (possibly after image registration between the tile images).

[0218] Another type of image would be a mask image; compare, for example, mask image 204 in FIG. 3. FIG. 3 shows a reference arrangement of the rods 309 (cf. FIG. 1) at 100% contrast (i.e., black and white), and also illustrates which rods are electrically connected via a corresponding short-circuit bridge. Mask image 204 thus provides geometric reference data. Segmentation masks could also be processed, which are associated with the respective camera images (i.e., have the same perspective) and separate the rods from the background. Instance segmentation masks can be used. To achieve a particularly robust separation of foreground and background, it would be conceivable to use a machine-learned segmentation model, which is not domain-specific, to obtain such a segmentation mask.

[0219] Referring again to FIG. 2: it is also shown there that the further coding branch 212 receives a feature vector 205 as input. For example, the feature vector 205 could be time series data, such as a time series Wa12439P / Be

[0220] 43 of nominal or measured process parameter values ​​of a process parameter of the gas phase deposition process, e.g. temperature, pressure, gas flow, etc.

[0221] By using the two coding branches 211, 212, it is possible to ensure that each of the coding branches 211, 212 is adapted to the specific structure of the input data 201, 202, 203 and 205 respectively, for example with regard to the number of layers, the size of the layers, the architecture, etc.

[0222] While the example in FIG. 2 shows two different coding branches 211, 212, it would in principle be possible to use only a single coding branch, for example coding branch 211. However, more coding branches could also be used.

[0223] The coding branches 211 and 212 each provide a latent feature vector 221 and 222, respectively. These two latent feature vectors are combined and then fed to a decoding branch 231. The decoding branch 231 then generates output data 235. In the example shown in FIG. 2, the output data 235 comprises two different output elements 236 and 237. Output element 236 is a mask image that, for example, localizes defects on the surface of the rods or specifies other particular surface properties. Thus, local properties of the rods are specified.

[0224] In addition to such image-based localization information, which marks specific areas in the images, the localization information can also be provided in other forms, for example as coordinates in a reference coordinate system or by means of an indicator that displays the Wa12439P / Be

[0225] 44 respective rods are indexed. The output element 237 comprises a measure that indicates the quality of the rods for the entire batch, i.e., a batch-global output element. Examples of the information content of such a batch-global output element 237 include, for example, a defect density of defects in the rods, a measure for the surface finish or morphology of the rods, a measure for the surface roughness of the rods, a deviation of an arrangement of the rods in the reaction chamber from a reference arrangement of the rods (see FIG. 3), for example, if rods are tilted or even in contact with each other.

[0226] Typically, the machine-learned model 210 can include multiple decoding branches for different types of output elements (not shown in FIG. 2 for simplicity).

[0227] Figure 4 is a flowchart of an exemplary procedure. The procedure in Figure 4 can be executed by one or more electronic data processing devices. For example, at least one processor can load and execute program code from one or more memories and then, based on the program code, execute the various boxes in Figure 4. Optional boxes and dependencies are shown with dashed lines in Figure 4.

[0228] The method of FIG. 4 is used to evaluate rods made of semiconductor material, for example, poly-SiC or polysilicon. These rods are produced using a manufacturing process in a gas-phase deposition reactor. Preferably, the gas-phase deposition reactor is a (Siemens) reactor for the production of polysilicon and / or for the production of poly-SiC. The gas-phase deposition reactor comprises a base plate and Wa12439P / Be

[0229] 45 a reactor shell which defines a reaction chamber, at least a supply line for the reaction gas into the reaction chamber, at least a discharge line for the gas removal from the reaction chamber and electrode holders arranged on the base plate for at least one support body, wherein, for example, at least an imaging device for generating 2D or 3D images is attached to an outer lower edge of the reactor shell.

[0230] The manufacturing process, i.e., the vapor deposition to produce the rods from semiconductor material, takes place in Box 3005. Box 3005 can therefore include, in particular, the control of one or more reactor components. For example, a heater can be controlled to maintain a specific temperature. Process gases can be supplied.

[0231] The semiconductor component contained in the reaction gas preferably comprises at least one silane, in particular a halosilane. The halosilane is preferably selected from the group consisting of the chlorosilanes of the general formulas H n SiC14-n, H m C16-mSi2 , (CHa ) nSiC14-n with n = 1 to 3 and m = 0 to 4. This embodiment is particularly suitable for the production of polysilicon (see US 2012 / 0100302 Al mentioned in the introduction). Preferably, the halosilane is TCS or a mixture of dichlorosilane and TCS. Such a reaction gas composition is particularly suitable for the production of polysilicon according to the Siemens process.

[0232] In another embodiment, the semiconductor component contained in the reaction gas can comprise at least one silane and at least one carbon component. Alternatively or additionally, the semiconductor component can also be an organosilane Wa12439P / Be

[0233] 46. ​​The semiconductor component can generally consist of only one organosilane, since both a carbon and a silicon source are contained in one molecule. This form is particularly suitable for the production of poly-SiC (see US 2023 / 0141427 Al mentioned in the introduction). The organosilane is preferably selected from the group of organochlorosilanes of the general formulas (CHa)nH m SiC14-nm with n = 0 to 3 and m = 0 or 1 and MenSiaCle-n with n = 1-5. The carbon component can be, for example, methane, ethane, propane, butane and combinations thereof.

[0234] A pump can set a specific pressure. Box 3005 can generate the corresponding control data, which is then transmitted to the controllers of such components. Alternatively or additionally, Box 3005 can also receive and store corresponding measurement data streams relating to the time dependency of process parameter values ​​for one or more process parameters.

[0235] In Box 3010, one or more images of the rods are captured. It can be particularly advantageous to capture multiple images from different perspectives. As explained above in relation to FIG. 1, one possibility is to move a camera along with the reactor hood when the hood is lifted from the reactor base plate. In addition to this longitudinal movement for lifting, the reactor hood can also be rotated to allow perspectives from different azimuth angles.

[0236] When capturing images in Box 3010, a corresponding camera of an imaging device can be controlled. The imaging device could also include a lighting Wa12439P / Be

[0237] 47

[0238] The device comprises one that can be controlled synchronously with the image acquisition. Preferably, a corresponding imaging device used in Box 3010 comprises one or more 2D cameras, 3D cameras, laser scanners, LiDAR (Light Detection and Ranging) cameras, infrared cameras, and combinations thereof. The cameras may be equipped with a fisheye and / or wide-angle lens. Images acquired in Box 3010 may, for example, contain depth information. A stereo camera or a time-lapse camera may be used for this purpose.

[0239] Generally, the best time to capture the images is immediately after lifting the reactor vessel and before removing (manipulating) the rods. This allows for an assessment of the rods without interrupting the actual removal process. After all, the reactor vessel must be completely lifted and possibly even removed laterally for rod removal anyway.

[0240] For example, the imaging device can determine the position of the two- or three-dimensional images using a distance sensor. Accordingly, it is preferred if the measuring device includes a distance sensor with which the distance to the base plate of the gas-phase deposition reactor and / or the distance to the semiconductor rods can be determined. Depending on the distance, image acquisition can be triggered and, if necessary, the focus adjusted. For example, a series of images can be triggered every 1, 5, or 10 cm increasing distance from the base plate. The distance sensor can be a laser rangefinder or a gyroscope. The distance can also be measured by triangulation using known points (or landmarks) on the reactor and / or Wa12439P / Be

[0241] 48 in the vicinity. In box 3010, a video recording could also be triggered, so that a sequence of images is captured.

[0242] In Box 3015, registration between different images and / or stitching between different images can optionally be performed. Registration can be image-based, meaning by comparing specific image features visible in different images. In particular, multimodal registration can be helpful if the different images show different perspectives of the batch of rods. Registration can also be based on prior knowledge regarding the perspective of the different images: for example, if the camera is moved along with the reactor hood, corresponding prior knowledge about the camera's positioning may be available. In other words, perspective information for the image perspectives may be available.By stitching the images, a mosaic image can be created which has a larger field of view than the individual images.

[0243] The registration and stitching described above are only exemplary operations that can be applied to the images from Box 3010. Alternatively or additionally, other image processing operations are conceivable, such as differentiating between images captured with different lighting configurations. Another example is the use of structured lighting, in which case a reconstruction algorithm for reconstructing the surface shape of the rods can be applied in Box 3015. The Wa12439P / Be

[0244] 49

[0245] The reconstruction algorithm can use prior knowledge about the structure of the structured lighting to reconstruct the surface shape.

[0246] Optionally, in Box 3020, a segmentation of certain structures in the images can be performed (for example, of the images captured in Box 3010, which are obtained as a result of Box 3015).

[0247] In Box 3025, the input data is then processed in a machine-learned model. An example of such a machine-learned model was discussed previously in connection with FIG. 2. Another machine-learned model can, for example, provide anomaly detection. This can fundamentally detect anomalies in the surface texture or morphology of the rods. Alternatively or additionally, it can also infer a process anomaly in the manufacturing process from Box 3005. Such a process anomaly can be assumed to be the root cause of a rod property, as observed in the images from Box 3010.

[0248] The machine-learned model can process image data, but also other data, such as time series data. The machine-learned model provides output data, which includes one or more output elements. These output elements can be used to evaluate the bars. For example, one or more of the output elements can provide a classification, such as whether certain defects are present or absent, or whether a specific predefined defect density is exceeded. The various output elements can relate not only to batch-global properties of the bars, but also to Wa12439P / Be

[0249] 50. Specific properties of the rods can also be specified. For example, a corresponding classification could be made for each individual rod, indicating whether certain defects are present or not, or whether a certain defect density is exceeded or not.

[0250] In principle, machine-learned models can solve not only classification tasks, but also regression tasks, either alternatively or additionally. For example, they could output a metric that quantifies the defect density or another quality criterion.

[0251] If necessary, the results from Box 3025 may be further processed in Box 3030. For example, it would be conceivable to aggregate certain output elements specific to individual bars, so that a batch-wide result is obtained for evaluating the bars.

[0252] Figure 4 further shows that the images 3010 can be processed not only in branch 3011 with boxes 3015, 3120, 3025, and 3030, but also in another branch 3101. In branch 3101, a 3D model of the batch of bars is generated: In the example of Figure 4, this is done in box 3105 based on the images available from box 3010, using photogrammetric reconstruction. However, other techniques for generating a 3D model are also conceivable. In particular, different measurement modalities than those used for branch 3011 can be employed in box 3105.

[0253] Box 3110 optionally augments the 3D model based on localization information obtained from the machine-learned model in Box 3025. For example, Wa12439P / Be

[0254] 51. Certain defects detected and located by the machine-learned model can be drawn into the 3D model, i.e., corresponding areas are marked in the 3D model. Nominal properties of the bars, e.g., surface roughness, etc., can also be noted in the 3D model.

[0255] Box 3210 can then be used to evaluate the results from Box 3030 and, optionally, the results from Box 3110. For example, Box 3210 could be used to check whether the rods meet certain minimum quality criteria and can be processed further. It could also be used to check whether a process anomaly occurred in the manufacturing process in Box 3005 that led to deviations of the rods from the standard. If necessary, corresponding process parameter values ​​could be adjusted for a subsequent manufacturing process. Reactor maintenance could also be requested.

[0256] Based on a result from Box 3210, the comminution of the rods can be controlled. For example, a fraction size can be set. Fraction size-dependent comminution means that semiconductor rods, for example, are subjected to different comminution processes depending on their surface properties. These comminution processes may be carried out with different crushing tools. Typically, there are five fraction size classes 0 to 4 (BGO to BG4), which are defined by the particle size of the fragments, where particle size is defined as the longest distance between two points on the surface of a silicon fragment. The fraction size classes group fractions with particle size ranges as follows.

[0257] BGO: 0, 1 to 9 mm Wa12439P / Be

[0258] 52

[0259] BG1: 1 to 18 mm BG2: 5 to 50 mm BG3: 20 to 65 mm BG4: 35 to 150 mm For example, the fracture size class could be determined depending on the evaluation from box 3210.

[0260] In Box 3215, one or more output elements of the machine-learned model from Box 3025 may be deemed sufficiently reliable to be accepted as the baseline truth. This could, for example, be based on corresponding user input received in Box 3220. If this is the case, corresponding output elements, along with their corresponding input elements, can be stored in a training dataset (Box 3225). Alternatively, instead of user input, validation could be based on a comparison between one or more parameters derived from the 3D model in Box 3105 and a corresponding output element obtained from the machine-learned model in Box 3025.If the comparison shows that the corresponding one or more output elements of the machine-learned model from Box 3025 are plausible in light of the corresponding one or more parameters derived from the 3-D model from Box 3105, then these one or more output elements could be stored as pseudo-labels in the training data set in Box 3225.

[0261] By populating the training dataset in Box 3225, continuous refinement of the machine-learned model is enabled by retraining it periodically based on additional training samples acquired in the field. In particular, this allows the machine-learned Wa12439P / Be to be refined.

[0262] 53

[0263] The model can be trained for several different arrangements of the rods in different reactor types if corresponding training datasets are collected for the different reactors and subsequently consolidated, for example, within the framework of so-called federated learning.

[0264] Figure 5 schematically illustrates an electronic data processing device 910 according to various examples. The electronic data processing device 910 comprises a processor 911, a memory 912, and a communication interface 913. The processor 911 can load and execute program code from the memory 912. When the processor 911 executes the program code, this causes the processor 911 to implement techniques as described above.

[0265] For example, the 911 processor

[0266] - control one or more components of a gas phase deposition reactor to carry out a manufacturing process for semiconductor rods,

[0267] - control an imaging device with a camera and / or other measuring components to capture images and / or other measurement data that characterize the semiconductor rods after the manufacturing process,

[0268] - execute one or more machine-learned models to perform an evaluation of the staffs,

[0269] - determine the comminution of the rods based on an output of the machine-learned model,

[0270] - Adjust one or more process parameters of a subsequent manufacturing process based on an output of the machine-learned model, etc. Wa12439P / Be

[0271] 54

[0272] In summary, techniques have been described that generally relate to the evaluation of semiconductor rods after a manufacturing process. These techniques generally involve inputting image data into a machine-learned model. The image data typically shows one or more rods as they appear after the manufacturing process. The machine-learned model provides an evaluation of the rod quality. For example, the corresponding output data may be indicative of a batch-global property of the batch of rods, such as a defect density of defects on the rods or an average surface quality of local surfaces of the rods.

[0273] Alternatively or additionally, one or more output elements can be indicative of a localized property that affects individual bars.

[0274] In particular, localization information can be provided that highlights specific defects in the images, such as cracks or contacts between adjacent bars.

[0275] The input data can optionally be determined based on geometric reference data. Optionally, depth information can be used to differentiate between various bars within the reactor. Optionally, perspective information can be considered for one or more images, so that, for example, the bars depicted in the respective image are positioned within the reference coordinate system.

[0276] Optionally, process parameter values ​​of the manufacturing process can also be taken into account in the evaluation.

[0277] Optionally, the tasks of locating the rods in the images and detecting defects can be performed using separate models. Wa12439P / Be

[0278] 55

[0279] Optionally, several machine-learned models can be used sequentially to solve different subtasks.

[0280] Naturally, the features of the previously described embodiments and aspects of the invention can be combined with one another. In particular, the features can be used not only in the described combinations, but also in other combinations or individually, without leaving the scope of the invention. For example, techniques have been described above in which at least one machine-learned model is used that processes image data. However, instead of image data, other measurement modalities, such as three-dimensional point clouds, can also be processed. The surface temperature of the rods could also be measured.

Claims

Wa12439P / Be 56 Patent claims 1. Computer-implemented method for evaluating rods made of semiconductor material, wherein the rods (309) are produced by a manufacturing process in a gas phase deposition reactor (300), the method comprising: - Controlling an imaging device with at least one camera for capturing one or more images of a batch of rods in a reaction chamber (305) of the gas-phase deposition reactor after completion of the manufacturing process for the batch of rods, and - Processing input data (201, 202, 203, 204, 205) in at least one machine-learned model to obtain output data, wherein the input data is based on one or more images, and wherein the output data includes one or more output elements for evaluating the bars.

2. Computer-implemented method according to claim 1, wherein the one or more images comprise several first images, wherein the several first images depict the reaction space (305) from several lateral perspectives (311, 312, 313) which are offset from each other at least along a longitudinal axis (306, z) of the reaction space (305).

3. Computer-implemented method according to claim 1 or 2, wherein the one or more images comprise several first images, the several first images representing the reaction space (305) from the several perspectives (311, Wa12439P / Be 57 312 , 313 ) image which maps the reaction space ( 305 ) from different azimuth angles ( cp ).

4. Computer-implemented method according to claim 3, wherein an azimuth angle (cp) of the multiple perspectives is selected to reduce the degree of occlusion of rods (309) in the images.

5. Computer-implemented method according to any one of the preceding claims, wherein the method further comprises: Performing a panorama stitching of the multiple first images to obtain a composite first image, where the input data is based on the composite first image.

6. Computer-implemented method according to claim 5, wherein the method further comprises: Performing an image registration of the first several images, with the panorama stitching being based on the image registration.

7. Computer-implemented method according to one of the preceding claims, wherein at least one of the at least one camera used to capture the multiple first images is arranged in the region of a lower edge of a removable and optionally rotatable reactor hood of the gas-phase deposition reactor, wherein the imaging device is controlled to capture the one or more images synchronously with the lifting of the reactor hood by a motorized lifting device, so that the multiple perspectives are captured by Wa12439P / Be 58 Movement of at least one camera together with the reactor hood is formed.

8. Computer-implemented method according to one of the preceding claims, wherein the one or more images comprise a second image depicting the reaction space from a zenith perspective.

9. Computer-implemented method according to one of the preceding claims, wherein the imaging device comprises an illumination device (320) which provides at least one planar illumination configuration.

10. Computer-implemented method according to one of the preceding claims, wherein the imaging device comprises a switchable lighting device which provides multiple lighting configurations, wherein the one or more images comprise one or more pairs of images which depict the arrangement of the rods in the reaction space from the same perspective and with different lighting configurations.

11. Computer-implemented method according to any one of the preceding claims, wherein the method further comprises: Performing a segmentation of the bars in the one or more image data sets to obtain segmentation information, with the input data still being based on the segmentation information. Wa12439P / Be 59 12. Computer-implemented method according to claim 11, wherein the segmentation is performed using a machine-learned segmentation model that is domain-unspecifically trained.

13. Computer-implemented method according to one of the preceding claims, wherein the input data are further based on geometry reference data specific to a reference arrangement of the rods defined by the respective gas phase deposition reactor.

14. Computer-implemented method according to claim 13, wherein the geometry reference data comprises a mask image that depicts the reference arrangement.

15. Computer-implemented method according to one of the preceding claims, wherein the input data are further based on measurement data with depth information indicating a topography of an arrangement of the rods in the reaction space.

16. Computer-implemented method according to one of the preceding claims, wherein the input data is further based on perspective information for perspectives of one or more images.

17. Computer-implemented method according to one of the preceding claims, wherein the input data is further stored on one or more Wal 2439P / Be 60 nominal or measured process parameter values ​​of the manufacturing process are based on .

18. Computer-implemented method according to one of the preceding claims, wherein the input data comprises image data, wherein the input data further comprises one or more feature vectors, wherein the image data and the one or more feature vectors are encoded in different coding branches (211, 212) of a deep neural network (210) and corresponding encoded representations are combined at the output of the coding branches (211, 212).

19. Computer-implemented method according to one of the preceding claims, wherein the at least one machine-learned model comprises a domain-specific trained machine-learned model, wherein the domain-specific trained machine-learned model is trained based on training data for several different arrangements of the rods in gas-phase deposition reactors of different types.

20. Computer-implemented method according to one of the preceding claims, wherein the one or more output elements comprise a quality measure determined by regression.

21. Computer-implemented method according to one of the preceding claims, wherein at least one of the one or more output- Wa12439P / Be 61 elements of the output data are indicative of a batch-global property of the rods.

22. Computer-implemented method according to claim 21, wherein the batch-global property of the rods is selected from the following group: a defect density of defects of the rods, a measure of the surface quality of the rods, a measure of the surface roughness of the rods, a deviation of an arrangement of the rods in the reaction space from a reference arrangement of the rods.

23. Computer-implemented method according to one of the preceding claims, wherein at least one of the one or more output elements of the output data is indicative of a local property of the rods.

24. Computer-implemented method according to at least one of claims 21 to 23, wherein the at least one of the one or more output elements, which is indicative of the local property of the rods, comprises associated localization information that marks an associated area in the one or more images.

25. Computer-implemented method according to at least one of claims 21 to 24, wherein the local property of the rods is selected from the following group: crack formation on a rod surface, contact between two adjacent rods, Wa12439P / Be 62 Rod diameter, inclination angle of the respective rod, surface finish of a local surface of the respective rod.

26. Computer-implemented method according to claim 24, wherein the method further comprises: Marking the areas indexed by the localization information in a model of the bars, which is optionally generated photogrammetrically.

27. Computer-implemented method according to one of the preceding claims, wherein the machine-learned model comprises anomaly detection.

28. Computer-implemented method according to one of the preceding claims, wherein at least one of the one or more output elements of the output data, which is indicative of a process anomaly of the manufacturing process, is assumed to be the root cause of a property of the rods.

29. Computer-implemented method according to any one of the preceding claims, wherein the method further comprises: Adjusting a process parameter value of the manufacturing process for a subsequent batch of rods, based on the output data. Wa12439P / Be 63 30. Computer-implemented method according to one of the preceding claims, wherein the method further comprises: - Validating the output data based on a comparison with one or more parameters of an alternatively created model of the bars and optionally based on user input, and - depending on the validation of the output data, optionally saving the output data in a training dataset for retraining the machine-learned model.

31. Electronic data processing device comprising at least one processor and at least one memory, wherein the at least one processor is configured to load and execute program code from the at least one memory, wherein when executing the program code, the at least one processor performs a method according to any of the preceding claims.

Images