Method for processing an image of a semiconductor structure and method for process characterization and process optimization by means of semantic data compression

Abstract

The present disclosure relates to a method of processing an image generated under an imaging method of a semiconductor structure of a solar cell. The corresponding method includes processing an image of the semiconductor structure by means of a trained neural network to generate a semantic representation of the image, wherein the data size of the semantic representation is smaller than the data size of the image. The method further includes storing the semantic representation corresponding to the semiconductor structure as an information carrier for determining information related to the performance of the solar cell and / or for crystal analysis of the semiconductor structure of the solar cell. Other aspects relate to a method of improving a semiconductor process, a method of evaluating a masked region in a spatially resolved quality image of a semiconductor structure, and a method of predicting at least one performance of a solar cell.

Description

Field of the Invention

[0001] The embodiments relate to methods used in the process of manufacturing solar cells. The examples relate to methods for processing images of semiconductor structures of solar cells generated in the case of using imaging methods, methods for improving semiconductor processes, methods for evaluating spatially resolved quality images of semiconductor structures, and methods for predicting at least one performance of solar cells. Background Art

[0002] In the production of semiconductor wafers and solar cells, it is necessary to monitor the quality of the processes and the materials used and to obtain information about the influence of the processes. This can be used to evaluate various parameters of solar cells or to classify solar cells into a certain quality class. For example, electrical measurements can be performed to determine the current-voltage parameters of solar cells.

[0003] It is also possible to use imaging measurement methods to measure wafers and solar cells. However, the measurement results are sometimes not comparable to each other due to the combination of varying spatial distributions and varying image structures. In addition, due to the correspondingly high image resolution, the corresponding data sets require a large storage capacity and may result in correspondingly long data processing times. In addition, it is difficult to correlate the results of different measurement systems with each other.

[0004] However, due to the high productivity in the mass production of solar cells, it is necessary to provide fast and efficient data processing to monitor the production process. Therefore, manual evaluation of measurement results is particularly likely to be inappropriate. Evaluation and classification with algorithms that quantify image structures according to purely human standards are usually time-consuming, error-prone, and generally unconvincing. The electrical contacting of individual solar cells can also be very time-consuming, so that quality inspection of solar cells cannot be carried out in mass production.

[0005] To classify solar cells, the measurement results of current-voltage characteristics can be recorded in a known manner. However, this method is time-consuming because it always requires the corresponding contacting of individual solar cells. For example, based on purely electrical measurements, the reasons for the low quality of solar cells can only be identified to a limited extent.

[0006] The object of the present invention is therefore to provide a method for the production of solar cells, which allows efficient monitoring of the quality of the semiconductor processes or the materials used and / or optimization of the semiconductor processes used. Summary of the Invention

[0007] Accordingly, a method for processing an image of a semiconductor structure of a solar cell generated by an imaging method is described. The method includes processing the image of the semiconductor structure by means of a trained neural network to generate a semantic representation of the image. The data size of the semantic representation generated according to the method is less than the data size of the image here. In addition, the method also includes storing the semantic representation assigned to the semiconductor structure. The semantic representation provides an information carrier for relevant image information, which can be used for the determination of the performance or parameters of the solar cell and / or for the crystal analysis of the semiconductor structure of the solar cell. Thus, for example, it is possible to use the semantic representation to predict solar cell parameters or to infer the materials used or the semiconductor process performed in the production of the solar cell based on crystal analysis.

[0008] The neural network shown in connection with the present invention can be executed and / or trained, for example, by a processor (such as a general-purpose CPU), an FPGA, an ASIC, or a similar computing device.

[0009] Correspondingly, the semiconductor structure of the solar cell whose image can be processed according to the method is, for example, a semiconductor wafer, a solar cell precursor such as a semiconductor structure that is only partially processed, or the finished solar cell itself. The relevant semiconductor structure can correspondingly include not only regions made of semiconductor materials, but also, for example, solar cell regions containing contact connection structures made of conductive materials. The semiconductor structure can also include, for example, semiconductor blocks (bricks) that can be divided into wafers.

[0010] The image of the semiconductor structure used can be generated by means of an imaging method or generated before the semantic representation is generated. Suitable methods include, for example, using electroluminescence measurement, photoluminescence measurement, or temperature recording. In addition, reflection signal measurement or transmission signal measurement can also be used to generate an image of the semiconductor structure. The image of the semiconductor structure can be, for example, a black-and-white image or a color image.

[0011] Accordingly, the method can include generating a corresponding image using one of the above methods or a suitable alternative method known to those skilled in the art. For example, before generating the semantic representation, the image can be reduced to the same resolution or a standard resolution in terms of image resolution, or the image coordinates can be corrected to uniformly position and scale the samples according to the same coordinates or standard coordinates in the image to improve image comparability and position-specific information of each image. Within the network, the input image can be preprocessed in different subnets and then combined at the same resolution. Data alignment / calibration can be performed prior (before generating the semantic representation) based on the same coordinate system.

[0012] For example, based on the image itself, it is impossible or very difficult to evaluate the quality or parameters of a semiconductor structure. Therefore, the present invention provides for using a trained neural network to generate a semantic representation of the image. The generation of the semantic representation can be referred to as data compression, for example. Here, semantically meaningful parameters can be extracted from the image by the neural network, or in other words, those parameters that are meaningful or persuasive in terms of material quality, solar cell quality, and / or process quality. Different from the specific optical image information of the image, the semantic representation can include abstract features of the image, which enable the inference of the parameters of the semiconductor structure. The neural network can be a convolutional neural network or other neural networks, such as a recurrent neural network.

[0013] The semantic representation can include the weight distribution of the neural network layer weights. Depending on the type of neural network used, different layers of the neural network can represent the semantic representation in terms of its depth. The layers used can represent a trade-off between the information content of the semantic representation and the data size of the semantic representation. For example, the last layer of the neural network can be used for the semantic representation, or a layer that is arranged closer to the output layer of the neural network than to the input layer of the neural network. It is also possible that the semantic representation can be based on the weight distribution of the weights of, for example, more than two adjacent layers of the neural network.

[0014] The network layer used for the semantic representation can be a layer from which parameters (such as one or more target values) can be predicted by means of a linear or non-linear classification method or a regression method. The semantic representation can be the value of a neural network layer before the final linear regression in the network.

[0015] In one example, a layer of a neural network having an encoder-decoder structure can be used, which layer is in the network bottleneck region, that is, for example, a layer having the smallest number of weights and thus the smallest data size. This allows for particularly efficient data compression, where the corresponding semantic representation includes all the image-related parameters in the case of a small data size. For example, the semantic representation has a plurality of vectors, each vector being assigned a solar cell parameter to be predicted.

[0016] Therefore, it is possible to achieve compressed storage of image information by generating the semantic representation. For example, the semantic representation has a data size of at most 10%, at most 1%, or at most 0.1% of the image data size. Thus, semantics (such as the persuasiveness of the semantic representation) are obtained from the image compression, that is, it is persuasive in terms of one or more quality parameters within the photovoltaic power generation value supply chain.

[0017] The semantic representation can be based on at least two images of the semiconductor structure of a solar cell processed by a neural network, where these images are generated by different imaging methods. In particular, three, four or more images can be used to generate the semantic representation. Thereby, for example, the prediction quality of parameter prediction or determination of semiconductor structure parameters based on the semantic representation can be improved. It is possible that two different images each contain relevant information. If, for example, there are different images of the semiconductor structure, the information about the semiconductor structure can be compressed and stored even more efficiently in a single semantic representation by means of the semantic representation, and the storage of multiple individual images can be dispensed with. In the case of relevant information, the method can be carried out more efficiently.

[0018] Furthermore, in the case of a combination of different input images and supplementary information, the semantics of the representation can also be improved. In the case of supplementary information, the semantic representation can become more persuasive and / or more efficient. Therefore, in a representation for efficient storage and analysis of measurement values, it is advantageous to input measurement data combinations into the network. For this, for example, these effects can be distinguished from each other, which, for example, occur differently in at least one of the recordings depending on the type of defect occurring in the semiconductor structure. In a solar cell output test, this can be, for example, a combination of thermographic and luminescence recordings, which is trained to predict current-voltage parameters (IV parameters). The semantic representation enables the distinction of many defects that are not visible only in one of the recordings. Thereby, the manual evaluation of many different images can be dispensed with, and instead, one semantic representation can be evaluated in a time-efficient manner. For example, a short circuit can be caused by poor edge insulation or a crack. The short circuit can be identified by an increased signal in the thermogram. The cause of this defect can be determined by the appearance in the luminescence measurement, where the two defect types represent different image signatures. In the case of material evaluation in an initial test, this can be a combination of photoluminescence recording and reflection or transmission measurement, for example, for distinguishing between recombination-active grain boundaries and non-recombination-active grain boundaries.

[0019] Therefore, the semantic representation provides an understandable or semantically meaningful display of the data that can be investigated during the production of solar cells. In order to obtain information about the process and the material, such understandable data preparation may be required. Therefore, the measurement data can be compressed and stored in a particularly suitable way by this method so that they can be compared with each other and relevant information can be drawn from them. In particular, relevant features (representations) can be extracted from the images or graphics (measurement data) so that, for example, quality changes and process changes of the material used in the production or the processes used can be determined.

[0020] In addition, due to the high production volume or high number of production pieces per unit time in semiconductor production or solar cell production, rapid compression of measurement data is required, which can be ensured by using a neural network. Semantic representations can be used, for example, to efficiently store image data. Semantic representations also allow quantitative data analysis including meaningful semantic features to monitor the crystallization and production processes. Accordingly, the representations can be used to predict parameters of solar cells. Such prediction allows determination of the parameters and / or quality grades of solar cells without contact measurement methods, thereby reducing the time required for characterizing the semiconductor structure of solar cells.

[0021] Advantageously, the method according to the invention can thus be used to efficiently store and / or compress data while making the comprehensible display and information content regarding the semiconductor structure accessible and improving the data values through the resulting comparability. Improvements to the method can allow prediction and / or classification of solar cell quality (such as classifying solar cells including corresponding semiconductor structures into a quality grade), differentiating process defects and material defects of semiconductor structures from each other, and / or analyzing the crystallization of semiconductor structures. Based on this, for example, semiconductor process can be monitored and / or optimized.

[0022] An advantage of the method of the invention can be the use of semantic representations in data analysis, which implicitly contain relevant image information regarding desired quality features. For example, multiple image data within the representation can also be semantically combined. In this case, data can be efficiently saved, efficiently compared with each other, and the process can be optimized. The method can be used to analyze and optimize the crystallization process, analyze and optimize the solar cell process, monitor the solar cell production process, and / or quickly determine the classification of solar cells by means of non-contact imaging methods.

[0023] To implement the method, for example, a neural network can be trained first. For this purpose, parameters of the semiconductor structure can be measured (for example, in each individual electrical measurement allowing determination of individual IV parameters), and the parameters should then be predicted using the neural network (for example, based on a single imaging measurement instead of a large number of individual electrical measurements). The corresponding parameters can be referred to as target parameters, for example. Accordingly, the method can include training a neural network using multiple images of each semiconductor structure or solar cell as input data and using many measurement performances of each solar cell assigned to the corresponding image as output data. The measured parameters can be used in combination with various different images of each semiconductor structure to train the neural network. A large amount of data can allow more detailed determination of, for example, the cause of parameter variations of solar cells.

[0024] As already mentioned, the stored semantic representations can be used to predict or determine characteristics of a semiconductor structure, such as parameters of a solar cell (IV parameters) or crystallization characteristics. For example, a method for analyzing the crystal structure of a semiconductor structure of a solar cell can be provided, and for this purpose, the method can include generating an image of the semiconductor structure of the solar cell by means of the imaging method mentioned. The method can also include generating a semantic representation based on the image and using a trained neural network and using the semantic representation to determine the crystallization characteristics of the semiconductor structure of the solar cell.

[0025] Thus, the semantic representation can be used to characterize a solar cell just as it is used to characterize a mere semiconductor wafer or a solar cell precursor.

[0026] One aspect of the present invention relates to a method for determining at least one performance of a solar cell. The method includes using a semantic representation of an image generated by means of an imaging method of a solar cell or a semiconductor structure of a solar cell, wherein the semantic representation is generated by means of a neural network, and the method further includes determining at least one performance of the solar cell based on the semantic representation.

[0027] In other words, the method can be used to determine parameters or performances of a solar cell based on the use of an image and a semantic representation of a semiconductor structure. The use of the semantic representation can allow for a more efficient determination of the parameters. The parameters can be determined in the case of using a neural network. For example, the neural network used to generate the semantic representation can be used.

[0028] The parameters or performances to be determined may relate to solar cell parameters of a solar cell. The at least one performance can for example include open circuit voltage, short circuit current, short circuit current density, efficiency, fill factor, dark saturation current and / or global solar cell parameters, such as the series resistance of a solar cell.

[0029] The series resistance can for example be determined by means of photoluminescence imaging of a solar cell, wherein the cell is partially shaded. Here, charge carriers generated in the illuminated part flow into the non-illuminated part. Thus, an internal current flows through the series resistance of the solar cell, which is visible in the luminescence image. A second record will be generated, wherein the illuminated area and the non-illuminated area are complementary to the first record. By combining the illuminated areas, the result is aggregated from the two records, wherein in particular calculation steps related to the illumination may be required. By using the semantic representation, the series resistance information can be presented and stored in a single semantic representation instead of two separate records or images.

[0030] At least one property to be determined according to the method relates, for example, to a spatially resolved quality criterion of a solar cell. For example, the dark saturation current can be predicted. The aim of the method can generally be to generate a semantic representation, refine the semantic representation and replace time-consuming spatially resolved measurements by only using them during training. The generation of the semantic representation can be carried out with the aid of an encoder-decoder architecture, where, for example, the input image (input) consists of online measurement structures during the solar cell process and the output image is a quality image of the solar cell. In a specific example, electroluminescence recordings and thermographic recordings can be compressed in relation to the prediction of the series resistance image and / or the dark saturation current image. In the combined evaluation of global and local quality values (by using a hybrid model including an encoder-decoder architecture as well as a regression / classification model), the generation of the representation can be accomplished from the network optimization with respect to all predicted parameters. This can be obtained, for example, by loss function weighting. The weights of the losses may vary with the learning process. In particular, combinations of representations from different predicted parameters can be merged, which can then be preferably used to predict another quality parameter. The weights of different loss functions can vary over time. For example, a partial representation can be derived from the short-circuit current prediction, while another prediction is derived from the open-circuit clamp voltage and the duty factor prediction. Additionally, these features can be combined and optimized in relation to the efficiency prediction.

[0031] For example, the method can include assigning a solar cell to a quality class based on at least one property determined by the method. The advantage is that classification can be based on an image, and thus the time-consuming contact measurements for classifying solar cells into different quality classes can be dispensed with.

[0032] For example, the method can be used to characterize solar cells, generally speaking semiconductor structures and / or wafers. At this time, solar cell parameters can be deduced from the properties or features of the wafer (such as cracks, grain boundaries, etc.).

[0033] In this context, examples relate to the characterization of solar cells. Images such as luminescence recordings (electroluminescence / photoluminescence), thermographic recordings, and reflection recordings can contain valuable information about the solar cell process, which is not sufficiently quantified for an empirical analysis of the images. For example, one objective in solar cell characterization may be to predict a parameter (prediction parameter) that represents some individual parameters. For example, in a convolutional neural network, a prediction model for predicting the IV parameter (IV: current-voltage) is trained for this purpose. The semantic representation of the data is obtained from one or more layers of the neural network and stored. This can be, for example, the layer before the last regression step for each prediction parameter in a multi-variate network for predicting all IV parameters. Alternatively, the prediction parameter can include a spatially resolved prediction. Correspondingly, the network can be optimized, for example, for the prediction of one or more spatially resolved quality parameters such as the dark saturation current. The prediction of the dark saturation current density, for example, enhances the detection of composite active structures that are related to the solar cell and have a strong material relationship. The prediction of, for example, a series resistance image enhances the representation of the resistance effect.

[0034] Alternatively, the prediction parameter can include a combined measurement. The prediction can be optimized for a combined prediction of multiple individual and / or spatially resolved measurements. In this case, the representations from different parts of the network can be combined. Thus, for example, the network can be designed such that different parts of the semantic representation are optimized for predicting spatially resolved information or global information and are combined and optimized for predicting another parameter, such as predicting the efficiency of a solar cell.

[0035] The semantic representation allows for a comparison of different samples of, for example, solar cells, since the distance between samples can be calculated from the semantic representation. Thus, the data can be analyzed by means of statistical methods, such as supervised and unsupervised machine learning methods. In this case, IV data can also be considered for the analysis. For example, specific defect phenomena can be identified by means of cluster analysis. Analysis methods such as hierarchical clustering methods, k-means clustering, EM algorithms, self-organizing maps, embedding techniques (such as t-SNE), or other methods can be considered for use in the analysis.

[0036] Phenomena can be determined by analyzing or characterizing semiconductor structures. Here, possible phenomena are: measurement errors during contact, short circuits in solar cells caused by electrodeposition, short circuits in solar cells caused by cracks, reduction of the series resistance caused by interrupted grid lines or incorrect contact of conductor tracks with the semiconductor structure of the solar cell, different material influencing factors affecting the efficiency and open-circuit voltage (such as the lifetime of excess charge carriers), distribution of composite active interference structures caused by misalignment, distribution of material defects caused by crucible contamination shortening the lifetime of excess charge carriers, and / or the influence of anti-reflection coatings or masking due to contact unit width on the short-circuit current.

[0037] For example, a phenomenon caused by semiconductor process variations may be a decrease in the series resistance of a solar cell. The analysis according to the method allows for the assignment of the cause of the presence of grid line interruptions to this phenomenon. Accordingly, this information can be used to eliminate the source of the cause, thereby improving the semiconductor process.

[0038] In other words, the analysis of semiconductor structures, i.e., solar cells or wafers, based on semantic representations allows for a more accurate or detailed view than based solely on IV parameters. This is due to the fact that semantic representations contain a large amount of additional information about the semiconductor structure that can be used to determine the IV parameters. However, the additional information provided allows, for example, also to determine the causes or phenomena that can explain the occurrence of the corresponding IV parameters.

[0039] An advantage of using semantic representations and neural networks to determine phenomena may also be that, due to the large amount of data and image information that can be processed, phenomena that cannot be revealed based on manual image evaluation can be determined. Thus, the method of the present invention can improve the characterization or analysis of semiconductor structures.

[0040] It is also feasible to use visualization techniques to analyze the representations. Thereby, for example, the effects or the causes of parameter variations can be presented in an understandable manner for manual evaluation or identification. The size of the representations can also be further reduced within the scope of embedding techniques, where the distances between the representations are preserved in the compressed data space. In the low-dimensional embedding space, for example, adjacent records and quality values can be visualized. In supervised learning methods, it is possible to consider using target parameters to group the representations.

[0041] For example, it is also feasible to use the method for wafer characterization, for example, for checking the material quality. In wafer characterization after the sawing process, photoluminescence absorption as well as transmission and reflection measurements with varying illumination structures, such as illumination angles, are generally considered as input values for the network. Photoluminescence absorption measures recombination-active defects here. Transmission absorption measures cracks and metal inclusions. Reflection measurements measure the grain structure.

[0042] Thus, semantic representations can contain information about semiconductor structures or wafers that is otherwise only available and stored in multiple individual images.

[0043] For example, some individual parameters of a wafer can be predicted (predicted parameter = individual parameter). The network is, for example, trained to predict a quality parameter, where the open-circuit voltage or the efficiency is a suitable parameter. Alternatively, spatially resolved measurements can be performed (predicted parameter = spatially resolved measurement). The network can be optimized with respect to the prediction of one or more spatially resolved quality parameters. The prediction of the dark saturation current density, for example, enhances the detection of the recombination-active structures associated with a solar cell. The detection of grain boundaries enhances the representation of the grain distribution. Alternatively, combined measurements can be performed (predicted parameter = combined measurement). The prediction can be optimized with respect to the combined prediction of a plurality of individual spatially resolved measurements. In this case, the representations from different parts of the network can be combined.

[0044] Phenomena that can be determined when analyzing the semantic representation of a wafer are, for example: material defects caused by recombination-active interference points, recombination-active grain boundaries, non-recombination-active grain boundaries, contamination caused by the crucible, shortening of the carrier lifetime caused by edge contamination, metal inclusions, and / or strong saw marks or types of grain distribution uniformity on the surface.

[0045] For example, it may be difficult to compare different semiconductor structures with each other solely based on images of the semiconductor structures. Therefore, according to the method, in one embodiment, it is provided to use the semantic representation to compare this semiconductor structure with another semiconductor structure. For example, this is done in the case of using the semantic representation corresponding to another semiconductor structure to determine information about the semiconductor structure and / or the semiconductor process for processing the semiconductor structure.

[0046] In other words, the use of the semantic representation according to the method allows both the comparison of the phenomena of different semiconductor structures with each other and the determination of the influence of the process on the semiconductor structure. For this purpose, the semantic representations of one or more semiconductor structures or samples can be used. It may be possible to compare the semantic representations of images based on different imaging methods. The comparison of semiconductor structures based on semantic representations allows for a quick and efficient comparison of relevant features (such as regarding the quality of a solar cell or solar cell parameters).

[0047] Accordingly, in one embodiment of the method, it is provided to process a semiconductor structure in a semiconductor process. The method includes: generating another image of the semiconductor structure after the semiconductor process, and using the other image, in addition to the said image, to generate a semantic representation. The method also includes using the semantic representation to determine the parameters of the semiconductor process.

[0048] Therefore, the method allows for a quick evaluation of a semiconductor process or some steps of a semiconductor process based on the use of semantic representations. In particular, it is possible to effectively determine what influence one semiconductor process or several semiconductor processes have on the parameters of a solar cell or a wafer.

[0049] Another aspect of the present invention relates to a method for characterizing a semiconductor process. The method includes using the semiconductor structures of respective solar cells and processing the semiconductor structures with predetermined process parameters during the semiconductor process. The method further includes generating respective semantic representations assigned to the semiconductor structures that include images of the respective semiconductor structures and using the semantic representations in combination with the predetermined process parameters to determine process phenomena.

[0050] Accordingly, respective images of the semiconductor structures can be generated before and after an operation of the semiconductor process, and respective semantic representations or the respective semantic representations can be generated based thereon. For example, there is a second model for determining process effects based on measurement results before and after one or more operations. Based on a comparison and / or processing and / or analysis of the semantic representations of the semiconductor structures before and after the operation, the operation can be characterized. In other words, the method can allow for determining the parameters that constitute an operation. For example, it can be identified what kind of influence a certain operation has on a certain solar cell parameter. By comparison, a prediction error that can represent the influence of the semiconductor operation can be determined.

[0051] For example, images of the semiconductor structures in respective different operations can be associated within the semantic representation range to predict process phenomena. The process phenomena may result from a comparison of a predicted value before the process with an actual quality value. For example, the predicted value before the process can also be compared with the predicted value after the process.

[0052] Based on the characterization of the semiconductor process, a method for improving or optimizing the semiconductor process or for process optimization can be provided. Correspondingly, the method can include adjusting the process parameters with the previously determined process phenomena. Both the method for characterizing the semiconductor structure and the method for optimizing the semiconductor process can be based on the respective similarities between the semantic representations.

[0053] When, for example, the semantic representation after an operation indicates that the quality of the solar cell parameters has decreased due to the operation, one or more process parameters can be adjusted to improve the quality of the solar cell parameters after the operation. For example, for this purpose, the corresponding process parameters can be set or adjusted based on prior knowledge. Alternatively, the process parameters can be continuously changed (reinforcement learning; for example, increased or decreased), and the respective changes in the semantic representation after the operation can be observed. Thereby, it can be found out which process parameter adjustment allows for improving the quality of the solar cell parameters (or other parameters of the semiconductor structure).

[0054] For example, it may be possible to output process error messages, such as a difference in the performance of a solar cell from the standard performance after a semiconductor process, even though the semantic representation of the solar cell semiconductor structure image generated before the semiconductor process corresponds to the standard representation. For example, the standard representation may represent the average semantic representation of such a semiconductor structure, which is assigned to a certain, e.g., high, quality grade. Thus, for example, process errors can be distinguished from material errors. If the semantic representation before the semiconductor process indicates that the semiconductor structure has high quality or high grade and the semantic representation after the semiconductor process indicates lower quality (e.g., sample or measurement values of the semiconductor structure itself can be additionally or alternatively used), then it is assumed that there is likely a process error, while, for example, basic material errors can be excluded.

[0055] There are other examples of possible methods for process characterization or process optimization of the determination. The semantic representation, for example, allows the characterization of solar cells at different stages of the manufacturing process. The representations for the process can be compared with each other. They vary with time, the selected process parameters of the pre-process, the material quality, etc. For process analysis, studies can be carried out taking into account other parameters such as process time, process parameters, the position in the process boat or the wafer position during crystallization.

[0056] It is also possible to analyze the crystallization process. The semantic representation of the wafer allows the comparison of crystallization processes. For this purpose, the semantic representations of wafers with different crystallizations can be compared with each other. In particular, samples from the same area of the brick can always be compared here. The crystallization similarity and the homogeneity of the process can be analyzed by the distance between the respective semantic representations. Deviations in different areas of the brick can also be identified.

[0057] For example, semiconductor wafers can be analyzed. The wafer representation (wafer semantic representation) allows, for example, the evaluation of the crystallization process. For this purpose, samples of the brick can be compared taking into account height information. For example, a small distance between adjacent samples indicates a homogeneous process. And a large gradient can indicate an event in the process. When a large difference appears between two successively created semantic representations, for example, a more careful examination of the process can be carried out in order to be able to exclude possible sources of error.

[0058] For example, the process of a solar cell can be analyzed. The performance of a solar cell (the semantic performance of a solar cell) allows for the comparison of the process of a solar cell. To this end, the performances of solar cells with natural or purposefully varied processes (different settings of the process parameters of a semiconductor process) can be compared with each other. To analyze process stability, the semantic performances can be compared with each other considering the process time instances. For example, a small distance between the semantic performances of temporally adjacent samples indicates a uniform process. A large gradient, on the other hand, can indicate an event in the process. This event is evaluated considering the phenomenon being analyzed. The performance can, for example, represent material effects.

[0059] A semiconductor process can, for example, be optimized. For example, in the optimization of the solar cell process, an attempt is made to improve the quality of the solar cell. To this end, a process adjustment (the effect / adjustment of process parameters) is carried out, which changes the current state in a way that improves the quality of the solar cell. In the example of reinforcement learning, it is crucial to know the effect of the state-action pairing for the optimization process. However, mere information about the current process configuration may not be sufficient to describe the current state instructively. The semantic performance of a network allows for additional information (e.g., different from mere IV parameters) that must be considered during process preference. For example, in addition to the process parameters, only four current-voltage parameters of a solar cell are available for conventional process optimization. The semantic performance can contain a large amount of additional information about the semiconductor structure that can affect the quality of the solar cell. This can, for example, be the material quality contained in the wafer performance. In the case of the semantic performance of a solar cell, the measurement records not only material-related phenomena but also process-related phenomena.

[0060] Generally speaking, one aspect of the method involves comparing the semantic performances of different semiconductor structures with each other. The similarity or difference between two semantic performances can be used to complete the characterization of a semiconductor structure (e.g., similarity or difference from a standard structure with a standard performance) or to complete process optimization.

[0061] Therefore, the analysis of a wafer precursor, a wafer, a solar cell precursor, or a solar cell can be achieved by this method. Accordingly, the processes of some individual stages of solar cell production can be optimized. In summary, therefore, the quality of the entire solar cell process can be improved by this method, for example.

[0062] Another aspect of the present invention relates to a method for evaluating an obscured region in a spatially resolved quality image of a semiconductor structure. The corresponding method includes using a semantic representation of a quality image of the semiconductor structure and processing the semantic representation in a neural network to generate first and second signal outputs. In addition, at least one local parameter is determined by analyzing the first signal output containing information about the unobscured region of the semiconductor structure, and at least one global parameter is determined by analyzing the second signal output containing information about the entire semiconductor structure. The method includes using at least one local parameter and at least one global parameter to determine at least one feature of the obscured region.

[0063] Thus, it is feasible to also predict the parameters of regions that are not shown in the image due to being covered or obscured based on this method. From the global information and the information of the unobscured region of the semiconductor structure, the obscured region can be inferred, for example. Thereby, for example, further measurements of the semiconductor structure can be avoided, but still the spatially resolved parameters of the region covered in the image or information about the covered region can be provided. For example, when the learning of the obscured region is associated with the semantic representation, the semantic representation can also be derived from the model for predicting the obscured region and the unobscured region, and this semantic representation is more valuable due to the additional information within the obscured region.

[0064] The prediction of this parameter or the simultaneous prediction of multiple parameters can be carried out by using the semantic representation and the neural network in a single step, because all the parameters can be determined at once based on the said image by means of the neural network. Thus, the high time requirement can be postponed, and this high time requirement only occurs, for example, when investigating the training data set and during the neural network training. In contrast, the determination of the semiconductor structure parameters based on the semantic representation can be carried out within tenths of a second, for example, in less than 100 milliseconds.

[0065] For example, this method can be used to achieve the prediction of the effects of semiconductor process steps. The influence of a process step can be determined by combining the measurement results before and after the process step. For example, the influence of the processing in a process step on the quality of a solar cell (reflection, IV parameters, etc.) can be studied. For this purpose, for example, data with different process characteristics or natural process variations are investigated. Then, a model or neural network for quality prediction can be trained with convincing measurement results before the actual process. Thereby, the prediction error can be determined. The prediction error represents, for example, the possible process influence of the subsequent process. Then, the combined measurement data investigated before and after the process step can be used to train a model for predicting the process influence.

[0066] With semantic representation, the main differences caused by the process can be advantageously obtained here. The characterization method can be used to identify different process results / phenomena. The local causes of the phenomena can also be identified by activating the map. Thus, the process representation can be improved and the location of the error source can be determined more precisely.

[0067] For example, the results of photoluminescence measurements in the state of the wafer (especially unprocessed or untreated wafers) particularly reflect the material properties. At this stage, for example, only a limited prediction of the quality of the solar cell can be achieved. Errors may occur when predicting the efficiency of the solar cell or other IV parameters. Therefore, the method provides that the model for predicting the open-circuit voltage is trained to determine the prediction error for the wafer photoluminescence record therefrom. Then, the neural network can be trained such that the wafer photoluminescence record and the solar cell record, such as the input of electroluminescence or photoluminescence on the solar cell, are combined to predict the process influence determined by the process or measurement and thus distinguish it from the material influence.

[0068] Another example relates to short-circuit current analysis. In a process after the surface structuring (texturing) of the semiconductor structure, the influence on the short-circuit current can be analyzed by measuring the diffuse reflection of the sample. The short-circuit current prediction error based on the diffuse reflection measurement can be the process influence of the subsequent process. After the anti-reflection coating, the reflection performance can be measured again at a selected wavelength. If this measurement result is combined with the measurement result after texturing as the input of the neural network for short-circuit current prediction, a model for predicting the process influence can be provided. The semantic representation of this model can be used to analyze and compare the anti-reflection coating process.

[0069] An example of the method relates to the prediction of at least one performance of a solar cell. The method accordingly includes generating at least one image of the semiconductor structure of the solar cell by means of an imaging method after one or different processes in the production process of the solar cell, wherein the image provides at least information about the recombination active defects of the semiconductor structure, about the process defects in the manufacturing, about the solar cell current input under different excitation conditions, or about the series resistance of the solar cell. The method also includes predicting the performance by means of a neural network trained with the measurement results of contact performance measurements, so as to predict the performance of the solar cell by means of non-contact measurements, wherein the performance includes at least one of the solar cell efficiency, the solar cell short-circuit current, the solar cell open-circuit voltage, or the solar cell duty factor.

[0070] By using a neural network to predict solar cell parameters, for example, electrical measurements can be dispensed with when determining the parameters, and the parameters can be determined based only on an image of a semiconductor structure. For example, the method allows, using a network or neural network, the measurement results of an imaging method to be used to characterize or analyze a semiconductor structure, rather than using electrical contact measurements or physical contact measurements of the semiconductor structure. This enables, for example, the duration for characterizing a semiconductor structure to be shortened. The neural network training may only need to be done once. For example, multiple parameters can be determined at once based on a single image. Thus, the method enables the characterization or quality determination of semiconductor structures in mass production, since strict time settings can be met. Description of the Drawings

[0071] Several examples of the device and / or method will now be described in detail by way of example only with reference to the drawings, wherein:

[0072] Figure 1 is a flowchart of a method for processing an image of a semiconductor structure;

[0073] Figure 2 is a flowchart of a method for determining the performance of a solar cell;

[0074] Figure 3 is a flowchart of a method for improving a semiconductor process;

[0075] Figure 4 is a flowchart of a method for evaluating a spatially resolved quality image;

[0076] Figure 5 is a flowchart of a method for predicting the performance of a solar cell;

[0077] Figures 6a to 6c is an architecture of a neural network with semantic representation;

[0078] Figure 7 is a schematic diagram showing a method for calibrating a spatially resolved quality measurement;

[0079] Figure 8 is a schematic diagram showing a method for evaluating a masked area in a spatially resolved quality image;

[0080] Figure 9 is a method for training a model for determining process impacts;

[0081] Figure 10 is a method for determining or analyzing process impacts. Detailed Description of the Embodiments

[0082] The examples will now be described in detail with reference to the drawings showing several examples. In the drawings, the size of lines, layers, and / or regions may be exaggerated for illustrative purposes.

[0083] Accordingly, although there are other examples applicable to various different modifications and alternative forms, only a few specific examples are shown in the figures and will be described explicitly below. However, this detailed description does not limit other examples to the specific forms described. Other examples may cover all modifications, equivalents, and alternatives falling within the scope of this document. Throughout the description of the figures, the same or similar reference numerals refer to the same or similar components, which can be implemented identically or in a modified form when compared with each other, but they provide the same or similar functions.

[0084] Obviously, when an element is referred to as "connected" or "joined" to another element, these elements can be connected or joined directly or through one or more intermediate elements. When using "or" to combine two elements A and B, this means that all possible combinations are disclosed, i.e., only A, only B, and both A and B, unless otherwise explicitly or implicitly defined. An alternative expression for the same combination is "at least one of A and B" or "A and / or B". With the necessary adaptations, this also applies to combinations of more than two elements.

[0085] The terms used herein to describe specific examples are not intended to limit other examples. When using the singular form such as "a", "an", "the" and the use of a single element is neither explicitly nor implicitly defined as mandatory, other examples can also use multiple elements to achieve the same function. When a function is described below as being achieved by multiple elements, other examples can use a single element or a single processing entity to achieve the same function. It is also self-evident that the terms "comprising", "including", "having" and / or "possessing", when used, clearly state the presence of the indicated features, integers, steps, operations, processes, elements, components and / or groups thereof, but do not preclude the presence or addition of one or more other features, integers, steps, operations, processes, elements, components and / or groups thereof.

[0086] Unless otherwise defined, all terms (including technical and scientific terms) used herein are used in their normal meaning in the field, which is an example.

[0087] Figure 1Method 100 for processing an image of a solar cell semiconductor structure generated using an imaging method is shown. Method 100 includes processing 110 the image of the semiconductor structure by means of a trained neural network to generate a semantic representation of the image. The data size of the semantic representation generated according to this method is less than the data size of the image here. Method 100 also includes storing 120 the semantic representation corresponding to the semiconductor structure. Here, the semantic representation provides an information carrier for relevant image information, and the information can be used for the determination of the performance of the solar cell and / or for the crystal analysis of the semiconductor structure of the solar cell.

[0088] The semiconductor structure can be, for example, a wafer for manufacturing a solar cell. The image can be a photoluminescence record of the wafer. In order to predict solar cell parameters based on a photoluminescence record on a cut wafer, the following method can be adopted. According to this method, certain artificially defined features are extracted. According to this exemplary method, the features are indirectly learned based on the prediction, and an intermediate representation is used to store / compress 120 the data and perform monitoring and classification.

[0089] In addition, according to the exemplary method 100, artificially defined features from electroluminescence measurements or photoluminescence measurements of the solar cell can be used to detect process errors. They are, for example, grid line interruptions, cracks, crystal structure defects, etc.

[0090] For example, a neural network, a convolutional neural network, or a deep learning method is used for this purpose. A deep learning method is a machine learning method that produces an end-to-end association between input data (such as an image) and output data (such as a quality grade or an image). Generally, the network is optimized to minimize a loss function. For example, this can be assigned to a class. For this, data can also be compressed by encoding the structure description in the network and then decoding it (encoder-decoder architecture). In the case of an autoencoder, the loss function is usually the difference between the reconstructed data and the measured data (input image) at this time. But other images can also be used.

[0091] A convolutional neural network learns to predict quality parameters based on one or more input images. During training, the convolutional neural network learns to predict the target parameters based on the pairing of the input data and the true measured values of the target parameters. In the application scenario (testing), the network is used to predict these parameters. In this case, only the quality parameters are still needed for evaluation. When passing through the network, the input data is gradually compressed more and more, so that data reduction, abstraction, and semantically valuable representations occur. The representations can be generated by the network for each sample and read from the network.

[0092] The described representation or semantic representation is composed of, for example, a feature vector or a multi-dimensional feature combination, which generally appears in a network terminal for classification or regression. For example, grading is equivalent to classification, and regression is equivalent to the prediction of continuous values. Then, the prediction of the quality parameter occurs based on the feature vector. It is not only persuasive regarding the quality parameter but also represents the compression of the structure in the image. When the spatially resolved structure is also predicted, the persuasiveness can be higher( Figure 3 ). In the case of an encoder-decoder architecture, for example, the semantic representation in the bottleneck region is obtained after the encoder or after the decoder at the network terminal. However, the representation can also be a combination of different features of the network.

[0093] One aspect of the present invention is to use the abstract representation of data for compressed storage of 120 data. In addition, the compressed representations can be combined to compare images. Then, based on the input and output data, crystallization can be analyzed, the process can be compared / monitored, material and process characteristics can be distinguished, and data can be classified.

[0094] There are many convolutional neural network architectures. The architectures may be distinguished in terms of whether they predict spatially resolved quality data (see Figure 6b ), or whether they predict one or a set of different parameters (classification model or regression model, see Figure 6a ). Basically, the models discussed herein consist of convolutional layers, non-linear units such as ReLU (Rectified Linear Unit), and downsampling pooling layers. Optionally, upsampling layers, normalization steps, and split and skip connections and similar layers can be part of the network. Data augmentation methods such as coordinate transformation, random edge truncation, or noise can be used during training to expand the data set and improve the model robustness. The so-called dropout layers can also be used during training to improve the model robustness. These layers can be combined in very different ways.

[0095] Figure 2 A flowchart of a method 200 for determining the performance of a solar cell is shown. The method 200 includes generating 210 a semantic representation of an image of a solar cell by means of an imaging method, wherein the semantic representation is generated by means of a neural network, and the method further includes determining 220 at least one performance of the solar cell based on the semantic representation.

[0096] Figure 3 A flowchart of a method 300 for improving a semiconductor process is shown. The method includes: using 310 the semiconductor structure of each solar cell, processing 320 the semiconductor structure with predetermined process parameters during the semiconductor process, generating 330 a respective semantic representation corresponding to the semiconductor structure from an image containing each semiconductor structure, using 340 the semantic representation in combination with the predetermined process parameters to determine process phenomena, and adjusting 350 the process parameters in the case of the previously determined process phenomena.

[0097] Also in combination with Figure 9 and Figure 10 provide additional examples and detailed descriptions of Method 100, Method 200, and / or Method 300.

[0098] Figure 4 A flowchart of a method 400 for evaluating a spatially resolved quality image is shown. The method 400 includes: using 410 the semantic representation of the quality image of a semiconductor structure, processing 420 the semantic representation in a neural network to generate a first signal output and a second signal output, determining 430 at least one local parameter by analyzing the first signal output containing information on the unmasked region of the semiconductor structure, determining 440 at least one global parameter by analyzing the second signal output containing information on the entire semiconductor structure, and using the at least one local parameter and the at least one global parameter to determine 450 at least one characteristic of the masked region.

[0099] Figure 5 A flowchart of a method 500 for predicting the performance of a solar cell is shown. The method 500 includes generating 510 at least one image of the semiconductor structure of the solar cell by means of an imaging method after one or different processes during the production of the solar cell, wherein the image contains at least information on the recombination active defects of the semiconductor structure, on process errors during the production process, on the solar cell current input under different excitation conditions, or on the solar cell series resistance, and the method further includes predicting 520 the performance by means of a neural network trained in the case of using the measurement results of contact performance measurements in order to predict the performance of the solar cell by means of non-contact measurements.

[0100] Figures 6a to 6c A neural network architecture with a semantic representation is shown. The shown architecture can be used to process 610 an image of a semiconductor structure. According to Figure 6a the example, the image 610 is input into the input end of the neural network 612 to generate a semantic representation 620. The semantic representation 620 can have various features 622 and the activation states 624 of the respective features. The semantic representation 620 can be processed 630 in the case of using a regression model or a classification model to generate output parameters 640, such as the IV parameters or the IV category of a solar cell.

[0101] According to Figure 6b , a CNN (CNN: Convolutional Neural Network) encoder 614 is used to process the image 610 to generate a representation 620. The semantic representation 620 can be processed in a CNN decoder 632 into, for example, initial data 642 for predicting position-resolved quality data.

[0102] Figure 6cshows the architecture of a hybrid model, which includes a CNN encoder 614, a CNN decoder 632, and a regression model 630. Using the hybrid model according to Figure 6c , not only can the spatially resolved quality data 642 be predicted, but also several parameters 640 or a set of parameters 640 (such as IV parameters) can be predicted.

[0103] Figure 7 shows a schematic diagram of a method for calibrating spatially resolved quality measurements. This method is based on the combination of a global loss function and a local loss function in network training. In the network, the spatially resolved prediction 710 of solar cell parameters is predicted. According to the analysis / empirical model 720, the value of a parameter 730 can be calculated from the spatially resolved quality parameters 710 from the current-voltage characteristic curve. When training the network, the model for predicting spatially resolved parameters is optimized in such a way that not only the local structure of the spatially resolved quality parameters (loss function 740) is retained, but also the prediction error of the global quality value is minimized (loss function 750). This calibration method can basically correspond to the points for evaluating the shaded areas.

[0104] According to Figure 7 , the input 610 is the measurement result at the solar cell and / or the solar cell precursor, such as the electroluminescence record and / or the photoluminescence record of the solar cell. The series resistance effect can be seen therein, but it cannot be immediately distinguished from other defects such as recombination defects. The network should predict an image of the series resistance. The measurement result 710a of the same sample is determined for network training. In addition, the global series resistance can be deduced from the global series resistance of the solar cell through an analysis model. It is determined according to the current-voltage characteristic. The global series resistance is the average value of the series resistance image. Now, the network is optimized for the prediction of the series resistance image and the global series resistance. Similarly, a model for predicting an image of the dark saturation current density or an image of the ratio of the global short-circuit current density to the dark saturation current density can also be predicted from the input image 610. Thus, the open-circuit voltage can be calculated according to the single diode model through an analysis model. When optimizing, the local prediction error of the ratio of the short-circuit current density to the dark saturation current density (loss function 740) is minimized, and the global open-circuit voltage error can be calculated according to the current-voltage characteristic curve (loss function 740).

[0105] Figure 8Schematic diagram showing a method for evaluating occluded regions in a spatially resolved quality image. The method is based on the combination of global and local loss functions in network training considering the occluded regions. In quality measurement, certain regions of a solar cell may be occluded, for example, due to the printing process, and thus they cannot be measured. The mask 800 causing the occlusion can be considered both when inputting into the network 614 and when training to minimize the loss function 840. The aim of the method is to determine the spatially resolved values of the occluded regions. The network consists of two output terminals 810, 830 (see Figure 7 ), and information about the occluded regions can also be added to the model. The concept now is to consider only the non-occluded regions in the locally resolved loss function 840. And the loss function 850 for predicting global parameters is calculated with respect to the entire image. This is particularly noteworthy when the input image contains information in the occluded regions because it was recorded before the occlusion.

[0106] The input to the network is the photoluminescence record 610 on the wafer or the luminescence measurement 610 of the solar cell. The network is trained to predict the ratio of the short-circuit current density to the dark saturation current density (loss function 840) and to predict the open-circuit voltage (loss function 850) through an additional analysis model. In the mask, the bus bar and grid line regions of the solar cell that cannot be measured in the optical measurement are marked. They are not considered in the loss function 840 and thus can be optimized through the loss function 850.

[0107] The optimal orientation of the samples for solar cell production can be derived from the model because the quality under metallization may also depend on the material quality. This applies to the entire training process where the mask follows or the rotation during training is always the same at this time. Thus, a model for predicting material quality is described. For example, the open-circuit voltage can be determined from the PL image of the wafer and the mask-related information, where the mask provides information about the printed image in the model.

[0108] Figure 9 Showing a method for training a model for determining process impacts, such as the first sub-method 900a.

[0109] The steps shown can be particularly used in a predetermined method that includes generating another image of a semiconductor structure after a semiconductor process, using the other image in addition to the said image to generate a semantic representation, and / or determining the parameters of the semiconductor process in the case of using the semantic representation.

[0110] The predetermined method can provide the required prediction model for predicting quality parameters in the first sub-method. In particular, prediction errors (such as process phenomena) may be relevant, which can be used in the second sub-method of the predetermined method (see Figure 10)。In the first sub-method, the records before the process and the quality after the process can be used. For example, this semantic manifestation is not absolutely necessary here. The process phenomenon may be the main result. For example, the process phenomenon corresponds to the prediction error.

[0111] For example, the first sub-method includes: generating at least one image of the semiconductor structure 901 before the process; and performing the process / process sequence 902; generating at least one image of the semiconductor structure 903 after the process or process sequence; optionally performing other processes 904; determining the quality data 905 for process optimization; creating a prediction model 906 for evaluating quality parameters using at least one image before the process; determining the process impact 907 by calculating the prediction error of the model; training the model 908 for determining the process impact based on at least one image before the process and at least one image after the process.

[0112] In addition, a method for predicting the prediction error (such as the process impact) can be provided from the first sub-method 900a. The model works, for example, using the images before and after the process. Then, this semantic manifestation can be used for analysis. For example, the process phenomenon corresponds to the parameters of the semiconductor process.

[0113] Figure 10 A corresponding second sub-method 900b for determining and / or analyzing the process impact is shown. The second sub-method 900b includes: generating at least one image of the semiconductor structure 909 before the process; generating at least one image of the semiconductor structure 910 after the process; calculating the semantic manifestation 911 according to a model for determining the measurements and process impact before and after the process; and optionally determining the process impact 912 by the model.

[0114] The first sub-method 900a and the second sub-method 900b can be executed especially in relation to method 100 for processing images generated in the case of using an imaging method for a solar cell semiconductor structure, which also wholly or partly includes processing the semiconductor structure (individual method steps) in a semiconductor process, in relation to method 200 and / or in relation to method 300 for improving the semiconductor process. Generally, some method steps that simply exemplarily correspond to a certain method can also be used in other shown methods in order to, for example, further improve the method.

[0115] In other examples, other aspects or improvement solutions of the present invention are described, which can bring advantages, for example, in combination with the foregoing aspects.

[0116] One example relates to the inspection of solar cells and a corresponding method for predicting current-voltage parameters (IV parameters). According to an example of the method, a prediction model is trained with the aid of a convolutional neural network, which has been optimized to predict one or more IV parameters (such as series resistance as well) from imaging parametric data that occur on samples during material input inspection, solar cell production, and solar inspection, and the semantic representation of the network can be derived from this model. In this case, the imaging measurement results can be incorporated into a matrix with different channels and / or can be input into different subnets and then combined.

[0117] In an example of the method, the semantic representation can be used to efficiently store the input data. The input data can here consist of any combination of imaging measurements, such as the presence of photoluminescence, electroluminescence, temperature recordings on the solar cell, and measurements on the wafer.

[0118] For example, a sample representation will be generated, which contains parts of the semantic representation and is analyzed in terms of process and material on this basis. In this case, the sample representation can also contain IV values and / or can be analyzed with respect to this. Thus, typical defect phenomena and process errors can be identified with the aid of machine learning methods such as clustering methods, regression methods, or classification methods. To identify the phenomena, these samples are classified with respect to the typical data distribution of the sample representation, whereby the samples can be classified. This can be done in an unsupervised manner (unsupervised learning of the neural network), for example based on the data distribution, and / or supervised (supervised learning of the neural network), based on other quality data such as IV data or labeled defect classes. In addition, information about material defects, process defects, and measurement errors such as contact errors can be obtained thereby. The prototypes of these "phenomenon classes" can be visualized and defect causes can be assigned.

[0119] Error detection (fault detection) can be provided, for example, with the aid of an error message signal. Despite similar sample representations, efficiency deviations also indicate, for example, anomalies in the sample or process errors. In addition, for similar descriptions containing the semantic representation, the process development can also be analyzed by comparing the determined efficiencies.

[0120] An example of the method relates to the prediction of solar cell parameters without contact measurements. It is particularly advantageous that a measurement system that requires time-consuming contact with the solar cell is not used. The input parameters can be: measurements for identifying recombination-active defects, such as PL recordings on wafers, solar cells, or solar cell precursors; measurements for identifying process errors on solar cells or solar cell precursors, such as PL recordings in each process step, IR transmission light measurements on wafers; measurements for determining current input, such as PL recordings (SunsPL) on solar cells under different excitation conditions; and / or measurements for determining series resistance, such as partially masked PL recordings on solar cells. In addition, measurements for determining quasi-static photoconductivity are also possible. The output parameters can be: parameters related to solar cell classification, in particular the short-circuit current density J sc derived from the IV parameters, efficiency, and open-circuit voltage V oc or quality grade.

[0121] It is possible to supplement other measurements or omit measurements. For example, color measurements for classification can be added.

[0122] An example relates to a method where a convolutional neural network is further optimized to predict spatially resolved quality criteria such as series resistance images or dark saturation current images. The semantic representation here can be a combination of features that are used on the one hand to predict one image and predict another image and are used in different combinations to predict certain IV parameters.

[0123] For example, a method is provided where the resulting semantic representation flows into the sample representation and thereby the process parameters are optimized. In this case, the optimization is achieved by parameterization based on the emerging representation.

[0124] An example relates to a method, such as a method for material and crystallization evaluation, where a convolutional neural network is trained to predict IV parameters or quality grades derived therefrom, and the measurement results used occur after crystallization on bricks and / or after slicing on wafers. This can particularly include measurements that allow the analysis of grain distribution, such as reflected light measurements, infrared transmission light measurements, measurements for determining charge carrier recombination activity / lifetime, such as photoluminescence measurements on wafers and bricks, or photoconductivity measurements, measurements for identifying inclusions, for example, by infrared (IR) transmission light methods on wafers, and measurements of resistance that can be measured on wafers and bricks. For example, silicon carbide inclusions can be measured in a spatially resolved manner by the transmission signal on bricks or the transmission signal on wafers.

[0125] One example also relates to a method in which the semantic representation resulting from this flows into the sample representation and is used, for example, in a similar manner to the above-described method to assess the crystallization. The sample representation allows sample comparisons. For the analysis, for example, silicon wafers can be distinguished from one another by certain crystallization criteria, such as brick origin, type of raw material used, crystallization method or seeding process, and the samples can be compared within one group by homogeneity, or similarities between classes can be compared with the aid of the sample representation. For the analysis, supervised and unsupervised learning methods can be used.

[0126] An example also relates to a method in which anomalies and improvements in a solar cell process are determined by the fact that a predicted solar cell quality does not match an expected solar cell quality, or samples with the same performance result in different solar cell results.

[0127] An example also relates to a method in which the convolutional neural network is also optimized to predict a spatially resolved quality criterion, such as a series resistance image or a dark saturation current image.

[0128] One example relates to the characterization of processes. In the analysis of processes or semiconductor processes, a space of semantic representations or feature spaces can be used to identify similar phenomena. For example, data analysis methods (such as clustering) can be used to identify unknown relationships in the data. In particular, the information goes beyond the information content of pure IV measurements. Different phenomena can be identified depending on the type of analysis. In crystallization analysis, for example, different material defects can be identified. In solar cell characterization, for example, material properties and / or process characteristics can be identified.

[0129] One example relates to the optimization of semiconductor processes. In process optimization, the distance between the representation and the representation derived therefrom can always be taken into account in connection with or in combination with additional information. The additional information can take into account process parameters or other information, such as, for example, the position of the sample or semiconductor structure in the brick in the optimization of the crystallization process, or the position of the solar cell in the carrier or carrier in connection with the process time and / or process parameters.

[0130] In summary, the present invention provides a method that utilizes semantic representations, which may allow efficient storage or analysis of parameters, for example by analyzing the semantic representation or comparing the semantic representation with one or more other semantic representations.

[0131] with one or more of the previously described examples and appendices Figure 1 The aspects and features described above may also be combined with one or more other examples to replace the same features of other examples or to additionally introduce the features into other examples.

[0132] In addition, an example may be or relate to a computer program having program code that, when run on a computer or processor, is used to perform one or more of the above methods. The steps, operations, or processes of the various above methods may be performed by a programmed computer or processor. An example may also cover a program storage device such as a digital data storage medium that is machine-readable, processor-readable, or computer-readable and encodes a machine-executable, processor-executable, or computer-executable program of the instructions. The instructions perform several steps or all steps of the above methods or cause them to be performed. The program storage device may be, for example, or include a digital memory, a magnetic storage medium such as disks and tapes, a solid state drive, or an optically readable digital data storage medium. Other examples may also cover a computer, a processor, or a control unit programmed to perform the steps of the above methods, or a (field) programmable logic array ((F)PLA = (field) programmable logic array) or a (field) programmable gate array ((F)PGA = (field) programmable gate array) programmed to perform the steps of the above methods.

[0133] Through the specification and the drawings, only the principles herein are illustrated. In addition, all examples listed herein are in principle clearly only for illustrative purposes to support the reader's understanding of the principles herein and the concepts contributed by the inventors for the improvement of the technology. All statements herein regarding the principles, aspects, and examples of this document, as well as the specific examples, include their equivalents.

[0134] A functional block referred to as an "apparatus" for performing a particular function may relate to a circuit designed to perform the particular function. Thus, an "apparatus for something" can be implemented as an "apparatus designed for or adapted to something", with components or circuits designed for or adapted to the respective tasks.

[0135] The functions of the various components shown in the figure, including each functional block referred to as "mechanism", "mechanism for providing a signal", "mechanism for generating a signal", etc., can be implemented in the form of dedicated hardware such as "signal providing device", "signal processing unit", "processor", "controller", etc., and in the form of hardware that can execute software in combination with the associated software. When provided by a processor, the function can be provided by a separate dedicated processor, a separate shared processor, or multiple independent processors, some or all of which are shareable. However, the term "processor" or "controller" is in no way limited to hardware that can only execute software, but can include digital signal processor hardware (DSP hardware; DSP = digital signal processor), network processors, application specific integrated circuits (ASICs), field programmable logic components (FPGA = field programmable gate array), read only memories (ROMs) for storing software, random access memories (RAMs), and non-volatile memories (storage). Other conventional and / or custom hardware may also be included.

[0136] A block diagram can represent, for example, a rough circuit diagram implementing the principles herein. In a similar manner, flowcharts, process diagrams, state transition diagrams, pseudocode, etc. can represent various processes, operations, or steps, which are represented, for example, substantially in a computer-readable medium and thus executed by a computer or processor, whether or not such computer or processor is explicitly shown. The methods disclosed in the specification or claims can be implemented by components having means for performing each of the steps of the method.

[0137] Obviously, the multiple steps, processes, operations, or functions disclosed in the specification or claims should not be construed as being in a prescribed order, unless explicitly or implicitly indicated otherwise, for example, for technical reasons. Thus, unless these steps or functions are non-interchangeable for technical reasons, they are not limited to a particular order by virtue of the disclosure of several steps or functions. In addition, in some examples, a single step, function, process, or operation may include and / or be divided into multiple sub-steps, sub-functions, sub-processes, or sub-operations. These sub-steps may be included, unless explicitly excluded, and are part of the disclosure of the respective single step.

[0138] In addition, the following claims are hereby incorporated into the detailed description, where each claim may stand on its own as a separate example. Although each claim may stand on its own as a separate example, it should be noted that although dependent claims in the claims may refer to a specific combination with one or more other claims, other examples may also include combinations of dependent claims with the subject matter of each other dependent claim or an independent claim. Unless explicitly stated that a certain combination is not intended, the use of these combinations is expressly recommended herein. In addition, the features of one claim should also be incorporated for each other independent claim, even if this claim does not directly refer to that independent claim.

Claims

1. A method for processing an image (610) of a semiconductor structure of a solar cell, the image (610) being generated using an imaging method, the method comprising: processing (110) at least two images (610) of the semiconductor structure by means of a trained neural network to generate a semantic representation (620) of the image (610), the at least two images (610) being generated using different imaging methods, wherein the data size of the semantic representation (620) is smaller than the data size of the image (610) input into the input end of the neural network; and storing (120) the semantic representation (620) corresponding to the semiconductor structure as an information carrier for information related to determining the performance of the solar cell and / or for crystal analysis of the semiconductor structure of the solar cell, wherein the information of the image (610) is compressed and stored (120) by generating the semantic representation (620), wherein the semantic representation (620) includes a weight distribution of the weights of the layers of the neural network.

2. A method for determining at least one performance of a solar cell, the method comprising: using (210) the semantic representation (620) of an image (610) of the solar cell generated using an imaging method, the semantic representation (620) being generated by means of a neural network; and determining (220) at least one performance of the solar cell based on the semantic representation (620).

3. The method according to claim 2, wherein, the performance is determined using a neural network.

4. The method according to claim 2 or 3, wherein, the at least one performance includes the open-circuit voltage, short-circuit current, short-circuit current density, efficiency, duty factor and / or series resistance of the solar cell.

5. The method according to claim 2, wherein, the at least one performance relates to a spatially resolved quality criterion of the solar cell.

6. The method according to claim 2, the method further comprising: assigning the solar cell to a quality class based on the at least one performance.

7. The method according to claim 1 or 2, wherein, the image (610) on which the semantic representation (620) is based is established using electroluminescence, photoluminescence or temperature absorption.

8. The method according to claim 1 or 2, the method comprising: comparing the semiconductor structure with another semiconductor structure using the semantic representation (620) in the case of using a semantic representation corresponding to another semiconductor structure to determine information about the semiconductor structure and / or information about a semiconductor process for processing the semiconductor structure.

9. The method according to claim 1 or 2, the method further comprising: processing the semiconductor structure in a semiconductor process; generating another image, the another image including the semiconductor structure after the semiconductor process; In addition to using the said image (610), the said another image is also used to generate a semantic representation (620); and parameters of the said semiconductor process are determined in the case of using this semantic representation (620).

10. A method for improving a semiconductor process, the said method comprises: using (310) the semiconductor structures of respective solar cells; processing (320) the said semiconductor structures with predetermined process parameters in a semiconductor process; generating (330) respective semantic representations (620) corresponding to the semiconductor structures of images (610) each containing a semiconductor structure, wherein the said semantic representations (620) are generated by the method according to claim 1; using (340) the said semantic representations (620) in combination with the said predetermined process parameters to determine process phenomena; and adjusting (350) the said process parameters in the case of using the process phenomena determined previously.

11. The method according to claim 10, the said method further comprises: outputting a process error message in the case of a deviation between the performance of the said solar cell and the standard performance occurring after a semiconductor process, even if the semantic representation (620) of an image generated before the said semiconductor process based on the said semiconductor structure of the said solar cell conforms to the standard representation.

12. A method for evaluating a spatially resolved quality image (610) of a semiconductor structure, the said method comprises: using (410) the semantic representation (620) of the quality image (610) of the said semiconductor structure generated by the method according to claim 1; processing (420) the said semantic representation (620) in a neural network to generate a first signal output and a second signal output; determining (430) at least one local parameter by evaluating the said first signal output containing information on an unshielded area of the said semiconductor structure; determining (440) at least one global parameter by evaluating the said second signal output containing information on the whole of the said semiconductor structure; and determining (450) at least one feature of a shielded area in the case of using the said at least one local parameter and the said at least one global parameter.

13. The method according to any one of claims 1, 2, 10, 12, wherein the said method is used for predicting at least one performance of a solar cell, and the said method further comprises: during the manufacture of the said solar cell, generating (510) at least one image (610) of the semiconductor structure of the said solar cell by means of an imaging method after one or each different process step, wherein the said image (610) contains at least information on recombination active defects of the said semiconductor structure, on process defects in the manufacture, on current input of the said solar cell under different excitation conditions or on the series resistance of the said solar cell; and predicting (520) the said performance by means of a neural network trained in the case of using the measurement results of a contact type performance measurement, so as to predict the said performance of the said solar cell by means of a non-contact measurement, Wherein, the performance includes at least one of the efficiency of the solar cell, the short-circuit current of the solar cell, the open-circuit voltage of the solar cell, and the duty factor of the solar cell.

Images