Parameter optimization system, parameter optimization method, and storage medium

Abstract

In the present invention, if a converted image generated by using a learning model that generates a converted image is utilized for measurement, a parameter of the learning model is optimized so that dimensional metrics can be improved. For this purpose, the present disclosure proposes a parameter optimization system for optimizing a parameter for generating converted images on the basis of input images, said system comprising a storage device for retaining a parameter optimization program, and a processor for reading in the parameter optimization program from the storage device and executing same. The processor executes: processing for calculating a conversion error between the converted images and training images inputted during learning; processing for calculating dimensional metrics of a pattern included in the training images on the basis of the converted images and the training images; processing for calculating a composite error on the basis of the conversion error and the dimensional metrics; training of a machine learning model for generating converted images so as to optimize the composite error; and processing for optimizing a parameter of the machine learning model (see fig. 1).

Description

Parameter optimization system, parameter optimization method, and storage medium

[0001] The present disclosure relates to a parameter optimization system, a parameter optimization method, and a storage medium that stores a parameter optimization program for optimizing parameters of a learning model that performs image conversion.

[0002] In recent years, attempts have been made to efficiently perform image measurement using AI (artificial intelligence) technology. For example, Patent Literature 1 discloses a technology for generating high-quality images using a learning model that generates a converted image based on an input image generated by an image generation tool. In particular, the learning model applies different learning to specific portions, such as edges of an object to be measured or inspected, in the converted image in order to improve the generation quality of the specific portions.

[0003] International Publication No. 2021 / 140620

[0004] However, in Patent Document 1, when training a learning model that outputs a converted image to be measured, only the conversion error between the converted image and the teacher image input during training is used. Therefore, dimension metrics are not utilized. Here, dimension metrics are index values ​​that include the measurement accuracy calculated from the dimension values ​​and the throughput from the generation of the input image by the image generation tool to the calculation of the measurement accuracy for the converted image. For this reason, even if the technology disclosed in Patent Document 1 is used, it may not be possible to generate a converted image with sufficiently improved dimension metrics.

[0005] On the other hand, deep learning using neural networks and the like requires that the errors used in learning be differentiable. Therefore, dimension metrics that correspond to non-differentiable operations, such as dimension values, cannot be used as errors. This makes it difficult to achieve the target values ​​of dimension metrics desired by the user, making it difficult to generate a learning model that performs image conversion suitable for practical use.

[0006] In view of this situation, the present disclosure proposes a technology for optimizing the parameters of a learning model to improve dimensional metrics when a converted image generated using the learning model is used for measurement.

[0007] In order to solve the above problems, the present disclosure proposes a parameter optimization system that optimizes parameters for generating a transformed image based on an input image, comprising a storage device that stores a parameter optimization program, and a processor that reads and executes the parameter optimization program from the storage device, wherein the processor performs the following processes: calculating a transformation error between the transformed image and a teacher image input during learning; calculating dimension metrics of a pattern contained in the teacher image based on the transformed image and the teacher image; calculating a composite error based on the transformation error and the dimension metrics; training a machine learning model that generates the transformed image so as to optimize the composite error; and optimizing the parameters of the machine learning model.

[0008] Further features related to the present disclosure will become apparent from the description of this specification and the accompanying drawings. Also, aspects of the present disclosure are achieved and realized by the elements and combinations of various elements and the aspects of the following detailed description and the appended claims. The description of this specification is merely exemplary and does not limit the scope or application of the claims of the present disclosure in any way.

[0009] The technology disclosed herein enables the generation of a learning model that generates a transformed image that is optimal for a dimension metric, thereby facilitating the generation of a transformed image that satisfies the dimension metric desired by the user.

[0010] FIG. 1 is a diagram showing an example of a schematic configuration of a parameter optimization system 10 according to a first embodiment. FIG. 2 is a flowchart for explaining details of the parameter optimization process (generation of an optimized trained model and generation of dimension metrics). FIG. 3 is a flowchart for explaining details of the conversion parameter optimization process (step S204) in the parameter optimization process (FIG. 2). FIG. 4 is a diagram showing an example of a configuration of an optimization condition setting GUI (Graphical User Interface) 301 that is displayed on the display screen of the output device 102 and that is used to input optimization conditions via the input device 101. FIG. 5 is a diagram showing an example of a configuration of an optimization result output GUI 401 that is displayed on the output device 102. FIG. 6 is a diagram showing an example of a configuration of an optimization condition setting GUI 301′ according to Modification 1. FIG. 7 is a diagram showing an example of a configuration of a parameter optimization system 10 according to Modification 1. FIG. 8 is a diagram showing an example of a schematic configuration of a learning model parameter optimization system 10 according to a second embodiment. FIG. 9 is a diagram showing an example of a schematic configuration of a parameter optimization system 10 according to a third embodiment. FIG. 10 is a flowchart for explaining details of the parameter optimization process according to the third embodiment.

[0011] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. In the accompanying drawings, functionally identical elements may be designated by the same numerals. Note that the accompanying drawings show specific embodiments in accordance with the principles of the present disclosure, but these are intended to help understand the present disclosure and are not to be used to interpret the present disclosure in any way as being limiting.

[0012] Although the present embodiment has been described in sufficient detail to enable those skilled in the art to practice the present disclosure, it should be understood that other implementations and forms are possible, and that changes in configuration and structure and substitutions of various elements are possible without departing from the scope and spirit of the technical ideas of the present disclosure. Therefore, the following description should not be interpreted as being limited thereto.

[0013] In this embodiment, the measurement of the dimensions of a circuit pattern will be described, taking the measurement of a circuit pattern of a semiconductor image as an example. However, the application of the technology disclosed herein is not limited to the measurement of semiconductor circuit patterns, but can be broadly applied to image measurement in general.

[0014] (1) First Embodiment The first embodiment relates to a technique for improving throughput by converting an input image to generate an image with higher image quality than the input image.

[0015] For example, to perform dimension measurement using a scanning electron microscope, the signals obtained by scanning the same field of view multiple times are averaged to generate a high-quality image with a high S / N ratio, and then measurement is performed. However, performing multiple scans requires a long imaging time, which reduces throughput. Therefore, in this embodiment, a low-quality image generated by performing fewer scans than when generating a high-quality image (an image with a quality that allows dimension measurement with a predetermined accuracy) is converted to generate an image equivalent to high quality. This eliminates the need to generate high-quality images captured at low speed, thereby improving throughput.

[0016] 1 is a diagram showing a schematic configuration example of a parameter optimization system 10 according to the first embodiment. The parameter optimization system 10 includes, for example, a calculation device 100 that calculates parameters, an input device 101 that is used when a user (operator) inputs target values ​​of dimension metrics (described later), an output device 102 that outputs (displays) optimization results, and a storage device 201 that stores high-quality images (e.g., SEM images) that have image quality that allows dimension measurement with a predetermined accuracy, and low-quality images (e.g., SEM images; it does not have to be unmeasurable) that have image quality inferior to the high-quality images.

[0017] A scanning electron microscope or the like can be used as an image generation tool to generate input images (high-quality images and low-quality images to be stored in the storage device 201) to be input for optimizing the parameters of the learning model. Furthermore, the input device 101 can be configured with a keyboard, a mouse, or the like. Furthermore, the output device 102 can be configured with a display or the like. The input device 101 is configured to receive target values ​​for each dimension metric input by the user (see FIG. 4 ), and output the optimized results to the output device 102.

[0018] Although not shown in FIG. 1, the arithmetic device 100 can be configured as a computer system including one or more CPUs (Central Processing Units) and GPUs (Graphics Processing Units), and various memories (storage devices) configured with RAMs (Random Access Memory), ROMs (Read Only Memory), etc., for storing various programs, parameters, etc.

[0019] The arithmetic device 100 uses a processor (such as a CPU) to read a parameter optimization program stored in a memory and load the program into the processor's internal memory, thereby realizing a parameter optimization function.

[0020] As shown in FIG. 1, the calculation device 100 includes, as the parameter optimization function, for example, an image conversion unit 1001, a conversion error calculation unit 1002, a dimension metrics calculation unit 1003, a composite error calculation unit 1004, a conversion parameter update unit 1005, and a learning parameter update unit 1006.

[0021] The image conversion unit 1001 receives an input image, such as an image generated using a scanning electron microscope or an image (low-quality image) stored in a storage device (storage medium) 201, and generates a converted image based on the input image. The image conversion unit 1001 includes a learning model in which conversion parameters (weights and biases) and hyperparameters are adjusted, and performs image conversion using this learning model. The learning model can be configured, for example, by a neural network, which is a type of machine learning model, a convolutional neural network, or a neural network with another structure.

[0022] The transformation parameter update unit 1005 adjusts the transformation parameters using the hyperparameters acquired from the learning parameter update unit 1006, and supplies the adjusted transformation parameters to the image conversion unit 1001. Here, the hyperparameters are parameters that determine the learning behavior of the learning model. The hyperparameters include various parameters used in the calculation (learning calculation) to determine each coefficient of the learning model, such as a learning rate that determines the update amount of the transformation parameters, a batch size, the structure and number of layers of the learning model, etc.

[0023] The learning parameter update unit 1006 adjusts the hyperparameters based on the composite error described below, and supplies the adjusted hyperparameters to the conversion parameter update unit 1005. As will be described later, the conversion parameters and hyperparameters are adjusted by repeatedly updating each parameter to minimize index values ​​called errors (conversion errors and composite errors) so that the learning model outputs desired results.

[0024] The conversion error calculation unit 1002 calculates the conversion error between the converted image generated from the low-quality image by the image conversion unit 1001 and the high-quality image used as the teacher image. More specifically, the conversion error calculation unit 1002 calculates the mean absolute error, mean square error, etc. as the conversion error, calculated from the pixel values ​​of the converted image output by the learning model and the pixel values ​​of the corresponding high-quality image. The mean absolute error can be calculated using the following equation (1). The pixel value of coordinates (i, j) in the converted image is represented by x_i,j, and the pixel value of coordinates (i, j) in the high-quality image is represented by y_i,j. N represents the number of pixels in the converted image or the high-quality image. The mean square error can be calculated by squaring the absolute value of equation (1).

[0025]

[0026] The dimension metrics calculation unit 1003 calculates dimension metrics of the converted image generated from the low-quality image by the image conversion unit 1001. The dimension metrics include, for example, measurement accuracy calculated by performing arbitrary calculations on dimension values ​​obtained by measuring the dimensions of a semiconductor circuit pattern, and throughput from input of an input image generated by a scanning electron microscope to calculation of the measurement accuracy for the converted image. Here, the dimension values ​​are obtained by a predetermined measurement algorithm calculating the distance between edges, roughness value, etc. using two or more edge coordinates (x_edge, y_edge) calculated based on the brightness gradient on the image. The measurement algorithm includes measurement parameters that control the measurement method, such as smoothing processing and edge detection thresholds. On the other hand, the throughput is affected by the calculation processing time of the neural network, which changes depending on the structure and number of layers of the learning model adjusted by hyperparameters.

[0027] Measurement accuracy can be calculated using a measurement index that statistically represents the similarity between the dimension values ​​obtained from the converted image and the high-resolution image. Examples include the average error between the dimension values ​​obtained from the converted image and the dimension values ​​obtained from the high-resolution image, sensitivity, coefficient of determination, and reproducibility. Throughput can be expressed, for example, as the number of images that can be processed per unit time, from image input (or generation of input images using an image generation tool such as a scanning electron microscope) to calculation of dimension values. Dimension metrics calculated by statistical processing generally do not involve the output of a neural network, as with dimension values, and therefore cannot be differentiated with transformation parameters using the gradient descent method described below. Therefore, they cannot be applied to deep learning using neural networks, etc. On the other hand, algorithms for optimizing hyperparameters do not have any restrictions on whether they can be differentiated. Therefore, in this disclosure, dimension metrics are included only in the error (compound error) when optimizing hyperparameters.

[0028] Generally, the error described above is a conversion error using the brightness value of the converted image output by the learning model. However, when using the generated converted image for measurement, conventional techniques, such as Patent Document 1, do not use dimension metrics as indicators. As a result, it is not possible to generate a converted image with sufficiently improved dimension metrics. Furthermore, deep learning using neural networks and the like requires the assumption that the errors used in learning are differentiable, making it difficult to use dimension metrics corresponding to non-differentiable operations, such as dimension values, as errors. As a result, the target value of the dimension metrics desired by the user cannot be reached, making it difficult to generate a learning model that performs image conversion suitable for practical use. Therefore, the present disclosure uses the composite error described below to adjust (optimize) hyperparameters, thereby enabling the generation of a learning model that generates a converted image optimal for the dimension metrics.

[0029] The composite error calculation unit 1004 calculates a composite error by combining, using a weighted sum, the conversion error L_image calculated by the conversion error calculation unit 1002 and the dimension metrics L_measure calculated by the dimension metrics calculation unit 1003. For example, the composite error can be calculated based on an arithmetic expression such as Equation (2).

[0030] α×L_image+β×L_measure...(2)

[0031] Here, α and β are weighting coefficients expressed as arbitrary real numbers. For example, if the weighting coefficient β of the dimension metric L_measure is set to a value larger than the weighting coefficient α of the conversion error L_image, the dimension metric will have a larger error than the conversion error, and it is possible to encourage updating to hyperparameters that place more emphasis on minimizing the dimension metric.

[0032] Examples of measurement indices included in the dimension metrics include the average error between the dimension values ​​obtained from the converted image and the dimension values ​​obtained from the high-resolution image, sensitivity, coefficient of determination, and reproducibility. Dimension metrics may also be configured to include multiple of these optimization indices. Below is a method for calculating each optimization index for n sets of converted images and high-resolution images. The average error can be calculated by averaging the absolute error between the dimension value X1 obtained from the converted image and the dimension value X2 obtained from the high-resolution image over n sets. Sensitivity can be expressed by the slope of an approximate line drawn by plotting the n sets of X1 and X2 on the x-axis and y-axis, respectively. The coefficient of determination represents the degree of variation of the n sets of X1 and X2 relative to the approximate line used to calculate sensitivity, and can be calculated from the sum of squared deviations of X1 and X2 for the n sets. Reproducibility can be determined by taking multiple photographs of the same field of view using a scanning electron microscope, averaging the images, and repeating this process multiple times to set the reproducibility to three times the standard deviation (3σ) of the dimensional values ​​(images within the 3σ range are considered to have reproducibility).

[0033] When multiple optimization indices are included in the dimension metrics, the optimization indices can be weighted together to form the dimension metrics, for example, as shown in equation (3).

[0034] γ × (average error) + δ × |1 − sensitivity| + ε × (1 − coefficient of determination) − ζ × throughput ... (3)

[0035] Here, the coefficients γ, δ, ε, and ζ are weighting coefficients expressed as arbitrary real numbers. For example, by setting the sensitivity coefficient δ to a value larger than the coefficients γ, ε, and ζ, the error in the sensitivity term is more strongly reflected in the optimization index, making it possible to update the hyperparameters to those that place more emphasis on improving sensitivity.

[0036] Furthermore, in the above-described formula (3), it is also possible to configure a dimension metric that prompts a specific optimization index, the target value of which is input via the input device 101, to satisfy the target value. For example, as in formula (4), when a specific optimization index falls below the target value, the residual from the target value can be added to the dimension metric.

[0037] γ × f (mean error, target value) + δ × g (|1 - sensitivity|, target value) + ε × h (coefficient of determination, target value) + ζ × i (throughput, target value) ... (4)

[0038] Here, f, g, h, and i represent arbitrary functions. Functions f and g are configured, for example, to output 0 if the value of their first argument (mean error, |1 - sensitivity|) is below the target value, and to output the residual calculated by (first argument) - (target value) if the value exceeds the target value. Functions h and i are configured, for example, to output 0 if the value of their first argument (coefficient of determination, throughput) is above the target value, and to output the residual calculated by (target value) - (first argument) if the value is below the target value. Alternatively, functions h and i may be configured to take a power or a radical root of (first argument) - (target value) or (target value) - (first argument), or to determine whether the target value is above or below a certain margin, or to use constrained optimization to reject (not adopt) learning models and hyperparameters if the dimension metrics are above or below the target value.

[0039] These weighting coefficients γ, δ, ε, and ζ are parameters that can be adjusted (set) according to the user's desired priority of the dimension metrics. By inputting each weighting coefficient in the priority coefficient input area 3013 included in the GUI 301 (see FIG. 4 ), the user can reflect the priority of the desired improvement of the dimension metrics in the composite error. For example, if the user desires a stronger improvement in the coefficient of determination than other dimension metrics, the user can input a larger value in the text box related to the coefficient of determination in the priority coefficient input area 3013 than in the text box related to the other dimension metrics. This makes the value of the coefficient of determination term in Equation (3) or (4) relatively larger than the terms for the other dimension metrics, thereby optimizing the parameters to more strongly improve the coefficient of determination. This allows the user to control the improvement of each dimension metric according to the desired priority.

[0040] The default values ​​of the weighting coefficients may be set to 1, for example, and may be adjusted by the user as appropriate. Alternatively, the default values ​​may be determined in advance through preliminary experiments using predetermined sample images so as not to deviate from the user's sense (so as not to become empirically unreasonable values). Furthermore, the dimension metrics may include not only the optimization indexes described above, but also other optimization index values.

[0041] The learning parameter update unit 1006 adjusts (optimizes) hyperparameters to suppress the composite error based on the composite error calculated by the composite error calculation unit 1004, and supplies the adjusted hyperparameters to the conversion parameter update unit 1005. For example, the learning parameter update unit 1006 optimizes the hyperparameters by adopting hyperparameters arbitrarily selected by a user from a set of possible hyperparameter values. Alternatively, for example, the learning parameter update unit 1006 may optimize the hyperparameters by sequentially selecting hyperparameters using a parameter search algorithm such as grid search, random search, or Bayesian optimization, and searching for a combination of hyperparameters that minimizes or maximizes a predetermined optimization index. The hyperparameters are set based on the above-described method. If the parameter optimization termination condition is not met after learning is completed, the hyperparameters are sequentially updated and set as parameters that determine the learning behavior when learning the learning model again. Adjusting (optimizing) the hyperparameters changes the performance and calculation processing time of image conversion by the image conversion model. Therefore, setting appropriate parameters and performing learning can improve dimension metrics. For example, when adjusting the number of layers of a learning model as a hyperparameter, a small number of layers shortens the calculation processing time but generates images of low quality. On the other hand, a large number of layers enables the generation of images that are faithful to the teacher image but increases the calculation processing time. The transformation parameter update unit 1005 adjusts (optimizes) the transformation parameters to suppress the transformation error based on the transformation error calculated by the transformation error calculation unit 1003, and supplies the adjusted parameters to the image conversion unit 1001. The transformation parameters are optimized by using, for example, a method called gradient descent, which involves partially differentiating a given error with the transformation parameters and sequentially updating them. In the above calculations, the transformation parameter update unit 1005 uses the hyperparameters calculated by the learning parameter update unit 1006.

[0042] 2 is a flowchart for explaining the details of the parameter optimization process (generation of an optimized trained model and generation of dimension metrics). The main actors in the operations at each step are the processing units (image conversion unit 1001, conversion parameter update unit 1005, learning parameter update unit 1006, etc.). However, the functions of these processing units are realized by loading a program into the internal memory of the processor, and therefore the operations at each step can be interpreted as operations mainly performed by the arithmetic device 100.

[0043] (i) Step S201 The image conversion unit 1001 acquires a pair of a low-quality image and a high-quality image stored in the storage device (storage medium) 201. Alternatively, the image conversion unit 1001 may acquire a pair of a low-quality image and a high-quality image captured using a scanning electron microscope.

[0044] (ii) Step S202 The learning parameter update unit 1006 receives from the input device 101 the target values ​​of the dimension metrics input by the user using the GUI screen as shown in FIG.

[0045] (iii) Step S203: The learning parameter update unit 1006 sets the hyperparameters of the learning model using a predetermined method. Examples of the method for setting the hyperparameters include a method in which the user arbitrarily selects from a set of possible hyperparameter values, as described above, or a method in which the hyperparameters are determined using a parameter search algorithm such as grid search, random search, or Bayesian optimization, or another search algorithm.

[0046] (iv) Step S204: The conversion parameter update unit 1005 performs a learning calculation using the hyperparameters set in step S203 to optimize the conversion parameters of the learning model (generate a learned model). Details of step S204 will be described later with reference to FIG. 3.

[0047] (v) Step S205 The image conversion unit 1001 applies the low-quality image to the trained model trained in step S204 to generate a converted image.

[0048] (vi) Step S206: The conversion error calculation unit 1002 calculates the conversion error using the high-quality image paired with the low-quality image and the converted image generated in step S205. The dimension metrics calculation unit 1003 calculates the dimension metrics. The dimension metrics can be calculated, for example, by performing image processing (e.g., edge extraction processing) on ​​the converted image generated in step S205 and performing statistical processing using measurement values ​​(multiple measurement values ​​obtained from multiple converted images) of a predetermined pattern (e.g., the width of a circuit pattern) in the image and throughput. The composite error calculation unit 1004 then calculates the composite error from the conversion error and the dimension metrics (see equation (2)).

[0049] (vii) Step S207: The learning parameter update unit 1006 determines whether the optimization termination condition is satisfied. Here, whether the optimization termination condition is satisfied may be determined, for example, by whether the dimension metrics (indicator values ​​obtained by statistical processing) calculated in step S206 satisfy the target values ​​set by the user in step S202, or by whether a predetermined number of optimizations have been performed.

[0050] If it is determined in step S207 that the optimization conditions are not satisfied (NO in step S207: for example, if the dimension metrics do not satisfy the target values), the process proceeds to step S208. That is, the hyperparameter values ​​and dimension metrics at that time are fed back as criteria for selecting hyperparameters. On the other hand, if the optimization termination conditions are satisfied (YES in step S208), the process proceeds to step S209.

[0051] (viii) Step S208: The learning parameter update unit 1006 updates the hyperparameters based on the combined error. After updating the hyperparameters, the process returns to step S203.

[0052] As in the method described in step S203, the learning parameter update unit 1006 may use a hyperparameter arbitrarily selected by the user from a set of possible hyperparameter values ​​as the update parameter, or may determine the update parameter using a parameter search algorithm such as grid search, random search, or Bayesian optimization or other search algorithm.

[0053] (ix) Step S209 The calculation device 100 outputs the optimized trained model and the dimension metrics to the output device 102 using a GUI such as that shown in FIG.

[0054] The above-described processing can optimize the parameters of the learning model. However, even when the optimization processing by the parameter optimization system 10 is performed, the target values ​​of the dimension metrics input by the user via the input device 101 may not be met. In this case, for example, the user may change the generation conditions (such as filter processing) of the image generation tool and re-execute the parameter optimization processing using an input image with better image quality (resulting in a slight sacrifice of throughput). For example, if the image generation tool is a scanning electron microscope, the generation conditions, such as the number of low-quality images averaged, filter processing, and autofocus processing, may be changed. This generates an input image with better image quality, thereby generating a converted image closer to a high-quality image, increasing the likelihood of improving the dimension metrics. Note that if the image quality of the input image is changed (changed to high image quality) in this way, when a converted image is actually generated using the trained model and measurement processing is performed, a low-quality image having the changed image quality level is input to the image conversion unit 1001 to generate a converted image to be measured.

[0055] Furthermore, in the first embodiment, the dimension metrics are calculated from the dimension values ​​obtained by measuring the dimensions of the semiconductor circuit pattern, but the dimension values ​​may also be, for example, overlay measurement values, which are the deviation of the center of gravity between layers of a multi-layer semiconductor, or dimension values ​​measuring any other shape within an image.

[0056] <Details of Conversion Parameter Optimization Process (Step S204)> FIG. 3 is a flowchart for explaining details of the conversion parameter optimization process (step S204) in the parameter optimization process (FIG. 2).

[0057] (i) Step S2041: The transformation parameter update unit 1005 acquires transformation parameters (initial values) set by the user. When setting the initial values ​​of the transformation parameters of the learning model, the user can set, for example, the initial values ​​of He, Xavier, or initial values ​​obtained using any other initialization method.

[0058] (ii) Step S2042 The image conversion unit 1001 inputs the low-quality image for learning into the learning model to generate a converted image.

[0059] (iii) Step S2043: The conversion error calculation unit 1002 calculates a conversion error using the high-quality image paired with the low-quality image for training and the converted image generated in step S2042. Examples of the conversion error include the mean absolute error (see formula (1)) and the mean square error.

[0060] (iv) Step S2044: The conversion parameter update unit 1005 updates (adjusts) the conversion parameters so as to suppress the conversion error based on the conversion error calculated in step S2043 and the hyperparameters set in step S203. As described above, the conversion parameters can be updated using, for example, the gradient descent method.

[0061] (v) Step S2045: The conversion parameter update unit 1005 determines whether a learning termination condition is met. The learning termination condition may be set based on whether the conversion error calculated in step S2043 has reached a predetermined number of learning iterations, or may be determined based on whether the transition of the conversion error during learning (learning curve) has reached its minimum value within a predetermined number of learning iterations.

[0062] If the learning termination condition is met (YES in step S2045), the learning process in S204 ends. On the other hand, if the learning termination condition is not met (NO in step S2045), the process returns to step S2042.

[0063] 4 is a diagram showing an example of the configuration of an optimization condition setting GUI (Graphical User Interface) 301 that is displayed on the display screen of the output device 102 and that is used to input optimization conditions via the input device 101. This GUI screen makes it possible to set optimization conditions that serve as the basis for optimization in the arithmetic device 100 of the parameter optimization system 10.

[0064] The optimization condition setting GUI 301 is composed of an optimization index selection area 3011, an optimization index input area 3012, and a priority coefficient input area 3013. The user can select a desired optimization index by checking the checkbox of the dimension metric that the user wants to use as the optimization index in the optimization index selection area 3011.

[0065] The user can set a target value for the optimization index by inputting the desired target value in the optimization index input field 3012. Furthermore, the user can adjust the strength of improvement of the dimension metric according to the value of the priority coefficient by inputting a priority coefficient (for example, the larger the value, the higher the priority) that indicates the priority of the dimension metric that the user wants to improve in the priority coefficient input field 3013.

[0066] The selected optimization index, the set target value of each optimization index, and the priority coefficient are transmitted to the calculation device 100, whereby the calculation device 100 can calculate the dimension metrics. Note that the dimension metrics shown in Fig. 4 are merely examples, and are not limited to the indices shown in Fig. 4 (average error, sensitivity, and coefficient of determination). Dimension metrics other than those shown in Fig. 4 may be added, and corresponding check boxes and text boxes may be added.

[0067] 5 is a diagram showing an example of the configuration of an optimization result output GUI 401 displayed on the output device 102. This GUI screen makes it possible to present to the user the optimization results obtained by the parameter optimization process of the arithmetic device 100 of the parameter optimization system 10.

[0068] The optimization result output GUI 401 is composed of a target achievement display area 4011, an optimal parameter display area 4012, and a parameter contribution rate display area 4013. The target achievement display area 4011 displays the target achievement rate of the optimal dimension metric output by the parameter optimization system 10 for the optimization index and its target value input in the optimization index input area 3012. For example, a dimension metric that has reached the target value indicates a target achievement rate of 100%. A dimension metric that has not reached the target value indicates the target achievement rate at that time. If the target value has not been reached, this means that even if parameter optimization was performed, it was difficult to improve the image quality of the given input image to a level that would achieve the target value. In this case, the user can adjust the imaging parameters of the scanning electron microscope, re-acquire a lower-quality image with better image quality, and perform parameter optimization again to try to achieve the target value.

[0069] The optimal parameter display area 4012 displays the hyperparameters optimized by the parameter optimization system 10. In addition, the parameter contribution rate display area 4013 displays how much the parameters targeted by the optimization index input in the optimization index input area 3012 contributed to the improvement of the combined error. This allows the user to know the parameters that contribute to the improvement of the dimension metrics, making it easier to determine whether to add or remove parameters targeted for parameter optimization.

[0070] <Modifications> (Modification 1) A parameter optimization system 10 according to Modification 1 of the first embodiment includes a loop for optimizing measurement parameters in addition to the components shown in Fig. 1. Modification 1 enables a user to optimize parameters of a learning model without manually adjusting measurement parameters so that the measurement parameters do not hinder improvement of dimension metrics.

[0071] (i) Modification of Optimization Condition Setting GUI Figure 6 is a diagram showing an example of the configuration of an optimization condition setting GUI 301' according to Modification 1. The optimization condition GUI 301' is configured to allow the user to select whether or not to include measurement parameters. The same components as those in Figure 4 are assigned the same reference numerals in Figure 6, and therefore, redundant explanations will be omitted below.

[0072] The optimization condition setting GUI 301′ includes a measurement parameter selection area 5001 as an additional configuration item. In the measurement parameter selection area 5001, the user can select measurement parameters to be included in the update targets (metrics) by checking the checkboxes of the measurement parameters that the user wants to optimize. In FIG. 6 , checking the “Smoothing” checkbox includes parameters for the smoothing process performed on the converted image during pattern measurement as optimization targets. Furthermore, checking the “Edge Threshold” checkbox includes a threshold for determining edges in the converted image during pattern measurement as optimization targets. Note that the measurement parameters shown in FIG. 6 are merely examples and are not limited to the measurement parameters (smoothing, edge threshold) shown in FIG. 6 . Measurement parameters other than those shown in FIG. 6 may be added, and corresponding checkboxes may be added.

[0073] (ii) Example of the configuration of the parameter optimization system 10 Fig. 7 is a diagram showing an example of the configuration of the parameter optimization system 10 according to the modified example 1. The parameter optimization system 10 according to the modified example includes a path for feeding back measurement parameters from the learning parameter update unit 1006 to the dimension metrics calculation unit 1003. The learning parameter update unit 1006 adds the measurement parameters for calculating the dimension values ​​to some of the hyperparameters, and includes them in the update target (optimization target).

[0074] When a measurement algorithm fails to calculate dimension values ​​(or produces inappropriate dimension values) without including the measurement parameter in the optimization target, including the measurement parameter in the optimization target (by checking the measurement parameter checkbox) can clarify difficult-to-see areas (e.g., emphasize edges), making it possible to measure the dimension values. Including measurement parameters in the optimization target in this way is effective when a measurement algorithm can hinder the improvement of dimension metrics. For example, to measure an image containing noise and making edges difficult to see, the image may be smoothed during measurement. If the measurement parameter that determines the degree of smoothing is small, the edges obtained from the image may vary, making it impossible to obtain appropriate dimension values. Variation in the edges results in large dimension values ​​or makes them unmeasurable, resulting in poor measurement accuracy. Therefore, parameter optimization can improve measurement accuracy and dimension metrics by adjusting the measurement parameters to improve measurement accuracy.

[0075] (Variation 2) According to Variation 2, the conversion parameter update unit 1005 (see FIG. 1 ) may perform data augmentation on one or more of the low-quality images and the high-quality images. The data augmentation is image processing that adds Gaussian noise, performs smoothing, or performs scaling using random parameters, and increases the variation between the low-quality images and the high-quality images, thereby improving the robustness of learning between the low-quality images and the high-quality images. In this case, the learning parameter update unit 1006 can add and adjust at least one of the parameters related to this data augmentation, i.e., the type of image processing to be performed, the number of image processing operations, or the maximum strength of the random parameters of the image processing (e.g., in the case of adding Gaussian noise, the maximum standard deviation of the noise to be added), as part of the hyperparameters.

[0076] According to the second modification, the dimension metrics shown in the above formula (3) or (4) can be easily improved, making it easier for the user to obtain a learning model that achieves the target values ​​of the desired dimension metrics.

[0077] (Variation 3) According to Variation 3, if the dimension metrics are differentiable, it is possible to adjust (optimize) the transformation parameters using the composite error. Therefore, the transformation parameter update unit 1005 may adjust the transformation parameters using the composite error output by the composite error calculation unit 1004. In this case, in the parameter optimization system 10 according to this variation, the path for transmitting the transformation error from the transformation error calculation unit 1002 to the transformation parameter update unit 1005 in FIG. 1 is replaced with a path for transmitting the composite error from the composite error calculation unit 1004 to the transformation parameter update unit 1005. In addition, in FIG. 3, the transformation parameter optimization process (step S204) is configured by including the following steps S2046 and S2047 between steps S2043 and S2044 in the parameter optimization process according to the first embodiment (see FIG. 2).

[0078] (i) Step S2046: The dimension metrics calculation unit 1003 calculates dimension metrics using the high-quality image paired with the low-quality image for learning and the converted image generated in step S2042. The dimension metrics calculation method can be the same as the method described in step S206.

[0079] (ii) Step S2047: The composite error calculation unit 1004 calculates a composite error using the conversion error calculated in step S2043 and the dimension metrics calculated in step S2046. According to Modification 3, the conversion parameters can also be adjusted to minimize the dimension metrics without adjusting (optimizing) the hyperparameters. Therefore, by improving the dimension metrics shown in the above formula (3) or (4), the user can obtain a learning model that achieves the desired target value of the dimension metrics. While Modification 3 can omit the adjustment (optimization) of the hyperparameters, the adjustment (optimization) of the hyperparameters makes it possible to optimize the conversion parameters using excellent hyperparameters, making it easier to achieve the desired target value of the dimension metrics than when only optimizing the conversion parameters.

[0080] <Technical Effects of the First Embodiment> As described above, Patent Document 1 discloses a method for generating a converted image or directly predicting dimensional values ​​that focuses on improving the image quality of a measurement area through a learning model. Meanwhile, according to the first embodiment, a converted image can be generated based on a compound error that includes an image conversion error and a user-desired dimensional metric (through statistical processing). Furthermore, by configuring the dimensional metric with elements desired by the user, it is possible to improve the user-desired dimensional metric and generate a converted image that improves the dimensional metric.

[0081] (2) Second Embodiment In addition to the features of the first embodiment, the second embodiment relates to a process in which image generation parameters are included in the optimization targets and the optimized image generation parameters are fed back to an image generation unit (e.g., an SEM device).

[0082] <Configuration Example of Parameter Optimization System> Fig. 8 is a diagram showing a schematic configuration example of a learning model parameter optimization system 10 according to the second embodiment. The same components as those in Fig. 1 are denoted by the same reference numerals in Fig. 8, and therefore, redundant explanations will be omitted below.

[0083] The parameter optimization system 10 includes an image generation unit 7001 that generates images to be input to a learning model using a scanning electron microscope instead of the storage device 201 (see FIG. 1 ) that stores SEM images. The parameter optimization system 10 is configured to include image generation parameters of the scanning electron microscope (image generation unit 7001) as targets for parameter optimization, and to feed back the image generation parameters from the learning parameter update unit 1006 to the image generation unit 7001. Note that the image generation unit 7001 may be a storage device that stores images generated in advance using a scanning electron microscope with each image generation parameter and the image generation parameters, and outputs the stored images.

[0084] Furthermore, the parameter optimization system 10 according to the second embodiment controls the image generation parameters of the input image to improve compound errors, and controls the image quality and image generation time of the low-quality image input to the learning model, thereby improving the image quality of the converted image and expanding the control range of the dimension metrics. For example, if the image quality of the low-quality image input to the learning model is improved, it becomes easier to generate a converted image close to a high-quality image, and improvements in image quality and measurement indexes can be expected. Furthermore, because the image generation parameters of the input image are controlled, the generation time of the input image also affects throughput. Therefore, if the generation time of the input image is shortened, improvements in throughput can be expected.

[0085] <Examples of Image Generation Parameters to be Optimized> Examples of image generation parameters that can be optimized in the image generation unit 7001 include the number of low-quality images averaged, the scanning speed (high-speed scanning or low-speed scanning), filter processing (smoothing filter processing or edge filter processing), and autofocus processing (image capture processing using a default value when autofocus adjustment is not performed). For example, increasing the number of low-quality images averaged, slowing down the scanning speed, or applying filter processing or autofocus processing can generate low-quality images with better image quality, but the time required to generate the image increases, thereby reducing throughput. On the other hand, decreasing the number of low-quality images averaged, speeding up the scanning speed, or not applying filter processing or autofocus processing can generate low-quality images with higher throughput, but the image quality of the low-quality image deteriorates, and the image quality and dimensional metrics of the converted image also deteriorate.

[0086] In this way, there is a trade-off between measurement indices (metric measurement indices: for example, mean error, sensitivity, coefficient of determination, reproducibility, etc.) and image quality and throughput. When target values ​​for both the measurement indices and throughput are input via the input device 101, the calculation device 100 in the parameter optimization system 10 performs optimization processing to output image generation parameters, conversion parameters, hyperparameters, and measurement parameters that are appropriately balanced so that each dimension metric (both the measurement indices and throughput) meets the target values.

[0087] Generally, there are a huge number of combinations of image generation parameters, transformation parameters, hyperparameters, and measurement parameters that can achieve measurement indices and throughput. Therefore, it is very difficult for a user to predict a desired combination (appropriate combination) in advance. However, by performing a parameter optimization process using the parameter optimization system 10 according to this embodiment, it is possible to derive appropriate combinations of various parameters.

[0088] <Processing after updating image generation parameters> If the image generation unit 7001 is configured as a scanning electron microscope, the scanning electron microscope acquires the image generation parameters fed back from the learning parameter update unit 1006, changes the shooting conditions according to the image generation parameters, and then re-captures the image. However, this is not limited to this, and a computer (not shown) may store image sets captured under multiple shooting conditions in a storage device (not shown) in advance, select an image set that matches the fed back image generation parameters, and input it to the learning model. Alternatively, if there is no image set that matches the fed back image generation parameters, the computer may instruct the scanning electron microscope to re-capture the image.

[0089] <Technical Effect of the Second Embodiment> According to the second embodiment, the user can optimize the parameters of a learning model without manually adjusting image generation parameters that will improve the dimension metrics. Furthermore, in the second embodiment, image generation parameters can also be included in the parameter optimization targets, and therefore the image generation parameters can also be displayed in the parameter contribution rate display area 4013. Therefore, an additional effect of the parameter contribution rate display area 4013 is that the user can know the image generation parameters that contribute to improving the dimension metrics, making it easier to adjust the image generation parameters that will improve the dimension metrics.

[0090] In the second embodiment, a scanning electron microscope is used as the image generation tool, but this is not limited to this. For example, an image generation tool such as an optical microscope, in which the image quality of the captured image can be controlled by image generation parameters, may also be used.

[0091] (3) Third Embodiment The third embodiment relates to a process of optimizing pre-processing and post-processing parameters of a learning model that generates a converted image from a low-quality image using a compound error instead of a hyperparameter.

[0092] <Configuration Example of Parameter Optimization System> FIG. 9 is a diagram showing a schematic configuration example of a parameter optimization system 10 according to a third embodiment. The parameter optimization system 10 is a system for optimizing pre-processing or post-processing parameters when inferring a transformed image using a trained learning model. In FIG. 9, the same components as those in FIG. 8 are assigned the same reference numerals, and therefore, redundant description will be omitted below. As shown in FIG. 9, the parameter optimization system 10 according to the third embodiment includes an inference parameter update unit 9001 that updates post-processing and pre-processing parameters applied to a low-quality image input to a learning model during inference or a transformed image to be inferred, instead of the transformation parameter update unit 1005 and the learning parameter update unit 1006 shown in FIG. 8.

[0093] <Contents of Post-Processing and Pre-Processing> (i) Post-processing refers to image processing performed on the converted image (processing performed after image conversion). Examples of this image processing include, but are not limited to, arithmetic processing such as Gaussian noise addition, smoothing, and brightness normalization, as well as synthesis processing with a second image different from the converted image. Examples of the second image different from the converted image include, but are not limited to, a high-quality image generated by simulation (e.g., an image created by a predetermined simulation from a design drawing of a semiconductor circuit pattern), a design drawing with noise superimposed, or a low-quality image. The second image can be input to the image conversion unit 1001 from the input device 101 or acquired by the image conversion unit 1001 from a storage device (not shown).

[0094] The inference parameter update unit 9001 calculates parameters related to the synthesis of a transformed image with a second image different from the transformed image, such as the ratio at which the two images are mixed (image mixing ratio) when the two images are mixed. This post-processing has the effect of changing the image quality of the transformed image (such as the amount of noise in the transformed image and the degree of smoothness of the transformed image). Because this post-processing changes the dimension values ​​calculated from the transformed image, optimizing the post-processing parameters makes it possible to improve the dimension metrics calculated from the dimension values.

[0095] (ii) Preprocessing refers to image processing performed on a low-quality image (processing performed before image conversion). Specifically, image processing performed on a low-quality image can include, similar to postprocessing, arithmetic operations such as Gaussian noise addition, smoothing, and brightness normalization, as well as synthesis with a second image different from the low-quality image. This preprocessing has the effect of changing the image quality of the low-quality image. This preprocessing changes the image quality and dimensional values ​​of the converted image calculated from the preprocessed low-quality image, so optimizing the preprocessing parameters can improve the dimensional metrics calculated from the dimensional values.

[0096] <Details of Parameter Optimization Process> Fig. 10 is a flowchart for explaining details of the parameter optimization process according to the third embodiment. In the third embodiment, inference of a converted image is performed using a trained learning model. Note that the same processes as those in Fig. 2 are assigned the same reference numerals in Fig. 10, and therefore, redundant explanations will be omitted below.

[0097] In the third embodiment, the arithmetic device 100 optimizes parameters related to the image processing, i.e., the type and number of image processing operations to be performed, and parameters required for the image processing (in the case of combining with a noisy image, the combining rate for combining the noisy image), as hyperparameters to suppress the compound error expressed by the above-mentioned formula (3) or (4). The parameter optimization process is composed of the same steps as the parameter optimization process according to the first embodiment (see FIG. 2 ), except for step S204. However, step S203 “setting hyperparameters” is replaced with step S1001 “setting pre-processing and post-processing parameters.” Furthermore, step S208 “updating hyperparameters” is replaced with step S1002 “updating pre-processing and post-processing parameters.” While hyperparameters are parameters that determine the learning behavior as described above, pre-processing and post-processing parameters are parameters used when inferring a converted image using a learning model for image conversion.

[0098] (i) Step S1001: The inference parameter update unit 9001 sets pre-processing and post-processing parameters (e.g., parameters for the image mixing ratio) of the learning model using a predetermined method. Examples of methods for setting the pre-processing and post-processing parameters include, as described above, a method in which the user arbitrarily selects from a set of possible pre-processing and post-processing parameter values, or a method in which the user determines the pre-processing and post-processing parameters using a parameter search algorithm such as grid search, random search, or Bayesian optimization, or another search algorithm.

[0099] (ii) Step S1002: The inference parameter update unit 9001 updates the pre-processing and post-processing parameters (parameters for the image blending ratio) based on the combined error (used as an index for optimization). After updating the pre-processing and post-processing parameters, the process returns to step S1001.

[0100] As with the method described in step S1001, the inference parameter update unit 9001 may use parameters arbitrarily selected by the user from a set of possible pre-processing and post-processing parameter values ​​as update parameters, or may determine update parameters using parameter search algorithms such as grid search, random search, or Bayesian optimization or other search algorithms.

[0101] <Technical Effects of the Third Embodiment> According to the third embodiment, it is possible to optimize pre-processing and post-processing parameters used during transformed image inference. Furthermore, it is possible to obtain a transformed image that further improves the dimension metrics of the transformed image inferred by the learning model. This makes it easier for the user to achieve the target values ​​of the desired dimension metrics.

[0102] (4) Summary (i) In the present embodiment (first to third embodiments), the arithmetic device (processor or computer) 100 executes a parameter optimization program to perform the following operations: calculating a conversion error between a converted image generated by applying an input image (low-quality image) to a machine learning model and a high-quality image (teacher image) input during training of the learning model; calculating dimension metrics of a pattern included in the teacher image based on the converted image and the teacher image; calculating a combined error based on the conversion error and the dimension metrics; training a machine learning model that generates the converted image so as to optimize the combined error; and optimizing parameters of the machine learning model. In the first embodiment, the parameter optimization process targets hyperparameters that define the behavior of machine learning and conversion parameters of the machine learning model that performs image conversion. In the second embodiment, the parameter optimization process targets image generation parameters in the image generation unit (image generation tool) 7001 in addition to the hyperparameters that define the behavior of machine learning and the conversion parameters of the machine learning model that performs image conversion. Furthermore, in the third embodiment, instead of the conversion parameters and hyperparameters of the first embodiment, the parameter optimization process targets pre-processing and post-processing parameters that are executed when inferring the converted image. In this way, it becomes possible to improve the dimension metrics desired by the user and generate a transformed image that improves the dimension metrics.

[0103] (ii) In this embodiment, the dimension metrics (dimension metrics of the converted image) include the measurement accuracy calculated using a measurement index that statistically represents the similarity between the dimension values ​​obtained from the converted image and the teacher image, and the throughput from the input of the input image (low-quality image) to the calculation of the dimension values ​​using the converted image. The dimension metrics are expressed as functions with the measurement accuracy and throughput as variables.

[0104] (iii) In the parameter optimization process, the arithmetic device 100 determines that the parameters have been optimized when the dimension metrics (expressed by Equation (3) or (4)) satisfy an optimization termination condition based on a target value (a value set by the user) (e.g., when the dimension metrics reach the target value or when a predetermined number of optimization processes have been performed) (see step S207). By repeatedly performing the optimization process, the dimension metrics can be optimized with a high probability. Note that if all of the set indicators do not reach their target values ​​even after the optimization process has been performed a predetermined number of times, an input image (low-quality image) with better image quality than the original image may be re-input and the optimization process may be performed again. Note that the arithmetic device 100 may output to the output device 102 a target attainment level indicating whether the dimension metrics have reached the set target value as a result of the optimization process.

[0105] Furthermore, the arithmetic device 100 alternately repeats learning of the machine learning model and parameter optimization (optimization of hyperparameters and conversion parameters, optimization of hyperparameters, conversion parameters, and image generation parameters, or optimization of pre-processing and post-processing parameters) until the combined error (see equation (2)) is optimized. The arithmetic device 100 adjusts the hyperparameters using a predetermined parameter search algorithm to optimize the combined error. As shown in equation (2), the combined error is expressed by multiplying the conversion error and the dimension metric by weighting coefficients (α and β), respectively, and then adding them together. By setting the weighting coefficient (β) for the dimension metric to be larger than the weighting coefficient (α) for the conversion error, it becomes possible to update the conversion parameters, hyperparameters, image generation parameters, and pre-processing and post-processing parameters to those that prioritize minimizing the dimension metric.

[0106] (iv) In this embodiment, the type of image processing, the number of image processing steps, or the maximum strength of image processing steps to be applied to at least one of the input image (low-quality image) and the teacher image (high-quality image) can be included in the hyperparameters that define the behavior of the machine learning. This allows these indicators to be optimized, thereby enabling the dimensional metrics to be positively improved. Therefore, users can obtain a machine learning model that achieves the target value of the desired dimensional metrics.

[0107] (v) According to the second embodiment, the parameter optimization system 10 may include an image generation tool (e.g., a scanning electron microscope) that generates an input image (low-quality image) instead of the storage device 201 that stores an image (SEM image). In this case, the calculation device 100 optimizes the combined error, thereby optimizing the image generation parameters used by the image generation tool to generate the input image. This makes it possible to build a machine learning model that generates an appropriate converted image while generating a low-quality image in real time using the image generation tool.

[0108] (vi) According to the third embodiment, as described above, the arithmetic device 100 optimizes parameters related to the inference of a transformed image using a machine learning model. Specifically, the arithmetic device 100 optimizes at least one of parameters of pre-processing before generating a transformed image or parameters of post-processing after generating a transformed image as parameters related to the inference of a transformed image. Specifically, the pre-processing is image processing performed on an input image (low-quality image) and includes one of Gaussian noise addition, smoothing, brightness normalization, or a process of combining the input image with an image different from the input image (a simulation image based on a blueprint). Furthermore, the post-processing is image processing performed on the transformed image and includes one of Gaussian noise addition, smoothing, brightness normalization, or a process of combining the transformed image with an image different from the transformed image (a simulation image based on a blueprint). This improves the dimensional metrics of the transformed image inferred by the machine learning model, allowing users to easily achieve their desired target values ​​for the dimensional metrics.

[0109] (vii) The functions of each embodiment can also be realized by software program code. In this case, a storage medium on which the program code is recorded is provided to a system or device, and the computer (or CPU or MPU) of the system or device reads the program code stored in the storage medium. In this case, the program code read from the storage medium itself realizes the functions of the above-mentioned embodiments, and the program code itself and the storage medium on which it is stored constitute the present disclosure. Examples of storage media for providing such program code include flexible disks, CD-ROMs, DVD-ROMs, hard disks, optical disks, magneto-optical disks, CD-Rs, magnetic tape, non-volatile memory cards, and ROMs.

[0110] Furthermore, an operating system (OS) or the like running on a computer may perform some or all of the actual processing based on instructions in the program code, thereby realizing the functions of the above-described embodiments.Furthermore, after the program code is read from a storage medium and written to a memory on the computer, a CPU (processor) or the like of the computer may perform some or all of the actual processing based on instructions in the program code, thereby realizing the functions of the above-described embodiments.

[0111] Furthermore, the program code of the software that realizes the functions of each embodiment may be distributed via a network and stored in a storage means such as a hard disk or memory of the system or device, or in a storage medium such as a CD-RW or CD-R, so that when used, the computer (or CPU or MPU) of the system or device reads and executes the program code stored in the storage means or storage medium.

[0112] The processes and techniques described herein are not inherently related to any specific device and can be implemented by a combination of components. Various types of general-purpose devices can also be added. A dedicated device may be constructed to perform the functions of each embodiment. Various functions can also be formed by appropriately combining multiple components disclosed in each embodiment. For example, some components may be omitted from all the components shown in each embodiment, or components from different embodiments may be appropriately combined.

[0113] Although specific embodiments are described in this disclosure, they are in all respects for the purpose of explanation (understanding the technology of the present disclosure) and not for the purpose of limitation. Those skilled in the art will recognize that there are many combinations of hardware, software, and firmware suitable for implementing the technology of the present disclosure. For example, the software described can be implemented in a wide variety of programming or scripting languages, such as assembler, C / C++, Perl, Shell, PHP, Java (registered trademark), etc.

[0114] Furthermore, in the above-described embodiment, the control lines and information lines are those that are considered necessary for the explanation, and not all control lines and information lines in the product are necessarily shown. All components may be interconnected.

[0115] In addition, other implementations of the present disclosure will become apparent to those skilled in the art from consideration of the embodiments. The specification and examples are exemplary only, with the scope and spirit of the technology of the present disclosure being indicated by the following claims.

[0116] 10 Parameter optimization system 100 Arithmetic device 101 Input device 102 Output device 201 Storage device 1001 Image conversion unit 1002 Conversion error calculation unit 1003 Dimension metrics calculation unit 1004 Composite error calculation unit 1005 Conversion parameter update unit 1006 Learning parameter update unit 7001 Image generation unit 9001 Inference parameter update unit

Claims

1. A parameter optimization system that optimizes parameters for generating a transformed image based on an input image, comprising: a storage device that holds a parameter optimization program; and a processor that reads and executes the parameter optimization program from the storage device, wherein the processor executes the following processes: calculating a transformation error between the transformed image and a teacher image input during learning; calculating dimensional metrics of a pattern contained in the teacher image based on the transformed image and the teacher image; calculating a combined error based on the transformation error and the dimensional metrics; and training a machine learning model that generates the transformed image so as to optimize the combined error, and optimizing parameters of the machine learning model.

2. A parameter optimization system as described in claim 1, wherein the processor includes in the dimension metrics the measurement accuracy calculated using a measurement index that statistically represents the similarity between the dimension values ​​obtained from the transformed image and the teacher image, and the throughput from the input of the input image to the calculation of the dimension values ​​using the transformed image.

3. The parameter optimization system according to claim 1, wherein the processor determines that the parameter is optimized when the dimension metric satisfies an optimization termination condition based on a set target value.

4. A parameter optimization system according to claim 1, wherein the processor alternately repeats training of the machine learning model and optimization of the parameters until the combined error is optimized.

5. A parameter optimization system according to claim 4, wherein the processor optimizes the composite error by adjusting hyperparameters that define the learning behavior of the machine learning model using a predetermined parameter search algorithm.

6. The parameter optimization system according to claim 1, wherein the processor trains the machine learning model using a gradient descent method.

7. The parameter optimization system according to claim 1, wherein the machine learning model includes at least a model using a neural network.

8. A parameter optimization system according to claim 1, wherein the processor optimizes hyperparameters that define the learning behavior of the machine learning model and parameters that define the dimension metrics.

9. A parameter optimization system as described in claim 8, wherein the processor includes in the hyperparameters the type of image processing, the number of image processing steps, or the maximum strength of the image processing steps to be applied to at least one of the input image or the teacher image, and optimizes the hyperparameters.

10. The parameter optimization system according to claim 1, wherein the composite error is expressed by the following formula: Composite Error = α x L_image + β x L_measure, where L_image is a transformation error, L_measure is a dimension metric, and α and β are weighting coefficients.

11. A parameter optimization system according to claim 2, wherein the dimension metric is expressed by the following formula: Dimension Metric = γ × L_bias + δ × |1-S| + ε × (1-R) ​​- ζ × T, where L_bias is the average error of the dimension value, S is the sensitivity, R is the coefficient of determination, T is the throughput, and γ, δ, ε, and ζ are weighting coefficients.

12. A parameter optimization system according to claim 3, wherein the processor outputs a target attainment level indicating whether or not the dimension metric has reached the set target value as a result of the optimization process.

13. The parameter optimization system of claim 1, wherein the processor optimizes parameters that define the dimensional metrics.

14. A parameter optimization system according to claim 1, further comprising an image generation tool that generates the input image, wherein the processor optimizes image generation parameters when the image generation tool generates the input image.

15. A parameter optimization system according to claim 1, further comprising an image generation tool that generates the input image, wherein the processor includes in the dimension metric a throughput from generation of the input image by the image generation tool to calculation of measurement accuracy for the converted image.

16. A parameter optimization system according to claim 1, wherein the processor optimizes parameters related to inference of the transformed image using the machine learning model.

17. A parameter optimization system as described in claim 16, wherein the processor optimizes at least one of parameters of pre-processing before generating the transformed image or parameters of post-processing after generating the transformed image as parameters related to the inference of the transformed image.

18. A parameter optimization process according to claim 17, wherein the pre-processing is image processing applied to the input image, and includes any one of Gaussian noise addition processing, smoothing processing, brightness normalization processing, or processing of synthesizing the input image with an image different from the input image, and the post-processing is image processing applied to the transformed image, and includes any one of Gaussian noise addition processing, smoothing processing, brightness normalization processing, or processing of synthesizing the transformed image with an image different from the transformed image, for the transformed image.

19. A parameter optimization method for optimizing parameters for generating a transformed image based on an input image, comprising: a processor that reads and executes a parameter optimization program from a storage device calculates a transformation error between the transformed image and a teacher image input during learning; the processor calculates dimension metrics of a pattern contained in the teacher image based on the transformed image and the teacher image; the processor calculates a combined error based on the transformation error and the dimension metrics; and the processor trains a machine learning model that generates the transformed image so as to optimize the combined error, and optimizes parameters of the machine learning model.

20. A storage medium carrying a program for causing a computer to execute a parameter optimization process that optimizes parameters for generating a transformed image based on an input image, the program causing the computer to execute the following processes: a process for calculating a transformation error between the transformed image and a teacher image input during learning; a process for calculating dimensional metrics of a pattern contained in the teacher image based on the transformed image and the teacher image; a process for calculating a combined error based on the transformation error and the dimensional metrics; and a process for training a machine learning model that generates the transformed image and optimizing parameters of the machine learning model so as to optimize the combined error.

Images