Artificial defect pattern generation and model training method and related device

Abstract

Embodiments of the present invention provide a method and related device for generating artificial defect patterns and model training. The provided method for training an artificial defect pattern generation model alternately trains the matrix model of the artificial defect pattern generation model and the pattern authenticity discrimination model. While ensuring that the pattern authenticity discrimination model has a high discrimination ability, it is achieved that the artificial defect patterns generated by the matrix model of the artificial defect pattern generation model are difficult to be recognized by the pattern authenticity discrimination model. Thus, the matrix generation accuracy of the matrix model of the trained artificial defect pattern generation model can be improved. Furthermore, a large number of artificial defect matrices meeting quality requirements can be generated by using the matrix model of the trained artificial defect pattern generation model.

Description

Technical Field

[0001] Embodiments of the present invention relate to the field of semiconductors, and in particular, to a method for generating a pattern, a method for network training, and related devices. Background Art

[0002] With the rapid development of semiconductor manufacturing technology, the feature size of the real pattern of integrated circuits during semiconductor processing is also continuously decreasing and has now dropped below 10 nm, and the pattern density of the real pattern of integrated circuits is also increasing.

[0003] During research and development, defective patterns will inevitably appear in the real patterns of integrated circuits. It is very difficult and time-consuming to manually search for and analyze these defective patterns. Therefore, a model can be used to speed up the process and increase the yield.

[0004] However, these models need to be trained according to corresponding defective pattern samples to find defective patterns. However, the proportion of defective patterns in the existing real patterns of integrated circuits is very small, and it is very difficult to extract defective pattern samples.

[0005] In order to obtain more defective pattern samples, the current method is usually to rotate and mirror the existing defective patterns, but this is far from enough for model training.

[0006] Therefore, how to generate a large number of artificial defective patterns with better quality has become a technical problem that needs to be urgently solved by those skilled in the art. Summary of the Invention

[0007] The technical problem solved by the embodiments of the present invention is how to generate a large number of artificial defective patterns with better quality.

[0008] To solve the above problems, embodiments of the present invention provide a method for training an artificial defective pattern generation model, including:

[0009] According to each random matrix, using the current matrix model of the artificial defective pattern generation model to be trained, each artificial defective matrix is obtained, and the artificial defective matrix is suitable for generating artificial defective patterns of integrated circuits;

[0010] Each of the artificial defective matrices and each of the obtained real defective matrices are respectively input into the same current pattern authenticity discrimination model to obtain the predicted artificial authenticity probability of each of the artificial defective pattern matrices and the predicted actual authenticity probability of each of the real defective matrices, wherein each of the real defective matrices is obtained based on each real defective pattern of the integrated circuit, the real defective matrix and the artificial defective matrix are matrices of the same type, and the value ranges of their elements are the same;

[0011] Using each of the predicted artificial authenticity probabilities, each of the predicted actual authenticity probabilities, the discrimination single threshold, the discrimination overall threshold, the matrix single threshold, and the matrix overall threshold, alternately fix one of the current matrix model and the current graphic authenticity discrimination model, and optimize the other to obtain a new current model of the other until the overall training requirement is met, and obtain the trained current matrix model, where the overall training requirement includes that the difference between the matrix objective functions of the matrix model before and after optimization is less than the matrix overall threshold and the difference between the discrimination objective functions of the discrimination model before and after optimization is less than the discrimination overall threshold, or the matrix overall training cycle reaches the predetermined matrix overall cycle and the discrimination overall training cycle reaches the predetermined discrimination overall cycle.

[0012] The training method of the artificial defect graphic generation model provided by the embodiment of the present invention trains the matrix model of the artificial defect graphic generation model and the graphic authenticity discrimination model through an alternating training method. While ensuring that the graphic authenticity discrimination model has a high discrimination ability, it is realized that the artificial defect graphics generated by the matrix model of the artificial defect graphic generation model are difficult to be recognized by the graphic authenticity discrimination model, so as to improve the matrix generation accuracy of the matrix model of the trained artificial defect graphic generation model. Furthermore, a large number of artificial defect matrices that meet the quality requirements can be generated by using the matrix model of the trained artificial defect graphic generation model, providing a basis for obtaining a large number of integrated circuit artificial defect graphics that meet the quality requirements, making the obtained integrated circuit artificial defect graphics have a relatively high similarity to the integrated circuit artificial defect graphics, having good applicability, and providing defect graphic samples for training the model used to find defective graphics in the real graphics of integrated circuits. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.

[0014] Figure 1 It is a schematic flowchart of the training method of the artificial defect graphic generation model provided by the embodiment of the present invention;

[0015] Figure 2 It is another schematic flowchart of the training method of the artificial defect graphic generation model provided by the embodiment of the present invention;

[0016] Figure 3 It is a schematic diagram of the real defect graphics and related matrices of integrated circuits of the training method of the artificial defect graphic generation model provided by the embodiment of the present invention;

[0017] Figure 4 Schematic diagram of an integrated circuit real defect pattern and a correlation matrix for the training method of the artificial defect pattern generation model provided by an embodiment of the present invention;

[0018] Figure 5 Schematic diagram of an integrated circuit real defect pattern and a correlation matrix for the training method of the artificial defect pattern generation model provided by an embodiment of the present invention;

[0019] Figure 6 Schematic flow chart of the artificial defect pattern generation method provided by an embodiment of the present invention;

[0020] Figure 7 Schematic diagram of the structure of the training device of the artificial defect pattern generation model provided by an embodiment of the present invention;

[0021] Figure 8 Schematic diagram of the structure of the artificial defect pattern generation device provided by an embodiment of the present invention;

[0022] Figure 9 Schematic diagram of the structure of the device provided by an embodiment of the present invention. Detailed implementation manners

[0023] As can be seen from the background art, the number of artificial defect patterns obtained by the existing methods is far from sufficient and the quality is poor.

[0024] To solve the above problems, an embodiment of the present invention provides a training method for an artificial defect pattern generation model, including:

[0025] According to each random matrix, using the current matrix model of the artificial defect pattern generation model to be trained, obtain each artificial defect matrix, and the artificial defect matrix is suitable for generating an integrated circuit artificial defect pattern;

[0026] Input each of the artificial defect matrices and each of the obtained real defect matrices into the same current pattern authenticity discrimination model, and obtain the predicted artificial authenticity probability of each artificial defect pattern matrix and the predicted actual authenticity probability of each real defect matrix, where each of the real defect matrices is obtained based on each integrated circuit real defect pattern, the real defect matrix and the artificial defect matrix are matrices of the same type, and the value ranges of their elements are the same;

[0027] Using each of the predicted artificial authenticity probabilities, each of the predicted actual authenticity probabilities, the single-discrimination threshold, the overall discrimination threshold, the single matrix threshold, and the overall matrix threshold, alternately fix one of the current matrix model and the current graphic authenticity discrimination model, and optimize the other to obtain a new current model of the other until the overall training requirements are met, and obtain the trained current matrix model, where the overall training requirements include that the difference between the matrix objective functions of the matrix model before and after optimization is less than the overall matrix threshold and the difference between the discrimination objective functions of the discrimination model before and after optimization is less than the overall discrimination threshold, or the overall matrix training cycle reaches the predetermined overall matrix cycle and the overall discrimination training cycle reaches the predetermined overall discrimination cycle.

[0028] The training method of the artificial defect graphic generation model provided by the embodiments of the present invention trains the matrix model of the artificial defect graphic generation model and the graphic authenticity discrimination model through an alternating training method. While ensuring that the graphic authenticity discrimination model has high discrimination ability, it is realized that the artificial defect graphics generated by the matrix model of the artificial defect graphic generation model are difficult to be recognized by the graphic authenticity discrimination model, so as to improve the matrix generation accuracy of the matrix model of the trained artificial defect graphic generation model. Furthermore, a large number of artificial defect matrices that meet the quality requirements can be generated by using the matrix model of the trained artificial defect graphic generation model, providing a basis for obtaining a large number of integrated circuit artificial defect graphics that meet the quality requirements, making the obtained integrated circuit artificial defect graphics have a relatively high similarity to the integrated circuit artificial defect graphics, having good applicability, and providing defect graphic samples for training the model used to find defect graphics in the real graphics of integrated circuits.

[0029] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0030] Please refer to Figure 1 , Figure 1 which is a schematic flowchart of the training method of the artificial defect graphic generation model provided by the embodiments of the present invention;

[0031] The embodiments of the present invention provide a training method of an artificial defect graphic generation model, including:

[0032] Step S101: According to each random matrix, use the current matrix model of the artificial defect graphic generation model to be trained to obtain each artificial defect matrix.

[0033] It can be understood that the random matrix satisfies a random distribution, such as various random distributions like uniform distribution, normal distribution, etc.

[0034] Based on each random matrix, the matrix model can be transformed to obtain each artificial defect matrix.

[0035] There are many variables in the matrix model, and these variables are continuously adjusted during the training process, so that the obtained artificial defect matrix has a relatively high degree of similarity to the real defect matrix.

[0036] It should be noted that during the training process, the parameters of the matrix model will be continuously adjusted. Therefore, before and after each adjustment, the matrix model will change. Therefore, the current matrix model refers to the matrix model with the current state parameters.

[0037] Step S102: Input each of the artificial defect matrices and each of the obtained real defect matrices into the same current graphic authenticity discrimination model, and obtain the predicted artificial authenticity probability of each of the artificial defect graphic matrices and the predicted actual authenticity probability of each of the real defect matrices.

[0038] In order to adjust the parameters of the matrix model, it is necessary to determine the quality of the obtained artificial defect matrix. Therefore, it is necessary to use the graphic authenticity discrimination model for discrimination. Whether the discrimination result of the graphic authenticity discrimination model is accurate also needs to be determined. For this purpose, it is also necessary to use the real defect matrix to obtain the predicted actual authenticity probability through the graphic authenticity discrimination model for reference. Therefore, after obtaining the artificial defect matrix, further input the obtained artificial defect matrix into the graphic authenticity discrimination model, use the graphic authenticity discrimination model to obtain the probability that the artificial defect graphic matrix is discriminated as a real defect, that is, the predicted artificial authenticity probability, and input the real defect matrix into the graphic authenticity discrimination model, use the graphic authenticity discrimination model to obtain the probability that the real defect graphic matrix is discriminated as a real defect, that is, the predicted actual authenticity probability.

[0039] It should be noted that the parameters in the current graphic authenticity discrimination model will change with training, and the current graphic authenticity discrimination models at different times are not necessarily the same.

[0040] Of course, based on the graphic authenticity discrimination model, each of the artificial defect matrices and each of the obtained real defect matrices can be input into the same current graphic authenticity discrimination model, and the acquisition order of the two is not restricted.

[0041] It can be understood that each of the real defect matrices is obtained based on each integrated circuit real defect graphic. The real defect matrix and the artificial defect matrix are of the same type, and the value ranges of their elements are the same.

[0042] Since each of the real defect matrices is obtained based on the real defect patterns of the integrated circuits, and the ultimate goal of the artificial defect matrix is to generate artificial defect patterns of the integrated circuits. Therefore, when the real defect matrix and the artificial defect matrix are not matrices of the same type, it is difficult to use the method based on which each of the real defect matrices is obtained to reversely generate artificial defect patterns of the integrated circuits. Thus, it is necessary to make the real defect matrix and the artificial defect matrix be matrices of the same type to ensure that the obtained artificial defect matrix can generate artificial defect patterns of the integrated circuits.

[0043] When the value ranges of the elements of the real defect matrix and the artificial defect matrix are different, it is possible that there are values of elements that only belong to the real defect matrix and values of elements that only belong to the artificial defect matrix.

[0044] If the values of the elements that only exist in the real defect matrix cannot appear in the artificial defect matrix, it is likely that the generated artificial defect patterns of the integrated circuits lack some features compared with the real defect patterns of the integrated circuits.

[0045] If the values of the elements that only exist in the artificial defect matrix cannot appear in the real defect matrix, it proves that the values of the elements that only exist in the artificial defect matrix do not correspond to the features of the real defect matrix and have no specific meaning. Therefore, the value ranges of the elements of the real defect matrix and the artificial defect matrix should be the same.

[0046] The pattern authenticity discrimination model is suitable for discriminating whether a matrix is a real defect matrix or an artificial defect matrix. The input of the pattern authenticity discrimination model is a matrix, and the output is the probability that the input matrix is a real defect matrix, or the probability that the input matrix is an artificial defect matrix. Obviously, the sum of the probability that a certain matrix is a real defect matrix and the probability that the matrix is an artificial defect matrix is 1. Therefore, for convenience, the matrix authenticity discrimination model can only output the probability of being one of the matrices.

[0047] There are many variables in the matrix authenticity discrimination model for discriminating matrices, and these variables can be changed as needed to improve the discrimination effect.

[0048] Step S103: Using each of the predicted artificial authenticity probabilities, each of the predicted actual authenticity probabilities, the single discrimination threshold, the overall discrimination threshold, the single matrix threshold, and the overall matrix threshold, alternately fix one of the current matrix model and the current pattern authenticity discrimination model, and optimize the other to obtain a new current model of the other until the overall training requirements are met, and obtain the trained current matrix model.

[0049] Among them, the overall training requirements include that the difference between the matrix objective functions of the matrix model before and after optimization is less than the overall matrix threshold and the difference between the discrimination objective functions of the discrimination model before and after optimization is less than the overall discrimination threshold, or the overall matrix training period reaches a predetermined overall matrix period and the overall discrimination training period reaches a predetermined overall discrimination period.

[0050] It should be noted that the training objectives of the discrimination model and the matrix model are preset. When the current matrix model and the current graphic authenticity discrimination model reach the corresponding training objectives, it is considered to meet the requirements.

[0051] Specifically, the overall matrix threshold and the overall discrimination threshold can be set according to the matrix objective function of the matrix model and the discrimination objective function of the discrimination model. When the difference between the matrix objective functions of the matrix model before and after optimization is less than the overall matrix threshold and the difference between the discrimination objective functions of the discrimination model before and after optimization is less than the overall discrimination threshold, it can be considered that the matrix model meets the requirements.

[0052] Of course, a predetermined overall matrix period and a predetermined overall discrimination period can also be set. When the overall matrix training period reaches the predetermined overall matrix period and the overall discrimination training period reaches the predetermined overall discrimination period, it can be considered that the matrix model meets the requirements.

[0053] Among them, there are various directions for optimizing the matrix model of the graphic authenticity discrimination model and the artificial defect graphic generation model. In a specific embodiment, the optimization direction when optimizing the graphic authenticity discrimination model is opposite to the optimization direction when optimizing the matrix model of the artificial defect graphic generation model.

[0054] In this way, the two optimization directions are opposite, so that the result obtained by optimizing the graphic authenticity discrimination model can be continuously optimized according to the artificial defect matrix generated by the matrix model of the artificial defect graphic generation model and the real defect matrix, and then the matrix model of the artificial defect graphic generation model can be optimized according to the result obtained by the graphic authenticity discrimination model, so that finally the matrix model of the artificial defect graphic generation model can be optimized to make the artificial defect matrix generated by it similar to the real defect matrix.

[0055] The training method of the artificial defect pattern generation model provided by the embodiments of the present invention trains the matrix model of the artificial defect pattern generation model and the pattern authenticity discrimination model in an alternating training manner. While ensuring that the pattern authenticity discrimination model has a high discrimination ability, it is realized that the artificial defect patterns generated by the matrix model of the artificial defect pattern generation model are difficult to be recognized by the pattern authenticity discrimination model. Thus, the matrix generation accuracy of the matrix model of the trained artificial defect pattern generation model can be improved. Furthermore, a large number of artificial defect matrices meeting the quality requirements can be generated by using the matrix model of the trained artificial defect pattern generation model, providing a basis for obtaining a large number of integrated circuit artificial defect patterns meeting the quality requirements, making the obtained integrated circuit artificial defect patterns have a relatively high similarity to the integrated circuit artificial defect patterns, having good applicability, and providing defect pattern samples for training the model used to find defect patterns in the real patterns of integrated circuits.

[0056] To further improve the effect of optimizing the current matrix model and the current pattern authenticity discrimination model, please refer to Figure 2 , Figure 2 which is another flowchart of the training method of the artificial defect pattern generation model provided by the embodiments of the present invention.

[0057] As shown in the figure, step S103 of the training method of the artificial defect pattern generation model provided by the embodiments of the present invention: Using each of the predicted artificial authenticity probabilities, each of the predicted actual authenticity probabilities, the single discrimination threshold, the overall discrimination threshold, the single matrix threshold, and the overall matrix threshold, alternately fixing one of the current matrix model and the current pattern authenticity discrimination model and optimizing the other to obtain a new current model of the other, until the overall training requirement is met. The steps for obtaining the trained current matrix model may include:

[0058] Step S201: Optimize the pattern authenticity discrimination model by using each of the predicted actual authenticity probabilities and each of the predicted artificial authenticity probabilities obtained from the artificial defect matrices acquired based on the current matrix model.

[0059] To realize the training of the pattern authenticity discrimination model, first, some artificial defect matrices can be generated only by using the current matrix model and the pattern authenticity discrimination model can be trained together with the real matrices.

[0060] It can be understood that when optimizing the graphic authenticity discrimination model, the variables in the current matrix model remain unchanged, and the current matrix model is only used to generate the artificial defect matrix. The variables in the graphic authenticity discrimination model are adjusted by using each of the predicted actual authenticity probabilities and each of the predicted artificial authenticity probabilities obtained from the artificial defect matrix acquired based on the current matrix model, so as to achieve the purpose of optimizing the graphic authenticity discrimination model.

[0061] Among them, the duration of each training of the graphic authenticity discrimination model does not need to be set too high. When the duration is set too high, the time required to reach the phased goal each time is relatively long. Moreover, when the discrimination ability of the graphic authenticity discrimination model is too good and it can extremely well distinguish the artificial defect matrix generated by the matrix model of the generation model, it is impossible to provide a suitable descending gradient for the matrix model, so it is very difficult to optimize the matrix model of the generation model according to the graphic authenticity discrimination model. Of course, if the discrimination ability of the graphic authenticity discrimination model is very poor, it cannot provide an effective gradient. At this time, it is necessary to optimize the graphic authenticity discrimination model until its discrimination ability meets certain requirements.

[0062] Step S202: Determine whether the predetermined number of single discrimination trainings is reached or whether the difference between the discrimination objective functions of the discrimination model before and after optimization is less than the single discrimination threshold. If so, execute step S203; if not, execute step S201.

[0063] Set the predetermined number of single discrimination trainings or the single discrimination threshold. When the number of trainings meets the predetermined number of single discrimination trainings or the difference between the discrimination objective functions of the discrimination model before and after optimization is less than the single discrimination threshold, it is considered that the graphic authenticity discrimination model meets the goal of this training round.

[0064] Step S203: Fix the graphic authenticity discrimination model to obtain a new graphic authenticity discrimination model, and use the new graphic authenticity discrimination model as the current graphic authenticity discrimination model.

[0065] It can be understood that fixing the graphic authenticity discrimination model means that in the training cycle of the next matrix model, the numerical values of the parameters in the graphic authenticity discrimination model will no longer be changed. Using the fixed graphic authenticity discrimination model at this time as the current graphic authenticity discrimination model to obtain the predicted artificial authenticity probability, further preparations for the training of the matrix model are made.

[0066] Step S204: Determine whether the overall discrimination requirement is met. If so, execute step S209; if not, execute step S205.

[0067] It can be understood that the overall discrimination requirement includes that the overall training period reaches a predetermined overall period or the difference between the discrimination objective functions of the discrimination model before and after optimization is less than the overall discrimination threshold.

[0068] A predetermined overall discrimination period or an overall discrimination threshold can be set. When the overall training period reaches the predetermined overall discrimination period or the difference between the discrimination objective functions of the discrimination model before and after optimization is less than the overall discrimination threshold, it can be considered that the graphic authenticity discrimination model meets the requirements. The predetermined overall discrimination period can be several times the number of predetermined single discrimination training times, so that the overall training period can reach the predetermined overall discrimination period after several trainings.

[0069] Step S205: Optimize the matrix model by using the respective predicted artificial authenticity probabilities obtained from the current matrix model and the current graphic authenticity discrimination model.

[0070] It can be understood that when optimizing the matrix model, the variables in the current graphic authenticity discrimination model remain unchanged, and the current graphic authenticity discrimination model only gives the predicted artificial authenticity probabilities of the artificial defect matrix generated by the matrix model.

[0071] Use each artificial defect matrix obtained from the current matrix model and its respective predicted artificial authenticity probabilities to adjust the variables in the matrix model, so as to achieve the purpose of optimizing the matrix model.

[0072] Step S206: Determine whether the number of single matrix trainings reaches a predetermined number or whether the difference between the matrix objective functions of the matrix model before and after optimization is less than the single matrix threshold. If so, execute Step S207; if not, execute Step S205.

[0073] Set a predetermined number of single matrix trainings or a single matrix threshold. When the number of trainings meets the predetermined number of single matrix trainings or the difference between the matrix objective functions of the matrix model before and after optimization is less than the single matrix threshold, it is considered that the matrix model meets the objective of this training round.

[0074] Step S207: Fix the matrix model to obtain a new matrix model, and use the new matrix model as the current matrix model.

[0075] It can be understood that fixing the matrix model means that the numerical values of the variables in the matrix model will not be changed during the optimization period of the next graphic authenticity discrimination model. During the optimization period of the next graphic authenticity discrimination model, the respective predicted artificial authenticity probabilities are obtained by using the artificial defect matrix obtained from the matrix model fixed at this time.

[0076] Step S208: Determine whether the overall generation requirement is met. If yes, execute Step S209; if no, execute Step S201.

[0077] It can be understood that the overall matrix requirement includes that the overall matrix training period reaches a predetermined overall matrix period or the difference between the matrix objective functions of the matrix model before and after optimization is less than the overall matrix threshold.

[0078] The predetermined overall matrix period or the overall matrix threshold can be set. When the overall matrix training period reaches the predetermined overall matrix period or the difference between the matrix objective functions of the matrix model before and after optimization is less than the overall matrix threshold, it can be considered that the graphic authenticity matrix model meets the requirements. The predetermined overall matrix period can be several times the number of predetermined single matrix training times, so that the overall matrix training period can reach the predetermined overall matrix period after several trainings.

[0079] When Step S208 is not satisfied, it proves that the matrix model has not met the overall generation requirement. However, it is difficult to achieve a good effect by training the matrix model alone. Therefore, it is necessary to retrain the graphic authenticity discrimination model, and then use the graphic authenticity discrimination model to continue training the matrix model. Therefore, continue from Step S201.

[0080] Step S209: Obtain the trained current matrix model.

[0081] When the matrix model meets the overall generation requirement, it can be considered that the matrix model is trained, and thus the matrix model can be used to generate integrated circuit artificial defect graphics.

[0082] The value ranges of the elements of the artificial defect matrix and the elements of the real defect matrix can be set as needed.

[0083] Therefore, in a specific embodiment, the elements of the artificial defect matrix and the elements of the real defect matrix each have only two values, namely the first value and the second value.

[0084] Please refer to Figure 3 , Figure 3 is a schematic diagram of an integrated circuit real defect graphic and related matrices of the training method of the artificial defect graphic generation model provided by an embodiment of the present invention. The integrated circuit real graphic is an integrated circuit design graphic ( Figure 3When it is the figure a) in [reference], since there is no gray value in the integrated circuit design pattern, there are only two situations at a certain point, belonging to a transmissive pattern or a masking pattern. Therefore, according to the characteristics of the integrated circuit design pattern, a method for obtaining the true defect matrix can be set. Thus, in a specific embodiment, the true integrated circuit pattern includes the integrated circuit design pattern; the step of obtaining the true defect matrices of the respective true defect patterns may include:

[0085] Classify the true defect pattern blocks into transmissive pattern blocks or masking pattern blocks, and make all the elements in the true defect matrix correspond to the true defect pattern blocks in the integrated circuit design pattern;

[0086] Assign the value of the element corresponding to the transmissive pattern block to the first value;

[0087] Assign the value of the element corresponding to the masking pattern block to the second value.

[0088] As Figure 3 shown in figure a) in [reference], the integrated circuit design pattern generally only includes the projected part and the masked part on the mask plate. Therefore, according to whether it is transmissive on the mask plate, each part of the true defect pattern block can be divided into a transmissive pattern block or a masking pattern block, and then all the elements in the true defect matrix are made to correspond to the true defect pattern blocks in the integrated circuit design pattern, that is Figure 3 each element in the matrix in figure b) in [reference]. Then assign the value of the element corresponding to the transmissive pattern block to the first value; assign the value of the element corresponding to the masking pattern block to the second value.

[0089] As Figure 3 shown in figure b) in [reference], the first value and the second value in the true defect matrix can respectively take 0 or 1. Also, since the value ranges of the elements of the true defect matrix and the artificial defect matrix are the same, correspondingly Figure 3 the values in the artificial defect matrix in [reference] d) are also the first value or the second value.

[0090] Similarly, after obtaining the artificial defect matrix, an artificial defect pattern can be generated using the above corresponding method.

[0091] Figure 3 The matrix in [reference] c) is a random matrix. The random matrix and the artificial defect matrix can be matrices of the same type or not of the same type. However, when the random matrix and the artificial defect matrix are matrices of the same type, the calculation is more convenient, and the artificial defect matrix can be obtained without deforming the random matrix.

[0092] The value range of the random matrix may be the same as or different from the value range of the true defect matrix or the artificial defect matrix. However, when the value range of the random matrix is the same as the value range of the true defect matrix or the artificial defect matrix, the calculation is more convenient, and the time for the matrix model to calculate the artificial defect matrix using the random matrix can be effectively saved.

[0093] By setting the values of the elements of the defect matrix and the elements of the true defect matrix to only two, namely the first value and the second value, not only can the training time and generation time be effectively saved, but also it can correspond to the characteristics of the integrated circuit design pattern, so that the generated integrated circuit artificial defect pattern can be closer to the integrated circuit design pattern.

[0094] Certainly, when the integrated circuit true pattern further includes an integrated circuit true image, the values of the elements of the artificial defect matrix and the elements of the true defect matrix can also be only two, namely the first value and the second value.

[0095] Therefore, in a specific embodiment, please refer to Figure 4 , Figure 4 FIG. is a schematic diagram of the integrated circuit true defect pattern and related matrices of the training method of the artificial defect pattern generation model provided by the embodiment of the present invention. The integrated circuit true pattern includes an integrated circuit true image. The step of obtaining each true defect matrix of each true defect pattern includes:

[0096] Corresponding all the elements in the true defect matrix to the true defect pattern blocks in the integrated circuit true image;

[0097] Classifying the true defect pattern blocks into white pattern blocks or black pattern blocks. The gray level of the white pattern blocks is less than the gray level threshold, and the gray level of the black pattern blocks is greater than the gray level threshold. The gray level threshold is less than the maximum gray level of the integrated circuit true image and greater than the minimum gray level of the integrated circuit true image;

[0098] Assigning the value of the element corresponding to the white pattern block to the first value;

[0099] Assigning the value of the element corresponding to the black pattern block to the second value.

[0100] As Figure 4As shown in Figure a, the real image of the integrated circuit includes two parts: a pattern and the blank area outside the pattern. There is a large gray-scale difference between the pattern and the blank area. Therefore, according to the gray-scale of each real defect pattern block, the real defect pattern blocks can be classified into white pattern blocks or black pattern blocks. The gray-scale of the white pattern blocks is less than the gray-scale threshold, and the gray-scale of the black pattern blocks is greater than the gray-scale threshold. The gray-scale threshold can be set according to the gray-scale at the edge of the pattern.

[0101] As Figure 4 shown in Figure b, the value of the element corresponding to the white pattern block is the first value, and the value of the element corresponding to the black pattern block is the second value. Similarly, the first value and the second value in the real defect matrix can be 0 or 1 respectively. Also, since the value ranges of the elements in the real defect matrix and the artificial defect matrix are the same, correspondingly Figure 4 the values in the artificial defect matrix in Figure d are also the first value or the second value.

[0102] Similarly, after obtaining the artificial defect matrix, the artificial defect pattern can be generated based on the random matrix using the above corresponding method.

[0103] Figure 4 The matrix in Figure c is a random matrix. The random matrix and the artificial defect matrix can be matrices of the same type or not. However, when the random matrix and the artificial defect matrix are matrices of the same type, the calculation is more convenient, and there is no need to transform the random matrix to obtain the artificial defect matrix.

[0104] The value range of the random matrix can be the same as or different from the value range of the real defect matrix or the artificial defect matrix. However, when the value range of the random matrix is the same as the value range of the real defect matrix or the artificial defect matrix, the calculation is more simple and convenient, and the time for the matrix model to calculate the artificial defect matrix using the random matrix can be effectively saved.

[0105] By setting the values of the elements of the defect matrix and the real defect matrix to only two, namely the first value and the second value, both the training time and the generation time can be effectively saved, and it can also correspond to the characteristics of the real image of the integrated circuit, so that the generated artificial defect pattern of the integrated circuit can be relatively similar to the real image of the integrated circuit.

[0106] Of course, the values of the elements of the artificial defect matrix and the real defect matrix can also be multiple. Therefore, in a specific implementation, the values of the elements of the artificial defect matrix and the real defect matrix are all integers, and are greater than or equal to the lower limit of the value and less than or equal to the upper limit of the value.

[0107] When the elements of the artificial defect matrix and the elements of the true defect matrix also take multiple values, they can well correspond to the gray levels of the graphic blocks in the true image of the integrated circuit.

[0108] Therefore, in a specific embodiment, please refer to Figure 5 , Figure 5 which is a schematic diagram of the true defect graphics of the integrated circuit and related matrices for the training method of the artificial defect graphics generation model provided by the embodiment of the present invention.

[0109] In a specific embodiment, the true integrated circuit graphics include the true image of the integrated circuit. The step of obtaining the respective true defect matrices of the respective true defect graphics includes:

[0110] Put all the elements in the true defect matrix in one-to-one correspondence with the true defect graphic blocks in the true image of the integrated circuit;

[0111] Assign the gray level value of the true defect graphic block to all the elements in the true defect matrix, and the range of the gray level value is less than or equal to the maximum gray level of the true image of the integrated circuit and greater than or equal to the minimum gray level of the true image of the integrated circuit.

[0112] As shown in Figure 5 a, each graphic block in the true image of the integrated circuit has its gray level value. Therefore, the gray level value of the true defect graphic block can be assigned to all the elements in the true defect matrix. For example, Figure 5 in b, the value of each element in the true defect matrix is the gray level value of its corresponding graphic block. The gray level value can take values from 0 to 255. Of course, other ranges can also be taken. Also, since the value ranges of the elements of the true defect matrix and the artificial defect matrix are the same, correspondingly Figure 5 the values in the artificial defect matrix in d also take values from 0 to 255.

[0113] Similarly, after obtaining the artificial defect matrix, an artificial defect graphic can be generated using the corresponding method as above.

[0114] Figure 5 The matrix in c is a random matrix. The random matrix and the artificial defect matrix can be matrices of the same type or matrices of different types. However, when the random matrix and the artificial defect matrix are matrices of the same type, the calculation is more convenient, and there is no need to transform the random matrix to obtain the artificial defect matrix.

[0115] The value range of the random matrix may be the same as or different from that of the real defect matrix or the artificial defect matrix. However, when the value range of the random matrix is the same as that of the real defect matrix or the artificial defect matrix, the calculation is more convenient, and the time for the matrix model to calculate the artificial defect matrix using the random matrix can be effectively saved.

[0116] It can be seen that by setting the values of the elements of the defect matrix and the elements of the real defect matrix, the values of the elements of the artificial defect matrix and the real defect matrix are both integers, greater than or equal to the lower limit of the value, less than or equal to the upper limit of the value, and corresponding to the gray values in the real image of the integrated circuit, so that the generated artificial defect pattern of the integrated circuit is closer to the real image of the integrated circuit.

[0117] The types of the matrix model and the graphic authenticity discrimination model can be selected according to needs. It is a convolutional neural network. Compared with other neural networks, the calculation is relatively simple, which can effectively save the training time and the generation time and improve the processing efficiency.

[0118] Among them, not only matrices with multiple rows and multiple columns can be selected from the artificial defect matrix and the real defect matrix. In a specific embodiment, in order to improve the processing efficiency, the artificial defect matrix and the real defect matrix are vectors.

[0119] When the artificial defect matrix and the real defect matrix are vectors, the calculation is more convenient compared with matrices with multiple rows and multiple columns, which can effectively save the training time and the generation time and improve the processing efficiency.

[0120] When the artificial defect matrix and the real defect matrix are vectors, since the vector cannot be directly corresponding to the graphic blocks of the real defect pattern of the integrated circuit, other methods can be used to generate the real defect matrix.

[0121] For example, the information of each bit can be set as the number of graphic blocks, the length or width of the graphic blocks, the corners in the graphic blocks, the included angles between the sides of each graphic, the distance between the graphic blocks, etc. These information can be set according to needs and converted into the real defect matrix by measuring the real graphic of the integrated circuit. Similarly, after obtaining the artificial defect matrix, the artificial defect pattern can be generated using the above corresponding method.

[0122] When the artificial defect matrix and the real defect matrix are vectors, it is difficult to process them using a convolutional neural network. Therefore, in a specific embodiment, the matrix model and the graphic authenticity discrimination model are fully connected neural networks. When the matrix model and the graphic authenticity discrimination model are fully connected neural networks, the effect of processing the case where the artificial defect matrix and the real defect matrix are vectors is better.

[0123] Of course, in order to obtain improved artificial defect graphics, the graphic generation model of the artificial defect graphic generation model can be further trained to generate artificial defect graphics using the artificial defect matrix. Therefore, in a specific embodiment, the training method of the artificial defect graphic generation model further includes:

[0124] Using the graphic generation model of the artificial defect graphic generation model to obtain the predicted real defect graphics of each of the real defect matrices;

[0125] Using each of the integrated circuit real defect graphics and each of the predicted real defect graphics to obtain a graphic similarity, and adjusting the parameters of the graphic generation model according to the graphic similarity until the image similarity meets the similarity threshold, to obtain the trained graphic generation model.

[0126] Among them, the graphic generation model can generate predicted real defect graphics using the real defect matrix, or generate integrated circuit artificial defect graphics using the artificial defect matrix. Therefore, using the trained graphic generation model, integrated circuit artificial defect graphics can be obtained according to the artificial defect matrix.

[0127] Moreover, the artificial defect matrix used can be a vector or a matrix independent of the distribution of the graphic blocks of the integrated circuit real defect graphics, that is, obtained using a random matrix. The meaning of each element of the artificial defect matrix used does not need to be set in advance, but is generated automatically through the training of the graphic generation model.

[0128] In this way, the necessary number of elements of the artificial defect matrix required by the graphic generation model is less. Compared with the method using graphic blocks, fewer elements can be used to express the information of the integrated circuit real defect graphics and the integrated circuit artificial defect graphics, thereby improving the training and generation efficiency of the artificial defect graphic generation model, and the quality of the obtained artificial defect graphics is also relatively high.

[0129] To solve the above problems, an embodiment of the present invention also provides an artificial defect graphic generation method. Please refer to Figure 6 , Figure 6 which is a schematic flowchart of the artificial defect graphic generation method provided by the embodiment of the present invention.

[0130] AsFigure 6 As shown in Figure 6 , the artificial defect pattern generation method provided by the embodiments of the present invention includes:

[0131] Step S301: According to each random matrix, use the matrix model of the artificial defect pattern generation model trained by the aforementioned artificial defect pattern generation model training method to generate a predicted artificial defect matrix.

[0132] Step S302: According to the predicted artificial defect matrix, use the pattern generation model of the artificial defect pattern generation model to obtain the integrated circuit artificial defect pattern.

[0133] Among them, the method of using the pattern generation model of the artificial defect pattern generation model to obtain the integrated circuit artificial defect pattern can either use the method of inversely inferring the integrated circuit artificial defect pattern according to the gray scale as described above, or use the trained artificial defect pattern generation model to generate the integrated circuit artificial defect pattern.

[0134] The artificial defect pattern generation method provided by the embodiments of the present invention can use the matrix model of the trained artificial defect pattern generation model to generate a relatively large number of artificial defect matrices that meet the quality requirements. When training the matrix model, by means of alternating training, train the matrix model of the artificial defect pattern generation model and the pattern authenticity discrimination model. While ensuring that the pattern authenticity discrimination model has a high discrimination ability, make the artificial defect patterns generated by the matrix model of the artificial defect pattern generation model difficult to be recognized by the pattern authenticity discrimination model, improve the matrix generation accuracy of the matrix model of the trained artificial defect pattern generation model, and thus use the matrix model of the trained artificial defect pattern generation model to generate a large number of artificial defect matrices that meet the quality requirements. Further use the pattern generation model to obtain a large number of integrated circuit artificial defect patterns that meet the quality requirements, so that the obtained integrated circuit artificial defect patterns have a relatively high similarity to the integrated circuit artificial defect patterns and have good applicability, and can provide defect pattern samples for training the model used to find defect patterns in the real patterns of integrated circuits.

[0135] Next, the training device and the artificial defect pattern generation device of the artificial defect pattern generation model provided by the embodiments of the present invention will be introduced. The training device and the artificial defect pattern generation device of the artificial defect pattern generation model described below can be considered as the functional module architectures that an electronic device (such as a PC) needs to set up to respectively implement the artificial defect pattern generation method provided by the embodiments of the present invention. The content of the artificial defect pattern generation model described below can be respectively corresponding and referred to the content of the artificial defect pattern generation method described above.

[0136] Please refer to Figure 7, an embodiment of the present invention further provides a training device for an artificial defect pattern generation model, including:

[0137] An artificial defect matrix acquisition unit 11, which, according to each random matrix, uses the current matrix model of the artificial defect pattern generation model to be trained to obtain each artificial defect matrix, and the artificial defect matrix is suitable for generating an integrated circuit artificial defect pattern;

[0138] A authenticity probability acquisition unit 12, which is suitable for inputting each of the artificial defect matrices and each of the acquired real defect matrices into the same current pattern authenticity discrimination model to obtain the predicted artificial authenticity probability of each of the artificial defect pattern matrices and the predicted actual authenticity probability of each of the real defect matrices. Among them, each of the real defect matrices is obtained based on each integrated circuit real defect pattern, the real defect matrix and the artificial defect matrix are matrices of the same type, and the value ranges of their elements are the same;

[0139] A matrix model acquisition unit 13, which is suitable for using each of the predicted artificial authenticity probabilities and each of the predicted actual authenticity probabilities to alternately fix one of the current matrix model and the current pattern authenticity discrimination model and optimize the other to obtain a new current model of the other until the overall training requirement is met, and obtain the trained current matrix model, where the overall training requirement includes that the overall training period of the OR matrix reaches a predetermined overall matrix period and the overall training period of the discrimination reaches a predetermined overall discrimination period.

[0140] Optionally, the matrix model acquisition unit is further suitable for using each of the predicted actual authenticity probabilities and each of the predicted artificial authenticity probabilities obtained from the artificial defect matrix based on the current matrix model to optimize the pattern authenticity discrimination model until a predetermined number of single discrimination training times is reached, fixing the pattern authenticity discrimination model to obtain a new pattern authenticity discrimination model, and using the new pattern authenticity discrimination model as the current pattern authenticity discrimination model;

[0141] When the overall discrimination requirement is not met, use each of the predicted artificial authenticity probabilities obtained by the current pattern authenticity discrimination model and each of the predicted actual authenticity probabilities to optimize the matrix model until a predetermined number of single matrix training times is reached, fix the matrix model to obtain a new matrix model, and use the new matrix model as the current matrix model. When the overall discrimination requirement is not met, turn to execute the optimization of the pattern authenticity discrimination model.

[0142] Optionally, the optimization direction when optimizing the pattern authenticity discrimination model is opposite to the optimization direction when optimizing the matrix model of the artificial defect pattern generation model.

[0143] Optionally, the elements of the artificial defect matrix and the elements of the true defect matrix have only two values, namely the first value and the second value, respectively.

[0144] Optionally, the integrated circuit real pattern includes an integrated circuit design pattern;

[0145] The training device of the artificial defect pattern generation model further includes: a true defect matrix acquisition unit 14 adapted to acquire the true defect matrices of the respective true defect patterns.

[0146] Optionally, the true defect matrix acquisition unit is adapted to classify the true defect pattern blocks into transmission pattern blocks or masking pattern blocks, and correspond all the elements in the true defect matrix to the true defect pattern blocks in the integrated circuit design pattern;

[0147] Assign the value of the element corresponding to the transmission pattern block to the first value;

[0148] Assign the value of the element corresponding to the masking pattern block to the second value.

[0149] Optionally, the integrated circuit real pattern includes an integrated circuit real image, and the true defect matrix acquisition unit is adapted to correspond all the elements in the true defect matrix to the true defect pattern blocks in the integrated circuit real image; classify the true defect pattern blocks into white pattern blocks or black pattern blocks, the gray level of the white pattern blocks is less than the gray level threshold, the gray level of the black pattern blocks is greater than the gray level threshold, the gray level threshold is less than the maximum gray level of the integrated circuit real image and greater than the minimum gray level of the integrated circuit real image; assign the value of the element corresponding to the white pattern block to the first value; assign the value of the element corresponding to the black pattern block to the second value.

[0150] Optionally, the elements of the artificial defect matrix and the true defect matrix are all integers, and are greater than or equal to the lower limit of the value and less than or equal to the upper limit of the value.

[0151] Optionally, the integrated circuit real pattern includes an integrated circuit real image, and the true defect matrix acquisition unit is adapted to correspond one by one all the elements in the true defect matrix to the true defect pattern blocks in the integrated circuit real image; assign all the elements in the true defect matrix to the gray level values of the true defect pattern blocks, and the range of the gray level values is less than or equal to the maximum gray level of the integrated circuit real image and greater than or equal to the minimum gray level of the integrated circuit real image.

[0152] Optionally, the matrix model and the pattern authenticity discrimination model are convolutional neural networks.

[0153] Optionally, the artificial defect matrix and the real defect matrix are vectors.

[0154] Optionally, the matrix model and the graphic authenticity discrimination model are fully connected neural networks.

[0155] Optionally, the random matrix and the artificial defect matrix are matrices of the same type.

[0156] Optionally, the random matrix and the artificial defect matrix have the same value range.

[0157] Optionally, it further includes: a graphic generation model training unit 15, adapted to use the graphic generation model of the artificial defect graphic generation model to obtain the predicted real defect graphics of each of the real defect matrices;

[0158] Using each of the integrated circuit real defect graphics and each of the predicted real defect graphics, obtain the graphic similarity, and adjust the parameters of the graphic generation model according to the graphic similarity until the image similarity meets the similarity threshold, to obtain the trained graphic generation model.

[0159] Please refer to Figure 8 , an embodiment of the present invention further provides an artificial defect graphic generation model, including:

[0160] A predicted artificial defect matrix acquisition unit 21, adapted to generate a predicted artificial defect matrix according to each random matrix, using the matrix model of the artificial defect graphic generation model trained by the training method of the artificial defect graphic generation model;

[0161] An artificial defect graphic acquisition unit 22, adapted to obtain an integrated circuit artificial defect graphic according to the artificial defect matrix, using the graphic generation model of the artificial defect graphic generation model.

[0162] An embodiment of the present invention further provides a storage medium, the storage medium stores a program adapted to train the training method of the artificial defect graphic generation model to implement the training method of the artificial defect graphic generation model, or stores a program adapted to generate the integrated circuit artificial defect graphic to implement the artificial defect graphic generation method.

[0163] An embodiment of the present invention further provides a device. The device provided by the embodiment of the present invention can load a program module architecture in the form of a program to implement the training method of the artificial defect graphic generation model or the artificial defect graphic generation method; this hardware device can be applied to an electronic device with specific data processing capabilities, and this electronic device can be: for example, a terminal device or a server device.

[0164] Therefore, please refer to Figure 9 , Figure 9Schematic diagram of the device provided by the embodiments of the present invention.

[0165] The device provided by the embodiments of the present invention includes: at least one memory 31 and at least one processor 32. The memory 31 stores one or more computer-executable instructions, and the processor 32 invokes the one or more computer-executable instructions to execute the training method of the artificial defect pattern generation model or the artificial defect pattern generation method.

[0166] It can be understood that the device may further include at least one communication interface 33 and at least one communication bus 34; the processor 32 and the memory 31 may be located in the same electronic device. For example, the processor 32 and the memory 31 may be located in a server device or a terminal device; the processor 32 and the memory 31 may also be located in different electronic devices.

[0167] In the embodiments of the present invention, the number of the processor 32, the communication interface 33, the memory 31, and the communication bus 34 is at least one, and the processor 32, the communication interface 33, and the memory 31 complete mutual communication through the communication bus 34; obviously, Figure 9 The schematic diagram of the communication connection of the shown processor 32, communication interface 33, memory 31, and communication bus 34 is only an optional way.

[0168] Optionally, the communication interface 33 may be an interface of a communication module, such as an interface of a GSM module; the processor 32 may be a central processing unit CPU, or a specific integrated circuit ASIC, or one or more integrated circuits configured to implement the embodiments of the present invention; the memory 31 may include a high-speed RAM memory or may also include a non-volatile memory, such as at least one disk memory.

[0169] It should be noted that the above device may further include other components (not shown) that may not be necessary for the disclosed content of the embodiments of the present invention; since these other components may not be necessary for understanding the disclosed content of the embodiments of the present invention, the embodiments of the present invention do not introduce them one by one.

[0170] The above-described embodiments of the present invention are combinations of the elements and features of the present invention. Unless otherwise mentioned, the elements or features may be considered optional. Each element or feature may be practiced without combination with other elements or features. Additionally, embodiments of the present invention may be constructed by combining some of the elements and / or features. The order of operations described in the embodiments of the present invention may be rearranged. Some configurations of any one embodiment may be included in another embodiment and replaced with corresponding configurations of another embodiment. It is obvious to those skilled in the art that claims that do not have an explicit citation relationship with each other in the appended claims may be combined into embodiments of the present invention or may be included as new claims in amendments after the filing of this application.

[0171] Embodiments of the present invention can be implemented by various means such as, for example, hardware, firmware, software, or a combination thereof. In a hardware configuration, the method according to an exemplary embodiment of the present invention can be implemented by one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), processors, controllers, microcontrollers, microprocessors, etc.

[0172] In a firmware or software configuration, embodiments of the present invention can be implemented in the form of modules, procedures, functions, etc. The software code can be stored in a memory unit and executed by a processor. The memory unit is located inside or outside the processor and can send data to and receive data from the processor via various known means.

[0173] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but rather to the broadest scope consistent with the principles and novel features disclosed herein.

[0174] Although the embodiments of the present invention are disclosed as above, the embodiments of the present invention are not limited thereto. Any person skilled in the art can make various changes and modifications without departing from the spirit and scope of the embodiments of the present invention. Therefore, the protection scope of the embodiments of the present invention should be determined by the scope defined by the claims.

Claims

1. A training method for an artificial defect pattern generation model, characterized in that, Including: According to each random matrix, using the current matrix model of the artificial defect pattern generation model to be trained, obtaining each artificial defect matrix, where the artificial defect matrix is suitable for generating integrated circuit artificial defect patterns; Inputting each of the artificial defect matrices and each of the obtained real defect matrices into the same current pattern authenticity discrimination model, obtaining the predicted artificial authenticity probabilities of each of the artificial defect pattern matrices and the predicted actual authenticity probabilities of each of the real defect matrices, where each of the real defect matrices is obtained based on each integrated circuit real defect pattern, the real defect matrix and the artificial defect matrix are matrices of the same type, and the value ranges of their elements are the same; Using each of the predicted artificial authenticity probabilities, each of the predicted actual authenticity probabilities, a discrimination single threshold, a discrimination overall threshold, a matrix single threshold, and a matrix overall threshold, alternately fixing one of the current matrix model and the current pattern authenticity discrimination model, and optimizing the other to obtain a new current model of the other until the overall training requirements are met, obtaining the trained current matrix model, where the overall training requirements include that the difference between the matrix objective functions of the matrix model before and after optimization is less than the matrix overall threshold and the difference between the discrimination objective functions of the discrimination model before and after optimization is less than the discrimination overall threshold, or the overall matrix training cycle reaches a predetermined matrix overall cycle and the overall discrimination training cycle reaches a predetermined discrimination overall cycle.

2. The training method for an artificial defect pattern generation model according to claim 1, characterized in that, The step of using each of the predicted artificial authenticity probabilities and each of the predicted actual authenticity probabilities to alternately fix one of the current matrix model and the current pattern authenticity discrimination model, and optimizing the other to obtain a new current model of the other until the overall training requirements are met, obtaining the trained current matrix model includes: Using each of the predicted actual authenticity probabilities and each of the predicted artificial authenticity probabilities obtained from the artificial defect matrix based on the current matrix model to optimize the pattern authenticity discrimination model until the difference between the discrimination objective functions of the discrimination model before and after optimization is less than the discrimination single threshold, or reaching a predetermined single discrimination training number, fixing the pattern authenticity discrimination model to obtain a new pattern authenticity discrimination model, and using the new pattern authenticity discrimination model as the current pattern authenticity discrimination model; When the overall discrimination requirements are not met, using each of the predicted artificial authenticity probabilities obtained from the current pattern authenticity discrimination model and each of the predicted actual authenticity probabilities to optimize the matrix model until the difference between the matrix objective functions of the matrix model before and after optimization is less than the matrix single threshold, or reaching a predetermined single matrix training number, fixing the matrix model to obtain a new matrix model, and using the new matrix model as the current matrix model. When the overall discrimination requirements are not met, turn to execute the optimization of the pattern authenticity discrimination model.

3. The training method for an artificial defect pattern generation model according to claim 2, characterized in that, The optimization direction when optimizing the pattern authenticity discrimination model is opposite to the optimization direction when optimizing the matrix model of the artificial defect pattern generation model.

4. The training method for an artificial defect pattern generation model according to claim 1, characterized in that, The elements of the artificial defect matrix and the elements of the true defect matrix each have only two values, namely the first value and the second value.

5. The training method for an artificial defect pattern generation model according to claim 4, characterized in that, The integrated circuit true pattern includes the integrated circuit design pattern; The step of obtaining the respective true defect matrices of the respective true defect patterns includes: Classifying the true defect pattern blocks into transmission pattern blocks or masking pattern blocks, and corresponding all the elements in the true defect matrix to the true defect pattern blocks in the integrated circuit design pattern; Assigning the value of the element corresponding to the transmission pattern block as the first value; Assigning the value of the element corresponding to the masking pattern block as the second value.

6. The training method for an artificial defect pattern generation model according to claim 4, characterized in that, The integrated circuit true pattern includes the integrated circuit true image, and the step of obtaining the respective true defect matrices of the respective true defect patterns includes: Corresponding all the elements in the true defect matrix to the true defect pattern blocks in the integrated circuit true image; Classifying the true defect pattern blocks into white pattern blocks or black pattern blocks, where the gray level of the white pattern blocks is less than the gray level threshold, the gray level of the black pattern blocks is greater than the gray level threshold, the gray level threshold is less than the maximum gray level of the integrated circuit true image, and greater than the minimum gray level of the integrated circuit true image; Assigning the value of the element corresponding to the white pattern block as the first value; Assigning the value of the element corresponding to the black pattern block as the second value.

7. The training method for an artificial defect pattern generation model according to claim 1, characterized in that, The values of the elements of the artificial defect matrix and the true defect matrix are all integers, and are greater than or equal to the lower limit of the value and less than or equal to the upper limit of the value.

8. The training method for an artificial defect pattern generation model according to claim 7, characterized in that, The integrated circuit true pattern includes the integrated circuit true image, and the step of obtaining the respective true defect matrices of the respective true defect patterns includes: One-to-one corresponding all the elements in the true defect matrix to the true defect pattern blocks in the integrated circuit true image; Assigning all the elements in the true defect matrix as the gray level values of the true defect pattern blocks, and the range of the gray level values is less than or equal to the maximum gray level of the integrated circuit true image and greater than or equal to the minimum gray level of the integrated circuit true image.

9. The training method for an artificial defect pattern generation model according to any one of claims 1-7, characterized in that, The matrix model and the graphic authenticity discrimination model are convolutional neural networks.

10. The training method for an artificial defect pattern generation model according to any one of claims 1-3, characterized in that, The artificial defect matrix and the true defect matrix are vectors.

11. The training method of the artificial defect pattern generation model according to claim 10, wherein, The matrix model and the graphic authenticity discrimination model are fully connected neural networks.

12. The training method of the artificial defect pattern generation model according to claim 1, wherein, The random matrix and the artificial defect matrix are matrices of the same type.

13. The training method of the artificial defect pattern generation model according to claim 1, wherein, The value ranges of the random matrix and the artificial defect matrix are the same.

14. The training method of the artificial defect pattern generation model according to any one of claims 1-8, 11-13, wherein, Further includes: Using the graphic generation model of the artificial defect graphic generation model to obtain the predicted true defect graphics of the respective true defect matrices; Using the respective integrated circuit true defect graphics and the respective predicted true defect graphics to obtain the graphic similarity, and adjusting the parameters of the graphic generation model according to the graphic similarity until the image similarity meets the similarity threshold, to obtain the trained graphic generation model.

15. An artificial defect pattern generation method, wherein, Includes: According to each random matrix, using the matrix model of the artificial defect graphic generation model trained by the training method of the artificial defect graphic generation model according to any one of claims 1-13, to generate a predicted artificial defect matrix; According to the predicted artificial defect matrix, an integrated circuit artificial defect pattern is obtained by using the pattern generation model of the artificial defect pattern generation model.

16. An artificial defect pattern generation method, wherein, Including: According to each random matrix, a predicted artificial defect matrix is generated by using the matrix model of the artificial defect pattern generation model trained by the training method of the artificial defect pattern generation model as described in claim 14. According to the predicted artificial defect matrix, an integrated circuit artificial defect pattern is obtained by using the pattern generation model of the artificial defect pattern generation model trained by the training method of the artificial defect pattern generation model as described in claim 14.

17. An artificial defect pattern generation model training device, wherein, Including: An artificial defect matrix acquisition unit, which obtains each artificial defect matrix according to each random matrix by using the current matrix model of the artificial defect pattern generation model to be trained, and the artificial defect matrix is suitable for generating an integrated circuit artificial defect pattern. A authenticity probability acquisition unit, which is suitable for inputting each of the artificial defect matrices and each of the obtained real defect matrices into the same current pattern authenticity discrimination model respectively, and obtaining the predicted artificial authenticity probability of each of the artificial defect pattern matrices and the predicted actual authenticity probability of each of the real defect matrices, wherein each of the real defect matrices is obtained based on each integrated circuit real defect pattern, the real defect matrix and the artificial defect matrix are matrices of the same type, and the value ranges of their elements are the same. A matrix model acquisition unit, which is suitable for using each of the predicted artificial authenticity probabilities and each of the predicted actual authenticity probabilities to alternately fix one of the current matrix model and the current pattern authenticity discrimination model, and optimize the other one to obtain a new current model of the other one until the overall training requirements are met, and obtaining the trained current matrix model, wherein the overall training requirements include that the overall matrix training period reaches a predetermined overall matrix period and the overall discrimination training period reaches a predetermined overall discrimination period.

18. An artificial defect pattern generation device, wherein, Including: A predicted artificial defect matrix acquisition unit, which is suitable for generating a predicted artificial defect matrix according to each random matrix by using the matrix model of the artificial defect pattern generation model trained by the training method of the artificial defect pattern generation model as described in any one of claims 1-14. An artificial defect pattern acquisition unit, which is suitable for obtaining an integrated circuit artificial defect pattern according to the artificial defect matrix by using the pattern generation model of the artificial defect pattern generation model.

19. An artificial defect pattern generation device, wherein, Including: A predicted artificial defect matrix acquisition unit, which is suitable for generating a predicted artificial defect matrix according to each random matrix by using the matrix model of the artificial defect pattern generation model trained by the training method of the artificial defect pattern generation model as described in claim 14. An artificial defect pattern acquisition unit, which is suitable for obtaining an integrated circuit artificial defect pattern according to the predicted artificial defect matrix by using the pattern generation model of the artificial defect pattern generation model trained by the training method of the artificial defect pattern generation model as described in claim 14.

20. A storage medium, wherein, The storage medium stores a program for a training method suitable for training an artificial defect pattern generation model to implement the training method of the artificial defect pattern generation model according to any one of claims 1-14, or stores a program suitable for generating an artificial defect pattern of an integrated circuit to implement the artificial defect pattern generation method according to claim 15 or 16.

21. A device, wherein, It includes at least one memory and at least one processor; the memory stores a program, and the processor calls the program to execute the training method of the artificial defect pattern generation model according to any one of claims 1-14 or execute the artificial defect pattern generation method according to any one of claims 15 or 16.

Images