Optical Proximity Correction Method and System, Mask, Equipment and Storage Medium

Abstract

The present invention provides an optical proximity correction method and system, a mask, a device, and a storage medium. The optical proximity correction method includes: providing an optical proximity correction model, where the optical proximity correction model includes a parameter set; optimizing the parameter set based on a machine learning method, and through machine learning training, screening out optimized parameters that have an impact on local feature size problems, so that the optimized parameters and parameter values can be adjusted for local feature size errors. The present invention can optimize the parameter set in the optical proximity correction model, enabling the optical proximity correction model to produce a better correction effect on local anomalies in graphic feature sizes and improving the efficiency of optical proximity correction.

Description

Technical Field

[0001] The present invention relates to the field of semiconductors, and particularly to an optical proximity correction method and system, a mask, a device, and a storage medium. Background Art

[0002] With the continuous reduction of feature sizes and the increasing complexity of patterns, optical proximity correction (OPC) technology has been widely applied to the mask design of each critical layer. Currently, the most widely used OPC method is the model-based OPC correction method. Specifically, by establishing certain types of physical models to simulate the interaction of light sources, optical components, light propagation, and light matter, the pattern profile and feature size on the wafer are predicted as accurately as possible.

[0003] A typical OPC model usually includes three sub-models: an optical model, a mask three-dimensional effect model (DDM model), and a photoresist model. The optical model simulates the image projected onto the resist, the DDM model simulates the light wave in the mask, and the photoresist model simulates the resist profile after development. Compared with the optical model and the DDM model, the physical mechanism of the photoresist model is unclear and difficult to quantify. Therefore, the process of adjusting the optimal parameters requires a large amount of data and computational effort. For example: the photoresist model usually requires adjusting dozens of parameters, and traditional methods simulate thousands of iterations (usually the number of iterations > 20,000), and then select relatively better iterations from them, and each iteration has its unique parameter setting.

[0004] However, the standard for the system to judge the quality of parameter settings is based on the model error. The model error is the difference between the overall feature sizes between the simulated feature size and the scanning electron microscope feature size, where the simulated feature size is the simulated feature size of the photoresist model, and the scanning electron microscope feature size is the feature size measured by the scanning electron microscope for the test pattern on the wafer.

[0005] Figure 1 It is a schematic diagram of the measurement of the feature size error in the anti-reagent model. Among them, the first strip pattern 011 and the second strip pattern 021 show the schematic diagram of the measurement of the feature size of the anti-reagent model in the prior art. In the prior art, the anti-reagent model can Figure 1 have a relatively good model error for the first feature size CD1, the second feature size CD2, and the third feature size CD3 in Figure 1There is a problem of abnormal graphics in the third strip-shaped graphic 010 and the fourth strip-shaped graphic 020. Since the graphics are irregular after the tops of the third strip-shaped graphic 010 and the fourth strip-shaped graphic 020 are formed on the wafer, there is a problem of abnormal value in the fourth characteristic dimension CD4 between the third strip-shaped graphic 010 and the fourth strip-shaped graphic 020. It is difficult to consider the defects caused by such local shapes in the adjustment of the photoresist model parameters in the downstream, which leads to the selected parameter settings optimized by the photoresist model simulation being easily unqualified, and further makes it difficult to optimize the above-mentioned graphic abnormal problems.

[0006] The current method has the problem that the results of optical proximity correction output are easily unqualified. Summary of the Invention

[0007] The problem solved by the present invention is to provide a method for an optical proximity correction model, an optical proximity correction system, and a mask plate to optimize the optical proximity correction effect.

[0008] To solve the above problems, the present invention provides a method for an optical proximity correction model, an optical proximity correction system, and a mask plate. The method for the optical proximity correction model includes: providing an optical proximity correction model, where the optical proximity correction model includes a parameter set; optimizing the parameter set based on a machine learning method, and through machine learning training, screening out the optimization parameters that have an impact on the local feature size problem, so that the optimization parameters and parameter values can be adjusted for the local feature size error.

[0009] Correspondingly, an embodiment of the present invention further provides an optical proximity correction system for performing optical proximity correction on the original layout graphics, including: a model establishment unit for establishing an optical proximity correction model; an optical proximity correction model for performing optical proximity correction, where the optical proximity correction model includes a parameter set; a correction unit, and the correction unit optimizes the parameter set based on a machine learning method, and through machine learning training, screens out the optimization parameters that have an impact on the local feature size problem, so that the optimization parameters and parameter values can be adjusted for the local feature size error.

[0010] Correspondingly, an embodiment of the present invention further provides a mask plate, and the graphics on the mask plate are obtained by the optical proximity correction system provided by the present invention.

[0011] Correspondingly, an embodiment of the present invention further provides a device, including: at least one memory and at least one processor, where the memory stores one or more computer instructions, and wherein the one or more computer instructions are executed by the processor to implement the optical proximity correction method provided by the present invention.

[0012] Accordingly, an embodiment of the present invention further provides a storage medium storing one or more computer instructions for implementing the optical proximity correction method provided by the present invention.

[0013] Compared with the prior art, the technical solution of the present invention has the following advantages:

[0014] By using the method of the optical proximity correction model provided by the present invention, through machine learning training, parameters that affect the local feature size problem are screened out, and the parameters and parameter values are adjusted for the local feature size error, which can optimize the parameter set in the optical proximity correction model, enabling the optical proximity correction model to have a better correction effect on the local abnormal problem of the graphic feature size, reducing the feature size error of the optical proximity correction model, and improving the efficiency of optical proximity correction, thereby improving the yield of chip production.

[0015] In an alternative solution, a decision tree classifier is used to calculate the adjustment range of each parameter value in the parameter set data, and the parameters in the parameter set are sorted according to the importance to form a parameter sequence. The top n parameters and their adjustment ranges in the parameter sequence are selected to establish an optimized parameter set, which is used for the test simulation of the optical proximity correction model to reduce the feature size error of the optical proximity correction model. The decision tree classifier is an advanced machine learning training method that can accurately find the parameters that have a greater impact on the local abnormal problem of the feature size, and can further effectively improve the correction effect on the local abnormal problem of the graphic feature size. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a schematic diagram of measuring the feature size error in the photoresist model;

[0017] Figure 2 is a step diagram of optimizing the parameter set in an embodiment of the optical proximity correction method of the present invention;

[0018] Figure 3 is Figure 2 a schematic diagram of defining the feature size measurement rule in the step of optimizing the parameter set;

[0019] Figure 4 is a schematic diagram of using a decision tree classifier to calculate the adjustment range of each parameter value in the parameter set data in an embodiment of the present invention;

[0020] Figure 5 is a schematic diagram of an embodiment of the optical proximity correction system of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] As described in the background art, when using the photoresist model of the prior art for optical proximity correction, for the patterns formed on the wafer, although the overall feature differences are within an acceptable range, local shapes are prone to defects. The abnormal problems caused by local shape defects are not considered in the adjustment of photoresist model parameters. This is likely to result in unqualified problems when setting the parameters selected for the simulation optimization of the photoresist model, thereby making it difficult to optimize the above-mentioned pattern abnormal problems.

[0022] The main method for currently adjusting the photoresist model parameter set [P1, P2, P3, P4, P5, …, Pn] is based on the model error function, and the iterative calculation method is used to test and simulate the parameter set. When adjusting the parameters, the parameter set will first be started with some default values, and each parameter in the parameter set has its unique adjustment range. The optical proximity correction process runs tens of thousands of iterations to reduce the model error by optimizing the parameter set.

[0023] Table 1 shows an example of the test simulation results of a photoresist model. The first column is the number of iterations, the second column is the model performance judgment criterion value, that is, the model error. The columns from P1 to P10 are the photoresist model parameter values. 23,000 iterations were calculated in Table 1. It can be seen that as the iterative calculation continues, the model error becomes smaller and smaller. After the 23,000th iteration, the model error is 5.44, and the parameter values of the parameter set [P1, P2, P3, P4, P5, …, P10] can make the optical proximity correction have acceptable overall feature differences. However, the model error at this time only reflects the overall feature size error of the photoresist model, and for Figure 1 the local shape defects shown, the existing test simulation results cannot reflect such defects, and it is still difficult to correct them. Moreover, the number of test simulation iterations shown in Table 1 is too large, time-consuming, and inefficient.

[0024] Number of iterations Model error P1 P2 P3 P4 P5 P6 P7 P8 P9 P10 1 44.70 200.00 0.44 115.99 0.25 124.01 0.46 37.46 182.93 113.96 0.29 2 8.08 168.90 0.39 168.90 0.39 118.90 0.54 118.90 118.90 118.90 0.54 3 7.80 273.78 0.39 168.90 0.39 118.90 0.54 118.90 118.90 118.90 0.54 … 23000 5.44 180.37 0.16 37.80 0.21 79.92 0.38 167.82 93.56 190.46 0.30

[0025] Table 1

[0026] Therefore, an embodiment of the present invention provides an optical proximity correction method to optimize the parameter set in the optical proximity correction model, enabling the optical proximity correction model to produce a better correction effect on the local abnormal problems of pattern feature sizes, improving the efficiency of optical proximity correction, and further improving the yield of chip production. The method of the optical proximity correction model in this embodiment includes:

[0027] Providing an optical proximity correction model, the optical proximity correction model including a parameter set;

[0028] Optimizing the parameter set based on a machine learning method, and through machine learning training, screening out the optimized parameters that have an impact on the local feature size problems, so that the optimized parameters and parameter values can be adjusted for the local feature size error.

[0029] By using the optical proximity correction method provided by the present invention, through machine learning training, parameters that affect the local feature size problem are screened out, and the parameters and their values are adjusted for the local feature size error, so as to optimize the parameter set in the optical proximity correction model, enabling the optical proximity correction model to produce a better correction effect on the local abnormal problem of the graphic feature size, reducing the feature size error of the optical proximity correction model, and improving the efficiency of optical proximity correction, thereby improving the yield of chip production.

[0030] Please refer to Figure 2 , which is a step diagram for optimizing the parameter set in an embodiment of the optical proximity correction method of the present invention. In this embodiment, based on the machine learning method, the steps for optimizing the parameter set include:

[0031] Step S1, defining a key feature size measurement rule;

[0032] Step S2, performing a first test simulation on the parameter set based on the model error function. The first test simulation includes multiple iterative calculations, and according to the key feature size measurement rule, parameter data in the parameter set is collected;

[0033] Step S3, defining an abnormal problem judgment rule; according to the abnormal problem judgment rule, an iteration mark is generated for each iteration, with the iteration mark without abnormal problems marked as normal and the iteration mark with abnormal problems marked as abnormal;

[0034] Step S4, using the parameter set data and the corresponding iteration mark of each iteration as the input data for machine learning;

[0035] Step S5, based on the input data, using a decision tree classifier to calculate the adjustment range of each parameter value in the parameter set data;

[0036] Step S6, sorting the parameters in the parameter set according to importance to form a parameter sequence, and selecting the top n parameters and their adjustment ranges in the parameter sequence to establish an optimized parameter set. The optimized parameter set is used for the second test simulation of the optical proximity correction model to determine the parameter values of the optimized parameter set and reduce the feature size error of the optical proximity correction model.

[0037] To make the above objects, features, and advantages of the present invention more obvious and understandable, the method of the optical proximity correction model in this embodiment will be further described below with reference to the accompanying drawings.

[0038] Specifically, referring to Figure 3 , for Figure 2 is a schematic diagram for defining the feature size measurement rule in step S1 shown. First, step S1 is executed to define a key feature size measurement rule.

[0039] The optical proximity correction method of this embodiment is used to improve the accuracy of the pattern on the mask corresponding to the wafer forming. The pattern on the mask includes Figure 3 the head-to-head region shown.

[0040] In this embodiment, a machine learning method is adopted to adjust the parameter set in the optical proximity correction model. In this embodiment, the model error function incorporates the size information of the pattern on the mask and its size calculation method. At each iteration, the local critical dimension anomaly problem in the head-to-head region is simultaneously simulated and calculated. The anomaly problem and the parameters of the parameter set are unified as the training data of machine learning, and an optical proximity correction model with the effect of excluding local critical dimension anomaly problems can be trained.

[0041] Compared with the prior art without a solution optimized for local abnormal critical dimensions, the result output by the optical proximity correction is better.

[0042] Specifically, in this embodiment, the steps of defining the critical dimension measurement rule include: incorporating the size information of the pattern on the mask and its size calculation method through the model error function, and measuring multiple points in the head-to-head region in the simulation. For example: measuring a center point at the center of the head-to-head region, and measuring multiple other points at equal distances or randomly distributed on both sides of the center point, and measuring the critical dimension between the two heads for each point.

[0043] Taking Figure 3 the shape of the head-to-head region shown as an example, measuring five points in the head-to-head region, where the fifth critical dimension CD0 between the two heads is measured at the center point, and the critical dimensions between the two heads of the four points measured at equal distances on both sides of the center point are the sixth critical dimension CDm, the seventh critical dimension CDj, the eighth critical dimension CDk, and the ninth critical dimension CDn respectively. The above critical dimensions are all critical dimensions reflecting Figure 3 the local anomaly problem of the head-to-head region shown.

[0044] In other embodiments, due to the different specific shapes of the head-to-head region anomaly problems, other positions and numbers of critical dimensions in the head-to-head region can also be simulated and measured, and the present invention does not limit this.

[0045] Execute step S2, perform a first test simulation on the parameter set based on the model error function. The first test simulation includes multiple iterative calculations, and according to the critical dimension measurement rule, collect the parameter data in the parameter set. Taking the parameter set data as the analysis data, it should be noted that in this embodiment, in addition to obtaining the parameter values of the parameter set for each iterative calculation, the corresponding critical dimensions are also obtained.

[0046] Specifically, please refer to Table 2, which shows the results of testing and simulating the optical proximity correction model in this embodiment. In this embodiment, the parameter set includes ten parameters, namely P1, P2, P3, P4, P5, … P10. However, the present invention does not limit the number of parameters in the parameter set. In addition, in this embodiment, the characteristic size results measured in the head-to-head region are added in the right column of the table, including the fifth characteristic size CD0, the sixth characteristic size CDm, the seventh characteristic size CDj, the eighth characteristic size CDk, and the ninth characteristic size CDn. As Figure 3 shown, the added fifth to ninth characteristic sizes can represent the pattern defects in the head-to-head region. However, the present invention does not limit the type and position of the added characteristic sizes. In other embodiments, other types and positions of characteristic sizes can also be added to optimize the local characteristic size abnormality problems in different regions.

[0047] It should be noted that 23,000 iterations of calculation are performed in this embodiment. In other embodiments, the number of iterations of calculation can also be set according to the requirements of simulation accuracy.

[0048]

[0049] Table 2

[0050] Step S3, define the abnormal problem judgment rule. According to the abnormal problem judgment rule, mark the iterations without abnormal problems as normal (for example: the mark is 0), and mark the iterations with abnormal problems as abnormal (for example: the mark is 1).

[0051] It should be noted that in this embodiment, the defined abnormal problem judgment rule is: if the characteristic size of the center point is less than the characteristic size of any other point, it is determined that there is no abnormal problem, and the iteration without abnormal problems is marked as normal, with the mark of 0; if the characteristic size of the center point is not less than the characteristic size of any other point, it is determined that an abnormal problem occurs, and the iteration with the abnormal problem is marked as abnormal, with the mark of 1.

[0052] Specifically, please refer to Table 3, which shows the results of testing and marking the optical proximity correction model in this embodiment. An iteration mark for each iteration is added in the right column of the table. In this embodiment, the fifth feature size CD0 is the feature size of the center point. As can be seen from Table 3, in the first iteration calculation, the fifth feature size CD0 is -1, which is less than or equal to the sixth feature size CDm, the seventh feature size CDj, the eighth feature size CDk, and the ninth feature size CDn in the first iteration calculation. Therefore, this embodiment determines that there is no abnormal problem in the first iteration calculation and marks it as 0. In the second iteration calculation, the fifth feature size CD0 is 195.52, which is greater than the sixth feature size CDm in the first iteration calculation, which is 185.33. Therefore, this embodiment determines that there is no abnormal problem in the second iteration calculation and marks it as 1.

[0053] Defining the abnormal problem judgment rule and marking the iterations with and without abnormal problems respectively serves to better provide input data for machine learning training. However, the specific manner in which the present invention provides input data for machine learning training is not limited.

[0054] It should be noted that the abnormal problem judgment rule in this embodiment is for local feature size abnormal problems, and the strictness of the judgment rule in this embodiment is moderate. While being able to exclude most local feature size abnormal problems, it ensures the reliability of the data. In other embodiments, more strict or more relaxed judgment rules can also be set according to the specific situation of the feature size abnormal problems.

[0055]

[0056] Table 3

[0057] Execute step S4, and use the parameter set data of each iteration and the corresponding iteration mark as the input data for machine learning.

[0058] It should be noted that to improve the training accuracy and efficiency of machine learning, in this embodiment, before using the parameter set data of each iteration and the corresponding iteration mark as the input data for machine learning, the parameter set data is subjected to data cleaning to remove the results with model errors exceeding the threshold. In other embodiments, the step of data cleaning may not be performed.

[0059] Specifically, an error threshold δ is defined for the model error value to remove the iteration results with model error > δ. In this embodiment, the value range of the error threshold δ is set between 2 and 30. If the error threshold δ is too large, there will be more iteration results with larger retained errors, which may reduce the reliability of the results of the first test simulation; if the error threshold δ is too small, it may lead to too little input data volume, affecting the accuracy of machine learning training. However, the present invention does not limit this. In other embodiments, the value range of the error threshold δ can be adjusted according to the number of iterations and the situation of local anomaly problems.

[0060] In this embodiment, after removing the iteration results with model error > δ, the remaining results are sampled to balance the number of iteration times with anomaly problems and the number of iteration times without anomaly problems in the sample, so as to improve the accuracy of machine learning training.

[0061] Specifically, the sampling method in this embodiment includes: Let the number of iteration times with anomaly problems before sampling be Np1, the number of iteration times without anomaly problems before sampling be Nn1, the number of iteration times with anomaly problems after sampling be Np2, and the number of iteration times without anomaly problems after sampling be Nn2.

[0062] In the case where Np1 is less than Nn1, if Nn1 ≤ α * Np1, all samples are taken; if Nn1 > α * Np1, then randomly sample all the iteration times with anomaly problems. After sampling, the number of anomaly iteration times Np2 is equal to the number of anomaly iteration times Np1 before sampling, and make the number of non - anomaly iteration times Nn2 after sampling = α * Np2.

[0063] In the case where Np1 is greater than Nn1, if Np1 ≤ α * Nn1, all samples are taken. If Np1 > α * Nn1, then randomly sample all the iteration times without anomaly problems. After sampling, the number of non - anomaly iteration times Nn2 is equal to the number of non - anomaly iteration times Nn1 before sampling, and make Np2 = α * Nn2.

[0064] Through the sampling method of this embodiment, when the difference between the number of iteration times with anomaly problems Np1 before sampling and the number of iteration times without anomaly problems Nn1 before sampling is not large, all samples are taken, ensuring the comprehensiveness and integrity of the sampled data. When the difference between the number of iteration times with anomaly problems Np1 before sampling and the number of iteration times without anomaly problems Nn1 before sampling is large, based on the item with the smaller number of samples in Np1 and Nn1, the quantity of the other item is adjusted so that the number of iteration times with anomaly problems Np2 after sampling and the number of iteration times without anomaly problems Nn2 after sampling maintain a gap of α times, thereby making the number of iteration times with anomaly problems in the sample and the number of iteration times without anomaly problems in the sample balanced.

[0065] Among them, α is the data balance threshold. In this embodiment, the value range of α is α ≤ 20, but the present invention does not limit this.

[0066] Using the sampling method of this embodiment, when the number of iterations Np1 with abnormal problems before sampling and the number of iterations Nn1 without abnormal problems before sampling differ greatly, making the number of iterations Np2 with abnormal problems after sampling and the number of iterations Nn2 without abnormal problems after sampling relatively balanced can effectively improve the training accuracy and efficiency of machine learning. In other embodiments, other methods can also be used for sampling.

[0067] Execute step S4, and use the parameter set data and the corresponding iteration mark of each iteration as the input data of machine learning.

[0068] Specifically, in this embodiment, each iteration is recorded as a feature vector with the parameter data as the independent variable and the parameter set mark as the dependent variable. As shown in Table 4, the parameter values of the parameter set [P1, P2, P3, P4, P5,..., P10] are used as the independent variable X, and the parameter set mark part is used as the dependent variable Y. The advantage of this is that it is convenient for data collection in machine learning training. In other embodiments, other input methods can also be used as the input data of machine learning.

[0069]

[0070] Table 4

[0071] Execute step S5, and use a decision tree classifier to calculate the adjustment range of each parameter value in the parameter set data according to the input data.

[0072] It should be noted that, in this embodiment, before using a decision tree classifier to calculate the adjustment range of each parameter value in the parameter set data, a random forest classifier is used to calculate the importance level of each parameter in the parameter set data according to the input data. The advantage of this is that it can be compared with the results of the decision tree classifier, further verifying the importance level of each parameter in the parameter set data and improving the accuracy and efficiency of machine learning.

[0073] Table 5 shows the importance of the parameter set [P1, P2, P3, P4, P5,..., P10] obtained by using a random forest classifier in this embodiment. It can be seen that, in this embodiment, the parameter P6 has the highest importance and the parameter P2 has the lowest importance.

[0074] Parameter P1 P2 P3 P4 P5 P6 P7 P8 P9 P10 Importance 0.0349 0.0127 0.0318 0.1347 0.1607 0.2890 0.0352 0.0736 0.0174 0.2100

[0075] Table 5

[0076] After calculating the importance level of each parameter in the parameter set data according to the input data using a random forest classifier, a decision tree classifier is used to calculate the adjustment range of each parameter value in the parameter set data according to the input data.

[0077] Specifically, referring to Figure 4 , a schematic diagram showing the calculation of the adjustment range of each parameter value in the parameter set data using a decision tree classifier in this embodiment is shown.

[0078] In this embodiment, the decision tree includes multiple levels, and each level includes one or more nodes. The steps of calculating the adjustment range of the parameter values in the parameter set data using a decision tree classifier include: according to the parameter set data and the corresponding iteration mark, using a decision tree classifier, with the parameters in the parameter set as nodes, obtaining an output result regarding the parameters and the parameter adjustment range.

[0079] Specifically, as Figure 4 shown, in this embodiment, first, the input data in the parameter set data includes 19,511 samples. These 19,511 samples are from the 23,000 - time iterative calculations in step S2. After data cleaning, 19,511 samples are left. The number of these samples is related to the settings of the model error function and the local feature size.

[0080] Using a decision tree classifier, it is set to be divided into multiple operation levels, and each level includes 1 to multiple nodes. Taking the parameter value range of a certain parameter in the parameter set [P1, P2, P3, P4, P5, …, P10] as alternative nodes, the adjustment range of each parameter in the parameter set [P1, P2, P3, P4, P5, …, P10] is determined.

[0081] Substitute the feature vector representing the parameter set data. The feature vector takes each iteration record as the independent variable with parameter data and the parameter set mark as the dependent variable. The nodes of the decision tree in this embodiment include three levels: the first level, the second level, and the third level. Among them, the first level includes one root node, the second level includes two nodes, and the third level includes four nodes. It should be noted that the levels of the decision tree can be freely set by the decision tree classifier. The more levels there are, the more accurate the final result will be. In this embodiment, setting it to three levels can already obtain accurate results, but the present invention does not limit this.

[0082] After being set to multiple operation levels, the decision tree classifier can calculate and set the parameters and parameter ranges of the nodes at each level through an autonomous operation method trained by machine learning, obtaining the result with the most significant difference between the samples labeled 0 and those labeled 1 within three levels. Furthermore, when the number of samples labeled 0 is the largest, the parameters and parameter ranges of the nodes at each level are obtained. That is to say, the decision tree classifier plays the role of automatically finding the optimal path in this embodiment to obtain the parameter setting method with the lowest probability of local feature size abnormality problems.

[0083] In this embodiment, the first level includes the first node 101, which is also the root node. In this embodiment, after the classification calculation of the decision tree classifier, the optimal decision is to use the decision condition parameter P6 ≤ 0.527 as the root node. For the first node 101 with P6 ≤ 0.527, 9152 samples are labeled 0 and 10359 samples are labeled 1. The left branch of the first node 101 in the figure indicates that if P6 ≤ 0.527 is true, 1052 samples are labeled 0 and 9731 samples are labeled 1. The right branch of the first node 101 in the figure indicates that if P6 ≤ 0.527 is false, 8100 samples are labeled 0 and 616 samples are labeled 1.

[0084] The second level includes the second node 102A and the third node 102B. After the calculation of the decision tree classifier, the second node 102A is the parameter P6 ≤ 0.354. The left branch of the first node 101 in the figure is under the condition of satisfying P6 ≤ 0.527, with 1052 samples labeled 0 and 9731 samples labeled 1.

[0085] The third node 102B is the parameter P10 ≤ 0.327. The right branch of the first node 101 in the figure is when P6 is greater than 0.527, with 8100 samples labeled 0 and 616 samples labeled 1. Therefore, if P6 ≤ 0.527 is false, this path is selected to enter the third node 102B, making the number of samples labeled 0 more and the number of samples labeled 1 less.

[0086] At the second level, according to the decision tree classifier, under the condition that P6 ≤ 0.527 is false, the situation without abnormal problems is greater than that with abnormal problems. The right branch of the third node 102B in the figure is when P10 ≤ 0.327 is false, and there are still 6094 samples left. Among them, 5941 samples are labeled 0 and 153 samples are labeled 1. The number of 0 samples in the other several branches at the second level is less than that of the right branch of the third node 102B. Therefore, this path is selected to enter the fourth node 103D, making the number of samples labeled 0 more and the number of samples labeled 1 less. The other branches are not shown.

[0087] Therefore, at the nodes of the third level, the situation without abnormal problems is much greater than that with abnormal problems.

[0088] The third level includes the fourth node 103D and the other three nodes not shown. The fourth node 103D has the parameter P5 ≤ 82.919. The right branch of the fourth node 103D in the figure has 5945 remaining samples when P6 ≤ 0.527 is false, P10 ≤ 0.327 is false, and P5 ≤ 82.919 is false. Among them, 5940 samples are labeled as 0 and 5 samples are labeled as 1.

[0089] In this way, through the decision tree classification of three levels, the parameter setting values of 5940 samples labeled as 0 and 5 samples labeled as 1 are obtained. The samples labeled as 0 are far more than those labeled as 1, that is, the probability of no local feature size abnormal problem is far greater than that of having a local feature size abnormal problem. That is to say, in this embodiment, P6, P10, and P5 are important parameters for excluding abnormal problems. When their parameter values are set such that P6 > 0.527, P10 > 0.327, and P5 > 82.919, and then the second test simulation is carried out, the probability of local feature size abnormal problems will be greatly reduced, and the local abnormal problems in the head-to-head area can be basically excluded in the simulation iterative calculation.

[0090] Therefore, by using the decision tree classifier, the important parameters and the parameter value adjustment ranges in the parameter set data can be calculated. Taking this embodiment as an example, the important parameters are P6, P10, and P5, and their adjustment ranges are P6 > 0.527, P10 > 0.327, and P5 > 82.919.

[0091] At the same time, due to the reduction of the number of parameters, the number of iterative calculations is greatly reduced, that is, the time-consuming of the simulation calculation is reduced, and the probability of local feature size abnormal problems in the head-to-head area is effectively reduced.

[0092] It should be noted that in this factual example, the important parameters obtained by the decision tree classifier are P6, P10, and P5, which are three nodes in three levels of the decision tree respectively. However, although the decision tree is divided into three levels, the number of important parameters is not necessarily 3. In fact, the three nodes in the three levels may also be the same parameter, but the adjustment range of the parameter becomes smaller and smaller from top to bottom in the three levels, which is related to the specific situation of the local feature size abnormal problem and the specific calculation of the decision tree classifier.

[0093] Next, as Figure 2 shown, in this embodiment, the matching degree between the output results of the decision tree classifier and the random forest classifier is measured;

[0094] If the output result of the decision tree classifier matches the output result of the random forest classifier, it is determined that the output result of the decision tree classifier is reliable;

[0095] If the output result of the decision tree classifier does not match the output result of the random forest classifier, the parameter set is re-adjusted based on the machine learning method, and the number of iterative calculations of the parameter set is increased in the step of performing the first test simulation on the optical proximity correction model.

[0096] Specifically, in this embodiment, the method for measuring the matching degree between the output result of the decision tree classifier and the output result of the random forest classifier includes: judging the matching degree between the top m nodes in the output result of the decision tree classifier and the top m important parameters in the output result of the random forest classifier, where the top m nodes in the output result of the decision tree classifier refer to the m nodes from the top down in the classification level of the decision tree, and the top m important parameters in the output result of the random forest classifier refer to the parameters ranked in the top m after calculating the importance levels of the parameters in the parameter set according to the random forest classifier. In this embodiment, m takes values in the range of 1% to 50% of the total number of parameters, but the present invention does not limit the value range of m.

[0097] As can be seen from Table 5, the top 3 important parameters in the output result of the random forest classifier are P5, P10, and P6, which are the same as the output result of the decision tree classifier. Therefore, in this embodiment, the output result of the decision tree classifier matches the output result of the random forest classifier, and it is determined that the output result of the decision tree classifier is reliable. In other embodiments, when the output result of the decision tree classifier does not match the output result of the random forest classifier, the parameter set is re-adjusted based on the machine learning method, and the number of iterative calculations of the parameter set is increased in the step of performing the test simulation on the optical proximity correction model.

[0098] It should be noted that this embodiment considers the output result of the decision tree classifier and the output result of the random forest classifier to improve the accuracy of the important parameters and their value ranges when adjusting the parameter set, but the present invention does not limit this. In other embodiments, only the decision tree classifier can also be used to achieve the effect of the present invention.

[0099] Execute step S6, sort the parameters in the parameter set according to the importance degree to form a parameter sequence, select the top n parameters and their adjustment ranges in the parameter sequence to establish an optimized parameter set, and the optimized parameter set is used for the second test simulation of the optical proximity correction model to determine the parameter values of the optimized parameter set and reduce the feature size error of the optical proximity correction model.

[0100] After re - determining the optimized parameter set according to the method of this embodiment, the second test simulation refers to the re - test simulation based on the model error function and the optimized parameter set. The second test simulation includes multiple iterative calculations. After the second test simulation, the exact numerical values of the parameters in the optimized parameter set can be further determined to further reduce the feature size error. When performing optical proximity correction subsequently, the second test simulation can be directly performed based on the model error function and the optimized parameter set.

[0101] Through such optimization, the number of parameters in the parameter set is reduced, the number of iterative calculations is reduced, the efficiency of the optical proximity correction model simulation is improved, and the data of the adjusted parameter set can better adapt to the judgment of local abnormal problems in the head - to - head area, resulting in a better correction effect on the local abnormal problems of the graphic feature size.

[0102] In this embodiment, the value of n ranges from 1% to 50% of the total number of parameters. For example, if the total number of parameters in this embodiment is 10 and the value of n is 3, but the present invention does not limit the value range of n. Generally, the value of n should be equal to the value of m in step S5, so as to make the best use of the results of machine learning. However, in order to improve the efficiency of the subsequent second test simulation, the value of n can also be less than the value of m.

[0103] By using the method of this embodiment, the parameter set in the optical proximity correction model can be optimized, the important parameters and the parameter value adjustment range in the parameter set data can be calculated, so that the optical proximity correction model can have a better correction effect on the local abnormal problems of the graphic feature size, reduce the number of iterative calculations, improve the efficiency of optical proximity correction, and thus improve the yield of chip production.

[0104] The present invention also provides an optical proximity correction system for performing optical proximity correction on the original layout pattern, including:

[0105] A model - building unit 200 for building an optical proximity correction model;

[0106] An optical proximity correction model 201 for performing optical proximity correction, and the optical proximity correction model 201 includes a parameter set;

[0107] A correction unit 202, and the correction unit 202 optimizes the parameter set based on the machine - learning method. Through machine - learning training, the optimized parameters that affect the local feature size problem are screened out, so that the optimized parameters and parameter values can be adjusted for the local feature size error.

[0108] In this embodiment, the correction unit 202 includes:

[0109] Critical feature size measurement rule module 203. The optical proximity correction system of this embodiment is mainly used to solve the problem of abnormal local feature sizes during the optical proximity correction process. Therefore, the critical feature size measurement rule module 203 of this embodiment measures multiple points in the head-to-head area during the simulation by means of the size information of the patterns on the built-in mask plate of the model error function and its size calculation method. For example: measure a central point in the center of the head-to-head area, and measure multiple other points at equal distances or randomly distributed on both sides of the central point, and measure the feature size between the two heads for each point.

[0110] Test simulation module 204. The test simulation module performs a first test simulation on the optical proximity correction model based on the model error function. The first test simulation includes multiple iterative calculations of the parameter set. According to the critical feature size measurement rule, the parameter set data is collected, and the parameter set data is used as analysis data. It should be noted that in this embodiment, in addition to obtaining the respective parameter values of the parameter set for each iterative calculation, the corresponding feature sizes are also obtained.

[0111] Abnormal problem judgment module 205. According to the abnormal problem judgment rule, the iterations without abnormal problems are marked as normal (for example: the mark is 0), and the iterations with abnormal problems are marked as abnormal (for example: the mark is 1).

[0112] It should be noted that in this embodiment, the defined abnormal problem judgment rule is: if the feature size of the central point is less than the feature size of any other point, it is determined that there is no abnormal problem, and the iterations without abnormal problems are marked as normal, with the mark being 0; if the feature size of the central point is not less than the feature size of any other point, it is determined that an abnormal problem has occurred, and the iterations with abnormal problems are marked as abnormal, with the mark being 1.

[0113] Machine learning module 206. The machine learning module is connected to the test simulation module and the abnormal problem judgment module, takes the parameter set data of each iteration and the corresponding iteration mark as the input data of machine learning, and uses a decision tree classifier to calculate the adjustment range of each parameter value in the parameter set data based on the input data.

[0114] Optimization module 207. The optimization module is connected to the machine learning module and the test simulation module, and is used to sort the parameters in the parameter set according to the importance to form a parameter sequence, select the top n parameters and their adjustment ranges in the parameter sequence to establish an optimized parameter set, and the optimized parameter set is used to perform a second test simulation on the optical proximity correction model based on the test simulation module to determine the respective parameter values of the optimized parameter set and reduce the feature size error of the optical proximity correction model.

[0115] After being optimized by the optical proximity correction system of this embodiment, important parameters and the adjustment ranges of parameter values in the parameter set data are calculated, reducing the number of parameters in the parameter set of the optical proximity correction model, improving the simulation efficiency of the optical proximity correction model, and enabling the adjusted parameter set data to be more adaptable to the judgment of local abnormal problems in the head-to-head area, resulting in a better correction effect on the local abnormal problems of the graphic feature size.

[0116] Therefore, the optical proximity correction system of the present invention can optimize the parameter set in the optical proximity correction model, enabling the optical proximity correction model to have a better correction effect on local abnormal problems of graphic feature size, improving the efficiency of optical proximity correction, and thus increasing the yield of chip production.

[0117] An embodiment of the present invention also provides a mask plate, and the graphics on the mask plate are obtained by the optical proximity correction system or the optical proximity correction method provided by the present invention.

[0118] The method of the optical proximity correction system or the optical proximity correction model of the present invention can optimize the parameter set in the optical proximity correction model, enabling the optical proximity correction model to have a better correction effect on local abnormal problems of graphic feature size, improving the efficiency of optical proximity correction, and thus increasing the yield of chip production.

[0119] An embodiment of the present invention also provides a device, and this device can implement the optical proximity correction method provided by the embodiment of the present invention by loading the above optical proximity correction method in the form of a program.

[0120] An optional hardware structure provided by the device in the embodiment of the present invention includes: at least one memory and at least one processor. The memory stores one or more computer instructions.

[0121] The processor and the memory can complete communication through one or more of a communication bus or a communication module interface.

[0122] Among them, the processor may be a central processing unit CPU, or a specific integrated circuit ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiment of the present invention.

[0123] The memory may include a high-speed RAM memory, and may also include non-volatile memory, such as at least one disk memory.

[0124] Among them, the memory stores one or more computer instructions, and the one or more computer instructions are executed by the processor to implement the optical proximity correction method provided by the embodiment of the present invention.

[0125] It should be noted that the above-mentioned implementation terminal device may further include other devices (not shown) that may not be necessary for the disclosed content of the embodiments of the present invention; in view of the fact that these other devices may not be necessary for understanding the disclosed content of the embodiments of the present invention, the embodiments of the present invention will not introduce them one by one.

[0126] The embodiments of the present invention further provide a storage medium, and the storage medium stores one or more computer instructions, and the one or more computer instructions are used to implement the optical proximity correction method provided by the embodiments of the present invention.

[0127] Although the present invention is disclosed as above, the present invention is not limited thereto. Any person skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention should be subject to the scope defined by the claims.

Claims

1. An optical proximity correction method for improving the problem of abnormal local feature sizes, characterized in that, Including: Providing an optical proximity correction model, the optical proximity correction model including a parameter set; Optimizing the parameter set based on a machine learning method. Through machine learning training, optimizing parameters that have an impact on local feature size problems are screened out, enabling the optimized parameters and parameter values to be adjusted for local feature size errors; Based on the machine learning method, the steps of optimizing the parameter set include: Defining a critical feature size measurement rule; Performing a first test simulation on the parameter set based on a model error function. The first test simulation includes multiple iterative calculations. According to the critical feature size measurement rule, parameter data in the parameter set is collected; Defining an abnormal problem judgment rule; according to the abnormal problem judgment rule, an iteration mark is generated for each iteration. An iteration mark without an abnormal problem is marked as normal, and an iteration mark with an abnormal problem is marked as abnormal; Using the parameter set data and the corresponding iteration mark of each iteration as input data for machine learning; Based on the input data, using a decision tree classifier to calculate the adjustment range of each parameter value in the parameter set data; Sorting the parameters in the parameter set according to importance to form a parameter sequence, and selecting the top n parameters and their adjustment ranges in the parameter sequence to establish an optimized parameter set. The optimized parameter set is used for a second test simulation of the optical proximity correction model to determine the parameter values of the optimized parameter set and reduce the feature size error of the optical proximity correction model.

2. The method according to claim 1, characterized in that, Before using the parameter set data and the corresponding iteration mark of each iteration as input data for machine learning, data cleaning is performed on the parameter set data to remove results with a model error exceeding a threshold.

3. The method according to claim 2, characterized in that, The steps of performing data cleaning on the parameter set data include: defining an error threshold δ for the model error value to remove iteration results with a model error > δ.

4. The method according to claim 3, characterized in that, The value range of the set error threshold δ is set between 2 and 30.

5. The method according to claim 1, characterized in that, The decision tree includes multiple levels, and each level includes one or more nodes; Using a decision tree classifier, the steps of calculating the adjustment range of parameter values in the parameter set data include: setting the number of levels included in the decision tree, and using a decision tree classifier based on the parameter data and the corresponding iteration mark, with the parameters in the parameter set as nodes, to obtain an output result regarding the parameters and the parameter adjustment range.

6. The method according to claim 1, characterized in that, Before using a decision tree classifier to calculate the adjustment range of each parameter value in the parameter set data, using a random forest classifier to calculate the importance level of each parameter in the parameter set data based on the input data.

7. The method according to claim 6, characterized in that, After using a decision tree classifier to calculate the adjustment range of each parameter value in the parameter set data, measuring the matching degree between the output result of the decision tree classifier and the output result of the random forest classifier; If the output result of the decision tree classifier matches the output result of the random forest classifier, it is determined that the output result of the decision tree classifier is reliable; If the output result of the decision tree classifier does not match the output result of the random forest classifier, readjust the parameter set based on the machine learning method, and in the step of performing the first test simulation on the optical proximity correction model, increase the number of iterative calculations of the parameter set.

8. The method according to claim 7, characterized in that, The method for measuring the matching degree between the output result of the decision tree classifier and the output result of the random forest classifier includes: judging the matching degree between the top m nodes in the output result of the decision tree classifier and the top m important parameters in the output result of the random forest classifier.

9. The method according to claim 8, characterized in that, The value of m ranges from 1% to 50% of the total number of parameters.

10. The method according to claim 1, characterized in that, In the step of using the parameter set data and the corresponding iteration mark of each iteration as the machine learning input data, the data of each iteration is recorded as a feature vector with the parameter data as the independent variable and the iteration mark as the dependent variable.

11. The method according to claim 3, characterized in that, After removing the iteration results with model error > δ, sampling is performed on the remaining results to balance the number of iteration times with abnormal problems and the number of iteration times without abnormal problems after sampling. The method of sampling includes: The number of iteration times with abnormal problems before sampling is Np1, the number of iteration times without abnormal problems before sampling is Nn1, the number of iteration times with abnormal problems after sampling is Np2, and the number of iteration times without abnormal problems after sampling is Nn2; When Np1 is less than Nn1, if Nn1 ≤ α * Np1, all sampling is performed; if Nn1 > α * Np1, random sampling is performed on all iteration times with abnormal problems. After sampling, the number of abnormal iteration times Np2 is equal to the number of abnormal iteration times Np1 before sampling, and the number of non - abnormal iteration times Nn2 after sampling is set to α * Np2; When Np1 is greater than Nn1, if Np1 ≤ α * Nn1, all sampling is performed; if Np1 > α * Nn1, random sampling is performed on all iteration times without abnormal problems. After sampling, the number of non - abnormal iteration times Nn2 is equal to the number of non - abnormal iteration times Nn1 before sampling, and Np2 is set to α * Nn2; Where α is the data balance threshold, and the value range of α is α ≤ 20.

12. The method according to claim 1, characterized in that, The method of optical proximity correction is used to improve the accuracy of the pattern on the mask corresponding to the wafer forming. The pattern on the mask includes a head - to - head region. The model error function incorporates the size information of the pattern on the mask and its size calculation method. The step of defining the key feature size measurement rule includes: measuring multiple points in the head - to - head region incorporated in the model error function through the built - in algorithm of the model error function. For each point, measure the feature size between the two heads.

13. The method according to claim 12, characterized in that,The step of measuring multiple points in the head - to - head region incorporated in the model error function includes: taking the center of the head - to - head region as the center point; taking points at equal distances on both sides of the center point or randomly distributing points; for each point, measure the feature size between the two heads.

14. The method according to claim 13, wherein The step of defining the abnormal problem judgment rule includes: if the feature size of the center point is less than or equal to the feature size of any other point, it is determined that there is no abnormal problem, and the iteration mark without abnormal problem is marked as normal; if the feature size of the center point is not less than the feature size of any other point, it is determined that an abnormal problem occurs, and the iteration mark with abnormal problem is marked as abnormal.

15. An optical proximity correction system for performing optical proximity correction on an original layout pattern, wherein Including: A model establishment unit for establishing an optical proximity correction model; An optical proximity correction model for performing optical proximity correction. The optical proximity correction model includes a parameter set; A correction unit, which optimizes the parameter set based on a machine learning method. Through machine learning training, it screens out the optimization parameters that have an impact on the local feature size problem, so that the optimization parameters and their values can be adjusted for the local feature size error. Among them, the correction unit includes: A critical feature size measurement rule module; A test and simulation module, which performs a first test and simulation on the optical proximity correction model based on a model error function. The first test and simulation includes multiple iterative calculations of the parameter set. According to the critical feature size measurement rule, it collects the parameter set data, and the parameter set data is used as analysis data. An abnormal problem judgment module, which marks the iterations without abnormal problems as normal and marks the iterations with abnormal problems as abnormal. A machine learning module, which is connected to the test and simulation module and the abnormal problem judgment module. It takes the parameter set data of each iteration and the corresponding iteration mark as the input data of machine learning, and uses a decision tree classifier to calculate the adjustment range of each parameter value in the parameter set data based on the input data. An optimization module, which is connected to the machine learning module and the test and simulation module. It is used to sort the parameters in the parameter set according to their importance to form a parameter sequence, select the top n parameters and their adjustment ranges in the parameter sequence to establish an optimization parameter set. The optimization parameter set is used to perform a second test and simulation on the optical proximity correction model using the test and simulation module to determine the values of the parameters in the optimization parameter set and reduce the feature size error of the optical proximity correction model.

16. A mask, wherein The pattern on the mask is obtained by an optical proximity correction system according to claim 15.

17. A device, wherein Including: At least one memory and at least one processor. The memory stores one or more computer instructions. Among them, the one or more computer instructions are executed by the processor to implement the optical proximity correction method according to any one of claims 1-14.

18. A storage medium, wherein The storage medium stores one or more computer instructions, and the one or more computer instructions are used to implement the optical proximity correction method according to any one of claims 1-14.

Images