CN112561873B - A Virtual Measurement Method of CDSEM Image Based on Machine Learning - Google Patents

Abstract

A CDSEM image virtual measurement method based on machine learning comprises a training set and a verification set generation step, a photoetching space image map and a CDSEM image alignment step, a neural network model is adopted, the photoetching space image map is used as input, the CDSEM image corresponding to the photoetching space image map is used as target output, and N1 groups of photoetching space image map-CDSEM image data pairs in the training set are traversed to finish training of the neural network model; and traversing N2 groups of the photoetching space image-CDSEM image data in the verification set to finish the verification of the neural network model. According to the method, the CDSEM image is generated by adopting machine learning through establishing the mapping between the post-OPC photomask pattern and the post-photoetching CDSEM image, and a verification model independent of an OPC model is obtained and used for confirming the quality of the photoetched pattern.

Description

A Virtual Measurement Method of CDSEM Image Based on Machine Learning

technical field

The invention belongs to the field of semiconductor integrated circuit production and manufacturing, and relates to a CDSEM image virtual measurement method based on machine learning.

Background technique

In the lithography process flow of semiconductor integrated circuit manufacturing, for a given pattern, when the lithography machine is focused and the dose is determined, the lithography aerial image (Aerial Image) of the photoresist on the wafer is also determined. When the photoresist is determined, the three-dimensional structure of the photoresist after development is determined, and at this time, the CDSEM image taken by a scanning electron microscope (Scanning Electronic Microscope, SEM) is also determined, and the CDSEM image is usually confirmed by this Ultimate photolithographic graphic quality.

Usually, in the photolithography process, in order to improve the quality of the pattern in the photolithography process, it is necessary to perform optical proximity correction (Optical Proximity Correction, OPC) on the pattern used to manufacture the mask. Model-based optical proximity effect correction is important to establish an accurate OPC model, which contains a series of parameters, which need to calibrate the OPC model through test graphics (for example, CDSEM images) experimental data, the more experimental data, the more accurate the model .

However, it is impossible for test patterns to cover all types, and the OPC model is not completely based on physical principles, and there are empirical components in the model, so the accuracy of the OPC model cannot be guaranteed. Currently, the test pattern quality after OPC is verified by using the simulation results of the same OPC model. Therefore, the verification quality depends on the accuracy of the OPC model itself for all patterns.

It is clear to those skilled in the art that the accuracy of the OPC model itself for all graphics cannot be guaranteed. Therefore, the industry is in great need of other models independent of the OPC model as a new means of verification to ensure the quality of graphic data after OPC.

In addition, lithography machines are divided into ultraviolet light source (UV), deep ultraviolet light source (DUV), and extreme ultraviolet light source (EUV). Random defects caused by stochastic effects are a big problem in defect detection in lithography, especially in EUV lithography. Moreover, due to the limitation of resolution, the pattern after EUV lithography needs to use an electron beam scanner (E-beam) to scan whether there are defects in the pattern after the lithography process on the wafer, and due to the design constraints of the electron beam scanner , The electron beam scanner cannot detect defects through the die-to-die inter-chip inspection mode, but can only detect defects through the die-to-database chip-to-database inspection mode. That is to say, the current EUV defect detection is realized by comparing the CDSEM image and the OPC target geometry, which is not the best method.

In the lithography process, the stability of lithography conditions plays a key role in the stability of the image quality on the wafer after lithography. Specifically, in the lithography process flow, for a given pattern, when the lithography machine is focused and the dose is determined, the lithography aerial image (Aerial Image) of the photoresist on the wafer is also determined. When the photoresist is determined, the three-dimensional structure of the photoresist after development is determined, and at this time, the CDSEM image taken by the scanning electron microscope is also determined.

Contents of the invention

In view of the problems in the monitoring technology of the stability of the above lithography conditions, the present invention proposes a machine learning-based CDSEM image virtual measurement method, by establishing the mapping of the photomask pattern after OPC and the CDSEM image after the lithography process. Machine learning generates CDSEM images to obtain a verification model independent of the OPC model, which is used to confirm the quality of the pattern after lithography.

To achieve the above object, the technical scheme of the present invention is as follows:

A kind of CDSEM image virtual measurement method based on machine learning, it comprises the steps:

Step S1: a step of generating a training set and a verification set; it includes:

Step S11: Provide a wafer, and preset the number of photolithography processes as K times; wherein, K is a positive integer greater than or equal to 1;

Step S12: Complete a lithography process flow on the wafer, use a scanning electron microscope to scan at different coordinates at M _i of the wafer after lithography, and obtain M _i CDSEM images; wherein, M _i is greater than or equal to A positive integer of 10, i is a value in 1, 2, 3...K;

Step S13: Calculate the lithographic spatial image with the same coordinates as the CDSEM image, and form a set of lithographic spatial image-CDSEM image data pairs with one of the CDSEM images and the corresponding lithographic spatial image, and finally obtain M Set the lithographic aerial image image-CDSEM image data pair, wherein the lithographic aerial image image includes at least one two-dimensional image at different depths of the photoresist;

Step S14: judging whether the group number of the lithographic aerial image map-CDSEM image data pair is equal to N, if not, execute step S12; if yes, execute step S15; wherein:

Step S15: Divide the N groups of photolithographic aerial image map-CDSEM image data pairs into a training set for model training and a verification set for verifying the model in proportion; The ratio of the number of groups of spatial image-CDSEM image data pairs is N1:N2, N=N1+N2;

Step S2: Aligning the photolithography aerial image and the CDSEM image;

Step S3: Using a neural network model, using the lithographic spatial image as input and the corresponding CDSEM image as the target output, traversing the N1 groups of lithographic spatial image-CDSEM image data in the training set The training of the neural network model is completed; the verification of the neural network model is completed by traversing the N2 groups of the lithographic aerial image map-CDSEM image data pairs in the verification set.

Further, described CDSEM image virtual measuring method based on machine learning, step S3 comprises:

Step S31: providing the neural network model;

Step S32: Taking the lithographic aerial image in the training set as input and the corresponding CDSEM image as a target output, traversing the lithographic aerial image-CDSEM image data pair in the training set, training the neural network model;

Step S33: traversing the lithography spatial image map-CDSEM image data pair in the verification set, verifying the neural network model, and calculating the loss function of the verification set;

Step S34: Judging whether the loss function is smaller than the set value, if yes, stop the training of the neural network model to obtain the final neural network model; if not, repeat steps S15 to S34; wherein, the neural network model It embodies the mapping between the lithography space image and the CDSEM image.

Further, the neural network model is a convolution-based deep convolutional neural network DCNN model or a generative confrontation network GAN model, using ReLU as an activation function; if the neural network model adopts the deep convolutional neural network In the DCNN model, the loss function is a mean square error loss function, and if the neural network model adopts the generative confrontation network GAN model, the loss function is a cross-entropy loss function.

Further, the group number N1 of the lithographic aerial image map-CDSEM image data pair in the training set is a multiple of 7, and the group number N2 of the photolithographic aerial image map-CDSEM image data pair in the verification set is 3 multiples of .

Further, the described CDSEM image virtual measurement method based on machine learning also includes:

Step S4: Based on the final neural network model, when a new lithographic spatial image is input, the final neural network model generates a corresponding virtual CDSEM image.

Further, the machine learning-based CDSEM image virtual measurement method also includes step S5: detecting the critical dimension of the virtual CDSEM image generated by the final neural network model, and determining the OPC optical dimension according to the critical dimension Whether the model requires calibration correction.

Further, the described CDSEM image virtual measurement method based on machine learning, the step S5 includes:

S51: Obtain the outline of the virtual CDSEM image, and find a key dimension through the outline;

S52: Judging whether the critical dimension meets the process requirements, and if not, performing calibration correction on the OPC optical model.

Further, the machine learning-based CDSEM image virtual measurement method also includes step S5': judging whether the random effect of a photolithography process is acceptable.

Further, the step S5' in the described CDSEM image virtual measuring method based on machine learning comprises:

S51' performs a lithography process flow with the lithography process conditions corresponding to the new lithography spatial image, and obtains an actual CDSEM image by measurement;

S52' compares the virtual CDSEM image generated by the final neural network model with the actual CDSEM image, and if the mean square error of the pixel values of the two meets the accuracy requirement, then it is determined that the random effect of the photolithography process is acceptable.

Further, the image size and resolution of the lithographic aerial image and the CDSEM image are the same.

It can be seen from the above technical scheme that the beneficial effect of a machine learning-based CDSEM image virtual measurement method of the present invention is that it establishes the mapping of the photomask pattern after OPC and the SEM image after the photolithography process, and generates SEM image, so as to achieve the purpose of confirming the quality of the pattern after lithography, is equivalent to establishing a verification model independent of the OPC model, that is, establishing a model that can generate a virtual reference SEM image from the geometric data after OPC.

Description of drawings

Fig. 1 shows the schematic flow chart of the CDSEM image virtual measurement method based on machine learning in the embodiment of the present invention

Fig. 2 shows the frame diagram of the CDSEM image after establishing the photolithography process based on machine learning in the embodiment of the present invention

Fig. 3 is a schematic diagram of the lithography spatial image obtained through strict optical model calculation for the OPC mask pattern in the embodiment of the present invention

Fig. 4 shows the schematic diagram of the experimental CDSEM image after the photolithography of the mask pattern after OPC in the embodiment of the present invention

Fig. 5 shows the schematic diagram of the virtual CDSEM image generated by the deep learning model after learning in the embodiment of the present invention

Detailed ways

Below in conjunction with accompanying drawings 1-5, the specific embodiment of the present invention will be described in further detail.

It should be noted that in the machine learning-based CDSEM image virtual measurement method disclosed in the present invention, in the photolithography process, the photolithography machine maps the pattern on the mask plate to the wafer coated with photoresist through EUV or UV, etc. (Wafer), in the actual lithography process flow, when the lithography process parameters (focus depth and dose) in the lithography machine are determined, the pattern on the wafer is also determined accordingly. The resulting CDSEM images were also identified. Therefore, when the process flow is determined, there is a certain correspondence between the CDSEM image and the photolithography process parameters.

Please refer to FIG. 1 . FIG. 1 is a schematic flowchart of a machine learning-based CDSEM image virtual measurement method in an embodiment of the present invention. As shown in Figure 1, this kind of CDSEM image virtual measurement method based on machine learning may include the following steps:

Step S1: a step of generating a training set and a verification set; it includes:

Step S11: Provide a wafer, and preset the number of photolithography processes as K times; wherein, K is a positive integer greater than or equal to 1;

Step S12: Complete a lithography process flow on the wafer, use a scanning electron microscope to scan at different coordinates at M _i of the wafer after lithography, and obtain M _i CDSEM images; wherein, M _i is greater than or equal to A positive integer of 10, i is a value in 1, 2, 3...K;

Step S13: Calculate the lithographic spatial image with the same coordinates as the CDSEM image, and form a set of lithographic spatial image-CDSEM image data pairs with one of the CDSEM images and the corresponding lithographic spatial image, and finally obtain M Set the lithographic aerial image image-CDSEM image data pair, wherein the lithographic aerial image image includes at least one two-dimensional image at different depths of the photoresist;

Step S14: judging whether the group number of the lithographic aerial image map-CDSEM image data pair is equal to N, if not, execute step S12; if yes, execute step S15; wherein:

Step S15: Divide the N groups of photolithographic aerial image map-CDSEM image data pairs into a training set for model training and a verification set for verifying the model in proportion; The ratio of the number of sets of aerial image image-CDSEM image data pairs is N1:N2, N=N1+N2.

Step S2: before performing model training, aligning the photolithography space image and the CDSEM image;

Step S3: Using a neural network model, using the lithographic spatial image as input and the corresponding CDSEM image as the target output, traversing the N1 groups of lithographic spatial image-CDSEM image data in the training set The training of the neural network model is completed; the verification of the neural network model is completed by traversing the N2 groups of the lithographic aerial image map-CDSEM image data pairs in the verification set.

That is to say, please refer to FIG. 2 , which is a functional block diagram of image-based off-line photolithography process stability control in an embodiment of the present invention. As shown in Figure 2, both the training set used for model training and the verification set used for verifying the model are obtained from multiple actual lithography processes (for example, lithography is performed 5 times, and the wafer coordinates of each scan are respectively 200, 300, 50, 150 and 300, 1000 CDSEM images will be finally obtained, ie N=1000). N groups of said lithography space image map-CDSEM image data are divided into training set for model training and verification set for verification model in proportion; the ratio of said training set and verification set is N1:N2, N=N1 +N2. Preferably, it may be carried out at a ratio of 7:3 between the training set and the verification set, wherein the training set includes 700 sets of lithography space image map-CDSEM image data pairs, and the verification set includes 300 sets of lithography space image data pairs. imagemap - CDSEM image data pair.

In an embodiment of the present invention, after having the lithographic spatial image map-SEM image data pair, for the mapping relationship between the two, it is possible to use a deep convolutional neural network (DCNN: deep convolutional neural networks) or a generative formula Adversarial networks (GAN: Generative adversarial networks) and other methods are used for derivation.

Please refer to Fig. 3, Fig. 4 and Fig. 5, Fig. 3 shows that in the embodiment of the present invention, the schematic diagram of the lithography spatial image obtained by calculating the OPC mask pattern through strict optical model, Fig. 4 shows the implementation of the present invention In the example, the schematic diagram of the experimental CDSEM image after the photolithography of the mask pattern after OPC, and FIG. 5 is a schematic diagram of the virtual CDSEM image generated by the learned deep learning model in the embodiment of the present invention.

In the embodiment of the present invention, since the coordinates of the actual pattern after lithography may deviate from the coordinates of the corresponding pattern on the reticle, before performing model training, it is necessary to perform step S2: combine the lithography space image with the The CDSEM image is aligned, and, preferably, the image size and resolution of the photolithographic aerial image and the CDSEM image are the same. The size of the image depends on the specific situation. In this example, it can be 512*512.

Step S3: Using a neural network model, using the lithographic spatial image as input and the corresponding CDSEM image as the target output, traversing the N1 groups of lithographic spatial image-CDSEM image data in the training set The training of the neural network model is completed; the verification of the neural network model is completed by traversing the N2 groups of the lithographic aerial image map-CDSEM image data pairs in the verification set.

Specifically, through the image-to-image (Image To Image) method, the main method is to generate a corresponding CDSEM image based on the lithographic spatial image in the photoresist after exposure, and use the lithographic spatial image as the input of the neural network model, The corresponding CDSEM image is used as the target output of the neural network model. Through continuous training and verification of the neural network model, and adjustment of the parameters of the neural network model, the mapping between the lithography space image and the CDSEM image is finally completed.

In an embodiment of the present invention, the step S3 of the CDSEM image virtual measurement method based on machine learning specifically includes:

Step S31: providing the neural network model;

Step S32: Taking the lithographic aerial image in the training set as input and the corresponding CDSEM image as a target output, traversing the lithographic aerial image-CDSEM image data pair in the training set, training the neural network model;

Step S33: traversing the lithography spatial image map-CDSEM image data pair in the verification set, verifying the neural network model, and calculating the loss function of the verification set;

Step S34: Judging whether the loss function is smaller than the set value, if yes, stop the training of the neural network model to obtain the final neural network model; if not, repeat steps S15 to S34; wherein, the neural network model It embodies the mapping between the lithography space image and the CDSEM image.

Further, the neural network model is a convolution-based deep convolutional neural network DCNN model or a generative confrontation network GAN model, using ReLU as an activation function; if the neural network model adopts the deep convolutional neural network In the DCNN model, the loss function is a mean square error loss function, and if the neural network model adopts the generative confrontation network GAN model, the loss function is a cross-entropy loss function.

Further, the deep convolutional neural network DCNN model may include an input layer, 13 intermediate layers and an output layer, the intermediate layers have the same structure, the convolution kernel size is 3*3, and the width of each layer is 64 or 128 feature maps, each convolution layer is followed by batch normalization, the input layer only performs convolution and activation operations, and the output layer only performs convolution operations.

With the above-mentioned neural network model, step S4 can be performed, that is, in the final neural network model, when a new lithographic spatial image is input, the final neural network model generates a corresponding virtual CDSEM image.

Step S4: Based on the trained neural network model, when a new lithographic spatial image is input, the neural network model generates a corresponding virtual CDSEM image; wherein, the lithographic spatial image is based on EUV The lithography spatial image at the same coordinates as the CDSEM image calculated by the mask pattern and process parameters, or the lithography spatial image is the post-OPC mask pattern calculated by the OPC optical model of the wafer Photolithography aerial image at the same coordinates as the CDSEM image.

In an embodiment of the present invention, when the new photolithographic aerial image is the photolithographic aerial image at the same coordinates as the CDSEM image obtained by calculating the mask pattern of the wafer after OPC through an optical model, it is also Step S5 is included: detecting the critical dimension of the virtual CDSEM image generated by the final neural network model, and determining whether calibration correction is required for the OPC optical model according to the critical dimension.

Further, the step S5 may include:

Detecting the critical dimension of the virtual CDSEM image generated by the final neural network model, and determining whether calibration correction is required for the OPC optical model according to the critical dimension.

In another embodiment of the present invention, when the new lithography aerial image is the lithography aerial image at the same coordinates as the CDSEM image calculated according to the EUV mask pattern and process parameters, further comprising the step S5': For the lithography spatial image of the EUV mask, use the neural network model generated above to generate a virtual CDSEM image as a reference CDSEM image for electron beam defect scanning, and use the reference CDSEM image to compare EUV lithography CDSEM image; when the pixel value mean square error between the CDSEM image after lithography and the reference CDSEM image meets the accuracy requirements, it is determined that there is no problem with the geometric data of the lithographic pattern generated after the EUV lithography, otherwise it is determined that the EUV The geometric data of the lithographic pattern generated after lithography has random defects.

Specifically, said step S5' may include:

S51' performs a lithography process flow with the lithography process conditions corresponding to the new lithography spatial image, and obtains an actual CDSEM image by measurement;

S52' compares the virtual CDSEM image generated by the final neural network model with the actual CDSEM image, and if the mean square error of the pixel values of the two meets the accuracy requirement, then it is determined that the random effect of the photolithography process is acceptable.

That is to say, in the model application stage, input the photolithographic spatial image (Aerial image), and the model outputs the corresponding photolithographic CDSEM virtual image. The CDSEM virtual image can be used as a standard independent of the OPC model, and the OPC model The simulated images after lithography are compared, and the OPC model whose mean square error is less than the preset accuracy is an acceptable OPC model, otherwise more experimental data will be used to recalibrate the OPC model.

The above are only preferred embodiments of the present invention, and the embodiments are not intended to limit the scope of patent protection of the present invention. Therefore, all equivalent structural changes made by using the description and accompanying drawings of the present invention should be included in the same way. Within the protection scope of the present invention.

Claims (7)

1. A CDSEM image virtual measurement method based on machine learning is characterized by comprising the following steps:

step S1: generating a training set and a verification set; it includes:

step S11: providing a wafer, and presetting the number of times of photoetching process as K times; wherein K is a positive integer greater than or equal to 1;

step S12: completing a photoetching process flow on the wafer once, and using a scanning electron microscope to perform photoetching on the wafer M _i Scanning at different coordinates to obtain M _i A CDSEM image; wherein M is _i Is a positive integer greater than or equal to 10, i is one value of 1,2,3 \8230k;

step S13: calculating a photoetching space image map with the same coordinate as the CDSEM image, forming a group of photoetching space image-CDSEM image data pairs by one CDSEM image and the corresponding photoetching space image map, and finally obtaining M groups of photoetching space image-CDSEM image data pairs, wherein the photoetching space image map comprises at least one two-dimensional image at different depths of the photoresist;

step S14: judging whether the group number of the photoetching space image map-CDSEM image data pairs is equal to N, if not, executing the step S12; if yes, executing step S15; wherein:

step S15: dividing N groups of photoetching space image map-CDSEM image data pairs into a training set for model training and a verification set for verifying a model in proportion; the group number proportion of the photoetching space image-CDSEM image data pairs in the training set and the verification set is N1: N2, and N = N1+ N2;

step S2: aligning the lithographic aerial image map and the CDSEM image;

and step S3: adopting a neural network model, taking the photoetching spatial image map as input, taking the CDSEM image corresponding to the photoetching spatial image map as target output, and traversing N1 groups of photoetching spatial image map-CDSEM image data in the training set to finish the training of the neural network model; traversing N2 groups of photoetching space image-CDSEM image data in a verification set to complete the verification of the neural network model;

and step S4: based on the neural network model, when a new photoetching spatial image is input, the final neural network model generates a virtual CDSEM image corresponding to the new photoetching spatial image;

step S5: detecting the key size of the virtual CDSEM image generated by the neural network model, and determining whether an OPC optical model needs to be calibrated and corrected according to the key size; alternatively, it is determined whether the random effect of a single photolithography process is acceptable.

2. The machine learning-based CDSEM image virtual measurement method of claim 1, wherein step S3 comprises:

step S31: providing the neural network model;

step S32: taking the photoetching spatial image map in the training set as an input, taking the CDSEM image corresponding to the photoetching spatial image map as a target output, traversing the photoetching spatial image map-CDSEM image data pair in the training set, and training the neural network model;

step S33: traversing the photoetching space image-CDSEM image data pairs in the verification set, verifying the neural network model, and calculating a loss function of the verification set;

step S34: judging whether the loss function is smaller than a set value or not, if so, stopping training the neural network model to obtain a final neural network model; if not, repeating the steps S15 to S34; wherein the neural network model embodies a mapping between a lithography aerial image map and the CDSEM image.

3. The machine-learning-based CDSEM image virtual measurement method according to claim 2, wherein the neural network model is a convolution-based Deep Convolution Neural Network (DCNN) model or a generative countermeasure network (GAN) model, and ReLU is used as an activation function; if the neural network model adopts the deep convolutional neural network DCNN model, the loss function is a mean square error loss function, and if the neural network model adopts the generative countermeasure network GAN model, the loss function is a cross entropy loss function.

4. The machine-learning-based CDSEM image virtual measurement method of claim 1, wherein the number of sets N1 of the lithography aerial image map-CDSEM image data pairs in the training set is a multiple of 7, and the number of sets N2 of the lithography aerial image map-CDSEM image data pairs in the verification set is a multiple of 3.

5. The virtual measurement method for CDSEM images based on machine learning as claimed in claim 1, wherein the step of detecting the critical dimension of the virtual CDSEM image generated by the neural network model in step S5 and determining whether the OPC optical model needs calibration correction according to the critical dimension comprises:

s51: acquiring the outline of the virtual CDSEM image, and finding out the critical dimension through the outline;

s52: and judging whether the critical dimension meets the process requirement, and if not, calibrating and correcting the OPC optical model.

6. The method for virtually measuring a machine learning based CDSEM image according to claim 1, wherein the step S5 of determining whether the random effect of the primary photolithography process is acceptable comprises:

s51', carrying out a photoetching process flow by using photoetching process conditions corresponding to the new photoetching spatial image, and measuring to obtain an actual CDSEM image;

s52' the virtual CDSEM image generated by the final neural network model is compared with the actual CDSEM image, and if the mean square error of the pixel values of the virtual CDSEM image and the actual CDSEM image meets the precision requirement, the random effect of the photoetching process is judged to be acceptable.

7. The machine-learning-based CDSEM image virtual measurement method of claim 1, wherein the image size and resolution of the lithography aerial image map and CDSEM image are the same.

Images