CN111310407A - Method for designing optimal feature vector of reverse photoetching based on machine learning - Google Patents

Abstract

An optimal feature vector design method for reverse photoetching solution based on machine learning comprises dividing a design target pattern into a plurality of grid units; calculating a set of feature functions { Ki (x, y) }, i ═ 1,2, … N1 according to imaging conditions; establishing a neural network model, and selecting training samples and verification samples required to be included in training; calculating a set of signals { Si (x, y) } for each of the grid cells using a set of feature functions { Ki (x, y) }; and taking the value of the strict reverse photoetching at the corresponding position as a target value of neural network training; during training, different input end dimensions N1, hidden layer numbers N2 and each hidden layer are adoptedNumber M of neurons in the hidden layer₁,M₂,…M_N2Training with training samples and validating with validation samples until finding a neuron with an input end dimension of N1, a number of hidden layers of N2, and a number of neurons per hidden layer of M is found₁,M₂,…M_N2The combined neural network model is satisfied. Therefore, the design method of the invention ensures that the neural network does not need a feature extraction layer any more, thereby simplifying the network architecture and shortening the training time.

Description

A method for optimal eigenvector design for reverse lithography based on machine learning

technical field

The invention belongs to the field of integrated circuit manufacturing, and in particular relates to a design method of an optimal feature vector for reverse lithography based on machine learning.

Background technique

Computational lithography plays a vital role in the semiconductor industry. When the semiconductor technology node shrinks to 14nm and below, the lithography technology is gradually approaching its physical limit. As a new type of resolution enhancement technology, Source Mask Optimization (SMO) can significantly improve the limit size. The overlapping process windows of semiconductor lithography effectively extend the life cycle of current conventional lithography technology. SMO is not only an important part of 193nm immersion lithography technology, but also an indispensable technology in EUV lithography.

The basic principle of the light mask co-optimization simulation calculation is similar to the model-based proximity effect correction. Move the edge of the mask pattern and calculate its deviation from the target pattern on the wafer, that is, the edge placement error. In the optimization model, disturbances of exposure dose, focus, and pattern size on the reticle are deliberately introduced, and the edge placement errors of the images on the wafer caused by these disturbances are calculated. Both the merit function and optimization are implemented based on edge placement errors. The result of the light mask co-optimization calculation includes not only a pixelated light source, but also a proximity effect correction made to the input design.

After source-mask co-optimization, inverse lithography has become the ultimate frontier of computational lithography. However, reverse lithography technology requires huge computing hardware resources and a long computing time, and it is still impractical to achieve strict full-chip reverse lithography solutions. Since the 3D effect of the extreme ultraviolet (EUV) mask is more obvious than that of the immersion lithography mask, in this case, if you also try to realize the full-chip strict reverse lithography solution of EUV, the amount of calculation is larger and the change more time-consuming and more difficult.

Inverse Lithography Technology (ILT) is considered to be a new generation of resolution enhancement technology for 45nm, 32nm and even 22nm lithography. A very promising technology to overcome this obstacle is to make full use of the increasingly mature machine learning technology based on neural network structure. Deep Convolutional Neural Networks (DCNNs), which obtain inverse lithography (ILT) solutions and are much faster than rigorous inverse lithography calculations.

However, in DCNN, in order to extract feature vectors with sufficient resolution and approximately complete representation ability, the feature vector extraction layer is very complicated and lacks real physical meaning. In order to extract feature vectors with sufficient resolution and approximately complete representation ability, the training of DCNN networks requires a large number of balanced samples, which makes training more difficult and time-consuming.

SUMMARY OF THE INVENTION

For achieving the above object, technical scheme of the present invention is as follows:

An optimal eigenvector design method for reverse lithography solution based on machine learning, for predicting/calculating the value of reverse lithography solution; the method comprises the following steps:

Step S1: dividing the design target pattern into N grid cells, wherein the size of the grid cells is determined by imaging conditions;

Step S2: Calculate the characteristic function set {Ki(x,y)}, i=1,2,...N1 according to the imaging conditions; wherein, the characteristic function set {Ki(x,y)} is a group of optimal optical functions A ruler, used to measure the surrounding environment of any grid unit in the design target pattern; the value of N1 is related to the requirement to characterize the completeness of the surrounding environment of the grid unit, and the N1 is the optical ruler Ki The number of (x,y);

Step S3: establishing a neural network model, the neural network model includes an input layer, a hidden layer and an output layer; wherein, the dimension of the input layer is equal to N1, the hidden layer has a total of N2 layers, and the size of each hidden layer is The number of neurons is M1, M2, ... MN2; wherein, the M1, M2, ... MN2 are the same, partially the same or different;

Step S4: training the neural network model needs to include training samples and verification samples, and the training samples and verification samples are randomly selected part of the target patterns in the design target patterns, using the feature function set {Ki(x,y) )} calculate the signal set {Si(x,y)} of each grid unit as the input of the neural network model of the grid unit, the signal set {Si(x,y)} characterizes one of the target patterns The surrounding environment of the grid unit, the signal set {Si(x,y)} is also called the feature vector; and the value of the strict reverse lithography at the corresponding position is used as the target value of the neural network training, that is, the same partial target pattern is used , using a strict reverse lithography algorithm to generate the best mask image as the original training target image for neural network training;

Step S5: During the training of the neural network model, different combinations of the input dimension N1, the number of hidden layers N2, and the number of neurons in each hidden layer M ₁ , M ₂ , . . . M _N2 are used. The training samples are used for training, and the verification samples are used for verification until a satisfactory combination with the input dimension N1, the number of hidden layers N2 and the number of neurons in each hidden layer M ₁ , M ₂ ,...M _N2 is found. where the satisfactory combination refers to, for each grid unit in the training set and the validation set, the difference between the predicted value of the neural network model and the value of the exact solution of reverse lithography The error is less than or equal to the predefined error specification;

Step S6: In the application realization stage, the designed wafer pattern is divided into grid cells, and {Si(x,y)} values are calculated for each of the grid cells, and the {Si(x,y) } value is input into the trained neural network model to obtain the value for predicting the reverse lithography solution.

Further, the reverse lithography solution is a machine learning-based reverse lithography solution, a machine learning-based optical proximity correction, or a machine learning-based lithography hot spot detection solution.

Further, when selecting training samples in step S4, firstly, the original training target image trained by the neural network is divided into a first type area and a second type area; wherein, the first type area is useful information for model training. Regions without obvious repeatability, the second type of regions are regions with obvious repeatability of useful information for model training.

Further, the step of dividing the original training target image trained by the neural network into the first type area and the second type area includes:

Find the maximum intensity value in the original training target image trained by the neural network;

Determine the intensity threshold for selecting seed pixels by multiplying the maximum intensity value found above by a coefficient;

create an auxiliary image of the same size as the original training target image, the intensity value of which is initially set to zero;

In the original training target image, find out the pixel position where the intensity value is greater than the seed threshold;

In the auxiliary image, the intensity value of the pixel position where the intensity value is greater than the seed threshold is found is set to a predetermined value, so that many small islands will be formed in the auxiliary image;

Through the image growth morphological operation, the islets are iterated for many times to finally form the first type area in the original training target image, and the remaining area in the original training target image is the second type area .

Further, in the training process of step S5, batch normalization technology is adopted, and dynamic adaptive sample weighting is adopted to improve the training quality.

Further, in the step S5, a stochastic gradient descent method is used to train the neural network model.

Further, the imaging conditions include exposure wavelength, numerical aperture, and exposure illumination modes and settings.

Further, the N1 is less than or equal to 200.

Further, the N2 is less than or equal to 6.

Further, the method further includes step S7: obtaining an image of the reverse lithography solution according to the value of the reverse lithography solution; identifying the main pattern area originally designed from the image of the reverse lithography solution; region, the auxiliary function region is located by a predefined intensity threshold to determine the optimal location of the whole-chip sub-resolution auxiliary pattern.

It can be seen from the above technical solutions that the present invention provides a method for designing an optimal feature vector for reverse lithography by machine learning. The physical-based feature vector design eliminates the need for a feature vector extraction layer in the DCNN neural network. Building the mapping function layer greatly simplifies the required neural network and greatly reduces the training time of the neural network. This neural network architecture can accelerate the generation of full-chip sub-resolution auxiliary patterns (SRAFs).

Description of drawings

Figure 1 shows a schematic diagram of the flow from design pattern to rigorous reverse lithography solution.

2 is a schematic flowchart of a method for designing an optimal feature vector for reverse lithography based on machine learning in an embodiment of the present invention

FIG. 3 is a schematic diagram showing that the design target pattern is divided into N1 grid cells in an embodiment of the present invention

FIG. 4 is a schematic diagram showing the measurement value set {Si(x,y)} under the feature function set {Ki(x,y)} as the input feature vector of reverse lithography based on the neural network model in the embodiment of the present invention

FIG. 5 is a schematic diagram of dividing an original training target image into a first-type area and a second-type area in an embodiment of the present invention

FIG. 6 is a schematic diagram showing the specific steps of dividing the original training target image trained by the neural network into the first type area and the second type area in an embodiment of the present invention

FIG. 7 is a schematic diagram of training a neural network model in an embodiment of the present invention

FIG. 8 to FIG. 14 are schematic diagrams showing the predicted values of the trained neural network model according to the embodiment of the present invention

FIG. 15 is a schematic diagram of determining the optimal position of the full-chip sub-resolution assist patterns (SRAFs) in an embodiment of the present invention

Detailed ways

The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings 1-15.

It should be noted that the optimal eigenvector design method for reverse lithography solution based on machine learning of the present invention is used to predict the value of the reverse lithography solution. Among them, the value of the inverse lithography solution can be applied to machine learning-based inverse lithography, machine learning-based optical proximity correction (OPC), machine learning-based lithography hot spot detection, etc. The optimal eigenvector design method can be used for Computational lithography for immersion lithography (reverse lithography, optical proximity correction, lithography hot spot detection), and can also be used for computational lithography for EUV lithography (reverse lithography, optical proximity correction, lithography hot spot detection).

Please refer to Figure 1, which shows an ideal schematic diagram from design pattern to rigorous reverse lithography solution. As shown in Figure 1, the left side of the figure is the design target pattern, and the right side is the ideal reverse lithography solution of the mask. It is clear to those skilled in the art that all machine learning-based computational lithography techniques, including machine learning-based inverse lithography techniques, need to solve how to characterize the environment near a point, and the response of a certain point (x, y) only depends on its The adjacent environment within the influence range is essentially a problem of eigenvector design. The unique method of the present invention extracts eigenvectors with sufficient resolution and approximate complete representation ability, and performs inverse lithography solutions based on machine learning. The optimal eigenvector design method , its physics-based feature vector design eliminates the need for the feature vector extraction layer in the DCNN neural network, and only needs to build a mapping function layer, which greatly simplifies the required neural network and greatly shortens the neural network training time.

Please refer to FIG. 2 , which is a schematic diagram of a method for designing an optimal feature vector for machine learning reverse lithography according to the present invention. As shown in the figure, the method includes the following steps:

Step S1: Divide the design target pattern into N grid cells, wherein the size of the grid cells is determined by imaging conditions; wherein, the imaging conditions include exposure wavelength, numerical aperture, and exposure illumination modes and settings.

It is clear to those skilled in the art that if the surrounding environment is simply divided into small units and the geometrical weights of the patterns in each unit are used as eigenvector elements, then the total amount of elements is:

Total Element Amount = ((2*Area of Influence)/Unit Step) ²

If it is assumed that the influence range=1000nm/edge, and the unit step size=10nm, then the total number of elements=(2000/10) ² =40000. This simple eigenvector design is inefficient.

In the embodiment of the present invention, the shape of the N grid cells is generally a square, and the side length of the square is determined by imaging conditions and is proportional to λ/(NA(1+σ _max )); where λ is the exposure wavelength , NA is the numerical aperture, and σ _max is the maximum angle of incidence of the exposure illumination. For example, in an embodiment of the present invention, the wafer design can be divided into grid cells, each cell can be 5nmx5nm (as shown in Figure 3), the grid cell size is determined by the imaging conditions, i.e. exposure wavelength and illumination Conditions are ok.

Furthermore, designing efficient eigenvectors while achieving optimal resolution, adequacy and efficiency requires consideration of the symmetric characteristics of the problem (imaging system).

Step S2: Calculate the characteristic function set {Ki(x,y)}, i=1,2,...N1 according to the imaging conditions; wherein, the characteristic function set {Ki(x,y)} is a group of optimal optical functions The ruler is used to measure the surrounding environment of any grid unit in the design target pattern; the value of N1 is related to the requirement of characterizing the completeness of the surrounding environment of the grid unit, and N1 is the value of the optical ruler Ki(x, y). number.

Step S3: establishing a neural network model, the neural network model includes an input layer, a hidden layer and an output layer; wherein, the dimension of the input layer is equal to N1, the hidden layer has a total of N2 layers, and the size of each hidden layer is The number of neurons is M ₁ , M ₂ , . . . M _N2 ; wherein, the M ₁ , M ₂ , . . . M _N2 may be the same, partially the same or different.

In the embodiment of the present invention, the established neural network model is related to the N1 and N2 layers, and the number of neurons included in each layer is M ₁ , M ₂ , . . . M _N2 . That is, in an embodiment of the present invention, it is possible to reset how many layers (eg, N2 layers) the neural network model includes, and how many neurons each layer includes (eg, , M ₁ , M ₂ ,...M _N2 ), preferably, the value of N1 is less than or equal to 140, the value of N2 is 5, and the number of neurons in the first layer of the neural network model is the same as that of N1 ( as shown in Figure 5).

Step S4: train the neural network model, the training needs include training samples and verification samples, and the training samples and verification samples are randomly selected part of the target patterns in the design target patterns, and then use the feature function set { Ki(x,y)} computes the signal set {Si(x,y)} for each said grid cell as input to the neural network model of the grid cell, said signal set {Si(x,y)} representing the surrounding environment of a grid cell in the target pattern, the signal set {Si(x,y)} is also called a feature vector; and the value of the strict reverse lithography at the corresponding position is used as the target value of the neural network training, that is Using the same partial target pattern, a rigorous inverse lithography algorithm is used to generate the optimal mask image as the original training target image for neural network training.

Specifically, computing the set of signal values for each grid cell can begin with an imaging equation.

Hopkin's partially coherent illumination imaging formula 1 is as follows:

where γ(x ₂ -x ₁ , y ₂ -y ₁ ) is the coefficient of mutual coherence between (x ₁ , y ₁ ) and (x ₂ , y ₂ ) on the object plane (ie, the mask plane), Determined by illumination; P(xx ₁ , yy ₁ ) is the impulse response function of the optical imaging system, which is determined by the pupil function of the optical system. More specifically, it is caused by the interference of a point light source of unit amplitude and zero phase at (x ₁ , y ₁ ) on the object plane (ie mask plane) at point (x, y) in the image plane complex amplitude. M(x ₁ , y ₁ ) is the complex transfer function at the point (x ₁ , y ₁ ) on the object plane (ie, the mask plane). Variables with an asterisk refer to the conjugate of the original variable, for example, P* is the conjugate of P and M* is the conjugate of M. According to Mercer's theorem, the above formula (1) can be transformed into a simpler formula:

表示函数Ki(x,y)与掩模传输函数M(x,y)之间的卷积运算；in,

Represents the convolution operation between the function Ki(x,y) and the mask transfer function M(x,y);

{α _i } and {K _i } are the eigenvalues and eigenfunctions of the following equations.

∫∫W(x ₁ ',y ₁ '; x ₂ ',y ₂ ')K _i (x ₂ ',y ₂ ')dx ₂ 'dy ₂ '=α _i K _i (x ₁ ',y ₁ ' ) (3a)

W(x ₁ ',y ₁ '; x ₂ ',y ₂ ')=γ(x ₂ '-x ₁ ',y ₂ '-y ₁ ')P(x ₁ ',y ₁ ')P ^* ( x ₂ ',y ₂ ') (3b)

The significance of the above formula (2) is that it shows that the partially coherent imaging system can be decomposed into a series of coherent imaging systems, and the coherent imaging systems are independent of each other. Although there are other ways to decompose a partially coherent imaging system into a series of coherent imaging systems, the above method has been shown to be the best, often referred to as optimal coherent decomposition.

Please refer to FIG. 4. FIG. 4 shows the measurement value set {Si(x,y)} under the feature function set {Ki(x,y)} in the embodiment of the present invention as the reverse lithography method based on the neural network model. Input feature vector diagram. As shown in the figure, the set of eigenfunctions {K _i (x,y)} is the best set of optical scales to characterize the surrounding environment of a point under given imaging conditions, and the set of eigenfunctions {K _i (x,y)} The set of measured values under {S _i (x, y)} is used as the input feature vector of inverse lithography based on the neural network model, and the set of signal values for each grid cell is calculated. During the training phase of the neural network model, the set of signal values (S1, S2...SN) are feature vectors, which are the input to the forward neural network.

Often, many training target images contain fairly large regions where there is significant repetition of information useful for model training. If these regions are fully placed in model training, the training sampling may be biased, which is detrimental to model accuracy. Therefore, regions that have useful information for model training without obvious repeatability can be selected first, such as the white regions shown in B in Figure 5. For the black area shown in B in Figure 5, we only select a small part for training, for example, in the black area, we can choose less than or equal to 10% of the samples. By doing this, model training sampling can be better balanced.

Specifically, when selecting training samples in step S4, firstly, the original training target image (as shown in Figure A in FIG. 5 ) trained by the neural network is divided into a first type area and a second type area; One type of area is the area that does not have obvious repetition of useful information for model training (the white area shown in Figure B in Figure 5), and the second type of area (the black area shown in Figure B in Figure 5) is: Regions with significant repeatability of information useful for model training.

In an embodiment of the present invention, the step of dividing the original training target image trained by the neural network into the first type area and the second type area includes:

Find the maximum intensity value in the original training target image (as shown in the left image in Figure 6) of the neural network training;

Determine the intensity threshold for selecting seed pixels by multiplying the maximum intensity value found above by a coefficient, for example, 0.05;

create an auxiliary image of the same size as the original training target image, the intensity value of which is initially set to zero;

In the original training target image, find out the pixel position where the intensity value is greater than the seed threshold;

In the auxiliary image, the intensity value of the pixel position where the intensity value is greater than the seed threshold is found is set to a predetermined value (for example, 1.0), in this way, many small islands will be formed in the auxiliary image (as shown in Fig. shown in the middle image in 6);

Through the image growth morphological operation, the islets are iterated for many times to finally form the first type area in the original training target image, and the remaining area in the original training target image is the second type area (as shown on the right in Figure 6).

Step S5: During the training of the neural network model, different combinations of the input dimension N1, the number of hidden layers N2, and the number of neurons in each hidden layer M ₁ , M ₂ , . . . M _N2 are used. The training samples are used for training, and the verification samples are used for verification until a satisfactory combination with the input dimension N1, the number of hidden layers N2 and the number of neurons in each hidden layer M ₁ , M ₂ ,...M _N2 is found. of the neural network model. Wherein, the satisfactory combination means that, for each grid unit in the training set and the validation set, the error between the predicted value of the neural network model and the value of the strict inverse lithography solution is less than or equal to a predefined error Norm; eg 10%.

In the embodiment of the present invention, the training target is an image of a strictly reverse lithography solution, and the training phase is divided into two phases. The first phase is to randomly select {Si(x,y)} for each training sub-batch, and , the network can be trained using the stochastic gradient descent (SGD) method. As an optimization algorithm commonly used in machine learning, gradient descent has three different forms: Batch Gradient Descent, Stochastic Gradient Descent, and Mini-Batch Gradient Descent. . Among them, the mini-batch gradient descent method is also commonly used for model training in deep learning.

That is, next, in the second stage, {S _i (x,y)} of all training sub-batches and strictly inverse lithography-solved images are taken as a single batch, using a neural network model, to generate {S _i (x) ,y)} corresponding strictly inverse lithography images to train the neural network model.

Preferably, after it is determined that the neural network model is a combination of the input dimension N1, the number of hidden layers N2, and the number of neurons in each hidden layer M ₁ , M ₂ , . . . M _N2 , He can be used. The initialization strategy configures the initial value of the network parameters during training, and the activation function of the network uses the ELU activation function, the Relu activation function, the tanh activation function or other activation functions. In addition, the optimization algorithm for training the neural network model can use Adam optimization algorithm; the optimized objective function can use MSE mean square error.

In the training process of the embodiment of the present invention, the trained neural network model can also be verified, and according to the verification result, it is judged whether the trained neural network model meets the design requirements. If the design requirements are not met, adjust the combination of the input dimension N1, the number of hidden layers N2, and the number of neurons in each hidden layer M ₁ , M ₂ , ... M _N2 to establish a revised neural network model, and then Perform the training steps; if the design requirements are met, the neural network model can be determined to be a trained neural network model. Wherein, meeting the design requirements means that, for each grid unit in the training set and the validation set, the error between the predicted value of the neural network model and the value of the strict inverse lithography solution is less than or equal to a predefined error specification .

Moreover, in the training process, batch normalization technology can also be used, and dynamic adaptive sample weighting can be used to improve the training quality.

In summary, the present invention makes full use of the increasingly mature machine learning technology based on the neural network structure, and uses the deep convolutional neural network (DCNN) to obtain the solution of the inverse lithography technology (ILT), and is more efficient than the strict inverse lithography. Calculations are much faster.

In the embodiment of the present invention, with the trained neural network model, the designed wafer pattern can be divided into grid units in the application implementation stage, and {Si(x,y) can be calculated for each grid unit } value, and input the {Si(x,y)} value into the trained neural network model to obtain the value that predicts the inverse lithography solution.

Step S6: In the application realization stage, the designed wafer pattern is divided into grid cells, and {Si(x,y)} values are calculated for each of the grid cells, and the {Si(x,y) } value is input into the trained neural network model to obtain the value for predicting the reverse lithography solution.

Please refer to FIGS. 8-14 . FIGS. 8 to 14 are schematic diagrams of predicting inverse lithography solution values by using a trained neural network model according to an embodiment of the present invention. As shown, these patterns are a good representation of the results of the reverse lithography of the present invention based on machine learning of physically optimized eigenvectors.

It should be noted that the neural network structure based on machine learning can accelerate the realization of full-chip reverse lithography. Starting from the inverse lithography solution, the optimal location of full-chip sub-resolution assist patterns (SRAFs) can be determined, where the advanced lithography process window can be maximized.

Please refer to FIG. 15. FIG. 15 is a schematic diagram of determining the optimal position of the sub-resolution assist patterns (SRAFs) on the whole chip according to an embodiment of the present invention. Specifically, the above method may further include step S7: obtaining an image of the reverse lithography solution according to the value of the reverse lithography solution; identifying the originally designed main pattern area from the image of the reverse lithography solution; in the remaining area In , the helper region is located by a predefined intensity threshold to determine the optimal location of the whole-chip sub-resolution helper pattern.

The above are only the preferred embodiments of the present invention, and the embodiments are not intended to limit the scope of the patent protection of the present invention. Therefore, any equivalent structural changes made by using the contents of the description and the accompanying drawings of the present invention shall also include within the protection scope of the present invention.

Claims (10)

1. an optimal eigenvector design method for carrying out reverse lithography solution based on machine learning, for predicting/calculating the value of reverse lithography solution; it is characterized in that, described method comprises the steps: Step S1: dividing the design target pattern into N grid cells, wherein the size of the grid cells is determined by imaging conditions; Step S2: Calculate the feature function set {K _i (x,y)}, i=1, 2,...N1 according to the imaging conditions; wherein, the feature function set {K _i (x, y)} is a set of optimal The optical ruler is used to measure the surrounding environment of any grid unit in the design target pattern; the value of N1 is related to the completeness of the surrounding environment of the grid unit, and the N1 is the optical scale. The number of rulers Ki(x,y); Step S3: establishing a neural network model, the neural network model includes an input layer, a hidden layer and an output layer; wherein, the dimension of the input layer is equal to N1, the hidden layer has a total of N2 layers, and the size of each hidden layer is The number of neurons is M ₁ , M ₂ , ... M _N2 ; wherein, the M ₁ , M ₂ , ... M _N2 are the same, partially the same or different; Step S4: training the neural network model needs to include training samples and verification samples, and the training samples and verification samples are randomly selected part of the target pattern in the design target pattern, using the feature function set {K _i (x, y)} computes the signal set {S _i (x,y)} of each said grid cell as the input to the neural network model of the grid cell, said signal set {S _i (x, y)} characterizing the target The surrounding environment of a grid unit in the pattern, the signal set {S _i (x, y)} is also called a feature vector; and the value of the strict reverse lithography at the corresponding position is used as the target value of the neural network training, that is, the same Part of the target pattern, using a strict reverse lithography algorithm to generate the best mask image, as the original training target image for neural network training; Step S5: During the training of the neural network model, different combinations of the input dimension N1, the number of hidden layers N2, and the number of neurons in each hidden layer M ₁ , M ₂ , . . . M _N2 are used. The training samples are used for training, and the verification samples are used for verification until a satisfactory combination with the input dimension N1, the number of hidden layers N2 and the number of neurons in each hidden layer M ₁ , M ₂ ,...M _N2 is found. until the neural network model of the neural network model; wherein, the satisfactory combination refers to, for each grid unit in the training set and the validation set, between the predicted value of the neural network model and the value of the strict inverse lithography solution The error is less than or equal to the predefined error specification; Step S6: In the application realization stage, the designed wafer pattern is divided into grid cells, and {Si(x,y)} values are calculated for each of the grid cells, and the {Si(x,y) } value is input into the trained neural network model to obtain the value for predicting the reverse lithography solution. 2 . The method according to claim 1 , wherein the inverse lithography solution is a machine learning-based inverse lithography solution, a machine learning-based optical proximity correction, or a machine learning-based lithography hot spot detection solution. 3 . 3. The method according to claim 1, wherein, when selecting training samples in step S4, firstly, the original training target image trained by the neural network is divided into a first type area and a second type area; wherein, The first type of area is an area that does not have obvious repetition of useful information for model training, and the second type of area is an area that has obvious repetition of useful information for model training. 4. The method according to claim 3, wherein the step of dividing the original training target image of the neural network training into the first type area and the second type area comprises: Find the maximum intensity value in the original training target image trained by the neural network; Determine the intensity threshold for selecting seed pixels by multiplying the maximum intensity value found above by a coefficient; create an auxiliary image of the same size as the original training target image, the intensity value of which is initially set to zero; In the original training target image, find out the pixel position where the intensity value is greater than the seed threshold; In the auxiliary image, the intensity value of the pixel position where the intensity value is greater than the seed threshold is found is set to a predetermined value, so that many small islands will be formed in the auxiliary image; Through the image growth morphological operation, the islets are iterated for many times to finally form the first type area in the original training target image, and the remaining area in the original training target image is the second type area . 5. The method according to claim 1, characterized in that, in the training process of step S5, batch normalization technology is adopted, and dynamic adaptive sample weighting is adopted to improve training quality. 6 . The method according to claim 1 , wherein in the step S5 , the neural network model is trained by using a stochastic gradient descent method. 7 . 7. The method of claim 1, wherein the imaging conditions include exposure wavelength, numerical aperture, and modes and settings of exposure illumination. 8 . The method according to claim 1 , wherein the N1 is less than or equal to 200. 9 . 9 . The method of claim 1 , wherein the N2 is less than or equal to 6. 10 . 10. The method according to claim 1, further comprising step S7: obtaining an image of the reverse lithography solution according to the value of the reverse lithography solution; identifying the original image from the reverse lithography solution image The designed main pattern area; in the remaining area, the auxiliary function area is located by a predefined intensity threshold to determine the optimal location of the whole-chip sub-resolution auxiliary pattern.

Images