Guide self-assembly photoetching guide template optimization method based on deep learning

Abstract

The invention discloses a guide self-assembly photoetching guide template optimization method based on deep learning. Firstly, a Latin hypercube sampling method is used for sampling in a certain guide template parameter range; and obtaining a training set and a test set according to the sampled parameters by using the guided self-assembly photoetching model. A deep neural network is trained by using a training set and a test set, and the network is directly deployed in guide template optimization of guide self-assembly lithography to replace a guide self-assembly lithography model in the prior art. One-time forward reasoning of the neural network is used for replacing a simulation model, the speed of solving the guide self-assembly photoetching guide template graph can be remarkably increased, and the reverse design efficiency is improved. According to the method, the deep neural network is trained by utilizing the three-dimensional oriented self-assembly photoetching physical simulation data, the three-dimensional information of the self-assembly structure is reserved in network output, the physical information loss is reduced while the speed is ensured, and the result is more accurate.

Description

Technical Field

[0001] The present invention relates to guided self-assembly lithography, and in particular to a method for optimizing a guided template in grapheme-guided self-assembly lithography based on a deep learning method. Background Art

[0002] Guided self-assembly lithography (DSA) technology is a new type of integrated circuit manufacturing graphic technology. The graphic epitaxy method uses the self-assembly of block copolymer materials in a guide template with a specific geometric structure to form the micro-nano structure required in integrated circuit manufacturing. For the patterning of the graphics required for integrated circuit manufacturing, a guide template of a specific shape can be made by photolithography. After filling the block copolymer material, a specific structure can be formed in the guide template through the self-assembly of the material. When the block copolymer material is certain, the shape of the guide template determines the structure of the graphics formed by the guided self-assembly. The guide template optimization method is to optimize the geometry of the guide template so that the block copolymer material can form a self-assembled graphic of a specific size, spacing, and without defects. The optimization method requires repeated calls to the guided self-assembly lithography model.

[0003] The guiding template graphic optimization technology for directed self-assembly proposed by Mentor Graphics Corporation of the United States (see prior art 1, Junjiang Lei, Le Hong, Yuansheng Ma, Guiding patterns optimizationfor directed self-assembly, Patent NO.: US10311165B2) uses a simulation model to calculate the template enhancement factor matrix (TEEF) for the parameterized guiding template geometry, which is used to approximately describe the relationship between the change in geometric parameters and the change in the structure of the hole pattern formed by self-assembly, thereby predicting the change in the self-assembly pattern when the given geometric parameters change, and thereby providing a basis for optimizing the geometric parameters. In the guided template optimization algorithm based on the linearized DSA model (see prior art 2, Azat Latypov, "Computational solution of inverse directed self-assembly problem," Proc. SPIE 8680, Alternative Lithographic Technologies V, 86800Z), the guided template optimization algorithm is modeled as a smooth constrained nonlinear optimization problem, and the linear form of the DSA simulation model is obtained using the weak non-uniform expansion method. The corresponding guided template graphics are solved according to the target structure while accelerating the calculation of the simulation model. In prior art 1, it is necessary to repeatedly call the two-dimensional guided self-assembly lithography model to calculate TEEF, and the simulation model consumes more computing resources and has a slow calculation time, especially for the calculation of nonlinear terms in TEEF, and the design efficiency of the guided template is low. In prior art 2, a linearized two-dimensional self-consistent field theory model is used for fast simulation, which loses some physical information of the self-assembly process and has low accuracy. In both prior art 1 and prior art 2, only the accuracy of the two-dimensional structure is considered, and the three-dimensional characteristics of the self-assembly structure cannot be reflected. Summary of the invention

[0004] The purpose of the present invention is to provide a method for optimizing a guided self-assembly lithography guide template based on deep learning. The method trains a guided self-assembly lithography model based on a deep learning method, replaces a common DSA lithography model, and accelerates the optimization of the guide template.

[0005] The technical solution of the present invention is as follows:

[0006] A method for optimizing a guided self-assembly lithography guide template based on deep learning, comprising the following steps:

[0007] S1) Set material parameters:

[0008] Determine the parameters of the block copolymer material and the guide template material, including: Block tool material separation strength

[0009] χN; Volume fraction of minority block f A ; Guide the interaction strength between the template sidewall and the block copolymer material

[0010] χN sidewall ; The interaction strength between the air layer on the top of the guiding template and the block copolymer material χN top ; Guide the interaction strength between the bottom substrate of the template and the block copolymer material χN bottom ; Radius of gyration R of block copolymer materials g .

[0011] S2) Set the boot template parameter range:

[0012] Determine the dimension m of the geometric parameter space of the guide template, including parameters such as length, width, and height; adopt curve parameterization methods, including spline curves, Bezier curves and other curves; determine the parameter ranges of the geometric parameterized guide template: According to the actual process requirements, determine the parameter ranges of the geometric parameterized guide template so that the parameter range can cover most of the design goals while meeting the manufacturability requirements and other constraints on the geometric shape of the guide template.

[0013] S3) Generate dataset:

[0014] According to the geometric parameter range determined in step S2, the Latin hypercube sampling method is used to fully sample n in the m-dimensional parameter space. train data points. The Latin hypercube sampling steps are as follows:

[0015] 1. Divide each dimension of the m-dimensional parameter space into n non-overlapping, equal-length train intervals to ensure that each interval has the same probability of being selected;

[0016] 2. Randomly sample a point from any interval in each dimension, and form an m-dimensional vector with the sampling points of all dimensions, each of which represents a guiding template geometry.

[0017] 3. Repeat step 2 above until every interval of every dimension of the m-dimensional parameter space has been sampled. The sampling interval of each dimension does not overlap with the previous sampling interval, so a total of n train different sampling points, corresponding to n train Different boot template geometries.

[0018] Also use Latin hypercube sampling to resample n test data points as the test set of the deep neural network. test<n train / 2, and there are no sampling points in the test set that are repeated with those in the training set.

[0019] According to the determined material parameters and sampled geometric parameters, the three-dimensional guided self-assembly lithography model is called to generate the self-assembly structure under each geometric parameter, and a data set including the guided template geometric parameters and the corresponding self-assembly structure is obtained.

[0020] S4) Training deep neural network:

[0021] The steps for training a deep neural network based on the generated training set data are as follows:

[0022] 1. Set up the deep neural network input. The deep neural network input can be a one-dimensional parameter vector, a three-dimensional guidance template voxel data, or other data that encodes the guidance template geometry information.

[0023] 2. Set the deep neural network output. The output target of the deep neural network is the three-dimensional structure formed by guided self-assembly in the corresponding guide template obtained through the guided self-assembly lithography model.

[0024] 3. Set the loss function to measure the difference between the neural network output and the target. The loss function can be in the form of mean square error (MCE), binary cross entropy (BCE), Dice coefficient, etc.

[0025] 4. Set up the structure of the deep neural network. The deep neural network structure uses the 3D-UNet structure as the basis of the neural network, which consists of an encoder connected to a decoder. The encoder compresses the input three-dimensional data into a latent space, and the decoder restores the features of the latent space to the input three-dimensional space. The network structure settings that need to be performed mainly include:

[0026] 4.1. Set up convolutional layers: Extract features through three-dimensional convolution operations between three-dimensional convolution kernels and features. The number of convolutional layers in the encoder is E, and the number of convolution kernels in each convolutional layer is [NE 1 ,NE 2 ,…,NE E ], size is N C ×N C ×N C , the convolution kernel step size is s C , the convolution kernel uses a circular padding method with a padding width of p C ; The number of convolutional layers in the decoder is D, and the number of convolutional kernels in each convolutional layer is [ND 1 ,ND 2 ,…,ND D ], size is N C ×N C ×N C, the convolution kernel step size is s C , the convolution kernel uses a circular padding method with a padding width of p C . Iterate the weight parameters of the 3D convolution kernel through the back-propagation algorithm;

[0027] 4.2. Set the pooling layer: The pooling method can be three-dimensional maximum pooling, three-dimensional average pooling, etc., which divides the data into several parallelepiped areas of size P, outputs the maximum value, average value or other statistics for each area, and realizes downsampling of the data;

[0028] 4.3. Set the activation layer: Apply a nonlinear activation function to the output features of the Batchnorm layer. The activation function can be a linear rectifier function (ReLU), a hyperbolic tangent function (tanh), etc. The activation layer can enhance the nonlinear characteristics of the deep neural network.

[0029] 4.4. Set up the deconvolution layer: The decoder upsamples the input features through a three-dimensional deconvolution operation. The number of deconvolution layers in the decoder is TC, and the number of convolution kernels in each deconvolution layer is [NTC 1 ,NTC 2 ,…,NTC TC ], size is N TC ×N TC ×N TC , the step length is s TC。 The weight parameters of the 3D deconvolution kernel are iterated through the back-propagation algorithm.

[0030] 4.5. Set Batchnorm layer: Apply standardization to the same batch output features of the convolution layer to speed up and simplify the training process. The calculation formula is:

[0031]

[0032] Where μ is the mean of the current batch feature; σ is the standard deviation of the current batch data; ε is a very small constant to ensure numerical stability; γ and β translate and scale the features, iterating through the back propagation algorithm.

[0033] 4.6. Set splicing: Splice the features at the same 3D feature size of the encoder and decoder together so that the decoder network can integrate features of different scales.

[0034] 5. Train the deep neural network. Set the learning rate and batch size. Based on the loss function, use the back propagation algorithm to update the neural network parameters of the convolutional layer, deconvolution layer, and batchnorm layer on the training set generated in step 3). The back propagation algorithm can be Adam, stochastic gradient descent (SGD), root mean square propagation algorithm (RMSProp), etc.

[0035] 6. Verify the deep neural network. The performance verification indicator can be the loss function used in the training process or other forms of performance verification indicators, such as recall rate, F1 score, area under the receiver operating characteristic curve (AUC), etc. Verify the performance of the neural network on the test set generated in step 3). If the performance of the neural network on the test set does not meet the specified requirements, increase the number of sampling points n of Latin cube sampling on the training set. train Increase the coverage of the dataset, resample, and retrain the deep neural network.

[0036] S5) Boot template optimization

[0037] The process of the guided template optimization method in the prior art is as follows Figure 3 As shown, the guided self-assembly lithography model is used to calculate the difference between the self-assembled structure formed by the guiding template and the target structure at each iteration, and the geometric shape of the guiding template is adjusted according to the difference through the optimization algorithm until the difference with the target is less than a set threshold.

[0038] The simulation model in the optimization process is replaced with the deep neural network trained in step S4). The material parameters are set to the material parameters of the data set generated in step S1), the target self-assembly pattern is set, and the execution Figure 3 The optimization process in the above example can obtain the guiding template graphics for generating the target structure. Since the optimization process requires repeated calls to the simulation model, the use of deep neural networks in the optimization process can significantly accelerate the entire optimization process while retaining sufficient simulation accuracy.

[0039] Compared with the prior art, the present invention has the following advantages:

[0040] 1. This method uses a deep neural network to train a guided self-assembly lithography model as a proxy model. Through a forward reasoning of the neural network, the guided self-assembly formation structure is predicted, which can significantly improve the speed of guided self-assembly lithography guide template graphic optimization and improve the efficiency of reverse design.

[0041] 2. This method uses three-dimensional physical simulation data to train a deep neural network. The three-dimensional information of the self-assembled structure is retained in the network output. Compared with other fast models, it loses less physical information and the results are more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a schematic diagram of the process of the present invention

[0043] Figure 2 This is a schematic diagram of the 3D-UNet network structure used in the present invention

[0044] Figure 3 A general flow chart of the guide template optimization method in the prior art

[0045] Figure 4 Schematic diagram of the self-assembly structure in the present invention

[0046] Figure 5 Schematic diagram of parameterized guide template in the example of the present invention

[0047] Figure 6 This is a graph showing the change of neural network performance evaluation index with training iterations in the example of the present invention.

[0048] Figure 7 This is a comparison diagram of the output of the neural network trained in the example of the present invention and the output of the simulation model

[0049] Figure 8 This is a flow chart of the guided template optimization method based on the adaptive evolution of the covariance matrix in the example of the present invention.

[0050] Fig. 9 It is a schematic diagram of the optimization result after the optimization algorithm deploys a fast model based on deep learning in an example of the present invention. DETAILED DESCRIPTION

[0051] The present invention will be further described below in conjunction with embodiments and drawings, but the protection scope of the present invention should not be limited by these embodiments.

[0052] A deep learning-based guided self-assembly lithography guide template optimization method, the method flow is as follows Figure 1 As shown, the method comprises the following steps:

[0053] 1) Set material parameters:

[0054] In guided self-assembly lithography, the parameters of block copolymer materials and guide template materials are determined according to the actual process conditions and directly affect the accuracy and effect of lithography, including: the separation strength of block tool materials χN = 15; the volume fraction of minority blocks f A =0.3, this block copolymer material will self-assemble on the unconstrained system to form a hexagonal stacked cylindrical phase; the interaction strength between the guiding template sidewall and the block copolymer material χN sidewall =17, indicating the guide template sidewalls that attract a few blocks; the interaction strength between the guide template top air layer and the bottom substrate and the block copolymer material χN top and χN bottom Set to 0, indicating a neutral air layer and substrate, which helps to generate a cylindrical phase that penetrates the guide template; the radius of gyration R of the block copolymer material g =5.6nm.

[0055] Under this parameter setting, after material self-assembly, the sidewall of the guiding template will absorb an affinity layer composed of a few blocks, and the remaining few blocks will form discrete columns or other defective shapes in the middle of the guiding template, such as Figure 4 shown.

[0056] 2) Set the boot template parameter range:

[0057] The guiding template is a key component in guided self-assembly lithography, and its geometric parameters directly affect the formation and accuracy of the final pattern.

[0058] The boot template definition method used in this embodiment is as follows Figure 5 As shown. The dimension of the geometric parameter space of the guide template is m=5: horizontal width CD-HX, vertical width CD-TY, middle horizontal width CD-TBL, middle vertical width CD-TBW, inflection point slope S. Five parameters determine 12 points on the two-dimensional plane, and the guide template is obtained by connecting the 12 points with a spline curve. According to the actual process requirements, the range of each parameter of the geometric parameterized guide template is determined as (unit: R g ): CD-HX∈[12,18],CD-TY∈[6.5,9],CD-TBW∈[(CD-TY-CD-TBW) / 2S,CD-HX-CD-TY],CD-TBW∈[0.4CD-TY,0.8CD-TY],S∈[1,2]. This parameter range can cover most of the design goals while meeting the manufacturability requirements and other constraints on the guide template geometry.

[0059] 3) Generate dataset:

[0060] According to the geometric parameter range determined in step 2, the Latin hypercube sampling method is used to fully sample n in the entire high-dimensional parameter space. train = 500 data points for training deep neural networks; sampling n test = 100 data points for testing the deep neural network. There are no repeated sampling points in the test set and the training set. The self-consistent field theory simulation model is called, and the material parameters set in step 1) and the sampled geometric parameters are used as input to generate the guided self-assembly structure under the corresponding input to obtain the data set.

[0061] 4) Training Deep Neural Networks:

[0062] The steps for training a deep neural network based on the generated training set data are as follows:

[0063] 1. Set the deep neural network input. The input is the voxel data of the three-dimensional guide template, that is, the internal space of the guide template is set to 1, and the rest of the space is set to 0;

[0064] 2. Set the output of the deep neural network. The output target of the deep neural network is the three-dimensional structure formed by guided self-assembly in the corresponding guide template obtained through the self-consistent field theory simulation model;

[0065] 3. Set the loss function to measure the difference between the neural network output and the target. Set it to the binary weighted cross entropy function, and the calculation formula is as follows;

[0066]

[0067] Here, α=2.8 is used to balance the class imbalance.

[0068] 4. Set the structure of the deep neural network. The deep neural network structure adopts the 3D-UNet structure, such as Figure 2 As shown in Figure 1. The network consists of an encoder connected to a decoder, where the encoder compresses the input three-dimensional data into a latent space, and the decoder restores the features of the latent space to the input three-dimensional space. The network structure mainly includes:

[0069] 4.1. Convolutional layer: extract features through 3D convolution operation between 3D convolution kernel and features. The number of convolutional layers in the encoder is E=8, the number of convolutional kernels in each convolutional layer is [32, 64, 64, 128, 128, 256, 256, 512], the size is 7×7×7, and the convolution kernel step size is s C =1, the convolution kernel uses a circular padding method with a padding width of p C =3; the number of convolutional layers in the decoder is D = 8, the number of convolutional kernels in each convolutional layer is [256, 256, 128, 128, 64, 64, 32, 1], the size is 7 × 7 × 7, and the convolution kernel step size is s C =1, the convolution kernel uses a circular padding method with a padding width of p C = 3. The weight parameters of the 3D convolution kernel are iterated through the back-propagation algorithm.

[0070] 4.2. Pooling layer: Select three-dimensional maximum pooling, divide the data into several parallelepiped regions of size P = 2, output the maximum value for each region, and achieve downsampling of the data;

[0071] 4.3. Activation layer: Apply a nonlinear activation function to the output features of the batchnorm layer. The activation function is selected as a linear rectification function, as shown in formula (1). The activation layer can enhance the nonlinear characteristics of deep neural networks.

[0072]

[0073] 4.4. Deconvolution layer: The decoder upsamples the input features through a three-dimensional deconvolution operation. The number of deconvolution layers in the decoder is TC=3, and the number of convolution kernels in each deconvolution layer is [512, 256, 128], the size is 2×2×2, and the stride is 2 。 The weight parameters of the 3D deconvolution kernel are iterated through the back-propagation algorithm.

[0074] 4.5.Batchnorm layer: Standardizes the output features of the same batch of the convolutional layer to speed up and simplify the training process. The calculation formula is:

[0075]

[0076] Where μ is the mean of the current batch feature; σ is the standard deviation of the current batch data; ε is a very small constant to ensure numerical stability; γ and β translate and scale the features, iterating through the back propagation algorithm.

[0077] 4.6. Splicing: Splice together the features at the same 3D feature size of the encoder and decoder so that the decoder network can integrate features of different scales.

[0078] 5. Train the deep neural network. Set the learning rate to 10 -6 , the batch size is set to 4. Based on the loss function, the back propagation algorithm is used to update the neural network parameters of the convolution layer, deconvolution layer, and batchnorm layer on the training set generated in step 3). The back propagation algorithm selects the root mean square propagation algorithm (RMSProp).

[0079] 6. Verify the deep neural network. The performance verification indicator is selected as the area under the receiver operating characteristic curve (AUC) and the like. Verify the performance of the neural network on the test set generated in step 3). Figure 6 The performance changes of the neural network on the test set during the training process are given. The performance of the neural network on the test set can reach AUC>0.999, which meets the requirements. Figure 7 Four sets of data from the test set were randomly selected to compare the output of the neural network and the simulation output of the self-consistent field theory model. The neural network can accurately predict the three-dimensional structure formed by self-assembly. While the self-consistent field theory simulation model takes 75 seconds to calculate, the neural network only takes 0.005 seconds, significantly improving the calculation time.

[0080] 5) Boot template optimization

[0081] Deploy a deep neural network in the guided template optimization method based on the covariance matrix adaptive evolutionary algorithm. The optimization method process is as follows Figure 8As shown. In each iteration, a new population is randomly sampled from the multivariate Gaussian distribution, and each sample in the population corresponds to a guiding template shape; the self-assembled structure of each sample is calculated through the self-consistent field theory model; the fitness is calculated through the loss function to measure the difference between the self-assembled structure of each sample and the target structure; the mean and covariance matrix of the multivariate Gaussian distribution are updated according to the fitness, and the process is repeated until the difference between the self-assembled structure of the sample and the target structure is less than the set threshold. Replace the self-consistent field theory model in the optimization process with the deep neural network ( Figure 8 The material parameters are set as the material parameters of the data set generated in step 1), the target self-assembly pattern is set, and the execution Figure 8 The optimization process in the above example can generate the guide template graph for the target structure. By using deep neural networks in the optimization process, the guide template optimization speed is increased by about 100 times while retaining sufficient simulation accuracy.

[0082] In order to verify the structure solved by the optimization algorithm based on deep learning acceleration, three different optimization objectives were tested, such as Fig. 9 shown. Fig. 9 The shapes of the guiding templates before and after optimization are given, and the self-assembled structures generated by each guiding template are verified using a self-consistent field theory model. Fig. 9 The optimization results show that when the optimization speed is increased by 100 times, the accuracy of the deep neural network trained based on the data set generated by the self-consistent field theory model is sufficient to replace the self-consistent field theory model in the optimization method, and the guiding template that can self-assemble to form the target structure and eliminate self-assembly defects is solved.

[0083] The above is only a specific embodiment of the present invention, which is only used to illustrate the technical solution of the present invention rather than to limit the present invention. Any technical solution that can be obtained by those skilled in the art through logical analysis, reasoning or limited experiments according to the concept of the present invention should be within the protection scope of the present invention.

Claims

1. A method for optimizing a guided self-assembly lithography guide template based on deep learning, characterized in that: The following steps are involved: S1. Setting parameters of the block copolymer material and the guide template material, including the separation strength of the block tool material χN, the volume fraction of the minority block fA, the interaction strength between the guide template sidewall and the block copolymer material χNsidewall, the interaction strength between the guide template top air layer and the block copolymer material χNtop, the interaction strength between the guide template bottom substrate and the block copolymer material χNbottom, and the gyration radius Rg of the block copolymer material; S2. Setting the parameter range of the guide template: determining the dimension m of the guide template geometric parameter space, and using a curve parameterization method to describe the shape of the guide template, and determining the parameter range of each parameter of the geometric parameterized guide template, so that the parameter range can cover most of the design goals, while meeting the manufacturability requirements and other constraints on the geometric shape of the guide template; S3. Generate a data set: According to the geometric parameter range determined in step S2., use the Latin hypercube sampling method to sample in the m-dimensional parameter space to obtain a training data set and a test data set, call the three-dimensional guided self-assembly lithography model to generate a self-assembled structure under each geometric parameter, and obtain a data set containing the geometric parameters of the guide template and the corresponding self-assembly structure; S4 training a deep neural network: using the training data set generated in step S3. to train a deep neural network, the deep neural network adopts a 3D-UNet structure, including an encoder, a decoder, a convolution layer, a deconvolution layer, a Batchnorm layer, etc., the network input is a guide template geometry parameter, the output is a predicted self-assembly structure, the difference between the network output and the target self-assembly structure is measured by a loss function, and the network parameters are optimized using a back propagation algorithm; S5. Guided template optimization: embed the trained deep neural network into the guided template optimization process, set the target self-assembly pattern, continuously adjust the geometric shape of the guided template through the optimization algorithm, use the deep neural network to predict the self-assembly structure in each iteration, and adjust the guided template according to the difference between the predicted result and the target structure until the difference is less than the set threshold, and obtain the guided template pattern that generates the target structure.

2. The method for optimizing the guided self-assembly lithography guide template based on deep learning according to claim 1, characterized in that: The step S3. uses Latin hypercube sampling to evenly divide intervals in the parameter space of each dimension, and randomly samples a point from each interval to form an m-dimensional vector representing a guiding template geometry, and repeats sampling until all intervals are covered to obtain a training data set and a test data set.

3. The method for optimizing guided self-assembly lithography guide template based on deep learning according to claim 1, characterized in that: In step S4, the encoder of the deep neural network includes multiple convolutional layers, each convolutional layer uses a three-dimensional convolution kernel to extract features, the decoder includes multiple deconvolution layers for restoring the features to three-dimensional space, and the Batchnorm layer is used to standardize the output features of the convolutional layer.

4. The method for optimizing guided self-assembly lithography guide template based on deep learning according to claim 1, characterized in that: The step S4. also includes the step of verifying the deep neural network, using a test data set to verify the performance of the neural network. If the performance does not meet the specified requirements, the number of sampling points in the training data set is increased and the deep neural network is retrained.

5. The method for optimizing guided self-assembly lithography guide template based on deep learning according to claim 1, characterized in that: In step S5., the optimization algorithm adopts a gradient descent method, a genetic algorithm or a simulated annealing algorithm, etc., and predicts the self-assembly structure by repeatedly calling a deep neural network, and adjusts the geometric shape of the guide template according to the prediction result.

Images