Reverse photoetching method based on multilayer perceptron

By performing iterative optimization of reverse lithography technology based on multi-layer perceptron machine learning model, the problems of high computational cost and high mask complexity in reverse lithography technology are solved, and fast and high-quality mask optimization is achieved, reducing the Runtime and complexity.

CN119045292AInactive Publication Date: 2024-11-29ZHEJIANG UNIV +1

Patent Information

Application Number
CN202411536852.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-10-31
Publication Date
2024-11-29
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The high computational cost, high mask complexity, and equipment limitation of reverse lithography technology make it difficult to apply on a large scale and to quickly optimize the mask.

Method used

Using a machine learning model based on a multi-layer perceptron, the graphics set and corresponding masks are trained to test the graphics set and the corresponding masks, intermediate parameters are generated and low-pass filtered, mapped to the grayscale mask, and iteratively optimized in combination with lithography simulation and loss function until The iteration end condition is met and an optimized mask is generated.

Benefits of technology

The running time of reverse lithography technology is shortened, the mask optimization effect is improved, the mask complexity is reduced, and the robustness of the optimization algorithm is improved through regularization loss function.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119045292A_ABST
    Figure CN119045292A_ABST
Patent Text Reader

Abstract

The invention provides a reverse photoetching method based on a multilayer perceptron. The reverse photoetching method comprises the following steps: acquiring a test pattern set and a mask corresponding to the test pattern set; constructing a machine learning model based on the multi-layer perceptron, taking the test pattern set and the corresponding mask as a training data set, and inputting the training data set into the machine learning model based on the multi-layer perceptron for training; providing an integrated circuit design layout (target graph), inputting the target graph into the trained model, and outputting intermediate parameters; and based on the intermediate parameters, carrying out iterative optimization by using a reverse photoetching technology algorithm, and when iteration is finished, outputting a final optimized mask. According to the reverse photoetching technology based on the multilayer perceptron, the operation time of the reverse photoetching technology can be shortened, and the mask optimization effect is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of semiconductor manufacturing, and in particular relates to a reverse lithography method based on a multi-layer perceptron. Background Art

[0002] Photolithography is the most critical and complex process in integrated circuit manufacturing. The quality and time consumption of photolithography greatly affect the cost of integrated circuit manufacturing. In advanced technology nodes, the size of the feature size is much smaller than the wavelength of the light source of the photolithography machine, which is inseparable from various resolution improvement technologies including optical proximity correction. Optical proximity correction occupies an important position in resolution improvement technology, including early rule-based optical proximity correction and subsequent model-based optical proximity correction technology. However, as the feature size of integrated circuits continues to shrink, traditional optical proximity correction can no longer meet the needs of photolithography, and engineers have begun to shift their attention to reverse lithography technology.

[0003] Inverse Lithography Technology (ILT) regards mask optimization as the inverse process of lithography, and the target image as the ultimate goal of optimization, and obtains the optimized mask through inverse calculation. Since inverse lithography has a higher degree of freedom, it can further improve the accuracy of imaging. In addition, since it can produce sub-resolution auxiliary graphics with more free shapes, its process variation bandwidth is also greatly improved compared with traditional optical proximity effect correction technology. However, inverse lithography has high computational cost, high mask complexity, equipment limitations and other problems, making it difficult to apply on a large scale. At present, the application of inverse lithography technology is mainly focused on the repair of lithography hotspots, the optimization of sub-resolution auxiliary graphics rules, and the optimization of masks for devices with high graphic repetition.

[0004] Therefore, how to speed up the reverse lithography technology while ensuring the quality of mask optimization has become an urgent problem to be solved. Summary of the invention

[0005] The purpose of the present invention is to provide a reverse lithography method based on a multi-layer perceptron to address the deficiencies of the prior art, so as to shorten the operation time of the reverse lithography technology and improve the effect of mask optimization.

[0006] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0007] S1, obtaining a test pattern set and its corresponding mask;

[0008] S2. Build a machine learning model based on a multi-layer perceptron, use the test graph set and the corresponding mask as a training data set, input them into the machine learning model based on the multi-layer perceptron for training; input the target graph into the trained model, and output the intermediate parameters;

[0009] The intermediate parameter takes any value between positive and negative infinity. Since the mask has only 0 and 1 values, it is impossible to directly perform gradient descent optimization in the parameter space. Therefore, the intermediate parameter is introduced in the present invention. After the intermediate parameter optimization is completed, the intermediate parameter is mapped and binarized to obtain a final mask with only 0 and 1 values.

[0010] S3, low-pass filtering the intermediate parameters, and then mapping the filtered parameter values ​​to between 0 and 1 to obtain a grayscale mask;

[0011] S4, performing lithography simulation on the test pattern set to obtain a lithography model; inputting the grayscale mask into the lithography model to obtain a photoresist pattern;

[0012] S5, calculating the loss of the target pattern and the photoresist pattern, back-propagating the loss value, obtaining the gradient of the intermediate parameter update, setting the update step size, and updating the intermediate parameters for iteration;

[0013] S6. According to the set iteration end condition, if the iteration end condition is met, the grayscale mask is converted into a binary mask to obtain a final optimized mask; otherwise, return to step S3 and perform iterative operation until the iteration end condition is met.

[0014] Furthermore, the above-mentioned test graphic set includes 1D graphics, 1.5D graphics and 2D graphics, and the feature size and spacing of the test graphic set should cover all situations in the integrated circuit design layout to ensure the quality of machine learning model training.

[0015] Furthermore, the photolithography model includes an optical model, a photoresist model and a mask model. The optical model, the photoresist model and the mask model are set according to process information and calibrated according to measured data, and their simulation results can represent actual exposure results.

[0016] Furthermore, the test pattern set and its corresponding mask are obtained by using an inverse lithography algorithm, which includes low-pass filtering operation, parameter mapping, lithography simulation, calculation loss, gradient propagation, and mask update. These operations are performed iteratively until convergence conditions are met. The loss may include L2 loss of the photoresist pattern and the target pattern, process variation bandwidth loss, and regularization loss. The convergence conditions include setting a fixed number of iterations, the root mean square of the gradient being less than a fixed threshold, etc.

[0017] Furthermore, the machine learning model based on the multi-layer perceptron adopts an encoder-decoder structure, wherein the encoder adopts three convolution modules and two multi-layer perceptron modules connected in series in sequence; the decoder is a reverse structure of the encoder, including two multi-layer perceptron modules and three convolution modules connected in series in sequence; and jump connections are adopted between the corresponding modules of the encoder and the decoder;

[0018] Each of the convolution modules includes an input layer, a convolution layer, a batch normalization layer, an activation function and an output layer connected in series in sequence; each of the multi-layer perceptron modules includes an input layer, a hidden layer and an output layer connected in series in sequence, and the hidden layer performs tokenization, full connection, layer normalization and activation operations in sequence.

[0019] Furthermore, when the method of the present invention is applied to actual production, the design layout is input as the target pattern, and the optimized mask pattern is output; specifically, the integrated circuit design layout is cut before input to adapt to the input size of the machine learning model. Then, the slices obtained by cutting are respectively optimized for mask optimization, and finally the optimized masks are spliced ​​to complete the mask optimization task of the integrated circuit design layout.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] The present invention provides a reverse lithography technology based on a multi-layer perceptron, which runs a reverse lithography technology algorithm on a test pattern set, and iteratively optimizes and generates an optimized mask corresponding to the test pattern set. The test pattern set and the corresponding mask are used as training sets of a machine learning model to train the machine learning model. The machine learning model is a lightweight encoder-decoder structure based on a multi-layer perceptron, which can quickly provide a high-quality initial solution after training, reducing the number of iterations required for subsequent reverse lithography technology, thereby shortening the time required for mask optimization. A low-pass filtering operation is introduced in the iterative process to reduce the complexity of the mask. Adding a regularization loss to the loss function is conducive to optimizing the gradient descent process to jump out of the local optimal solution and improving the robustness of the optimization algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a schematic diagram of the process of reverse lithography technology based on multi-layer perceptron in an example of the present invention;

[0023] Figure 2 is a flowchart of the training and deployment of a machine learning model in an example of the present invention;

[0024] Figure 3 It is a schematic diagram of the structure of the machine learning model in the example of the present invention;

[0025] Figure 4 It is a schematic diagram of the structure of the convolution block in the example of the present invention;

[0026] Figure 5 It is a schematic diagram of the structure of the multi-layer perceptron block in the example of the present invention. DETAILED DESCRIPTION

[0027] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0028] Figure 1 It is a flow chart of the reverse lithography technology based on the multi-layer perceptron in the example of the present invention.

[0029] Please refer to Figure 1 , the reverse lithography correction technology based on multi-layer perceptron includes:

[0030] Step S10: training and deploying a machine learning model based on a multi-layer perceptron, inputting the target graph into the machine learning model based on the multi-layer perceptron, and the machine learning model inferring to obtain intermediate parameters;

[0031] Step S20: performing a low-pass filtering operation on the intermediate parameters and mapping the parameters to between 0 and 1 to obtain a grayscale mask;

[0032] Step S30: inputting the grayscale mask into the photolithography simulation model, performing photolithography simulation under different process conditions, and obtaining a series of photoresist patterns;

[0033] Step S40: Calculating the loss based on the loss function according to the target pattern and the photoresist pattern obtained in the previous step;

[0034] Step S50: back-propagate the loss value to obtain the gradient of parameter update, set the update step size, and update the parameters;

[0035] Step S60: According to the set iteration end condition, if the iteration end condition is met, the grayscale mask is converted into a binary mask to obtain a final optimized mask; otherwise, return to step S20 and iterate until the iteration end condition is met.

[0036] Next, please combine Figures 2 to 5 Each step is analyzed and explained. Figure 2 It is a flowchart process of training and deploying a machine learning model in an example of the present invention; Figure 3 It is a schematic diagram of the structure of the machine learning model in the example of the present invention; Figure 4 It is a schematic diagram of the structure of the convolution block in the example of the present invention; Figure 5 It is a schematic diagram of the structure of the multi-layer perceptron block in the example of the present invention.

[0037] Please continue to refer to Figure 1, executing step S10: training and deploying the machine learning model, inputting the target graph into the machine learning model, and the machine learning model inferring to obtain initial parameters;

[0038] Please refer to Figure 2 In this embodiment, the training and deployment process of the machine learning model based on the multi-layer perceptron includes:

[0039] Step S101, designing a test pattern set according to process information and design layout;

[0040] In this embodiment, the test pattern set includes 1D patterns, 1.5D patterns and 2D patterns, and the feature sizes and spacings of the test pattern set should cover all situations in the integrated circuit design layout.

[0041] Step S102, according to the test pattern set, using the reverse lithography technology algorithm to iteratively optimize and generate a corresponding optimized mask to form a data set;

[0042] In this embodiment, the inverse lithography algorithm includes low-pass filtering, parameter mapping, lithography simulation, loss calculation, gradient propagation, and mask update, and these operations are performed iteratively until convergence conditions are met.

[0043] The filtering operation is low-pass filtering, the purpose of which is to reduce the complexity of the mask. Since the manufacture of the mask has mask rules such as the minimum feature size and the minimum spacing, reducing the complexity of the mask during the optimization process helps to generate a mask with stronger manufacturability. Low-pass filtering operations include Gaussian filtering, mean filtering, and median filtering. Preferably, Gaussian filtering is used as the mask filtering method, and the range of filtering and the strength of the smoothing effect can be controlled by controlling the size of the filter kernel and the size of the standard deviation.

[0044] The parameter mapping operation maps the intermediate parameters in the positive and negative infinity range to the mask transmittance between 0 and 1. Since the gradient descent needs to be performed in the positive and negative infinity range, the intermediate parameters are introduced and the gradient descent optimization is performed on the intermediate parameters. The transmittance of the grayscale mask ranges from 0 to 1, so the positive and negative infinity parameter range needs to be mapped to between 0 and 1, and the mapping function is required to be differentiable. Further, the mapping function includes a Sigmoid function and a Tanh function. Preferably, a Sigmoid function is used as the mapping function.

[0045] The lithography model used in the lithography simulation includes an optical model, a photoresist model and a mask model. The optical model, the photoresist model and the mask model are set according to process information and calibrated according to measured data, and the simulation result can represent the actual exposure effect.

[0046] The loss function includes L2 loss of photoresist pattern and target pattern, process variation bandwidth loss and regularization loss, etc. L2 loss can minimize the difference between photoresist pattern and design pattern and improve pattern accuracy. Process variation bandwidth loss helps to reduce the variation range of photoresist during process fluctuation and widen the process window. Regularization operation helps the optimization process to jump out of the local optimal solution, enhance the robustness of the optimization algorithm and improve the optimization effect. Among them, the regularization loss adopts a coefficient attenuation strategy, and the coefficient attenuation strategy includes equal interval attenuation, exponential attenuation, cosine attenuation, etc. Preferably, exponential attenuation is used as the coefficient attenuation strategy.

[0047] The convergence condition includes setting a fixed number of iterations, the root mean square of the gradient being less than a fixed threshold, etc. Preferably, setting a fixed number of iterations is used as the convergence condition.

[0048] Step S103, using the above training set to train a machine learning model;

[0049] Please refer to Figure 3 In this embodiment, the machine learning model is an encoder-decoder structure, including multiple downsampling modules and upsampling modules. Each upsampling module or downsampling module is a convolution module or a multi-layer perceptron module, and a jump connection is used between the corresponding upsampling modules and downsampling modules. Preferably, three consecutive convolution modules are used as the downsampling model, two multi-layer perceptron modules are used as the downsampling module, two multi-layer perceptron modules are used as the upsampling block, and three convolution modules are used as the upsampling model.

[0050] Please refer to Figure 4 , in this embodiment, the convolution module performs continuous two-dimensional convolution, batch normalization and activation operations.

[0051] Please refer to Figure 5 In this embodiment, the multi-layer perceptron module performs tokenization, full connection, layer normalization and activation operations.

[0052] Step S104, deploying the machine learning model. After the machine learning model is deployed, the parameters can still be updated through online learning methods to further improve the prediction effect of the machine learning model.

[0053] The process of inputting the target graphic and inferring the initial function includes cutting the integrated circuit design layout to adapt to the input size of the machine learning model.

[0054] Step S20: performing a low-pass filtering operation on the intermediate parameters and mapping the parameters to between 0 and 1 to obtain a grayscale mask;

[0055] The purpose of the low-pass filtering operation is to reduce the complexity of the mask. Since the manufacture of the mask has mask rules such as the minimum feature size and the minimum pattern spacing, reducing the complexity of the mask during the optimization process helps to generate a mask with stronger manufacturability. Low-pass filtering operations include Gaussian filtering, mean filtering, and median filtering. Preferably, Gaussian filtering is used as the mask filtering method, and the range of filtering and the strength of the smoothing effect can be controlled by controlling the size of the filter kernel and the size of the standard deviation.

[0056] The parameter mapping operation maps the intermediate parameters in the positive and negative infinity range to the mask transmittance parameters between 0 and 1. Since the gradient descent needs to be performed in the positive and negative infinity range, the intermediate parameters are introduced. The transmittance of the grayscale mask ranges from 0 to 1, so the positive and negative infinity parameter range needs to be mapped to between 0 and 1, and the mapping function is required to be differentiable. Further, the mapping function includes a Sigmoid function and a Tanh function. Preferably, a Sigmoid function is used as the mapping function.

[0057] Step S30: inputting the grayscale mask into the photolithography simulation model, performing photolithography simulation under different process conditions, and obtaining a series of photoresist patterns;

[0058] The lithography model used in the lithography simulation includes an optical model, a photoresist model and a mask model. The optical model, the photoresist model and the mask model are set according to process information and calibrated according to measured data, and the simulation results can represent the actual exposure results.

[0059] Step S40: Calculating the loss based on the loss function according to the target pattern and the photoresist pattern obtained in the previous step;

[0060] The loss function includes L2 loss of photoresist pattern and target pattern, process variation bandwidth loss and regularization loss, etc. L2 loss can minimize the difference between photoresist pattern and design pattern and improve pattern accuracy. Process variation bandwidth loss helps to reduce the variation range of photoresist during process fluctuation and widen the process window. Regularization operation helps the optimization process to jump out of the local optimal solution, enhance the robustness of the optimization process and improve the optimization effect. Among them, the regularization loss adopts a coefficient attenuation strategy, and the coefficient attenuation strategy includes equal interval attenuation, exponential attenuation, cosine attenuation, etc. Preferably, exponential attenuation is used as the coefficient attenuation strategy.

[0061] Step S50: The loss value is back-propagated to obtain the gradient of parameter update, the update step size is set, and the parameters are updated; the update step size can be constant or decaying. Preferably, the update step size is decaying. The decaying update step size is beneficial to jump out of the local optimal solution in the early stage of optimization and stabilize the convergence in the later stage of optimization.

[0062] Step S60: According to the set iteration end condition, if the iteration end condition is met, the grayscale mask is converted into a binary mask to obtain the final optimized mask; otherwise, return to step S20 and iterate until the iteration end condition is met. When the mask optimization task of all slices is completed, the optimized masks are finally spliced, that is, the mask optimization task of the integrated circuit design layout is completed.

[0063] In this embodiment, the iteration end condition is set to a fixed number of iterations; if only a mask needs to be obtained quickly, the number of iterations is set to 20 times; if a high-quality mask needs to be obtained, the number of iterations is set to 100 times.

[0064] In order to illustrate that the method of the present invention can shorten the time required for mask optimization, conventional inverse lithography technology algorithm A2 inverse lithography technology (A2-ILT), multi-level inverse lithography technology (Multi-Level Inverse Lithography Technology, Multi-Level-ILT) and Swin Transformer-based inverse lithography technology (Swin Transformer-based Inverse Lithography Technology, SwinT-ILT) are used to compare with the method of the present invention, and the results are shown in Table 1 below:

[0065] Table 1 Mask optimization results of different reverse lithography methods

[0066]

[0067] It can be seen from Table 1 that the method of the present invention reduces the number of iterations required for the subsequent reverse lithography technology, thereby shortening the time required for mask optimization.

[0068] The above description is merely a preferred embodiment of one or more embodiments of the present specification and is not intended to limit one or more embodiments of the present specification. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of one or more embodiments of the present specification shall be included in the scope of protection of one or more embodiments of the present specification.

Claims

1. A reverse lithography method based on a multi-layer perceptron, characterized in that: The method comprises the following steps: S1, obtaining a test pattern set and its corresponding mask; S2. Build a machine learning model based on a multi-layer perceptron, use the test graph set and the corresponding mask as a training data set, input them into the machine learning model based on the multi-layer perceptron for training; input the target graph into the trained model, and output the intermediate parameters; S3, low-pass filtering the intermediate parameters, and then mapping the filtered parameter values ​​to between 0 and 1 to obtain a grayscale mask; S4, performing lithography simulation on the test pattern set to obtain a lithography model; inputting the grayscale mask into the lithography model to obtain a photoresist pattern; S5, calculating the loss of the target pattern and the photoresist pattern, back-propagating the loss value, obtaining the gradient of the intermediate parameter update, setting the update step size, and updating the intermediate parameters for iteration; S6. According to the set iteration end condition, if the iteration end condition is met, the grayscale mask is converted into a binary mask to obtain a final optimized mask; otherwise, return to step S3 and perform iterative operation until the iteration end condition is met.

2. The reverse lithography method based on multi-layer perceptron according to claim 1, characterized in that: The method comprises the following steps: the test graphic set comprises 1D graphics, 1.5D graphics and 2D graphics.

3. The reverse lithography method based on multi-layer perceptron according to claim 1, characterized in that: The photolithography model includes an optical model, a photoresist model and a mask model.

4. The reverse lithography method based on multi-layer perceptron according to claim 1, characterized in that: The low-pass filtering includes Gaussian filtering, mean filtering and median filtering.

5. The reverse lithography method based on multi-layer perceptron according to claim 1, characterized in that: The loss function uses the L2 loss of the photoresist pattern and the target pattern, the process variation bandwidth loss and the regularization loss.

6. The reverse lithography method based on multi-layer perceptron according to claim 1, characterized in that: The machine learning model based on the multi-layer perceptron adopts an encoder-decoder structure, wherein the encoder adopts three convolution modules and two multi-layer perceptron modules connected in series in sequence; the decoder is the reverse structure of the encoder, including two multi-layer perceptron modules and three convolution modules connected in series in sequence; and the corresponding modules of the encoder and the decoder adopt a jump connection; Each of the convolution modules includes an input layer, a convolution layer, a batch normalization layer, an activation function and an output layer connected in series in sequence; each of the multi-layer perceptron modules includes an input layer, a hidden layer and an output layer connected in series in sequence, and the hidden layer performs tokenization, full connection, layer normalization and activation operations in sequence.

7. A system for implementing the method according to any one of claims 1 to 6, characterized in that: include: The image acquisition module is used to acquire the target image, input the target into the trained model, and output the intermediate parameters; The reverse lithography module performs low-pass filtering, parameter mapping, lithography simulation, loss calculation, gradient propagation, mask update on the target graphics or intermediate parameters, and outputs the optimized mask.

8. A computer-readable storage medium having a computer program stored thereon, wherein when the program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

9. An electronic device, comprising a processor and a memory, wherein the memory stores a computer program, and when the processor executes the computer program, the method according to any one of claims 1 to 6 is implemented.

Citation Information

Patent Citations

  • Lithography model calibration

    CN111627799A

  • Logistic regression model training method and device, computer equipment and storage medium

    CN113255927A

  • Neural network based mask synthesis for integrated circuits

    CN114365042A

  • Method for improving consistency of mask pattern generation

    CN114981724A

  • Steel fuel tank for vehicle

    KR102640046B1

Cited By

  • Mask pattern optimization method, medium, product and equipment

    CN121254557A