Optical proximity correction method and photoetching mask generation model

Abstract

The invention provides a photoetching mask generation model for optical proximity correction. The photoetching mask generation model comprises an encoder consisting of coordinate convolution modules and a decoder consisting of convolution modules, the encoder performs spatial feature extraction on the input feature map to obtain the specific position of each pixel coordinate in the input design layout; the decoder maps the output feature of the encoder back to the original input dimension to generate a prediction output; a coordinate channel is added in an input feature map to obtain a modified input feature map, the convolution kernel dimension of a coordinate convolution module, receiving the modified input feature map, of the encoder is correspondingly increased, and the added coordinate channel is used for improving the capacity of the encoder for obtaining spatial position information of a design layout. The invention also provides an optical proximity correction method which is trained by using the photoetching mask generation model for optical proximity correction provided by the invention so as to improve the correction precision.

Description

Technical Field

[0001] The present invention belongs to the technical field of integrated circuit manufacturing, and in particular relates to an optical proximity correction method and a photolithography mask generation model. Background Art

[0002] As a core process in the manufacture of very large-scale integrated circuits, the performance of photolithography technology directly determines the integration and yield of chips. With the continuous evolution of semiconductor manufacturing processes, device feature sizes have been reduced by a ratio of 0.7 from generation to generation (i.e., the critical dimensions of each technology node generation are 70% of the previous generation), significantly increasing device density per unit area and reducing the cost of individual transistors. However, when technology nodes reach 90 nanometers and below, traditional photolithography processes face severe challenges: due to the physical differences between the light source wavelength, numerical aperture (NA), and mask design, optical imaging systems struggle to accurately reproduce the design layout, resulting in significant distortion of the developed pattern on the wafer surface.

[0003] To address this issue, Optical Proximity Correction (OPC) technology was developed. This technology compensates for diffraction and proximity effects during the photolithography process by correcting the graphics of the chip design layout during the mask design phase (such as adding auxiliary features and phase-shift structures), thereby ensuring that the developed pattern is highly consistent with the target design. Early rule-based OPC methods (Rule-Based OPC) were gradually phased out due to their inability to adapt to complex process conditions. Model-Based OPC, while offering higher accuracy, is limited by its high computational complexity (requiring massive photolithography simulations) and long optimization cycles (usually taking hours to days), making it difficult to meet the dual efficiency and cost requirements of advanced processes.

[0004] In recent years, breakthroughs in graphics processing units (GPUs) and deep learning technologies have brought new opportunities to the OPC field. The parallel computing capabilities of GPUs can efficiently process large amounts of data in lithography simulations (such as mask images and optical propagation matrices), while deep neural networks (DNNs) directly model the nonlinear mapping relationship between the design layout and the correction mask through an end-to-end learning mechanism. This fusion solution not only significantly reduces computing resource consumption (by 1-2 orders of magnitude compared to traditional methods), but also achieves dual optimization of the global geometric consistency and local critical dimensional stability of the correction effect through multi-dimensional evaluation indicators, providing a feasible path for OPC technology innovation in advanced processes.

[0005] However, the existing technology based on deep learning-assisted optical proximity correction has the following problems: although the traditional convolutional neural network can share unified convolution kernel parameters at different positions in the image, it is not ideal in obtaining the position information of the mask pattern structure. Due to the limitation of traditional convolution operations, the spatial position representation cannot be converted into coordinates in Cartesian space, and the exact position of the structural features in the layout cannot be perceived, resulting in low mask correction accuracy.

[0006] Therefore, it is necessary to develop a new photolithography mask generation model for optical proximity correction, which can improve the photolithography mask correction accuracy. Summary of the Invention

[0007] The present invention provides a photolithography mask generation model for optical proximity correction, which improves the layout space position acquisition capability by introducing an additional coordinate channel.

[0008] The present invention also provides an optical proximity correction method based on the combination of loss calculation and lithography simulation to improve the accuracy of the lithography mask generation model.

[0009] Other purposes and advantages of the present invention can be further understood from the technical features disclosed in the present invention.

[0010] In order to achieve one or part or all of the above-mentioned purposes or other purposes, a technical solution of the present invention provides a photolithography mask generation model for optical proximity correction, wherein the photolithography mask generation model includes an encoder composed of a coordinate convolution module and a decoder composed of a convolution module; the encoder performs spatial feature extraction on the input feature map to obtain the specific coordinate position of each pixel in the input design layout; the decoder maps the output features of the encoder back to the original input dimension to generate a predicted output; a coordinate channel is added to the input feature map to obtain a modified input feature map, and the convolution kernel dimension of the coordinate convolution module of the encoder that receives the modified input feature map is increased accordingly, and the added coordinate channel is used to improve the ability of the encoder to obtain the spatial position information of the design layout.

[0011] Two additional coordinate channels are generated in the input feature map as the x-direction coordinate position and the y-direction coordinate position of the original input; the x-direction coordinate position of the pixel represents the relative position of each pixel in the width direction of the image, and the y-direction coordinate position of the pixel represents the relative position of each pixel in the height direction of the image.

[0012] The normalized x- and y-coordinate positions of each pixel in the original input are used to represent the x- and y-coordinate positions of the original input.

[0013] The newly added coordinate channel is connected to the original feature map in the channel dimension.

[0014] Each coordinate convolution module of the encoder includes two progressive sub-modules, and the two progressive sub-modules of each coordinate convolution module of the encoder realize the increasing of feature channels; the coordinate convolution modules of adjacent layers of the encoder perform downsampling to realize the decreasing of feature resolution; each convolution module of the decoder includes two progressive sub-modules, and the two progressive sub-modules of each convolution module of the decoder realize the decreasing of feature channels; the convolution modules of adjacent layers of the decoder upsample to realize the increasing of feature resolution; a deep learning module is arranged at the bottom of the encoder and decoder; jump connections are made between the same layers of the encoder and decoder; residual connections are made between the sub-modules of the same layers of the encoder and decoder.

[0015] Each submodule of the encoder and the decoder includes a 3×3 convolution kernel, a batch normalization layer, and a ReLU activation function.

[0016] The coordinate convolution modules of adjacent layers of the encoder are downsampled using a maximum pooling operation, and the window size of the maximum pooling operation is 2×2.

[0017] Another technical solution of the present invention provides an optical proximity correction method, including training a photolithography mask generation model for optical proximity correction described above, including: inputting a chip design layout into the photolithography mask generation model to generate an inferred mask; comparing the inferred mask with the target mask, and using a Dice loss function to quantify the difference between the inferred mask and the target mask, using the calculated Dice loss of the inferred mask and the target mask as the training loss, and optimizing the parameters of the photolithography mask generation model through back propagation; during the optimization process, using a stochastic gradient descent algorithm to update the network weights, gradually reducing the loss function value until the verification set error no longer decreases; simulating photolithography on the inferred mask and the target design mask to obtain an inferred layout and a target layout; evaluating the obtained inferred layout and target layout, and evaluating the photolithography mask generation model based on the difference between the inferred layout and the target layout.

[0018] The training learning rate of the photolithography mask generation model is 0.001 and the batch size is 4.

[0019] The L2 norm and PV band difference are used as evaluation indicators for the difference between the inferred layout and the target layout.

[0020] Compared with the prior art, the beneficial effects of the present invention mainly include: 1. The optical proximity correction photolithography mask generation model of the present invention, by embedding a coordinate convolution module, enables the convolution module in the photolithography mask generation model to accurately obtain the exact position of the structural features in the layout. 2. The photolithography mask generation model of the present invention effectively solves the problem of gradient disappearance during training by introducing the residual connection technology, accelerates the convergence speed, and at the same time reduces the feature resolution through the maximum pooling operation, reduces the loss of computing resources, and improves the computing speed. 3. The residual connection technology introduced by the present invention and the introduction algorithm of the residual connection can further improve the model's ability to process complex layout data by learning the residual between the input and output, which not only improves the expression ability of the model, but also enhances the stability of training. 4. The solution of the present invention is improved on the existing encoder-decoder structure, so the improved photolithography mask generation model can be directly integrated into the existing photolithography process flow, and has better compatibility.

[0021] In order to make the above and other objects, features and advantages of the present invention more clearly understood, preferred embodiments are given below with reference to the accompanying drawings for detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the specific embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0023] Figure 1 This is a network architecture diagram of the photolithography mask generation model of the present invention.

[0024] Figure 2 Schematic diagram of adding coordinate channels to the input feature map of the present invention.

[0025] Figure 3 Schematic diagram of model training in the optical proximity correction method of the present invention.

[0026] Figure 4 Schematic diagram of a photolithography system in the prior art. DETAILED DESCRIPTION

[0027] The foregoing and other technical aspects, features, and functions of the present invention are clearly presented in the following detailed description of a preferred embodiment with reference to the accompanying drawings. Directional terms such as up, down, left, right, front, and back, used in the following embodiments, are intended solely to refer to the directions in the accompanying drawings. Therefore, the directional terms used are for illustrative purposes only and are not intended to limit the present invention.

[0028] Example 1

[0029] Embodiment 1 provides a photolithography mask generation model for optical proximity correction, the photolithography mask generation model including an encoder composed of a coordinate convolution module and a decoder composed of a convolution module; the encoder performs spatial feature extraction on an input feature map, i.e., an input chip design layout, to obtain the specific coordinate position of each pixel in the chip design layout; the decoder maps the output feature of the encoder back to the original input dimension to generate a predicted output; a coordinate channel is added to the input feature map to obtain a modified input feature map, and the convolution kernel dimension of the coordinate convolution module of the encoder that receives the modified input feature map is correspondingly increased, and the added coordinate channel is used to improve the encoder's ability to obtain spatial position information of the design layout.

[0030] A photolithography mask generation model for optical proximity correction provided in Example 1 is inputted with a chip design layout, wherein the photolithography mask generation model is trained to correct the pattern of the input chip design layout to obtain a target mask structure that meets the requirements for photolithography use (that is, the pattern of the photolithography mask is obtained after correction of the pattern of the chip design layout).

[0031] The training of the photolithography mask generation model can be carried out through machine learning technology. A large number of chip design layout structures and corresponding target mask structures are input into the photolithography mask generation model, where the chip layout structure serves as the network input and the target mask structure serves as the training target. The training can adopt a supervised learning method, use the loss function to calculate the error between the network output and the training target, and use the back propagation algorithm to adjust the network parameters.

[0032] The chip design layout is input into the trained photolithography mask generation model, and the photolithography mask generation model corrects the chip design layout and outputs it to obtain the corrected photolithography mask pattern (that is, the target mask structure).

[0033] The following is a detailed explanation of the photolithography mask generation model for optical proximity correction in Example 1 with reference to the accompanying drawings:

[0034] like Figure 1 The optical proximity correction (OPC) mask generation model of the present invention improves upon the classic U-Net architecture, employing an encoder-decoder network structure. The encoder identifies spatial features in the input design layout, known as the contraction path. The decoder upsamples the features extracted by the encoder to gradually restore resolution and generate pixel-level predictions, known as the expansion path. Skip connections are used between the decoder and encoder to combine low-level features from the encoder with high-level features from the decoder, preserving detailed information.

[0035] The encoder includes multiple layers of coordinate convolution modules, and the decoder includes multiple layers of convolution modules. Each layer of coordinate convolution modules and convolution modules includes two progressive sub-modules, each of which includes a 3×3 convolution kernel (to perform a 3×3 convolution operation), a ReLU activation function, and a batch normalization layer, where the batch normalization layer is placed between the convolution kernel and the activation function.

[0036] The two progressive submodules of each coordinate convolution module of the encoder realize the increment of feature channels, see Figure 1 Taking the coordinate convolution module at the top of the encoder as an example, the image features input by the two progressive sub-modules are 512×512×3 and 512×512×16 respectively, to achieve the increase of feature channels (512×512 represents the resolution, 3 and 16 represent the number of feature channels).

[0037] The two progressive submodules of each convolutional module of the decoder implement the reduction of feature channels, e.g. Figure 1 In the convolution module, the two progressive sub-modules of the bottom convolution module input image features of 128×128×256 and 128×128×128 respectively, achieving the reduction of feature channels.

[0038] Figure 1 The network architecture encoder includes a 3-layer coordinate convolution module, and the decoder correspondingly includes a 3-layer convolution module. The coordinate convolution module and the convolution module in Example 1 are illustrated with three layers as an example. In actual design, multiple layers can be designed as needed. Located at the bottom of the network architecture is the deep learning module. The deep learning module consists of three convolution modules and each convolution module has no sub-modules. The deep learning module further increases the number of feature channels for deep learning of features. The three convolution modules used for deep learning, each convolution module includes a 3×3 convolution kernel, a ReLU activation function, and a batch normalization layer, wherein the batch normalization layer is placed between the convolution kernel and the activation function.

[0039] The coordinate convolution modules between adjacent layers of the encoder perform downsampling to reduce the image resolution layer by layer without changing the number of channels. Figure 1 The coordinate convolution module output of the first layer is 512×512×16, and the coordinate convolution module output of the second layer is 256×256×16, and the resolution is changed from 512×512 to 256×256. The coordinate convolution modules of adjacent layers of the encoder are downsampled using the maximum pooling operation, and the window size of the maximum pooling operation is 2×2.

[0040] The convolution modules between adjacent layers of the decoder perform upsampling to restore the image resolution layer by layer. For example Figure 1The third-layer convolution module and the second-layer convolution module increase the image resolution from 128×128 in the third layer to 256×256.

[0041] The decoder and encoder use skip connections to combine the low-level features of the encoder with the high-level features of the decoder to retain detailed information. Figure 1 The solid line connection between the encoder and decoder at the same layer is the skip connection method.

[0042] There are residual connections between the submodules of each layer of the encoder and the submodules of each layer of the decoder. Figure 1 The dotted line connecting the encoder and decoder at the same layer in the figure is the residual connection. Residual connections can further deepen the information exchange between the encoder and decoder, effectively solving the vanishing gradient problem during training and accelerating convergence.

[0043] In order to improve the ability of the network architecture of the present invention to extract the specific position of each pixel coordinate in the input design layout, the present invention improves the input feature map, and adds two coordinate channels to the input feature map as the x-direction coordinate position and the y-direction coordinate position of the original input. Specifically, for the input feature map, two additional coordinate channels are first generated, which are used to represent the normalized coordinate position of each pixel in the x-direction (the relative position of each pixel in the width direction of the image) and the y-direction (the relative position of each pixel in the height direction of the image). The normalized coordinate position of the pixel in the x-direction and the y-direction here refers to the relative position ratio of the pixel point in the image height and width, and the ratio is in the range of [0, 1]. The normalized coordinate position is used to provide unified and standardized position information for algorithms of images of different sizes; the newly added coordinate channel is connected to the original feature map in the channel dimension to form an enhanced feature representation, and then the enhanced feature representation is subjected to a standard convolution operation, so that the network can simultaneously learn spatial features and position information.

[0044] When adding coordinate channels, you can use the coordinate convolution module to add coordinate channels. For details, see the attached Figure 2 In the input feature map, the x-coordinate augmentation module and the y-coordinate augmentation module are used to add channels to the input feature map to represent the x- and y-coordinate positions of the original input, respectively. The input feature map has C channels, and after channel augmentation, the number of channels in the input feature map is C + 2. The input feature map with augmented channels undergoes standard convolution operations, allowing the network to learn both spatial features and positional information.

[0045] Example 2

[0046] A second embodiment provides an optical proximity calibration method. The optical proximity calibration method provided in the second embodiment trains the optimized photolithography mask generation model of the first embodiment. The training results are evaluated and the photolithography mask generation model is optimized based on a combination of loss calculation and photolithography simulation. Simultaneously, the design layout is modified based on the trained photolithography mask generation model to generate a target mask pattern.

[0047] like Figure 3 As shown, the training of the photolithography mask generation model of the second embodiment includes: inputting the chip design layout into the photolithography mask generation model to generate an inferred mask; comparing the inferred mask with the target mask, and using the Dice loss function (a loss function based on the Dice coefficient, used to measure the similarity of two sets) to quantify the difference between the inferred mask and the target mask, using the calculated Dice loss of the inferred mask and the target mask as the training loss, and optimizing the model parameters through back propagation; performing simulated lithography on the inferred mask and the target design mask to obtain the inferred layout and the target layout; evaluating the obtained inferred layout and target layout, and evaluating the photolithography mask generation model based on the difference between the inferred layout and the target layout. During the optimization process, the stochastic gradient descent algorithm can be used to update the network weights, gradually reducing the loss function value, and training continues until the validation set error no longer decreases. By continuously learning the differences between the inferred layout and the target layout, the system uses training data to continuously optimize the parameters of the lithography mask generation model and realizes real-time updates of the model parameters, which helps to generate an inferred layout that is closer to the graphic features of the target layout, thereby significantly improving the correction capability and accuracy of the lithography mask generation model.

[0048] To assist the optical proximity correction method in the second embodiment, when performing simulated lithography, an optical imaging model and a photoresist approximation model are used to perform lithography simulation. Figure 4 The figure shows a typical photolithography system in the prior art. Light from a light source passes through a focusing lens, a mask, a projection lens, and an imaging lens before reaching the photoresist on the wafer surface for exposure. A secondary developer then develops the surface, resulting in the desired pattern. The optical imaging model and photoresist approximation model were created based on existing photolithography systems.

[0049] The optical imaging model used in Example 2 is the Hopkins optical imaging model, and the mathematical expression of its K-th order approximate model is: Where k represents the number of kernels in the optical imaging model, h_k is the amplitude impulse response of the lens, M is the pixel matrix of the mask pattern, In the present invention, the approximate model sets K=24.

[0050] The photoresist approximation model is a mathematical framework used to describe the interaction between the intensity of light propagating from a lens system and the photoresist during the development process. This model effectively simplifies complex physical and chemical phenomena, making computer simulation of the photolithography process possible. The spatial image intensity is transmitted to the surface of a photoresist-coated wafer and evaluated against the photoresist's predetermined sensitivity threshold. When the spatial image intensity I(x,y) on the wafer surface exceeds this threshold I_th, a series of chemical reactions are triggered within the photoresist, ultimately leading to its dissolution during subsequent development. Specifically, the approximate model of a constant-threshold photoresist can be expressed as follows: when I(x,y)>I_th, the photoresist is removed; when I(x,y)≤I_th, the photoresist remains.

[0051] The lithography mask generation model is trained with a learning rate of 0.001 and a batch size of 4.

[0052] The L2 norm (Euclidean distance, L2 norm measures the overall geometric difference between two structures and is used to quantify the overall shape deviation) and PV band difference (Process variation band, PV band difference measures the degree of local variation of key dimensions and is used to capture the size fluctuation of key areas and evaluate local errors) are used as evaluation indicators for the difference between the inferred layout and the target layout.

[0053] Example 2 uses the L2 paradigm and PV band difference as evaluation metrics. These metrics are based on the difference between the developed structure of the target pattern and the developed structure of the inference mask pattern. These metrics comprehensively assess the effectiveness of optical proximity correction, ensuring that the corrected mask can produce the desired pattern in the actual lithography process. Example 2 combines loss calculation with lithography simulation to evaluate the training results and optimize the lithography mask generation model to improve its accuracy.

[0054] Based on the optical proximity correction method of Example 2, the photolithography mask generation model in Example 1 is trained, and based on the trained photolithography mask generation model, the design layout is corrected to generate an inferred mask and a target mask. After simulated lithography, the inferred layout and the target layout are obtained; the obtained inferred layout and the target layout are evaluated based on the L2 paradigm and PV band difference as evaluation indicators. It is found that the photolithography mask generation model in this application is compared with the traditional deep learning methods (U-Net, U-Net++, U-Net+++, Res-Net, R2U-Net and Attention R2U-Net). It is found that the L2 paradigm of the present invention is reduced by 2%, 1%, 2%, 7%, 7% and 1% respectively, and the PV band difference is reduced by 3%, 2%, 3%, 12%, 6% and 6% respectively. It shows that the difference between the inferred mask generated by the photolithography mask generation model of the present invention and the target mask is small, and the correction accuracy of the photolithography mask generation model is high.

[0055] The above describes in detail the optical proximity correction method and photolithography mask generation model provided by the present invention. Specific examples are used herein to illustrate the structure and operating principles of the present invention. The above embodiments are merely intended to facilitate understanding of the method and core concepts of the present invention. It should be noted that those skilled in the art may make various improvements and modifications to the present invention without departing from the principles of the present invention, and such improvements and modifications fall within the scope of protection of the claims.

Claims

1. A photolithography mask generation model for optical proximity correction, characterized in that: The photolithography mask generation model includes an encoder composed of coordinate convolution modules and a decoder composed of convolution modules. The encoder extracts spatial features from the input feature map to obtain the specific coordinate position of each pixel in the input chip design layout. The decoder maps the encoder output features back to the original input dimensions to generate a predicted output. A coordinate channel is added to the input feature map to obtain a modified input feature map. The convolution kernel dimension of the coordinate convolution module of the encoder receiving the modified input feature map is increased accordingly. The added coordinate channel is used to improve the ability of the encoder to obtain the spatial position information of the design layout.

2. The photolithography mask generation model for optical proximity correction according to claim 1, wherein: Generate two additional coordinate channels in the input feature map as the x-direction coordinate position and the y-direction coordinate position of the original input; The x-coordinate position of a pixel represents the relative position of each pixel in the width direction of the image, and the y-coordinate position of a pixel represents the relative position of each pixel in the height direction of the image.

3. The photolithography mask generation model for optical proximity correction according to claim 2, wherein: The normalized x- and y-coordinate positions of each pixel in the original input are used to represent the x- and y-coordinate positions of the original input.

4. The photolithography mask generation model for optical proximity correction according to claim 3, wherein: The newly added coordinate channel is connected to the original feature map in the channel dimension.

5. The photolithography mask generation model for optical proximity correction according to claim 1, wherein: Each coordinate convolution module of the encoder includes two progressive submodules, and the two progressive submodules of each coordinate convolution module of the encoder realize the increment of feature channels; The coordinate convolution modules of adjacent layers of the encoder perform downsampling to achieve a decrease in feature resolution; Each convolution module of the decoder includes two progressive submodules, and the two progressive submodules of each convolution module of the decoder realize the decrease of feature channels; The convolution modules of adjacent layers of the decoder are upsampled to achieve an increase in feature resolution; A deep learning module is provided at the bottom of the encoder and decoder; There is a jump connection between the same layer of the encoder and the decoder; The submodules of the same layer of the encoder and decoder are residually connected.

6. The photolithography mask generation model for optical proximity correction according to claim 5, wherein: Each submodule of the encoder and the decoder includes a 3×3 convolution kernel, a batch normalization layer, and a ReLU activation function.

7. The photolithography mask generation model for optical proximity correction according to claim 5, wherein: The coordinate convolution modules of adjacent layers of the encoder are downsampled using a maximum pooling operation, and the window size of the maximum pooling operation is 2×2.

8. An optical proximity correction method, characterized in that: The method comprises training a photolithography mask generation model for optical proximity correction according to any one of claims 1 to 7, comprising: inputting a chip design layout into the photolithography mask generation model to generate an inferred mask; Comparing the inferred mask with the target mask, and quantifying the difference between the inferred mask and the target mask using the Dice loss function, using the calculated Dice loss between the inferred mask and the target mask as the training loss, and optimizing the parameters of the lithography mask generation model through back propagation; During the optimization process, the stochastic gradient descent algorithm is used to update the network weights and gradually reduce the loss function value until the validation set error no longer decreases; Performing simulated photolithography on the inferred mask and the target design mask to obtain an inferred layout and a target layout; The obtained inferred layout and target layout are evaluated, and the photolithography mask generation model is evaluated based on the difference between the inferred layout and the target layout.

9. The optical proximity correction method according to claim 8, wherein: The training learning rate of the photolithography mask generation model is 0.001 and the batch size is 4.

10. The optical proximity correction method according to claim 8, wherein: The L2 norm and PV band difference are used as evaluation indicators for the difference between the inferred layout and the target layout.

Images