Electron beam exposure charge effect correction method based on neural network, medium and system
By using neural network architecture and data optimization methods, the error and dependency problems of traditional polynomial fitting methods in charge effect correction are solved, achieving higher accuracy charge effect correction and improving the accuracy of graphic placement.
Patent Information
- Application Number
- CN202511051878.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-29
- Publication Date
- 2025-11-07
AI Technical Summary
In existing technologies, traditional polynomial fitting methods are not effective in reducing the influence of charge effects during electron beam exposure, and the solution of the unit charge response function has errors. The parameter solution relies too much on the exposure experiment, which leads to model instability.
A neural network-based approach is adopted, which models the charge density distribution through the U-Net network architecture and uses 1×1 convolutional neural network blocks to correct mechanical installation and environmental errors. The unit charge response function is optimized by combining computer simulation data pre-training and real data fine-tuning.
It improves the accuracy of charge effect correction, reduces graphic placement errors, enhances the stability and robustness of the model, and reduces deviations caused by mechanical installation, etc.
Smart Images
Figure CN120909075A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application mainly relates to the field of semiconductor technology, and in particular to a method for correcting charging effect in electron beam exposure based on neural network, medium and system. BACKGROUND
[0002] With the development of semiconductor technology to more advanced nodes, higher and higher requirements are put forward for the pattern placement accuracy in the electron beam exposure process. When the electron beam irradiates non-conductive or low-conductive materials, the incident electrons interact with the materials, causing the accumulation of electric charge in the photoresist to form a local electric field, which will cause the trajectory of the incident electron beam to be deflected, and then affect the pattern placement accuracy. Therefore, the charging effect is the largest error source affecting the pattern placement accuracy, and how to reduce the influence of the charging effect has become an important research direction to improve the pattern placement accuracy.
[0003] At present, the methods for reducing the influence of the charging effect on the substrate surface mainly include two categories. One method is to coat a charge dissipation layer (CDL) on the surface, but this method has the following shortcomings: 1. Adding a new coating layer may introduce a new source of pollution; 2. The basic chemical properties of the charge dissipation layer are acidic, so the coating layer may not be compatible with the existing process; 3. Users must purchase additional coating equipment, increasing production costs; Due to the above defects of the method of coating the charge dissipation layer, another type of mathematical modeling method based on the development of this method is driven, which studies and models the formation of the surface charge density of the substrate and the deflection of the electron beam during electron beam exposure, so that the deflection amount can be pre-compensated, and then the placement accuracy during pattern exposure is improved.
[0004] According to the research, the current research based on mathematical modeling method still stays in the stage of rough fitting using traditional polynomials. This method tends to use a white box model to model the charging effect based on the understanding of physics. First, the charge density is modeled and solved based on information such as exposure dose and layout density, and then the unit charge response function in the exposure environment is used to solve the position shift. For the solution of the polynomial parameters, the current public method uses the data of the actual exposure experiment to solve the model parameters by fitting. This method has the following shortcomings: 1. The effect of the traditional polynomial fitting method is poor Currently, there are two aspects that limit the modeling effect of the method. On the one hand, from the physical point of view, the understanding of the physical behavior of the charging effect is still insufficient, and the formation reason and mechanism of the charging effect cannot be well explained from the physical point of view; on the other hand, from the mathematical point of view, the fitting ability of the traditional polynomial is very limited, and cannot simulate complex physical behavior, which often depends on modeling based on artificial experience, and such models often deviate from the real physical phenomenon. The above two aspects lead to the fact that the traditional method based on polynomial fitting often cannot get good results, and cannot meet the requirements of the increasingly reduced semiconductor nodes for pattern placement error.
[0005] 2. Error in solving unit charge response function In solving the pattern shift caused by the charging effect, an important step is to solve the unit charge response function. The function is usually solved by simplifying the actual exposure environment, and then using mirror charge method or finite element analysis method to simulate the electric field caused by unit charge. This method often has errors. Due to unknown factors such as mechanical installation, actual exposure environment, and other factors affecting the electric field distribution, the unit charge response function obtained by simulation often deviates from the actual situation.
[0006] 3. Parameter solving excessively depends on exposure experiment In the current disclosed method, the solving of model parameters often excessively depends on the data obtained by exposure experiment for parameter fitting. However, due to the complexity of the model and the fact that the data collected by the actual exposure experiment often contains unknown noise, there is a risk of falling into local optimal solution and non-convergence in parameter solving. SUMMARY
[0007] In view of the technical problems existing in the prior art, the present application provides an electron beam exposure charging effect correction method based on neural network with high correction precision, medium and system.
[0008] To solve the above technical problems, the technical solution provided by the present application is: A kind of electron beam exposure charging effect correction method based on neural network, including steps: S1, obtaining dose map after proximity effect correction And pre-computed layout density map , generate unit area exposure dose map ; S2, divide the exposure layout into exposure area and non-exposure area; use U-Net network to model charge density distribution in the exposure area, output charge density distribution; use single-layer network or Gaussian diffusion function to obtain charge density distribution in the non-exposure area; S3, input the unit charge response function obtained by simulation into a 1*1 convolutional neural network block to learn a non-linear mapping relationship to correct mechanical installation and environmental errors; S4, calculate the electron beam deflection and generate a graphical offset map based on the charge density distribution in step S2 and the corrected unit charge response function in step S3.
[0009] Preferably, in step S2, the U-Net network architecture comprises an encoder, a decoder and a skip connection; the encoder is composed of several convolutional modules, each of which comprises two convolutional layers and an average pooling layer, wherein the convolutional layers are used to extract features of the exposure plate dose, and the average pooling layer is used for down-sampling to gradually reduce the spatial resolution of the plate dose map while increasing the dose information of the features; the decoder is also composed of multiple convolutional modules, and the decoder gradually restores the spatial resolution of the charge density distribution through up-sampling operations, and maps the dose and charge density relationship features extracted by the encoder to a more manifold space; the skip connection part directly splices the feature map of the encoder part with the corresponding level feature map of the decoder part, and transmits the detailed spatial information retained in the encoder to the decoder to help the decoder better restore the details of the charge density distribution.
[0010] Preferably, the unit charge response function in step S3 is obtained by simplifying the model and by mirror charge method or finite element analysis simulation from the actual exposure environment and physical structure.
[0011] Preferably, steps S1-S4 are implemented through a neural network model; specifically, computer simulation data is used for pre-training of the neural network model, and then real data is used for fine-tuning to ensure stable and robust convergence of the model parameters.
[0012] Preferably, the specific process of obtaining computer simulation data is as follows: PMMA is selected as the resist material, with a thickness of 0.4 um, and chromium Cr is selected as the substrate material; the scanning area is 1.0 um x 1.0 um, and the interval of adjacent incident points is d=0.05, 0.1, 0.2 um; the exposure conditions are: acceleration voltage is 50KV, normalized exposure dose is used, scanning mode is zigzag, and scanning speed is standard television scanning speed; in simulation, these parameters, materials and substrate structures are settable, and the influence of charging effect under different process conditions is studied by changing these settings using simulation means.
[0013] Preferably, the steps of electron beam lithography charging effect simulation include: Divide the electron beam dose into a three-dimensional space grid; Calculate the deflection of the electron beam in the vacuum exposure cavity; Monte Carlo method is used to calculate the spatial distribution of the charge deposited in the substrate; wherein the necessary simulation information is inputted including exposure sequence, simulation dose, material structure information.
[0014] Preferably, the exposure dose map per unit area .
[0015] The application further discloses a computer program product comprising a computer program which, when executed by a processor, performs the steps of the method as described above.
[0016] The application further discloses a computer readable storage medium having stored thereon a computer program which, when executed by a processor, performs the steps of the method as described above.
[0017] The application further discloses a neural network-based electron beam exposure charging effect correction system, comprising a memory and a processor connected to each other, wherein the memory stores a computer program which, when executed by the processor, performs the steps of the method as described above.
[0018] Compared with the prior art, the application has the following advantages: The application uses a neural network method to replace the current traditional charging effect correction modeling method based on a polynomial, uses the powerful fitting capability of the neural network for data, and can further reduce the pattern placement error caused by the charging effect. Meanwhile, the actual exposure data is used to optimize the unit charge response function obtained through simulation, reduces the deviation caused by mechanical installation and the like, and makes the response function more suitable for actual experiments. Finally, in order to ensure the stable and robust convergence of the model parameters, the application uses computer simulation data to pre-train the neural network, and then uses real data to fine-tune the parameters to optimize the parameters. BRIEF DESCRIPTION OF DRAWINGS
[0019] Figure 1 The neural network architecture diagram for the application applied to the charging effect correction.
[0020] Figure 2 The substrate simulation structure diagram in the application.
[0021] Figure 3 The charging effect simulation method flowchart in the application.
[0022] Figure 4 The charging effect correction method flowchart in the embodiment of the application. DETAILED DESCRIPTION
[0023] The application will be further described below in combination with the accompanying drawings and specific embodiments.
[0024] AsFigure 1 and Figure 4 As shown in the method for correcting the charging effect of the electron beam exposure based on the neural network provided by the embodiments of the present application, the method comprises the steps of: S1、 Figure 1 The basic architecture of the neural network applied to the correction of the charging effect is constructed based on the basic principle of the charging effect, and the neural network is used to compensate for the shortcomings of the existing methods by taking advantage of the neural network. The input of the model is the dose map after the proximity effect correction and the layout density map calculated in advance According to the two variables, the exposure dose per unit area is obtained The output of the model is the pattern offset map, and the offset is the offset caused by the charging effect during actual exposure.
[0025] S2, according to the basic principle of the charging effect, the exposure layout can be generally divided into an exposure area and a non-exposure area, and the exposure area has more complex physical processes than the non-exposure area. The charging effect in the exposure area usually involves the coupling of complex secondary electron emission, electron and hole capture, breakdown effect and other effects. This complex coupling phenomenon brings great challenges to modeling. In the non-exposure area, the formation of the charging effect is relatively simple. The current data research shows that due to the long-range charge offset in the non-exposure area, it is generally believed that this effect is caused by the fog effect electron caused by the secondary electron hitting the top of the exposure chamber. The mechanism of the fog effect electron is generally considered to be relatively simple, and in the scene where the accuracy requirement is not high, a single-layer network or a Gaussian diffusion function can be used to represent the influence of the fog effect electron.
[0026] According to this action principle, in Figure 1In the network architecture, the exposure dose map is divided into exposure regions and non-exposure regions using layout information, the complexity of the network architecture is adjusted according to the complexity of its physical principle, the complexity of the exposure region is high, and the complexity of the non-exposure region is low. The charge density modeling of the exposure region uses a U-Net network architecture, which is a classic convolutional neural network architecture. The network structure mainly includes an encoder, a decoder and a skip connection. The encoder part is composed of several convolution modules, each of which contains two convolution layers and an average pooling layer. The convolution layer is used to extract the features of the exposure layout dose, and the average pooling layer is used for down-sampling, gradually reducing the spatial resolution of the layout dose map while increasing the dose information of the features. The decoder part is also composed of multiple convolution modules, but unlike the encoder part, the decoder gradually recovers the spatial resolution of the charge density distribution through up-sampling operations, mapping the dose and charge density relationship features extracted by the encoder to a more manifold space. Common up-sampling methods include deconvolution and bilinear interpolation. The skip connection part is one of the core innovations of the U-Net network, which directly splices the feature maps of the encoder part with the corresponding level feature maps of the decoder part. In this way, the detailed spatial information retained in the encoder can be passed to the decoder, helping the decoder better recover the details of the charge density distribution. At the same time, this network structure has good adaptability to small data sets and relatively small demand for training data, which can also to some extent alleviate the characteristics of the charging effect data not easy to obtain.
[0027] S3, input the simulated unit charge response function into a 1*1 convolutional neural network block to learn the nonlinear mapping relationship to correct mechanical installation and environmental errors, and obtain a corrected unit charge response function; The unit charge response function is a key link from the charge density distribution to the electron beam deflection. Usually, this response function is obtained by simplifying the model simulation from the actual exposure environment and physical structure. The main methods for solving include the image charge method and the finite element analysis method. However, this kind of method cannot model the actual mechanical installation error and the unknown influence source easily introduced in the electric field, etc. These errors will eventually cause the simulated unit charge simulation function to deviate from the actual response function, and this error will further accumulate to the final pattern deflection.
[0028] This invention, based on the simulated unit charge response function, introduces a 1x1 convolutional neural network block. This module increases the nonlinearity of the response function, allowing the network to automatically learn more complex feature mapping relationships from the data. However, these mapping relationships should not vary excessively; in practical applications, regularization is preferred to limit their variation. The nonlinear mapping relationships learned by this convolutional neural network block often represent error factors that simulation modeling cannot account for. By adding this module, the charge effect correction model can better reflect reality.
[0029] Both polynomial-based modeling methods and the neural network-based method of this invention require data for parameter solving. In previous methods, parameter solving often relied too heavily on actual exposure experiments. However, the data obtained from exposure experiments is not only limited in quantity but also contains a certain amount of noise. This leads to parameter optimization easily getting trapped in local optima and failing to converge, posing challenges to the correction of the charge effect. To alleviate this problem, this invention uses computer-simulated data for neural network pre-training, followed by fine-tuning with real data to ensure stable and robust convergence of the model parameters.
[0030] Specifically, in order to simulate the actual situation during actual electron beam exposure, such as Figure 2 As shown, in this invention, PMMA with a thickness of 0.4 μm was selected as the photoresist material during simulation, and chromium (Cr) was chosen as the substrate material. The scanning area was 1.0 μm x 1.0 μm, and the spacing between adjacent incident points was selectable as d = 0.05, 0.1, or 0.2 μm. Exposure conditions included an accelerating voltage of 50 kV, a normalized exposure dose, a zigzag scanning method, and a standard television scanning speed. During simulation, these parameters, materials, and substrate structure were all adjustable. By changing these settings, simulation methods could be used to study the effects of charge generation under different process conditions.
[0031] The steps for simulating the charge effect in electron beam lithography are as follows: Figure 3 As shown, it is mainly divided into two parts: 1) deflection of the electron beam in the vacuum exposure chamber; 2) spatial distribution of charge deposition in the substrate.
[0032] First, input the necessary simulation information, including exposure sequence, simulation dosage, material structure, etc. Step 1: Define the electron beam dose and three-dimensional spatial grid; Step 2: Calculate the deflection of the electron beam within the vacuum exposure chamber; Step 3: The Monte Carlo method is used to calculate the spatial distribution of charge deposition within the substrate.
[0033] Through this simulation means, a large amount of charging effect data that cannot be obtained through actual exposure experiments can be obtained, and the data can meet the needs of neural network training, and the parameters in the network converge to the optimal solution.
[0034] However, the simulation data inevitably has errors compared with the actual data, so finally the application will use the data obtained through actual exposure to fine-tune the parameters, dynamically allocate weights according to the charge accumulation degree in the exposure experiment, so that the network model parameters converge to the optimal solution. In actual use, the application will preprocess the exposure data, identify obviously incorrect data, and give higher weights to data with greater charging effect in the exposure experiment, so that the model parameter solving is more robust.
[0035] S4, based on the charge density distribution in step S2 and the corrected unit charge response function in step S3, calculate the electron beam deflection quantity and generate a graphical offset map.
[0036] The application uses the powerful fitting ability of neural networks to data to build a neural network structure suitable for charging effect correction on the basis of the existing charging effect correction model based on polynomials, better models the offset caused by the charging effect, and ultimately achieves the purpose of improving the charging effect correction capability. The network structure is mainly based on a convolutional neural network, and uses a neural network to break through the blind area of the existing charging effect mechanism and the error in modeling the unit charge response function, so that the existing model better represents the graphical offset caused by the actual charging effect. In view of the fact that the parameter solving is too dependent on the exposure experiment, computer simulation data is used for neural network pre-training, and then real data is used for fine-tuning to ensure stable and robust convergence of the model parameters; the data obtained through actual exposure experiments is used to fine-tune the unit charge response function through an optimization method, so as to obtain a distribution closer to the actual situation.
[0037] The application uses a charging effect correction method based on a neural network, uses the powerful fitting ability of a neural network to data to improve the placement error correction accuracy of the current polynomial method; the unit charge response function is optimized using data obtained through actual exposure experiments, thereby improving the fitting degree of the unit charge response function obtained through the simulation method and the actual response function; the computer simulation and actual exposure experiment are combined, the computer simulation data is used for neural network pre-training, and then the real data is used for fine-tuning to ensure stable and robust convergence of the model parameters.
[0038] The present application adopts a neural network method to replace the current traditional charge effect correction modeling method based on a polynomial, uses the powerful fitting capability of the neural network for data, and can further reduce the graphic placement error caused by the charge effect. Meanwhile, the actual exposure data is used to optimize the unit charge response function obtained by simulation, reduces the deviation caused by mechanical installation, and makes the response function more suitable for actual experiments. Finally, in order to ensure the stable and robust convergence of the model parameters, the present application adopts a computer simulation data to pre-train the neural network, and then uses real data to fine-tune the parameters. Through the above three aspects of improvement, the charge effect correction scheme is further optimized.
[0039] The embodiment of the present application further discloses a computer program product, comprising a computer program, which executes the steps of the method as described above when run by a processor. The embodiment of the present application also discloses a computer readable storage medium, which stores a computer program, which executes the steps of the method as described above when run by a processor. The embodiment of the present application also discloses a neural network-based electron beam exposure charge effect correction system, comprising a memory and a processor connected to each other, wherein the memory stores a computer program, and the computer program executes the steps of the method as described above when run by the processor.
[0040] The present application can realize all or part of the processes in the above-mentioned embodiment methods, and can also be completed by computer program instruction related hardware. The computer program can be stored in a computer readable storage medium, and the computer program can realize the steps of the above-mentioned method embodiment when executed by a processor. The computer program includes computer program code, which can be in the form of source code, object code, executable files or some intermediate forms. The computer readable storage medium includes any entity or device capable of carrying computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. The memory is used to store computer programs and / or modules, and the processor realizes various functions by running or executing the computer programs and / or modules stored in the memory, and calling the data stored in the memory. The memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one magnetic disk storage device, a flash memory device, or other volatile solid-state storage devices, etc.
[0041] Noun explanation: CDL (Charge Dissipation Layer): Charge Dissipation Layer; PMMA (Polymethyl methacrylate): a type of photoresist; U-Net: a typical convolutional neural network architecture; Convolution: convolution; Skip Connection: skip connection; Avg pooling: average pooling; Up Sampling: up sampling; Scan Area: scan area; Normalized Dose in TV Scan: Normalized Dose in TV Scan.
[0042] The above is only the preferred embodiment of the present application, and the protection scope of the present application is not limited to the above-mentioned embodiments only. Any technical solution falling within the concept of the present application shall fall within the protection scope of the present application. It should be noted that, for ordinary skilled persons in the art, some improvements and refinements without departing from the principles of the present application shall be considered as the protection scope of the present application.
Claims
1. A neural network-based charged effect correction method for electron beam exposure, characterized in that, The method comprises the steps of: S1, obtaining a dose map corrected for proximity effects and a pre-computed layout density map , generating an exposure dose map per area ; S2, dividing the exposure layout into exposed regions and non-exposed regions; using a U-Net network architecture to model the charge density distribution of the exposed regions, and outputting the charge density distribution; using a single-layer network or a Gaussian diffusion function to obtain the charge density distribution of the non-exposed regions; S3, inputting the unit charge response function obtained by simulation into a 1x1 convolutional neural network block to learn the non-linear mapping relationship to correct mechanical installation and environmental errors; S4, based on the charge density distribution in step S2 and the corrected unit charge response function in step S3, calculating the electron beam deflection and generating a graphical offset map.
2. The neural network-based e-beam exposure charging effect correction method of claim 1, wherein, In step S2, the U-Net network architecture includes an encoder, a decoder, and a skip connection; the encoder is composed of several convolution modules, each of which includes two convolution layers and an average pooling layer, wherein the convolution layers are used to extract the dose features of the exposure layout, and the average pooling layer is used for downsampling to gradually reduce the spatial resolution of the layout dose map while increasing the dose information of the features; The decoder is also composed of multiple convolution modules, and the decoder gradually restores the spatial resolution of the charge density distribution through upsampling operations, maps the dose and charge density relationship features extracted by the encoder to a more manifold space; the skip connection part directly splices the feature map of the encoder part with the corresponding level feature map of the decoder part, transmits the detailed spatial information retained in the encoder to the decoder, and helps the decoder to better restore the details of the charge density distribution.
3. The neural network-based e-beam exposure charging effect correction method of claim 1, wherein, The unit charge response function in step S3 is obtained by simplifying the model and simulating it through image charge method or finite element analysis from the actual exposure environment and physical structure.
4. The neural network-based e-beam exposure charging effect correction method according to claim 1 or 2 or 3, characterized in that, Steps S1-S4 are implemented through a neural network model; specifically, computer simulation data is used to pre-train the neural network model, and then real data is used to fine-tune to ensure stable and robust convergence of the model parameters.
5. The neural network-based e-beam exposure charging effect correction method of claim 4, wherein, The specific process of obtaining computer simulation data is as follows: PMMA is selected as the resist material, with a thickness of 0.4um, and chromium Cr is selected as the substrate material; the scanning area is 1.0umx 1.0um, and the interval between adjacent incident points is d=0.05, 0.1, 0.2um; the exposure conditions are: acceleration voltage is 50KV, normalized exposure dose is used, scanning mode is zigzag, and scanning speed is standard television scanning speed; in simulation, these parameters, materials and substrate structure are settable, and the influence of charging effect under different process conditions is studied by changing these settings using simulation means.
6. The neural network-based e-beam exposure charging effect correction method of claim 5, wherein, The steps of electron beam lithography charging effect simulation include: Divide the electron beam dose into a three-dimensional grid; Calculate the deflection of the electron beam in the vacuum exposure cavity; Calculate the spatial distribution of charge deposition in the substrate by Monte Carlo method; The necessary simulation information includes exposure sequence, simulation dose, and material structure information.
7. The neural network-based e-beam exposure charging effect correction method according to claim 1 or 2 or 3, characterized in that, Exposure dose per unit area map .
8. A computer program product comprising a computer program, characterized in that, The computer program, when executed by a processor, performs the steps of the method of any one of claims 1-7.
9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by a processor, performs the steps of the method of any one of claims 1-7.
10. A neural network-based charged effect correction system for electron beam exposure, comprising a memory and a processor connected to each other, wherein the memory has stored thereon a computer program, characterized in that, The computer program, which when executed by a processor, performs the steps of the method as claimed in any one of claims 1-7.