Electron beam lithography dose correction method and device based on machine learning and electronic equipment

Abstract

The invention provides an electron beam lithography dose correction method and device based on machine learning and electronic equipment, and the method comprises the steps: obtaining real-time process environment parameters of electron beam lithography equipment and a target graphic layout to be processed by a user, and calculating target energy deposition distribution of discrete pixel grids in the target graphic layout; predicting a point spread function key parameter group according to the real-time process environment parameters based on a pre-trained machine learning model; and correcting a dose modulation layout corresponding to the target graphic layout according to the point spread function key parameter group and the target energy deposition distribution. According to the method, online, real-time and high-precision proximity effect correction is realized, the proximity effect is effectively inhibited, and the precision and consistency of pattern processing are improved; meanwhile, the calculation time is shortened from several hours to several minutes, and the exposure dose can be dynamically adjusted in real time, so that a technical basis is provided for electron beam lithography with high resolution and complex patterns.

Description

Technical Field

[0001] This invention relates to the field of electron beam lithography technology, and in particular to an electron beam lithography dose correction method, apparatus, and electronic device based on machine learning. Background Technology

[0002] Electron beam lithography is a high-precision micro / nano fabrication technique used in semiconductor manufacturing and nanodevice fabrication. It forms fine patterns by exposing a focused electron beam onto photoresist. However, electrons are scattered within the photoresist and substrate material, causing energy to diffuse to neighboring areas and triggering the proximity effect. This effect can lead to defects such as uncontrolled linewidth, blurred edges, feature adhesion, or breakage, severely impacting processing resolution and pattern fidelity.

[0003] To correct for the proximity effect, existing technologies mainly employ two types of methods: one is the empirical dose correction method based on the Gaussian model, which is fast in calculation but simplifies the model and is difficult to accurately reflect the real energy deposition in multilayer heterogeneous materials, resulting in insufficient correction accuracy; the other is the all-physics simulation method, which is highly accurate but involves huge computational loads and takes hours or even days, failing to meet the needs of real-time and efficient industrial applications.

[0004] Therefore, existing technologies face the contradiction of "high precision comes at the expense of low efficiency, and high efficiency comes at the expense of low precision," lacking an intelligent dose correction scheme that combines high precision and high computational efficiency, making it difficult to achieve rapid and accurate proximity effect correction for complex patterns. Summary of the Invention

[0005] This invention provides a machine learning-based electron beam lithography dose correction method, apparatus, and electronic device to overcome the shortcomings of existing electron beam lithography dose correction methods that cannot balance accuracy and efficiency. It achieves online, real-time, and high-precision proximity effect correction, effectively suppresses proximity effects, and improves the accuracy and consistency of pattern processing. At the same time, it reduces the calculation time from several hours to several minutes and can dynamically adjust the exposure dose in real time, thus providing a technical foundation for high-resolution, complex pattern electron beam lithography.

[0006] On one hand, the present invention provides a machine learning-based electron beam lithography dose correction method, comprising: acquiring real-time process environment parameters of the electron beam lithography equipment and the target pattern layout to be processed by the user, and calculating the target energy deposition distribution of discrete pixel grids in the target pattern layout; predicting key parameter sets of the point spread function based on a pre-trained machine learning model according to the real-time process environment parameters; wherein, the machine learning model is obtained by training and optimization based on a training sample set composed of process environment parameter samples and their corresponding point spread function parameter samples; and correcting the dose modulation layout corresponding to the target pattern layout according to the key parameter sets of the point spread function and the target energy deposition distribution; the dose modulation layout includes the theoretical dose to be applied to each discrete pixel grid.

[0007] Furthermore, constructing the training sample set specifically includes: identifying key process parameters affecting electron beam scattering behavior; simulating the electron beam scattering trajectory and energy loss process under the key process parameters using Monte Carlo simulation, and establishing an energy deposition distribution map; fitting a point spread function based on the energy deposition distribution map, and extracting the point spread function parameters; using the key process parameters as the process environment parameter sample, and the point spread function parameters as the point spread function parameter sample, to construct the training sample set.

[0008] Furthermore, the key process parameters include electron beam parameters and material parameters; the electron beam parameters include electron beam acceleration voltage and the number of simulated electrons; the material parameters include resist layer thickness, resist material type, and substrate material type.

[0009] Furthermore, the key process parameters, the energy deposition distribution map, the point diffusion function parameters, and various graphic structures are organized into a structured database and stored. Correspondingly, the step of calculating the target energy deposition distribution of the discrete pixel grid in the target graphic layout includes: discretizing the target graphic layout into multiple discrete pixel grids; matching the closest graphic structure in the structured database based on the geometric features of the target graphic layout, and extracting the energy deposition distribution map corresponding to the closest graphic structure; and determining the energy deposition distribution map corresponding to the closest graphic structure as the target energy deposition distribution corresponding to the multiple discrete pixel grids.

[0010] Furthermore, training and optimizing the machine learning model includes: taking the process environment parameter samples as input, the predicted point spread function parameters as output, and the difference between the predicted point spread function parameters and the point spread function parameter samples as training loss, and iteratively optimizing the machine learning model; wherein, the machine learning model selects different models according to data features.

[0011] Further, the step of correcting the dose modulation pattern corresponding to the target pattern pattern based on the key parameter set of the point spread function and the target energy deposition distribution includes: combining the dose modulation pattern with the target pattern pattern to generate a target exposure file; loading the target exposure file into the electron beam lithography equipment and performing exposure.

[0012] Secondly, the present invention also provides a machine learning-based electron beam lithography dose correction device, comprising: an energy deposition distribution calculation module, used to acquire real-time process environment parameters of the electron beam lithography equipment and the target pattern layout to be processed by the user, and to calculate the target energy deposition distribution of discrete pixel grids in the target pattern layout; a point spread function key parameter group prediction module, used to predict the point spread function key parameter group based on a pre-trained machine learning model and according to the real-time process environment parameters; wherein, the machine learning model is obtained by training and optimization based on a training sample set composed of process environment parameter samples and their corresponding point spread function parameter samples; and a dose modulation layout correction module, used to correct the dose modulation layout corresponding to the target pattern layout according to the point spread function key parameter group and the target energy deposition distribution; the dose modulation layout includes the theoretical dose to be applied to each discrete pixel grid.

[0013] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the electron beam lithography dose correction method based on machine learning as described above.

[0014] Fourthly, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the electron beam lithography dose correction method based on machine learning as described above.

[0015] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the machine learning-based electron beam lithography dose correction method as described above.

[0016] This invention provides a machine learning-based electron beam lithography dose correction method. It acquires real-time process environment parameters of the electron beam lithography equipment and the target pattern to be processed by the user, calculates the target energy deposition distribution of discrete pixel grids in the target pattern, and then predicts key parameter sets of the point spread function based on a pre-trained machine learning model according to the real-time process environment parameters. The machine learning model is trained and optimized using a training sample set composed of process environment parameter samples and their corresponding point spread function parameter samples. Therefore, based on the key parameter sets of the point spread function and the target energy deposition distribution, the dose modulation pattern corresponding to the target pattern is corrected. The dose modulation pattern includes the theoretical dose to be applied to each discrete pixel grid. This method, by introducing a machine learning model and training it into a highly efficient simulator, achieves online, real-time, and high-precision proximity effect correction, effectively suppressing the proximity effect and improving the accuracy and consistency of pattern processing. Simultaneously, it reduces computation time from several hours to several minutes and allows for real-time dynamic adjustment of the exposure dose, thus providing a technical foundation for high-resolution, complex pattern electron beam lithography. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is a schematic flowchart of the electron beam lithography dose correction method based on machine learning provided in an embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram of the simulation results of Monte Carlo simulation provided in an embodiment of the present invention.

[0020] Figure 3 This is a schematic diagram of the point spread function extraction process provided in an embodiment of the present invention.

[0021] Figure 4 This is a schematic diagram illustrating the fitting effect of the point spread function provided in an embodiment of the present invention.

[0022] Figure 5 This is a schematic diagram illustrating the effect of the machine learning model predicting the key parameter set of the point spread function provided in the embodiment of the present invention.

[0023] Figure 6 This is a schematic diagram of the structure of the electron beam lithography dose correction device based on machine learning provided in an embodiment of the present invention.

[0024] Figure 7This is a schematic diagram of the physical structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0026] It is important to note that electron beam lithography is a high-precision micro-nano fabrication technology widely used in semiconductor manufacturing, nanodevice fabrication, and other fields. It is a core technology for manufacturing high-resolution photomasks, nanodevices, and integrated circuit prototypes. The core principle of electron beam lithography is to use a high-energy electron beam to expose photoresist, forming fine patterns. However, the proximity effect caused by the scattering behavior of electrons in the resist (such as photoresist) and substrate material is the fundamental physical bottleneck limiting the final fabrication fidelity. This effect can lead to fatal defects such as uncontrolled linewidth, blurred edges, and feature adhesion.

[0027] Specifically, when a high-energy electron beam is incident on a photoresist surface, electrons undergo elastic and inelastic scattering with the atomic nuclei and electrons in the material, causing a deflection of the electrons' motion direction and energy loss. This scattering behavior leads to the following phenomena: the energy of the electron beam is not only deposited near the incident point but also diffuses into the surrounding area, resulting in an increased exposure dose in adjacent areas. In high-resolution pattern processing, the proximity effect can cause blurred pattern edges, uneven linewidth, and even defects such as bridging or breakage. Furthermore, due to the presence of the scattering effect, the energy distribution of the electron beam is uneven, making it difficult to precisely control the dose in the exposed area. The severity of the proximity effect is closely related to factors such as electron beam energy, photoresist thickness, and substrate material properties.

[0028] To address this issue, the industry has developed various electron beam lithography dose correction techniques, mainly including traditional dose correction methods based on empirical formulas and simple all-physical simulation model methods, but all of them have significant limitations.

[0029] Traditional dose correction methods often employ Gaussian distribution models to describe electron beam energy deposition. These methods simplify the complex scattering process into a superposition of several Gaussian functions, offering the advantage of fast computation. However, the models are overly simplified and cannot accurately describe the true energy deposition distribution of electrons in multilayer heterogeneous materials. These methods typically neglect the physical properties of the photoresist and substrate materials (such as density, atomic number, and thickness), leading to insufficient dose correction accuracy, especially when feature sizes reach the deep submicron scale. Furthermore, the parameters of empirical formulas are usually obtained through experimental fitting, lacking universality and difficulty in adapting to different materials and processing conditions.

[0030] Full-physics simulation models can faithfully simulate the scattering path and energy loss of every electron, and are currently recognized as the "gold standard" with extremely high accuracy. However, their fatal flaw is the enormous computational burden. For complex patterns, existing methods have high computational complexity, making real-time dose correction difficult. They typically require multiple experimental verifications and adjustments, which are time-consuming and fail to meet the demands of high-efficiency manufacturing. Performing full-chip-level Monte Carlo simulation correction on a slightly complex pattern often takes hours or even days, completely failing to meet the real-time, high-efficiency computational requirements of industrial production. Furthermore, they rely heavily on manual intervention and adjustments, lacking automated and intelligent dose correction capabilities.

[0031] Therefore, the current field of proximity effect correction in electron beam lithography equipment faces a core contradiction: high-precision methods (physical simulation) are too slow, while fast methods (empirical models) lack sufficient accuracy. There is an urgent need for a new technical solution that combines the advantages of both, achieving both speed and accuracy.

[0032] In view of this, the present invention proposes a machine learning-based electron beam lithography dose correction method, specifically, Figure 1 A schematic flowchart of the electron beam lithography dose correction method based on machine learning provided in an embodiment of the present invention is shown.

[0033] like Figure 1 As shown, the method includes: S110, acquiring real-time process environment parameters of the electron beam lithography equipment and the target pattern layout to be processed by the user, and calculating the target energy deposition distribution of the discrete pixel grids in the target pattern layout; S120, based on a pre-trained machine learning model, predicting the key parameter set of the point spread function according to the real-time process environment parameters; wherein, the machine learning model is obtained by training and optimizing a training sample set composed of process environment parameter samples and their corresponding point spread function parameter samples; S130, correcting the dose modulation layout corresponding to the target pattern layout according to the key parameter set of the point spread function and the target energy deposition distribution; the dose modulation layout includes the theoretical dose to be applied to each discrete pixel grid.

[0034] The following will provide a detailed description of steps S110-S130 and related steps.

[0035] S110: Obtain the real-time process environment parameters of the electron beam lithography equipment and the target pattern layout to be processed by the user, and calculate the target energy deposition distribution of the discrete pixel grid in the target pattern layout.

[0036] It is easy to understand that in order to perform electron beam lithography dose correction, it is first necessary to obtain the real-time process environment parameters of the electron beam lithography equipment, and at the same time obtain the target pattern layout to be processed by the user.

[0037] Real-time process environment parameters refer to a set of physical and process variables that directly affect electron scattering behavior and energy deposition distribution during the current processing. These parameters directly determine the shape of the point spread function (PSF). Real-time process environment parameters include at least electron beam parameters and material parameters.

[0038] The electron beam parameters include the electron beam accelerating voltage (unit: keV), for example from 10keV to 100keV in 5keV increments, and may also include beam current intensity (unit: pA or nA), scanning speed (unit: nm / s or μm / ms), and focusing conditions (such as beam spot size and astigmatism correction status).

[0039] Material parameters include the thickness of the resist (photoresist) layer, the type of resist material, and the type of substrate material. The resist layer thickness ranges from 50 nm to 500 nm, with a step size of 50 nm. The resist material type includes polymer-based positive resists (such as PMMA and ZEP520), chemically amplified resists (such as MA-P1275), negative resists (such as SU-8), inorganic resists (such as HSQ), and metal oxide-based resists (such as HfO2 nanoresists), which can be encoded by their chemical composition and density. The substrate material type includes single-crystal silicon (Si), silicon dioxide (SiO2), silicon nitride (Si3N4), gallium arsenide (GaAs), and quartz, etc.

[0040] The target pattern layout refers to the final micro / nano structure pattern that the user expects to form on the substrate, supporting formats such as GDSII, OASIS, and CIF. After receiving the target pattern layout, it is analyzed and digitized to generate a two-dimensional discretized pixel grid matrix, where each pixel corresponds to a minimum exposure unit for electron beam scanning and is marked as belonging to an area requiring exposure, for example, "1" indicates exposure and "0" indicates no exposure.

[0041] Subsequently, based on the geometric features such as the shape and size of the target graphic layout, corresponding energy distribution data is extracted from a pre-built structured database as the target energy deposition distribution for the discrete pixel grid in the target graphic layout. The pre-built structured database includes various typical graphic structures, different key process parameters, high-fidelity energy deposition distribution maps obtained through Monte Carlo simulations under different key process parameters, and fitted point spread function parameters. The target energy deposition distribution is in the form of a two-dimensional array, where each element corresponds to the expected energy deposition density on a discrete pixel grid.

[0042] Based on obtaining the real-time process environment parameters of the electron beam lithography equipment and the target pattern layout to be processed by the user in step S110, and calculating the target energy deposition distribution of the discrete pixel grid in the target pattern layout, step S120 is further executed.

[0043] S120, Based on a pre-trained machine learning model, predict the key parameter set of the point spread function according to the real-time process environment parameters; wherein, the machine learning model is obtained by training and optimizing a training sample set composed of process environment parameter samples and their corresponding point spread function parameter samples.

[0044] To accurately describe the proximity effect during the interaction between the electron beam and the resist, and to perform effective dose correction accordingly, it is necessary to accurately predict the key parameters of the point spread function. To this end, this embodiment proposes a pre-trained machine learning model, which is equivalent to the high efficiency of Monte Carlo simulation / simulation, capable of accurately predicting the corresponding set of key point spread function parameters for any given new process environment parameters in a short time.

[0045] Machine learning models can be built based on artificial neural networks, support vector machines, or random forests. They are obtained by training and optimizing samples of process environment parameters and their corresponding point spread function (PFD) parameters. During training, the process environment parameter samples are used as input, the predicted PFD parameters are used as output, and the difference between the predicted PFD parameters and the PFD parameter samples is used as the training loss to iteratively optimize the machine learning model.

[0046] For a machine learning model trained to convergence, the real-time process environment parameters of the electron beam lithography equipment are input into the model. The model completes the forward propagation calculation in a short time and directly predicts and outputs a high-precision set of key parameters of the point spread function under the current real-time process environment parameters. The key parameters of the point spread function include, but are not limited to, the forward scattering radius, the backscattering radius, and the energy ratio of the backscattering radius to the forward scattering radius, which characterize the physical scattering law of the electron beam.

[0047] For example, in one embodiment, the electron beam acceleration voltage on the current production line is 50 keV, PMMA resist is used, the substrate is a silicon wafer, and the ambient temperature and humidity are 25°C and 40%RH, respectively. These parameters are input into a pre-trained machine learning model to quickly predict the key parameter set of the point spread function applicable to this operating condition, such as a lateral spread radius of approximately 10 nm and a longitudinal attenuation length of 5 nm.

[0048] After predicting the key parameter set of the point spread function based on the pre-trained machine learning model and the real-time process environment parameters in step S120, step S130 is further executed.

[0049] S130, based on the key parameter set of the point spread function and the target energy deposition distribution, correct the dose modulation layout corresponding to the target graphic layout; the dose modulation layout includes the theoretical dose that needs to be applied to each discrete pixel grid.

[0050] After obtaining the key parameters of the point spread function and the target energy deposition distribution, dose inversion and modulation calculations are further performed to generate the final dose modulation pattern, thereby achieving high-fidelity pattern processing. Specifically, deconvolution operations are performed on the key parameters of the point spread function and the target energy deposition distribution to eliminate the influence of scattering effects on pattern accuracy, obtain the corrected electron beam dose distribution, and form the dose modulation pattern.

[0051] In a dose modulation pattern, each pixel value represents the electron beam exposure dose that should be applied at that location. Generally, isolated areas require a higher dose due to the lack of neighboring exposure contributions; dense areas require a lower dose to avoid overexposure due to the superposition of proximity effects; special structures (such as corners and tips) can introduce local enhancement or suppression factors.

[0052] The corrected dose modulation pattern is then combined with the target pattern to generate a new exposure file with precise dose modulation information. This exposure file is directly imported into the pattern generator of the electron beam lithography equipment to control the exposure time and beam intensity of each pixel, enabling spatially variable dose writing.

[0053] In this embodiment, by acquiring the real-time process environment parameters of the electron beam lithography equipment and the target pattern to be processed by the user, and calculating the target energy deposition distribution of discrete pixel grids in the target pattern, a pre-trained machine learning model is used to predict the key parameter set of the point spread function based on the real-time process environment parameters. The machine learning model is trained and optimized using a training sample set composed of process environment parameter samples and their corresponding point spread function parameter samples. Therefore, based on the key parameter set of the point spread function and the target energy deposition distribution, the dose modulation pattern corresponding to the target pattern is corrected. The dose modulation pattern includes the theoretical dose to be applied to each discrete pixel grid. This method, by introducing a machine learning model and training it into a highly efficient simulator, achieves online, real-time, and high-precision proximity effect correction, effectively suppressing the proximity effect and improving the accuracy and consistency of pattern processing. Simultaneously, it reduces the computation time from several hours to several minutes and allows for real-time dynamic adjustment of the exposure dose, thus providing a technical foundation for high-resolution, complex pattern electron beam lithography.

[0054] Based on the above embodiments, the following will further describe in detail the process of constructing the training sample set for the machine learning model.

[0055] The construction of the training sample set specifically includes: identifying key process parameters affecting electron beam scattering behavior; simulating the electron beam scattering trajectory and energy loss process under key process parameters using Monte Carlo simulation, and establishing an energy deposition distribution map; fitting a point spread function based on the energy deposition distribution map and extracting the point spread function parameters; and constructing the training sample set by using the key process parameters as process environment parameter samples and the point spread function parameters as point spread function parameter samples.

[0056] It is easy to understand that, in order to achieve fast and accurate prediction of key parameters of the point spread function, a high-quality training sample set needs to be pre-constructed for training and optimizing the machine learning model. This training sample set consists of process environment parameter samples and their corresponding point spread function parameter samples.

[0057] First, key process parameters that significantly influence the electron beam scattering behavior in the resist / substrate are identified. These key process parameters collectively determine the spatial distribution characteristics of energy deposition. These key process parameters include at least electron beam parameters and material parameters.

[0058] The electron beam parameters include the electron beam accelerating voltage (unit: keV), for example from 10keV to 100keV, with a step size of 5keV, and may also include the simulated number of electrons, beam current intensity (unit: pA or nA), scanning speed (unit: nm / s or μm / ms), and focusing conditions (such as beam spot size and astigmatism correction status).

[0059] Material parameters include the thickness of the resist (photoresist) layer, the type of resist material, and the type of substrate material. The resist layer thickness ranges from 50 nm to 500 nm, with a step size of 50 nm. The resist material type includes polymer-based positive resists (such as PMMA and ZEP520), chemically amplified resists (such as MA-P1275), negative resists (such as SU-8), inorganic resists (such as HSQ), and metal oxide-based resists (such as HfO2 nanoresists), which can be encoded by their chemical composition and density. The substrate material type includes single-crystal silicon (Si), silicon dioxide (SiO2), silicon nitride (Si3N4), gallium arsenide (GaAs), and quartz, etc.

[0060] Subsequently, for each set of key process parameters, Monte Carlo electron scattering simulation was used to simulate a large number of (e.g., 10) processes. 6 Up to 10 8 The scattering trajectory and energy loss process of electrons in the resist / substrate multilayer structure defined by this set of key process parameters form a three-dimensional energy deposition point cloud data within the resist. This three-dimensional energy deposition point cloud data is projected onto a two-dimensional plane and spatially statistically analyzed to establish an energy deposition distribution map. The energy deposition distribution map accurately reflects the spatial response of energy deposition induced by exposure of a single point source under specific process conditions and is the physical prototype of the point spread function.

[0061] In this simulation, material properties (such as density and atomic number) and electron beam parameters (such as energy and beam intensity) are considered to generate a high-precision scattering energy distribution map. Elastic scattering, inelastic scattering, and energy loss processes of electrons are taken into account to generate a high-precision energy deposition distribution map. A multilayer material structure model is constructed based on the physical properties of the photoresist and substrate materials (such as density, atomic number, and thickness). The accuracy of the scattering model is ensured through experimental verification and parameter optimization. Energy deposition distribution maps are generated based on the simulation results and stored in a database for subsequent calculations.

[0062] Accordingly, Figure 2 A schematic diagram of the Monte Carlo simulation results provided in this embodiment of the invention is shown, intuitively illustrating the energy deposition behavior during the interaction between the electron beam and the material. Figure 2 In the diagram, the horizontal axis represents the spatial range, approximately ±3000 nm; the vertical axis represents the vertical depth, approximately 0-3613.9 nm. The red lines represent the trajectories of primary electrons in the PMMA resist layer, while the blue lines represent the propagation paths of secondary electrons and backscattered electrons in PMMA and Si.

[0063] Furthermore, based on the aforementioned energy deposition distribution map, mathematical modeling and parameter fitting were performed to extract an accurate point spread function that can be used for dose correction. This point spread function is a key parameter describing the electron beam scattering effect, and its accurate extraction and modeling are crucial for optimizing the electron beam dose distribution.

[0064] Accordingly, Figure 3 A schematic diagram illustrating the point spread function extraction process provided in an embodiment of the present invention is shown. Figure 3 As shown, firstly, the two-dimensional plane is divided into several concentric annular regions with equal spacing, with the electron beam incident point as the origin. Then, within each annular region, the total energy loss of all electrons within that region is calculated, and this value is taken as the electron energy deposition in that region. Next, the number of forward-scattered electrons (near the center) is large and the data is dense, but due to the small number of electrons in the initial stage, it may cause local sparsity. Therefore, spline interpolation is performed on this part to smooth and complete the data. At the same time, the backscattered electrons have a wide propagation range and complex path, resulting in high noise in the data in the far-distance region. Therefore, the backscattered part is smoothed (e.g., by moving average or Gaussian filtering) to reduce the impact of noise. Finally, multiple candidate point spread function models (e.g., single Gaussian, double Gaussian, exponential decay, etc.) are used to fit the processed energy distribution, and the model with the best fitting effect is selected as the final point spread function.

[0065] Figure 4 This diagram illustrates the fitting effect of the point spread function provided in an embodiment of the present invention, demonstrating a comparison between energy deposition data obtained through Monte Carlo simulation in electron beam lithography and fitting curves from various mathematical models. Figure 4 In the diagram, the horizontal axis represents the distance from the electron incident point on a logarithmic scale, covering a range from 1 nm to 1000 nm; the vertical axis represents the point diffusion function value, i.e., the energy deposition intensity per unit area, also on a logarithmic scale.

[0066] The black scatter plots (smoothed data) represent the energy deposition data points obtained from the original Monte Carlo simulation after preprocessing (such as smoothing and denoising). The red curve (double Gaussian model fitting) represents the double Gaussian model fitting result. The blue curve (triple Gaussian model fitting) represents the triple Gaussian model fitting result. The green curve (aya model fitting) represents the aya model fitting result. The blue curve (multinomial fitting) represents the eighth-order polynomial fitting result.

[0067] The point spread function can be described by the sum of multiple Gaussian functions. For example, this embodiment uses a double Gaussian model to characterize the point spread function, and its mathematical form is: in, Radial distance, Forward scattering radius, The backscattering radius is These parameters, representing the energy ratio of backscattered to forward scattered energy, collectively and precisely describe the characteristics of electron beam scattering. The corresponding values ​​for each set of key process parameters are extracted from the point spread function. , which is a set of point spread function parameters.

[0068] The results of the above process are organized into a training sample set. Each training sample in the training sample set represents a complete simulation experiment, and its format is: Input features (process environment parameter samples: key process parameters) -> Output labels (point spread function parameter samples: fitted point spread function parameters) ).

[0069] In addition, the data involved in the above process can be organized into a structured database. Similarly, each row of this structured database represents a complete simulation / simulation experiment, but it specifically includes key process parameters, energy deposition distribution map, point spread function parameters and corresponding graphical structures, ultimately forming a high-quality structured database containing thousands to tens of thousands of records with a precise mapping relationship "from process environment to physical scattering law".

[0070] Based on the above embodiments, the training and optimization process of the machine learning model will be described in detail below.

[0071] Training and optimizing the machine learning model includes: taking process environment parameter samples as input, predicting point spread function parameters as output, and using the difference between the predicted point spread function parameters and the point spread function parameter samples as training loss, iteratively optimizing the machine learning model; wherein the machine learning model is a random forest or a gradient boosting decision tree.

[0072] It is easy to understand that the constructed training sample set is first standardized to improve the stability and convergence speed of model training.

[0073] Then, select a suitable machine learning algorithm based on the characteristics of the data: 1) Deep learning models, such as convolutional neural networks, recurrent neural networks, or graph neural networks, are suitable for processing high-dimensional nonlinear data; 2) Traditional machine learning algorithms, such as support vector machines and random forests, are suitable for small to medium-sized datasets. Since the input and output in this embodiment are both structured numerical vectors, models such as random forests and gradient boosting decision trees are preferred, as these models are more suitable and efficient than convolutional neural networks used for image generation.

[0074] During the model training and optimization process, the input of the model is a one-dimensional numerical vector describing the process environment (process environment parameter sample), such as [voltage, adhesive thickness, substrate atomic number, ...]. The corresponding model output is a set of point spread function parameters describing the physical scattering law of the electron beam (point spread function parameter sample), such as [α, β, η, ...].

[0075] Specifically, during model training and optimization, the constructed training sample set is proportionally divided into a training set, a validation set, and a test set. The machine learning model is trained to learn the complex nonlinear mapping from input (process environment parameter samples) to output (predicted point spread function parameters). Root mean square error, mean absolute error, and coefficient of determination are used as loss functions or evaluation metrics, and the model is trained on the training set. During training, the validation set is used to monitor model performance, prevent overfitting, and perform hyperparameter tuning. After training, an independent test set is used to evaluate the performance of the final model. This model achieves high efficiency in Monte Carlo simulation, enabling accurate prediction of the corresponding key parameter set of the point spread function for any given new process environment parameters within a short time.

[0076] Correspondingly, Figure 5 The illustration shows the effect of the machine learning model provided in the embodiment of the present invention in predicting the key parameter set of the point spread function. It demonstrates the comparative analysis process of the predicted value and the true value, which can be used to evaluate the accuracy of the machine learning model in predicting the key parameter set of the point spread function.

[0077] exist Figure 5 In the diagram, the horizontal axis represents the true values ​​(i.e., the point spread function parameter samples obtained through Monte Carlo simulation), and the vertical axis represents the predicted values ​​(i.e., the key parameter set of the point spread function output by the trained machine learning model). A red asterisk (*) indicates the machine learning model's prediction for each sample, and the light blue dot represents the "perfect prediction point," the ideal situation where the predicted value equals the true value. The blue dashed line represents the true value fitting line, and the black straight line represents the predicted value fitting line, which is the linear regression result for all predicted points.

[0078] according to Figure 5 As can be seen, almost all the predicted points in the figure fall on the true value fitting line, indicating that the machine learning model has extremely high prediction accuracy and stability.

[0079] Based on the above embodiments, the following will further describe in detail the calculation process of the target energy deposition distribution of discrete pixel grids in the target graphic layout, and the correction process of the final dose modulation layout.

[0080] As can be easily understood, as described in the above embodiments, a structured database is pre-built and stored. Each row of this structured database represents a complete simulation / simulation experiment, which specifically includes key process parameters, energy deposition distribution maps, point spread function parameters, and corresponding graphical structures. It is a high-quality database containing thousands to tens of thousands of records with a precise mapping relationship "from process environment to physical scattering laws".

[0081] When calculating the target energy deposition distribution of discrete pixel grids in a target graphic layout, for the acquired target graphic layout to be processed by the user (e.g., GDSII or OASIS format), it is first discretized into a high-density pixel grid. Then, based on the geometric features of the target graphic layout (such as geometry and size), the closest graphic structure is matched in the structured database, and the energy deposition distribution map corresponding to the closest graphic structure is extracted. Finally, the energy deposition distribution map corresponding to the closest graphic structure is determined as the target energy deposition distribution corresponding to multiple discrete pixel grids.

[0082] The total energy actually deposited on each discrete pixel grid It is applied to all discrete pixel grids. dosage After being scattered by the point spread function, it contributes to the discrete pixel grid. The sum of energy at each point forms the target energy deposition distribution. Specifically, it can be represented as a large system of linear equations. Among them, matrix The elements represent the points spread function predicted by the machine learning model, determined by a set of key parameters from a discrete pixel grid. To discrete pixel grid The energy ratio, Represents all pixels The theoretical dose that should be applied above.

[0083] Proximity effect correction is achieved by solving the deconvolution of the aforementioned large system of linear equations. A target energy deposition distribution is defined. (Usually 1 in the graphic area and 0 in the non-graphic area), then solve. This solution process can be achieved through efficient algorithms such as deconvolution, Fourier transform, or iterative solution. For example, the extracted key parameter set of the point spread function is deconvolved with the target energy deposition distribution of the target processing pattern (target graphic layout) to eliminate the influence of scattering effects on the pattern accuracy, thereby obtaining the corrected electron beam dose distribution, i.e., the dose modulation pattern, and ensuring the clarity of the pattern edges.

[0084] Final calculated dose modulation pattern This process involves determining the precise exposure dose required for each discrete pixel grid, combining this dose modulation pattern with the target pattern pattern to generate a new exposure file with accurate dose modulation information. This exposure file is then loaded into an electron beam lithography apparatus for exposure. Experiments are conducted to verify the corrected dose modulation pattern, optimize processing parameters, and ensure that the processing accuracy meets design requirements.

[0085] In some other embodiments, the key points and beneficial technical effects of the machine learning-based electron beam lithography dose correction method provided by the present invention are illustrated.

[0086] The electron beam lithography dose correction method based on machine learning proposed in this invention constructs a massive database containing the mapping relationship between "key process parameters - energy deposition distribution map - point spread function parameters - pattern structure" through large-scale, high-precision Monte Carlo physical simulation. Based on this massive database, a machine learning model is trained to learn and replicate the predictive ability of the Monte Carlo model, becoming its highly efficient simulator. Before actual processing, this lightweight and fast machine learning model is used to replace the traditional time-consuming Monte Carlo simulation, calculating the optimal correction dose for any pattern in real time and accurately.

[0087] The electron beam lithography dose correction method based on machine learning provided in this invention dynamically adjusts the exposure dose by accurately simulating electron beam scattering behavior, effectively suppressing proximity effects and significantly improving the accuracy and consistency of electron beam lithography master fabrication. Compared with traditional methods, this invention has advantages such as high precision, high efficiency, intelligence, and wide applicability, and is particularly suitable for high-resolution, complex pattern micro / nano fabrication. Through experimental verification and practical application, this invention demonstrates significant advantages in pattern clarity, linewidth uniformity, and processing efficiency, and has important industrial application value.

[0088] The electron beam lithography dose correction method based on machine learning proposed in this invention has the following four significant advantages.

[0089] Firstly, high-precision energy distribution simulation: The Monte Carlo method is used to simulate the scattering behavior of electron beams in photoresist and substrate materials, accurately describing the electron trajectory and energy deposition distribution, overcoming the limitations of traditional methods. By considering the physical properties of the material (such as density, atomic number, thickness, etc.), a high-precision target energy deposition distribution is generated, providing a reliable theoretical basis for dose correction.

[0090] Secondly, dynamic dose adjustment: Based on the geometry and size of the target pattern, combined with the target energy deposition distribution, the exposure dose of each pixel is dynamically adjusted to effectively suppress the proximity effect. By introducing a proximity effect correction factor, the sharpness of the pattern edges and the consistency of linewidth are ensured.

[0091] Thirdly, the intelligent dose correction system integrates machine learning technology to automate and intelligently correct dose, improving processing efficiency. An integrated dose correction system is developed to seamlessly integrate scattering simulation, dose calculation, and optimization algorithms. It supports real-time communication with existing electron beam lithography equipment to achieve automated dose correction.

[0092] Fourthly, broad applicability: This method is applicable to a variety of materials and processing scenarios, including semiconductor lithography, nanoimprinting, and micro / nano fabrication. It is suitable for various materials and processing conditions, meeting the requirements for high-resolution, complex pattern fabrication. By adjusting the scattering model parameters, it can adapt to different electron beam energies, photoresist thicknesses, and substrate material properties.

[0093] The electron beam lithography dose correction method based on machine learning proposed in this invention brings the following beneficial technical effects.

[0094] First, a technological breakthrough combining accuracy and speed: For the first time, high-precision physical simulation is combined with high-efficiency machine learning, breaking the bottleneck of the traditional method where accuracy and speed cannot be simultaneously achieved. The accuracy of the calibration results approaches the Monte Carlo "gold standard," while the computational speed is increased by orders of magnitude, realizing a "fast and accurate" micro-nano fabrication method.

[0095] Secondly, it boasts strong universality and adaptability: When the processing technology (such as changing the resist) changes, traditional methods require a significant amount of time to recalibrate the experiments. This invention, however, only requires inputting the new process parameters into the pre-trained model to instantly obtain a new, high-precision correction scheme, demonstrating extremely strong adaptability.

[0096] Finally, it significantly reduces R&D and production costs: by replacing online, repetitive, time-consuming calculations and a large number of experimental trials with offline, one-time simulation calculations, it greatly saves equipment time, manpower and material costs, and provides core algorithm support for realizing a fully automated and intelligent electron beam lithography equipment process control system.

[0097] Corresponding to the machine learning-based electron beam lithography dose correction method described in the above embodiments, this invention also proposes a machine learning-based electron beam lithography dose correction device. Specifically, Figure 6 A schematic diagram of the structure of the electron beam lithography dose correction device based on machine learning provided in an embodiment of the present invention is shown.

[0098] like Figure 6As shown, the device includes: an energy deposition distribution calculation module 610, used to acquire real-time process environment parameters of the electron beam lithography equipment and the target pattern layout to be processed by the user, and calculate the target energy deposition distribution of discrete pixel grids in the target pattern layout; a point spread function key parameter group prediction module 620, used to predict the point spread function key parameter group based on a pre-trained machine learning model and the real-time process environment parameters; wherein, the machine learning model is obtained by training and optimization based on a training sample set composed of process environment parameter samples and their corresponding point spread function parameter samples; and a dose modulation layout correction module 630, used to correct the dose modulation layout corresponding to the target pattern layout based on the point spread function key parameter group and the target energy deposition distribution; the dose modulation layout includes the theoretical dose to be applied to each discrete pixel grid.

[0099] In this embodiment, the energy deposition distribution calculation module 610 acquires the real-time process environment parameters of the electron beam lithography equipment and the target pattern to be processed by the user, and calculates the target energy deposition distribution of discrete pixel grids in the target pattern. Then, the point spread function key parameter group prediction module 620 predicts the point spread function key parameter group based on a pre-trained machine learning model and the real-time process environment parameters. The machine learning model is trained and optimized using a training sample set composed of process environment parameter samples and their corresponding point spread function parameter samples. Therefore, the dose modulation pattern correction module 630 corrects the dose modulation pattern corresponding to the target pattern based on the point spread function key parameter group and the target energy deposition distribution. The dose modulation pattern includes the theoretical dose to be applied to each discrete pixel grid. This device, by introducing a machine learning model and training it into a highly efficient simulator, achieves online, real-time, and high-precision proximity effect correction, effectively suppressing the proximity effect and improving the accuracy and consistency of pattern processing. Simultaneously, it reduces the computation time from several hours to several minutes and allows for real-time dynamic adjustment of the exposure dose, thus providing a technical foundation for high-resolution, complex pattern electron beam lithography.

[0100] It should be noted that the machine learning-based electron beam lithography dose correction device provided in the embodiments of the present invention can be referred to in correspondence with the machine learning-based electron beam lithography dose correction method described in the above embodiments, and will not be repeated here.

[0101] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7As shown, the electronic device may include: a processor 710, a communications interface 720, a memory 730, and a communication bus 740, wherein the processor 710, the communications interface 720, and the memory 730 communicate with each other via the communication bus 740. The processor 710 can call logical instructions in the memory 730 to execute a machine learning-based electron beam lithography dose correction method. This method includes: acquiring real-time process environment parameters of the electron beam lithography equipment and the target pattern layout to be processed by the user, and calculating the target energy deposition distribution of discrete pixel grids in the target pattern layout; predicting key parameter sets of the point spread function based on a pre-trained machine learning model according to the real-time process environment parameters; wherein the machine learning model is obtained by training and optimization using a training sample set composed of process environment parameter samples and their corresponding point spread function parameter samples; and correcting the dose modulation layout corresponding to the target pattern layout according to the key parameter sets of the point spread function and the target energy deposition distribution; the dose modulation layout includes the theoretical dose to be applied to each discrete pixel grid.

[0102] Furthermore, the logical instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0103] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the electron beam lithography dose correction method based on machine learning provided by the above methods. The method includes: acquiring real-time process environment parameters of the electron beam lithography equipment and the target pattern layout to be processed by the user, and calculating the target energy deposition distribution of discrete pixel grids in the target pattern layout; predicting key parameter sets of the point spread function based on a pre-trained machine learning model according to the real-time process environment parameters; wherein the machine learning model is obtained by training and optimization based on a training sample set composed of process environment parameter samples and their corresponding point spread function parameter samples; and correcting the dose modulation layout corresponding to the target pattern layout according to the key parameter sets of the point spread function and the target energy deposition distribution; the dose modulation layout includes the theoretical dose to be applied to each discrete pixel grid.

[0104] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the machine learning-based electron beam lithography dose correction method provided by the methods described above. This method includes: acquiring real-time process environment parameters of the electron beam lithography equipment and a target pattern layout to be processed by the user, and calculating the target energy deposition distribution of discrete pixel grids in the target pattern layout; predicting key parameter sets of the point spread function based on a pre-trained machine learning model according to the real-time process environment parameters; wherein the machine learning model is obtained by training and optimization using a training sample set composed of process environment parameter samples and their corresponding point spread function parameter samples; and correcting the dose modulation layout corresponding to the target pattern layout according to the key parameter sets of the point spread function and the target energy deposition distribution; the dose modulation layout includes the theoretical dose to be applied to each discrete pixel grid.

[0105] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0106] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A machine learning-based electron beam lithography dose correction method, characterized in that, include: The system acquires real-time process environment parameters of the electron beam lithography equipment and the target pattern layout to be processed by the user, and calculates the target energy deposition distribution of the discrete pixel grid in the target pattern layout. Based on a pre-trained machine learning model, the key parameter set of the point spread function is predicted according to the real-time process environment parameters; wherein, the machine learning model is obtained by training and optimizing a training sample set consisting of process environment parameter samples and their corresponding point spread function parameter samples. Based on the key parameter set of the point spread function and the target energy deposition distribution, the dose modulation layout corresponding to the target graphic layout is corrected; the dose modulation layout includes the theoretical dose that needs to be applied to each discrete pixel grid.

2. The electron beam lithography dose correction method based on machine learning according to claim 1, characterized in that, Constructing the training sample set specifically includes: Determine the key process parameters that affect electron beam scattering behavior; Based on Monte Carlo simulation, the electron beam scattering trajectory and energy loss process under the key process parameters are simulated to establish an energy deposition distribution map; The point diffusion function is fitted based on the energy deposition distribution map, and the point diffusion function parameters are extracted; The key process parameters are used as process environment parameter samples, and the point spread function parameters are used as point spread function parameter samples to construct the training sample set.

3. The electron beam lithography dose correction method based on machine learning according to claim 2, characterized in that, The key process parameters include electron beam parameters and material parameters; The electron beam parameters include the electron beam accelerating voltage and the simulated number of electrons; The material parameters include the thickness of the resist layer, the type of resist material, and the type of substrate material.

4. The electron beam lithography dose correction method based on machine learning according to claim 2, characterized in that, The key process parameters, the energy deposition distribution map, the point diffusion function parameters, and various graphical structures are organized into a structured database and stored. Accordingly, the step of calculating the target energy deposition distribution of the discrete pixel grid in the target graphic layout includes: The target graphic layout is discretized into multiple discrete pixel grids; Based on the geometric features of the target graphic layout, the closest graphic structure is matched in the structured database, and the energy deposition distribution map corresponding to the closest graphic structure is extracted. The energy deposition distribution map corresponding to the closest graphic structure is determined as the target energy deposition distribution corresponding to the multiple discrete pixel grids.

5. The electron beam lithography dose correction method based on machine learning according to claim 1, characterized in that, Training and optimizing the machine learning model includes: The machine learning model is iteratively optimized by taking the process environment parameter samples as input, the predicted point spread function parameters as output, and the difference between the predicted point spread function parameters and the point spread function parameter samples as training loss. The machine learning model selects different models based on data characteristics.

6. The electron beam lithography dose correction method based on machine learning according to any one of claims 1-5, characterized in that, The step of correcting the dose modulation pattern corresponding to the target pattern based on the key parameter set of the point spread function and the target energy deposition distribution, followed by: The dose modulation pattern is combined with the target graphic pattern to generate a target exposure file; The target exposure file is loaded into the electron beam lithography equipment, and exposure is performed.

7. A machine learning-based electron beam lithography dose correction device, characterized in that, include: The energy deposition distribution calculation module is used to acquire the real-time process environment parameters of the electron beam lithography equipment and the target pattern layout to be processed by the user, and to calculate the target energy deposition distribution of the discrete pixel grid in the target pattern layout; The point spread function key parameter group prediction module is used to predict the point spread function key parameter group based on the real-time process environment parameters using a pre-trained machine learning model; wherein, the machine learning model is obtained by training and optimizing a training sample set consisting of process environment parameter samples and their corresponding point spread function parameter samples. The dose modulation pattern correction module is used to correct the dose modulation pattern corresponding to the target graphic pattern based on the key parameter set of the point spread function and the target energy deposition distribution; the dose modulation pattern includes the theoretical dose that needs to be applied to each discrete pixel grid.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the machine learning-based electron beam lithography dose correction method as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the machine learning-based electron beam lithography dose correction method as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the machine learning-based electron beam lithography dose correction method as described in any one of claims 1 to 6.

Images