Photomask pattern optimization method and system, storage medium and electronic equipment

Abstract

The invention provides a photomask pattern optimization method and system, a storage medium and electronic equipment. The method comprises the following steps: acquiring an initial layout corresponding to a target process layer and environment information of the initial layout in a preset neighborhood range; inputting the initial layout and the environment information into a teacher model after meta-learning training, and outputting a first initial mask solution by the teacher model; migrating the first initial mask solution knowledge to a student model to generate a compressed student model; and taking a second initial mask solution output by the compressed student model as an initial value of optical mask optimization iteration, executing iterative calculation, and generating a final mask pattern. According to the method, the parameter quantity and the calculation quantity of the optical model can be reduced, rapid convergence is realized, and the optimization efficiency and precision of the photomask pattern are improved.

Description

Technical Field

[0001] This application belongs to the field of integrated circuit manufacturing technology, and in particular relates to a photomask pattern optimization method, system, storage medium and electronic device. Background Technology

[0002] Currently, OPC (Optical Proximity Correction) technology is mainly used in the photomask pattern optimization process. In model-based optical proximity correction, a simulated target pattern is typically used as the initial photomask pattern. The exposure pattern is calculated using an optical model and photoresist chemical reactions. Through multiple iterations of optimization, a photomask pattern that meets the requirements is finally obtained. OPC primarily compensates for imaging distortion caused by optical diffraction by adjusting the edges of the photomask pattern or adding polygons, ensuring that the silicon wafer circuitry is consistent with the target pattern.

[0003] However, using simulated target patterns as initial photomask patterns increases the number of iterations, and sometimes fails to converge due to significant differences between the initial and final photomask patterns. Furthermore, a key aspect of this method is establishing accurate lithography models, including optical and photoresist models. To achieve high computational speeds, a large amount of experimental data is typically required for fitting to ensure model accuracy. However, too many test patterns result in excessive data volume, significantly limiting efficiency, runtime, and accuracy. Summary of the Invention

[0004] This application discloses a photomask pattern optimization method, system, storage medium, and electronic device, which can reduce the number of parameters and computational load of optical models and achieve fast convergence, thereby improving the optimization efficiency and accuracy of photomask patterns.

[0005] Other objectives and advantages of this application can be further understood from the technical features disclosed herein.

[0006] To achieve one, some, or all of the above objectives or other objectives, in a first aspect, this application provides a photomask pattern optimization method, the method comprising: Obtain the initial layout corresponding to the target process layer and the environmental information of the initial layout within a preset neighborhood; The initial layout and environmental information are input into the teacher model trained by meta-learning, and the teacher model outputs the first initial mask solution. The knowledge of the first initial mask solution is transferred to the student model to generate a compressed student model; The second initial mask solution output by the compressed student model is used as the initial value for the photomask optimization iteration. Iterative calculations are performed to generate the final mask pattern.

[0007] In one implementation, the meta-learning training process of the teacher model includes: The teacher model is trained on multiple different process tasks using a meta-learning algorithm to determine initial parameters that are sensitive to task distribution.

[0008] In one implementation, the teacher model is trained on multiple different process tasks using a meta-learning algorithm to determine initial parameters sensitive to task distribution, including: M process tasks are randomly selected from the multiple different process tasks as the support set, and the remaining process tasks are used as the query set. For each process task, the current model parameters are fine-tuned using the support set to determine the specific parameters for that process task; The query set is used to calculate the loss function for the specific parameters of each process task, and the initial parameters are updated by gradient descent. Repeat the above steps until the loss function converges, and obtain the task distribution-sensitive initial parameters.

[0009] In one implementation, the step of inputting the initial layout and environmental information into the trained teacher model and having the teacher model output the initial mask solution includes: inputting the initial layout and environmental information into the teacher model based on the task distribution-sensitive initial parameters, fixing the bottom feature extraction layer of the task distribution-sensitive initial parameters, fine-tuning the top decision layer until convergence, and outputting the first initial mask solution.

[0010] In one implementation, the transfer of the initial mask solution knowledge of the teacher model to the student model includes: constructing a mask-level loss function, an edge gradient loss function, and a lithographic evaluation index loss function based on the soft target and intermediate features output by the teacher model; and jointly training the student model until the student model reproduces the initial mask solution distribution of the teacher model within a preset accuracy threshold.

[0011] In one implementation, the number of iterations in the iterative calculation is less than a set number of iterations.

[0012] In one implementation, the student model has a lower single-inference latency than the teacher model.

[0013] Secondly, this application provides a photomask pattern optimization system, the system comprising at least: The meta-learning training module is used to train the teacher model on multiple process tasks to learn initial parameters that are sensitive to task distribution. The knowledge distillation module is used to transfer the knowledge of the first initial mask solution of the teacher model to the student model, generating a compressed student model. The iterative optimization module is used to perform photomask optimization iterations starting from the second initial mask solution output by the student model, and output the final mask pattern. A control module is used to coordinate the above modules to execute the photomask pattern optimization method as described in any of the first aspects.

[0014] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the photomask pattern optimization method described in any of the first aspects.

[0015] Fourthly, this application provides an electronic device including one or more processors; Memory; One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the one or more processors, the one or more applications being configured to perform the photomask graphics optimization method as described in any of the first aspects.

[0016] The aforementioned optimization method, system, computer-readable storage medium, and electronic device for photomask patterns have at least the following technical effects: 1) By training the teacher model on multiple different process tasks using a meta-learning algorithm, initial parameters sensitive to task distribution are learned. When applying a new task, only minor fine-tuning is needed to generate a high-quality initial mask solution. Furthermore, the initial mask solution output by the compressed student model is used as the initial value for photomask optimization iteration, resulting in fewer iterations than conventional iterations. This significantly reduces the number of parameters and computational load of the optical model and achieves rapid convergence, improving optimization efficiency and accuracy. 2) The output distribution of the student model gradually approximates the teacher model until a preset accuracy threshold is met. Moreover, the student model inherits the generalization ability of the teacher model, requiring only minor fine-tuning for new process tasks without retraining. This adapts to the rapid iteration requirements of advanced processes, enhancing the flexibility and adaptability of the model.

[0017] To make the above and other objects, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the specific embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is an iterative flowchart of photomask pattern optimization technology in related technologies.

[0020] Figure 2 This is a flowchart illustrating the photomask pattern optimization method described in this application.

[0021] Figure 3 This is an iterative flowchart of the photomask pattern optimization method described in this application.

[0022] Figure 4 This is a flowchart illustrating the photomask pattern optimization method described in this application.

[0023] Figure 5 This is a flowchart illustrating the photomask pattern optimization method described in this application.

[0024] Figure 6 This is a block diagram of the architecture of the photomask pattern optimization system described in this application. Detailed Implementation

[0025] The foregoing and other technical contents, features, and effects of this application will be clearly presented in the following detailed description of a preferred embodiment with reference to the accompanying drawings. The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate; this is merely a method of distinguishing objects with the same attributes in the embodiments of this application.

[0026] refer to Figure 1 The proposed process for optimizing photomask patterns in OPC (Optimal Photomask Processing) involves using a simulated target image as the initial photomask pattern. This target image is then substituted into the optical model. Auxiliary images are used to adjust edge distortions in the initial photomask pattern, such as linewidth and corner radius. Through multiple iterative calculations, optimization is gradually achieved, ultimately yielding a photomask pattern that meets the requirements. However, this method suffers from problems such as excessive model data volume and numerous iterations, significantly limiting optimization efficiency and accuracy. Furthermore, if migration to a new process platform is required, rapid migration of the model from the old platform is not feasible, necessitating a large amount of new platform data for training, resulting in low model flexibility and adaptability.

[0027] To solve the above technical problems, refer to Figure 2 and Figure 3 The photomask pattern optimization method provided in this application embodiment specifically includes the following steps: Step S101: Obtain the initial layout corresponding to the target process layer and the environmental information of the initial layout within a preset neighborhood.

[0028] For example, the initial layout is the physical layout data of the target process layer exported from an EDA tool, including the geometric parameters of the circuit pattern (i.e., the initial photomask pattern), such as linewidth, spacing, and pattern topology. The environmental information covers key parameters across the entire photolithography process chain, including equipment parameters, dielectric properties, and environmental control parameters. The equipment parameters include light source wavelength, numerical aperture, and illumination mode; the dielectric properties include photoresist photosensitivity parameters and etching process parameters; and the environmental control parameters include cleanliness, temperature, humidity, and vibration frequency.

[0029] Step S102: Input the initial layout and environment information into the teacher model trained by the meta-learning algorithm, and output the first initial mask solution from the teacher model.

[0030] In some embodiments, the meta-learning training process of the teacher model includes: training the teacher model on multiple different process tasks using a meta-learning algorithm to determine initial parameters sensitive to task distribution. The task distribution-sensitive initial parameters obtained through meta-learning can quickly adapt to various process tasks.

[0031] In some embodiments, the process of inputting the initial layout and environmental information into a trained teacher model and having the teacher model output a first initial mask solution includes: inputting the initial layout and environmental information into the teacher model based on the task distribution-sensitive initial parameters, fixing the bottom feature extraction layer of the task distribution-sensitive initial parameters, fine-tuning the top decision layer until convergence, and outputting the first initial mask solution.

[0032] After training, the teacher model can master the initial mask solution for different process tasks. The initial mask solution includes the mask edge smoothness of the photomask pattern, the preservation rules of key graphic features, the etching parameter mapping, defect detection features, etc.

[0033] Step S103: Transfer the knowledge of the first initial mask solution to the student model to generate a compressed student model.

[0034] Specifically, the first initial mask solution output by the teacher model is close to the final mask pattern. Knowledge distillation is used to transfer the knowledge of the first initial mask solution of the teacher model to the student model as a soft target to guide the training of the student model. The purpose is to extract and pass on the generation logic of high-quality initial mask solutions, the feature mapping relationship sensitive to task distribution, and the decision-making experience of complex scenarios to the more streamlined student model. At the same time, redundant parameters are removed through model compression technology, and finally a lightweight student model with accuracy close to the teacher model and a significantly reduced size is obtained, which is suitable for real-time deployment requirements of production lines.

[0035] Step S104: Using the second initial mask solution output by the compressed student model as the initial value for the photomask optimization iteration, perform iterative calculations to generate the final mask pattern.

[0036] refer to Figure 3 The second initial mask solution (i.e. the initial mask solution of the photomask pattern shown in the figure) output by the compressed student model is used as the initial value for photomask optimization iteration. Through fast iterative calculation, the number of iteration convergence cycles can be reduced compared with the default unified initial value method. That is, the number of iteration calculation cycles is less than the conventionally set number of iteration cycles, while still being able to generate the final mask pattern that meets the preset process window and mask rule check.

[0037] The student model is deployed in an online photomask correction system. The single inference latency of the student model is lower than that of the teacher model, thereby improving the overall correction throughput and energy efficiency without reducing the quality of the final mask image.

[0038] Based on the above description, the photomask pattern optimization method described in this application involves training the teacher model on multiple different process tasks using a meta-learning algorithm to learn initial parameters sensitive to task distribution. When applying a new task, only minor fine-tuning is needed to generate a high-quality initial mask solution. Furthermore, the initial mask solution output by the compressed student model is used as the initial value for photomask optimization iteration, resulting in fewer iterations than in conventional iterations. This significantly reduces the number of parameters and computational load of the optical model, achieves rapid convergence, and improves optimization efficiency and accuracy.

[0039] In some embodiments, reference Figure 4 The teacher model is trained on multiple different process tasks using a meta-learning algorithm to determine initial parameters sensitive to task distribution, specifically including the following steps: Step S201: Randomly select M process tasks from the multiple different process tasks as a support set, and use the remaining process tasks as a query set; the multiple different process tasks may come from the collected relevant historical process task data.

[0040] Step S202: For each process task, fine-tune the current model parameters using the support set to determine the specific parameters for that process task.

[0041] Step S203: Calculate the loss function of the specific parameters for each process task using the query set, and update the initial parameters using gradient descent.

[0042] Step S204: Repeat the above steps until the loss function converges to obtain the task distribution-sensitive initial parameters.

[0043] Based on the task-distribution-sensitive initial parameters, the initial layout and environmental information are input into the teacher model. The bottom-level feature extraction layer with the task-distribution-sensitive initial parameters is fixed, and the top-level decision layer is fine-tuned until convergence, outputting an initial mask solution close to the final mask pattern. Thus, through a meta-learning algorithm, the model acquires the ability to quickly adapt to new process tasks and improves the sensitivity of the initial parameters to key features in different process tasks, reducing the sample size required for fine-tuning.

[0044] For example, the teacher model, through a meta-learning algorithm, grasps the common patterns of different GAA (Gate-All-Around) etching tasks, obtaining initial parameters sensitive to device type, etching parameters, and process effect distribution, enabling rapid adaptation of etching parameters for new devices. Specifically, it collects M+N historical process task data, such as GAA devices with various channel widths and stacking layers. Each process task includes complete data on device structural parameters, etching parameter combinations, and process effect indicators, such as etching depth error, sidewall angle deviation, and linewidth uniformity. The differences between different process tasks lie in the device structural parameters and the target process effect threshold. During each meta-training iteration, M etching tasks are randomly selected as the support set, and the remaining N are used as the query set. The initial parameter θ of the initialized teacher model adapts to the common characteristics of different GAA etching tasks. For each etching task in the support set, the current model parameters θ are fine-tuned using the device structure, etching parameters, and process effect data for that task, resulting in task-specific parameters θ_i (where i is the task index in the support set). The fine-tuning process minimizes the deviation loss between the model's predicted etching parameters and the actual process effect. The fine-tuned parameters θ_i are then applied to similar tasks in the query set, such as GAA devices with the same stacking layer number, to calculate the process effect loss of the predicted etching parameters. The loss function is a weighted sum of etching depth error, sidewall angle deviation, and linewidth uniformity. The initial parameters θ are updated using gradient descent. This process of support set extraction, fine-tuning, query set loss calculation, and initial parameter update is repeated until the overall loss of the query set converges, ultimately yielding initial parameters sensitive to the distribution of GAA etching tasks.

[0045] For example, the teacher model learns multiple defect detection tasks at different process nodes and wafer layers through a meta-learning algorithm, mastering common features of defect detection, such as edge and texture anomalies, and obtaining initial parameters sensitive to the distribution of defect tasks. When faced with a new detection task, high-precision detection can be achieved with only a small amount of data fine-tuning based on these initial parameters. Specifically, M+N defect detection task data points at different process nodes and wafer layers are collected. Each task includes a wafer surface image and defect annotations. The defect types for each task cover several common defects, such as type, location, and size. The difference between the tasks lies in that the process node determines the defect size, and the wafer layer determines the background texture, such as regular texture in the lithography layer and irregular texture in the etching layer. During each meta-training iteration, M detection tasks are randomly selected as the support set, and the remaining N are used as the query set. Within each task, a small number of samples are further divided as support samples for that task, and the rest are used as query samples. The initial parameter η of the teacher model is initialized, which can adapt to detection tasks with different defect sizes and different background textures. For each detection task in the support set, the initial parameter η is fine-tuned using multiple support samples from that task to obtain the task-specific detection parameter η_k, where k is the task index in the support set. The fine-tuning loss is the defect classification cross-entropy loss plus the bounding box regression loss. The specific detection parameter η_k is applied to the samples in the query set for that task to calculate the detection accuracy loss. After summing the query losses of all support set tasks, the initial parameter η is updated using gradient descent to make η more closely resemble the distribution characteristics of all tasks. When a new defect detection task is added, a small number of samples from that task are input into the teacher model based on task distribution-sensitive initial parameters. Only 1-2 rounds of fine-tuning are needed to obtain a high-precision detection model, eliminating the need for training from scratch.

[0046] In some embodiments, reference Figure 5 The transfer of the first initial mask solution knowledge from the teacher model to the student model specifically includes the following steps: Step S301: Based on the soft target and intermediate features output by the teacher model, construct mask-level loss, edge gradient loss and lithography evaluation index loss; Step S302: Jointly train the student model until the student model reproduces the initial mask solution distribution of the teacher model within a preset accuracy threshold.

[0047] That is, by leveraging the soft objectives of the teacher model, such as probabilistic decision information, and intermediate features, such as the ability to extract abstract features, deep knowledge distillation is achieved, enabling the student model to not only replicate the output of the teacher model but also inherit its task distribution-sensitive mask generation logic. Simultaneously, for the process characteristics of photomasks, three types of loss functions—mask level, edge gradient, and lithography evaluation index—are designed to form joint constraints, ensuring that the output of the student model not only conforms to the teacher distribution but also meets the actual requirements of the lithography process. Ultimately, high-precision reproduction is achieved in a lightweight student model.

[0048] For example, the soft target is the pixel-level probability distribution of the mask output by the teacher model, such as the confidence of each pixel as the effective mask region. After smoothing with a temperature coefficient, it retains the decision hesitation information of the teacher model in blurred areas, such as the transition area of ​​the mask edge, allowing the student model to learn more refined generation logic rather than simply copying pixel results. After being trained by meta-learning, the deep intermediate features of the teacher and student models have the ability to capture key information of the process task, such as the adaptation relationship between the layout and the mask. Furthermore, by constraining the intermediate features of the student model to align with the teacher model, it is ensured that the student model learns the teacher's feature extraction paradigm, rather than merely imitating the surface output.

[0049] Through the above methods, the output distribution of the student model gradually approaches that of the teacher model until it meets the preset accuracy threshold. Furthermore, the student model inherits the generalization ability of the teacher model, requiring only minor adjustments for new process tasks without retraining. This adapts to the rapid iteration requirements of advanced processes and enhances the flexibility and adaptability of the model.

[0050] It should be understood that at least some steps in the flowcharts involved in the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but may be executed at different times. The execution order of these steps or stages is not necessarily sequential, but may be executed in turn or alternately with other steps or at least some of the steps or stages in other steps.

[0051] Based on the same inventive concept, this application also provides a photomask pattern optimization system 100, see reference. Figure 6 The photomask pattern optimization system 100 includes at least: The meta-learning training module 11 is used to train the teacher model on multiple process tasks to learn initial parameters that are sensitive to task distribution. The knowledge distillation module 12 is used to transfer the initial mask solution knowledge of the teacher model to the student model to generate a compressed student model; The iterative optimization module 13 is used to perform photomask optimization iteration starting from the initial mask solution output by the student model, and output the final mask pattern. Control module 14 is used to coordinate the above modules to execute the photomask pattern optimization method as described in any of the above items.

[0052] The solution to this problem is similar to the solution described in the above methods, and will not be repeated here.

[0053] In one embodiment, this application also provides a computer-readable storage medium storing program code, which is invoked by a processor to execute any of the above-described photomask pattern optimization methods.

[0054] In one embodiment, this application also provides an electronic device comprising one or more processors; a memory; and one or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the one or more processors, and the one or more applications are configured to perform the photomask patterning optimization method described above.

[0055] It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from the principles thereof, and such improvements and modifications also fall within the scope of protection of the claims of this application. It should be understood that certain features of this disclosure described in the context of individual embodiments for clarity may also be provided in combination in a single embodiment. Conversely, various features of this disclosure described in the context of individual embodiments for clarity may also be provided individually or in any suitable combination or as part of any other described embodiment of this disclosure.

Claims

1. A method for optimizing photomask patterns, characterized in that, The method includes: Obtain the initial layout corresponding to the target process layer and the environmental information of the initial layout within a preset neighborhood; The initial layout and environmental information are input into the teacher model trained by meta-learning, and the teacher model outputs the first initial mask solution. The knowledge of the first initial mask solution is transferred to the student model to generate a compressed student model; The second initial mask solution output by the compressed student model is used as the initial value for the photomask optimization iteration. Iterative calculations are performed to generate the final mask pattern.

2. The method of claim 1, wherein, The meta-learning training process of the teacher model includes: The teacher model is trained on multiple different process tasks using a meta-learning algorithm to determine initial parameters that are sensitive to task distribution.

3. The method of claim 2, wherein the step of optimizing the photomask pattern is performed by a computer program. The teacher model is trained on multiple different process tasks using a meta-learning algorithm to determine initial parameters sensitive to task distribution, including: M process tasks are randomly selected from the multiple different process tasks as the support set, and the remaining process tasks are used as the query set. For each process task, the current model parameters are fine-tuned using the support set to determine the specific parameters for that process task; The query set is used to calculate the loss function for the specific parameters of each process task, and the initial parameters are updated by gradient descent. Repeat the above steps until the loss function converges, and obtain the task distribution-sensitive initial parameters.

4. The method of claim 2, wherein the step of optimizing the photomask pattern is performed by a computer program. The step of inputting the initial layout and environmental information into the trained teacher model and having the teacher model output the first initial mask solution includes: inputting the initial layout and environmental information into the teacher model based on the task distribution-sensitive initial parameters, fixing the bottom feature extraction layer of the task distribution-sensitive initial parameters, fine-tuning the top decision layer until convergence, and outputting the first initial mask solution.

5. The photomask pattern optimization method according to claim 1, characterized in that, The step of transferring the initial mask solution knowledge of the teacher model to the student model includes: constructing a mask-level loss function, an edge gradient loss function, and a lithographic evaluation index loss function based on the soft target and intermediate features output by the teacher model; and jointly training the student model until the student model reproduces the initial mask solution distribution of the teacher model within a preset accuracy threshold.

6. The method of claim 1, wherein: The number of iterations in the iterative calculation is less than the set number of iterations.

7. The photomask pattern optimization method according to claim 1, characterized in that, The student model has a lower single-inference latency than the teacher model.

8. An optical mask pattern optimization system, characterized by, The system includes at least: The meta-learning training module is used to train the teacher model on multiple process tasks to learn initial parameters that are sensitive to task distribution. The knowledge distillation module is used to transfer the knowledge of the first initial mask solution of the teacher model to the student model, generating a compressed student model. The iterative optimization module is used to perform photomask optimization iterations starting from the second initial mask solution output by the student model, and output the final mask pattern. A control module is used to coordinate the above modules to execute the photomask pattern optimization method as described in any one of claims 1 to 7.

9. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by the processor, it implements the photomask pattern optimization method according to any one of claims 1 to 7.

10. An electronic device, comprising: Includes one or more processors; Memory; One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the one or more processors, the one or more applications being configured to perform the photomask graphics optimization method as described in any one of claims 1 to 7.

Images