SRAF graph generation method and system based on large language model

Abstract

The invention discloses an SRAF graph generation method and system based on a large language model, and the method comprises the steps: carrying out the preprocessing of multi-modal sample data collected in advance, converting the sample data into word vectors which can be recognized by the large language model, the sample data comprises an SRAF graph, a target layout pattern for generating the SRAF graph and a process description in a text form; inputting the preprocessed training data into a large language model for training to obtain an SRAF generation model capable of automatically generating an SRAF graph according to the target layout pattern and the process description; and inputting the preprocessed pattern and process description of the target layout into the SRAF generation model, and automatically generating an SRAF graph of the target layout. By adopting the technical method provided by the invention, the generation efficiency of the SRAF graph can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of semiconductor manufacturing, and in particular to a SRAF pattern generation method and system using large language models (LLM). BACKGROUND

[0002] Lithography is a crucial process in modern semiconductor manufacturing, used to accurately transfer designed patterns onto the surface of a semiconductor substrate. The process flow includes the following key steps. First, a layer of photoresist is uniformly coated on the surface of the substrate to be processed, forming a thin film sensitive to light. Then, the designed mask plate is installed in the lithography equipment and accurately aligned to ensure that the pattern can be projected to the target position. Next, through the exposure module of the lithography equipment, a light source of a specific wavelength is directed to the surface of the photoresist, and the light selectively transmits energy through the pattern area of the mask plate, forming a specific exposure area in the photoresist. After development, the exposed or unexposed parts of the photoresist (depending on the type of photoresist) are dissolved or solidified, forming a pattern structure on the substrate surface corresponding to the mask pattern. After development, further processing such as etching, ion implantation or material deposition can be performed to modify the substrate material. Finally, the remaining photoresist is removed, completing the lithography process. By repeating the above steps, complex micro-nano structures can be built layer by layer to achieve high-precision semiconductor device manufacturing.

[0003] However, as the process technology node advances to 5nm, 3nm and other advanced processes, lithography technology faces increasingly severe challenges. During the projection of the mask pattern to the wafer, due to the optical proximity effect (OPE), there is a significant deviation between the actual pattern formed and the target layout. In order to reduce this deviation, optical proximity correction (OPC) has become a standard process, and the generation of sub-resolution assist features (SRAF) is particularly critical. The design of SRAF optimizes the phase distribution of light energy transmission by introducing small auxiliary patterns (which are not directly imaged) to enhance the stability of the target pattern within the process window, reducing edge placement error (EPE) caused by focus deviation and exposure dose fluctuation.

[0004] In a standard mask optimization flow, the generation of SRAFs needs to be performed in coordination with OPC, mask rule check (MRC), and lithography compliance check (LCC) to ensure that the final mask meets the requirements of the process window and manufacturing rules. However, current SRAF generation techniques have limitations in several aspects. For example, rule-based methods rely on pre-defined rules to determine the position, size, and shape of SRAFs, which are efficient and suitable for simple design patterns, but perform poorly under complex designs and variable lithography conditions. Rule tables usually rely on the professional experience of engineers, which not only has a long development cycle and high maintenance cost, but also lacks flexibility and automation capabilities. Model-based methods use optical simulation or inverse lithography technology (ILT) to guide the generation of auxiliary patterns, which have significant advantages in precision and adaptability to complex designs. However, this method has a very high computational cost, especially in advanced process nodes, which consumes a lot of resources. In addition, the SRAF patterns generated by this method may have consistency problems under similar layout configurations, making it difficult to meet the high reliability requirements in mass production. ML-based methods predict the probability of the existence of SRAFs in target layouts by extracting features from historical designs, and then generate auxiliary patterns. Although machine learning has advantages in computational efficiency, its performance is highly dependent on the quality and uniformity of the training data. For different process layers or lithography conditions, the model usually needs to be retrained, limiting its versatility and flexibility. When faced with diverse design requirements, this method's performance is not mature. In summary, existing SRAF generation methods face significant limitations in efficiency, flexibility, consistency, and versatility. Especially in advanced processes, the complexity of mask design and process window requirements are further enhanced, and traditional methods are difficult to fully meet these demands. Therefore, it is urgent to explore new SRAF generation methods that combine artificial intelligence technology. SUMMARY

[0005] The purpose of the present application is to overcome the defects of the prior art, and to provide a large language model-based SRAF pattern generation method and system that can improve the generation efficiency of SRAF patterns.

[0006] In the embodiments of the present application, a large language model-based SRAF pattern generation method is provided, which includes:

[0007] Pre-process a pre-collected multi-modal sample data, convert it into a word vector recognizable by a large language model, and the sample data includes an SRAF pattern and a target layout pattern and a text form process description used to generate the SRAF pattern;

[0008] Input the pre-processed training data into the large language model for training to obtain an SRAF generation model capable of automatically generating an SRAF pattern according to a target layout pattern and a process description;

[0009] Input the pre-processed target layout pattern and process description into the SRAF generation model to automatically generate an SRAF pattern of the target layout.

[0010] In the embodiment of the application, the topological shape of the SRAF pattern is a rectangle, and the coordinates of the center point and the length and width are used to represent it.

[0011] In the embodiment of the application, the process description of the target layout includes design rules, exposure conditions and SRAF constraints, and the pre-processed training data further includes an optical intensity map, which is obtained according to the target layout pattern and the lithography model of the exposure conditions.

[0012] In the embodiment of the application, a visual encoder is used to convert the target layout pattern and the optical intensity map of the target layout pattern into a word vector, the visual encoder includes a visual conversion module, a visual-word vector alignment module and a projection module, the visual-word vector alignment module is stacked by a multi-layer self-attention mechanism module, a cross-attention module and a multi-layer perception forward propagation module, and the projection module is obtained by combining a multi-layer perception machine.

[0013] In the embodiment of the application, a visual encoder is used to convert the target layout pattern and the optical intensity map of the target layout pattern into a word vector, including:

[0014] Divide each input image into multiple blocks, and convert each block into an image feature representation using the visual conversion module;

[0015] Convert the image feature representation using the visual-word vector alignment module to obtain an aligned image feature representation;

[0016] Convert the dimension of the aligned image feature representation using the projection module to output a visual word vector.

[0017] In the embodiment of the application, the training data further includes a text instruction input.

[0018] In the embodiment of the application, an embedding layer is used to convert the text instruction input and the process description of the target layout into a word vector.

[0019] In the SRAF generation model training process in the embodiment of the application, a large language model is trained through a supervised fine-tuning method, the SRAF generated by the SRAF generation model according to target layout data and process description in sample data is compared with the corresponding SRAF in the sample data, and model parameters are updated according to the comparison result.

[0020] In the embodiment of the application, the SRAF pattern generation method based on a large language model further comprises:

[0021] A new SRAF pattern is automatically generated by taking a new target layout and process description as input, the generated SRAF pattern is simulated using a lithography simulation model, the matching degree of the generated mask pattern with the target layout pattern in the lithography process is verified, and the SRAF generation model parameters are fine-tuned according to the matching degree.

[0022] In the embodiment of the application, a SRAF pattern generation system based on a large language model is also provided, which automatically generates SRAF patterns by using the SRAF pattern generation method based on a large language model described above.

[0023] Compared with the prior art, the SRAF pattern generation method and system based on a large language model of the application realize automatic optimization of masks by combining optical simulation results with the multi-modal reasoning ability of a large language model, ensure that the pattern formed after lithography simulation is highly consistent with the target layout, and meet the stability requirements within the process window; the powerful reasoning ability of a large language model significantly reduces the calculation time and resource consumption in the SRAF generation process, greatly reduces the involvement of artificial labor, and improves the overall optimization efficiency; through large-scale data training, it has good adaptability and can cover various target layouts and process nodes, realizing cross-layer and cross-scene generalization design; supports multi-modal input, including graphic data, text description and lithography conditions, and under large-scale data training, the model can cope with different design complexity and variable lithography process environment. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 is a flowchart of the SRAF pattern generation method based on a large language model of the embodiment of the application.

[0025] Figure 2 is a schematic diagram of SRAF pattern annotation of the embodiment of the application.

[0026] Figure 3 is a structural schematic diagram of the visual feature encoder VPG of the embodiment of the application.

[0027] Figure 4 is a schematic diagram of process description text of the embodiment of the application.

[0028] Figure 5 is a schematic diagram of system instruction input of an embodiment of the present application.

[0029] Figure 6 is a schematic diagram of multi-modal data input to a large language model of an embodiment of the present application.

[0030] Figure 7 is a schematic diagram of large model supervised fine-tuning of an embodiment of the present application. DETAILED DESCRIPTION

[0031] As Figure 1 shown, in an embodiment of the present application, a SRAF pattern generation method based on a large language model is provided, which includes steps S1-S5. The following will be described respectively.

[0032] Step S1, collect training sample data.

[0033] Use a lithography tool to perform SRAF design on a series of target layouts, to generate high-quality SRAF pattern results as target samples of training data. In order to simplify the difficulty of model learning, the topological shape of the SRAF pattern is limited to a rectangular representation, and the coordinates of the center point and the length and width are used in the sample, as Figure 2 shown. The sample data is multi-modal data, which includes SRAF patterns and target layout patterns and process descriptions for generating the auxiliary patterns. By using these data, a real and diverse lithography environment can be provided for the training model, so that the generated SRAF can be applicable in different situations.

[0034] Step S2: pre-process the pre-collected multi-modal sample data, and convert it into a word vector that can be recognized by a large language model.

[0035] It should be noted that since the collected sample data is multi-modal data including graphics and text, different processing is required for different modal data. The encoding of the graphic data is to convert the graphic data of the target layout into a high-dimensional feature representation, thereby providing accurate information for further reasoning of the large language model. As Figure 3As shown, in the embodiment of the present application, the target layout pattern and the light intensity map are converted into word vectors by using a visual encoder, the visual encoder includes a vision conversion module (Vision Prompt Transformer, VIT), a visual-word vector alignment module, and a projection module (Project), the visual-word vector alignment module is stacked by a multi-layer self-attention mechanism module (Self Attention), a cross-attention module (Cross Attention), and a multi-layer perception forward propagation module (Feed Forward), and the projection module is obtained by combining a multi-layer perception machine.

[0036] In the embodiment of the present application, the target layout pattern and the light intensity map of the target layout pattern are converted into word vectors by using a visual encoder, which includes:

[0037] Each input image is divided into multiple blocks (Patches), and each block is converted into an image feature representation by using the vision conversion module;

[0038] The image feature representation is converted by using the visual-word vector alignment module to obtain an aligned image feature representation;

[0039] The aligned image feature representation is dimensionally converted by using the projection module to output a visual word vector.

[0040] In addition, the process description of the target layout (including design rules, exposure conditions, etc.) needs to be represented in text form, and these process descriptions usually include design rules (such as minimum spacing, minimum width, etc.), exposure conditions (such as exposure dose, focal length, etc.), and SRAF constraints (layout constraints, spacing constraints, shape constraints). These process descriptions need to be converted into text format and provided as part of the input to the large language model, and the input format can be as shown. Figure 3 The text form of the process description can be converted into a word vector by using an embedding layer (Embedding Layer) configured according to the large language model. Further, in order to make the large model interactive, in the embodiment of the present application, the training data further includes a text instruction input, which is converted into a word vector by using an embedding layer.

[0041] In addition, it should be noted that the preprocessed training data also includes a light intensity map, which is obtained by a photolithography model that meets the exposure conditions in the process conditions according to the target layout pattern. In semiconductor manufacturing, a photolithography model is used to simulate the optical effects in the photolithography process. The light intensity map is an important output of the photolithography model, which shows the distribution of light on the photoresist layer. The generation process of the light intensity map usually includes the following steps:

[0042] Target pattern provision: First, the pattern of the target layout, which is the final pattern formed on the wafer, needs to be provided;

[0043] Lithography system parameter setting: According to the actual lithography system, parameters such as light source, mask, lens, etc. are set, which will affect the propagation and distribution of light;

[0044] Light intensity calculation: Using the lithography model, the intensity distribution of light on the photoresist layer is calculated;

[0045] Light intensity map generation: According to the calculation results, a light intensity map is generated. The light intensity map shows the distribution of light on the photoresist layer, usually represented by a grayscale or color map.

[0046] Through the above method, the large language model can process various modal data by fusing different modal inputs, and can learn task-related semantic information from different information, helping the model to get a better optical proximity effect correction method.

[0047] Step S3: input the preprocessed training data into the large language model for training to obtain an SRAF generation model that can automatically generate SRAF patterns according to target layout patterns and process conditions.

[0048] First, select a pre-trained large language model as the base model, such as the GPT series, etc. Most large language models are currently based on the Transformer architecture, such as Figure 6 , which uses a multi-head attention mechanism as the main feature extraction module, combined with other nonlinear transformation modules for multi-layer stacking. These pre-trained models have learned language processing and reasoning capabilities on large-scale datasets and can serve as the basis for the SRAF pattern generation task. Then, as shown in Figure 7 , during the training of the SRAF generation model, the supervised fine-tuning (SFT) method is used to train the large language model with Ground Truth sample data as a reference. In this process, the input target layout data and process description will generate the model's prediction results, which are then compared with the actual Ground Truth results to update the model parameters to ensure that the generated results meet the design requirements.

[0049] Step S4: experimental verification, input a new target layout and process description to automatically generate a new SRAF pattern, use a lithography simulation model to simulate the generated SRAF, verify the matching degree of the generated mask pattern with the target layout pattern in the lithography process, and fine-tune the parameters of the SRAF generation model according to the matching degree.

[0050] After the model training is completed, enter the SRAF generation and verification phase. Use the trained LLM to input the new target layout into the model to generate the corresponding SRAF result. At this time, the target layout and process description are input, and the model will generate a new SRAF pattern based on the knowledge trained before. Then, the generated SRAF is simulated using the lithography simulation model to verify the matching degree of the generated mask pattern with the target layout in the lithography process. Through simulation, it can be checked whether the generated SRAF pattern meets the design requirements and predicts its effect in actual manufacturing. If the error between the generated SRAF pattern and the target layout exceeds the preset threshold, the system will automatically feedback, adjust the input data or fine-tune the model until the error reaches an acceptable level.

[0051] Step S5: SRAF pattern generation output, input the preprocessed pattern of the target layout and the process description into the SRAF generation model to automatically generate the SRAF pattern of the target layout. The generated SRAF pattern can be output in a format suitable for subsequent processing, and includes simulation verification results for quality checking.

[0052] Further, the embodiment of the present application also provides a SRAF pattern generation system based on a large language model, which automatically generates SRAF patterns by using the above-mentioned SRAF pattern generation method based on a large language model.

[0053] In summary, by using the SRAF pattern generation method and system based on a large language model of the present application, by combining the optical simulation results with the multi-modal reasoning ability of the large language model, the automatic optimization of the mask is realized, ensuring that the pattern formed after lithography simulation is highly consistent with the target layout, while meeting the stability requirements within the process window; by using the powerful reasoning ability of the large language model, the calculation time and resource consumption in the SRAF generation process are significantly reduced, the manual participation is greatly reduced, and the overall optimization efficiency is improved; through large-scale data training, good adaptability is achieved, which can cover multiple target layouts and process nodes, realizing cross-layer and cross-scene generalization design; multi-modal input is supported, including graphic data, text description and lithography conditions, and under large-scale data training, the model can cope with different design complexity and variable lithography process environment.

[0054] The above only describes the preferred embodiments of the present application and should not be used to limit the present application. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A method for generating SRAF patterns based on a large language model, characterized in that, The method comprises: preprocessing a pre-collected multi-modal sample data to convert it into a word vector recognizable by a large language model, the sample data including an SRAF pattern and a target layout pattern and a text form process description used to generate the SRAF pattern; inputting the preprocessed training data into the large language model for training to obtain an SRAF generation model capable of automatically generating an SRAF pattern according to a target layout pattern and a process description; inputting the preprocessed target layout pattern and process description into the SRAF generation model to automatically generate an SRAF pattern of the target layout.

2. The SRAF pattern generation method based on a large language model of claim 1, wherein, The SRAF pattern has a rectangular topological shape, and is represented by a center point coordinate and a length and a width.

3. The SRAF pattern generation method based on a large language model of claim 1, wherein, The process description of the target layout includes design rules, exposure conditions and SRAF constraints, and the preprocessed training data further includes an optical intensity map obtained according to the target layout pattern by a photolithography model under the exposure conditions.

4. The SRAF pattern generation method based on a large language model of claim 3, wherein, The target layout pattern and the optical intensity map of the target layout pattern are converted into word vectors by a visual encoder, the visual encoder including a visual conversion module, a visual-word vector alignment module and a projection module, the visual-word vector alignment module being stacked by a multi-layer self-attention mechanism module, a cross-attention module and a multi-layer perception forward propagation module, and the projection module being obtained by a multi-layer perception machine.

5. The SRAF pattern generation method based on a large language model of claim 4, wherein, The target layout pattern and the optical intensity map of the target layout pattern are converted into word vectors by a visual encoder, comprising: dividing each input image into multiple blocks, and converting each block into an image feature representation by the visual conversion module; converting the image feature representation by the visual-word vector alignment module to obtain an aligned image feature representation; performing dimension conversion on the aligned image feature representation by the projection module to output a visual word vector.

6. The SRAF pattern generation method based on a large language model of claim 1, wherein, The training data further includes a text instruction input.

7. The SRAF pattern generation method based on a large language model of claim 6, wherein, The text instruction input and the process description of the target layout are converted into word vectors by an embedding layer configured corresponding to the large language model.

8. The SRAF pattern generation method based on a large language model of claim 1, wherein, During the training of the SRAF generation model, the large language model is trained by a supervised fine-tuning method, the SRAF generation model is compared with the corresponding SRAF pattern in the sample data according to the SRAF pattern generated by the target layout data and the process description in the sample data, and the model parameters are updated according to the comparison result.

9. The SRAF pattern generation method based on a large language model of claim 1, wherein, Further comprising: inputting a new target layout and process description as input to automatically generate a new SRAF pattern, simulating the generated SRAF pattern by a photolithography simulation model, verifying the matching degree of the generated mask pattern with the target layout pattern in the photolithography process, and fine-tuning the parameters of the SRAF generation model according to the matching degree.

10. A system for generating SRAF patterns based on a large language model, comprising: It automatically generates an SRAF pattern by the large language model-based SRAF pattern generation method according to any one of claims 1-9.

Images