Optical proximity effect correction method and system based on diffusion strategy
By adopting an optical proximity effect correction method based on the diffusion Transformer model, the flexibility and efficiency issues of optical proximity effect correction under high precision and multi-constraint conditions in the prior art are solved, and efficient optimization and robustness improvement of lithography patterns are achieved.
Patent Information
- Application Number
- CN202511228924.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-29
- Publication Date
- 2025-11-21
AI Technical Summary
Existing optical proximity effect correction techniques struggle to achieve a comprehensive breakthrough in flexibility and efficiency under high precision and multiple constraints. In particular, they suffer from high computational costs and strong data dependence in complex process nodes and large-scale tasks. Furthermore, multi-objective optimization cannot simultaneously satisfy imaging quality, design rule checking, and process window robustness.
An optical proximity effect correction method based on the diffusion Transformer model is adopted. By preprocessing and iteratively training multimodal sample data, a diffusion strategy is used to generate mask patterns. Combined with multi-scale pyramid representation and visual encoder, the imaging accuracy, mask complexity and process window robustness of the lithography pattern are optimized.
It improves the precision and robustness of photolithography patterns, reduces computational costs, enhances the convergence speed and processing power of models, adapts to large-scale complex tasks, and achieves efficient optimization of multiple objectives.
Smart Images
Figure CN120997343A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of semiconductor manufacturing, and more specifically to a method and system for correcting the optical proximity effect based on a diffusion strategy. Background Technology
[0002] With the continuous advancement of semiconductor manufacturing technology, chip dimensions have entered the deep submicron and even nanometer scales. At this point, the Optical Proximity Effect (OPE) has become one of the main bottlenecks limiting lithography accuracy and manufacturing yield. The OPE primarily stems from the diffraction characteristics of the optical system, the complexity of the mask pattern, and the non-ideals of the exposure process, leading to a significant deviation between the actual pattern on the silicon wafer and the design target. This deviation is particularly pronounced in designs with extremely small feature sizes or complex geometries, severely impacting chip performance, power consumption, and manufacturing yield.
[0003] To address the optical proximity effect problem, Optical Proximity Correction (OPC) technology has become a core component of photolithography. Existing OPC methods can be broadly categorized as follows: **Rule-based methods:** These methods correct the design pattern by predefining a series of geometric rules, such as adding auxiliary lines or adjusting edge shapes. These methods are computationally efficient, easy to implement in engineering, and suitable for simple designs and larger process nodes. However, their correction effectiveness relies on empirical knowledge, and the flexibility of the rules is limited, making it difficult to adapt to the complex geometries and higher precision requirements of emerging process nodes. **Iterative optimization methods:** These methods repeatedly calculate deviations in the target pattern and gradually adjust the design until specific convergence conditions are met. These methods typically rely on optimization algorithms (such as gradient descent) and loss function design, achieving high correction accuracy. However, iterative optimization is computationally expensive, especially when handling industrial-scale tasks, where efficiency is insufficient for practical needs. Furthermore, these methods require accurate modeling of complex photolithography processes, and model errors may affect the reliability of the final results. **Machine learning-based methods:** With the rise of artificial intelligence, machine learning-based OPC methods have received widespread attention. By training neural network models, machine learning methods can quickly generate corrected results and alleviate the computational bottleneck of traditional methods to some extent. In particular, deep learning models can handle a large number of complex geometric structures and quickly optimize masks.
[0004] However, the practical application of such methods still has limitations: (1) Strong data dependence: Deep learning models require a large amount of high-quality training data, and the acquisition cost of this data is high; (2) Single optimization objective: Existing models can usually only optimize imaging accuracy, and it is difficult to consider multiple objectives such as design rule checking (MRC), mask complexity and process window robustness at the same time; (3) OPC optimization is non-convex optimization, that is, it is difficult to obtain the optimal solution; at the same time, there is a multi-modal problem of solutions, that is, different solutions yield the same index.
[0005] In industrial practice, optical proximity correction needs to simultaneously meet multiple constraints, such as high-precision imaging quality, stringent design rule checks (MRC), and stable robustness within the process window. Furthermore, the complexity of mask design and manufacturing costs must also be optimized in actual production. These multi-objective contradictions become increasingly prominent as design scale increases and process nodes shrink, making it difficult for existing technologies to achieve a comprehensive breakthrough in accuracy, efficiency, and flexibility. Summary of the Invention
[0006] The purpose of this invention is to provide an optical proximity effect correction method and system based on a diffusion strategy, thereby improving the accuracy and process robustness of photolithography patterns.
[0007] In this embodiment of the invention, an optical proximity effect correction method based on a diffusion strategy is provided, characterized by comprising:
[0008] The pre-collected multimodal sample data is preprocessed and converted into word vectors that can be recognized by the large model. The sample data includes multimodal conditional data and OPC-corrected mask patterns. The multimodal conditional data includes an initial mask pattern in graphic format, an initial imaging error distribution map obtained from optical simulation, and process parameters and target design rule constraints in text format.
[0009] The preprocessed multimodal sample data is sequentially input into the diffusion model for iterative training;
[0010] The current multimodal conditional data is input into the trained diffusion model to obtain the OPC-corrected mask pattern.
[0011] In this embodiment of the invention, the mask pattern is represented by dividing points in a clockwise direction, starting from the upper left corner, i.e., [(X1,Y1,L1),(X2,Y2,L2),(X3,Y3,L3),...,(X n ,Y n ,L n ]], where n is the number of segments, X and Y are the center points of each segment, and L represents the length of each segment.
[0012] In this embodiment of the invention, the diffusion model is a Transformer-based diffusion model.
[0013] In this embodiment of the invention, the diffusion model includes multiple diffusion Transformer units, which are stacked together by a Self Attention module, a Cross Attention module, and a Feed Forward module.
[0014] In this embodiment of the invention, a visual encoder is used to convert the initial mask pattern, the OPC-corrected mask pattern, and the initial imaging error distribution map into word vectors.
[0015] In this embodiment of the invention, an embedding layer is used to convert process parameters and target design rule constraints into word vectors.
[0016] In this embodiment of the invention, the optical proximity effect correction method based on a diffusion strategy further includes:
[0017] For image-formatted data, a multi-scale pyramid representation is constructed. The visual encoder encodes the features of the images at different scales and then fuses the encoded information from each scale to output word vectors.
[0018] In this embodiment of the invention, the optical proximity effect correction method based on a diffusion strategy further includes:
[0019] The OPC mask pattern generated by the diffusion model is corrected to correct boundaries that do not conform to manufacturing rules;
[0020] The imaging effect and design rule compliance of the modified OPC mask pattern were verified using industry-standard photolithography simulation tools.
[0021] In this embodiment of the invention, an optical proximity effect correction system based on a diffusion strategy is also provided, which automatically generates an OPC-corrected mask pattern using the aforementioned optical proximity effect correction method based on a diffusion strategy.
[0022] Compared with existing technologies, the optical proximity effect correction method and system based on diffusion strategy of the present invention, by introducing a diffusion model, generates patterns with inherent diversity, which can effectively solve the multi-mode problem of solutions, thereby accelerating model convergence and reducing learning difficulty. It can achieve efficient optimization among multiple objectives, including the imaging accuracy of lithographic patterns, mask complexity, and robustness of process windows. Through the stepwise generation mechanism of diffusion strategy, different indicators can be optimized by learning from a large amount of data. By introducing the generation process of diffusion strategy, computational efficiency can be greatly improved, especially when facing large-scale data and complex tasks, showing high processing power and lower computational cost. The model inference process based on diffusion strategy can be parallelized, adapted to high-performance hardware, and significantly improves overall robustness and efficiency. Attached Figure Description
[0023] Figure 1 This is a flowchart of an optical proximity effect correction method based on a diffusion strategy according to an embodiment of the present invention.
[0024] Figure 2 This is a schematic diagram of mask segmentation representation according to an embodiment of the present invention.
[0025] Figure 3 This is a schematic diagram of word vector representation extraction from visual images according to an embodiment of the present invention.
[0026] Figure 4 This is a schematic diagram of the extraction of process parameter terms according to an embodiment of the present invention.
[0027] Figure 5 This is a schematic diagram of the architecture of the diffusion Transformer model according to an embodiment of the present invention.
[0028] Figure 6 This is a schematic diagram of the diffusion model iteration process according to an embodiment of the present invention.
[0029] Figure 7 This is a schematic diagram of multi-scale visual word vector expression feature extraction according to an embodiment of the present invention. Detailed Implementation
[0030] like Figure 1 As shown in the figure, an optical proximity effect correction method based on a diffusion strategy is provided in this embodiment of the invention, which includes steps S1-S5. These are described below.
[0031] Step S1: Collect training sample data.
[0032] Commercial lithography tools (such as Calibre) are used to perform OPC processing on a series of target layouts to obtain boundary segmentation schemes and generate high-quality OPC-corrected mask patterns, which serve as sample data for training. The sample data includes multimodal conditional data and OPC-corrected mask patterns. The multimodal conditional data includes initial mask patterns in graphic format, initial imaging error distribution maps obtained from optical simulations, and process parameters and target design rule constraints in text format. This data covers data collected under different process nodes, exposure parameters, and various combinations of target layouts. Ensuring data diversity and the generalization ability of the training model, this data provides a realistic and diverse lithography environment for the training model, enabling its applicability in different situations.
[0033] To simplify the model learning process, the mask pattern is represented by segmentation lines starting from the top left corner and proceeding clockwise, i.e., [(X1,Y1,L1),(X2,Y2,L2),(X3,Y3,L3),...,(X n ,Y n ,L n ]], where n is the number of divisions, X and Y are the center points of each division segment, and L represents the length of each division segment. Figure 2 As shown.
[0034] Step S2: Preprocess the pre-collected multimodal sample data and convert it into word vectors that can be recognized by the large model.
[0035] It should be noted that since the collected sample data is multimodal data including both graphics and text, different processing methods are required for different modalities. Encoding the graphics data involves transforming the graphic data of the target layout into a high-dimensional feature representation, thereby providing accurate information for further inference in large models. For example... Figure 3 As shown, in this embodiment of the invention, a visual encoder is used to convert the initial mask pattern, the OPC-corrected mask pattern, and the initial imaging error distribution map into word vectors.
[0036] like Figure 4 As shown, process parameter data and target design rule constraints are typically textual data. To represent these parameters as a unified "word representation," an embedding layer is used to extract features from this textual data.
[0037] Step S3: Input the preprocessed multimodal sample data into the diffusion model for iterative training.
[0038] It should be noted that the diffusion model can adopt the commonly used diffusion-based Transformer model, such as... Figure 5 As shown, it includes multiple DiT Blocks, each composed of a stacked Self Attention, Cross Attention, and FeedForward modules. Furthermore, to effectively utilize the advantages of large-scale data, open-source model weights can be used for initialization. The model as a whole adopts a multimodal fusion structure, which can combine visual encoders and embedding layer configurations to achieve comprehensive modeling of optical effects and process constraints. Through this fusion of multimodal information, the diffusion model can optimize patterns from different dimensions.
[0039] like Figure 6 As shown, the model's training process and the diffusion of input samples are illustrated. In each diffusion step, the denoising network performs the next optimization based on the mask pattern A obtained from the previous optimization path and the current multimodal input data O. Each time the model optimizes, it uses the mask pattern A obtained from the previous optimization path as input. Each diffusion step fine-tunes the optimization path and guides the generated OPC optimization pattern towards the target direction based on the loss function. This stepwise generation mechanism of the diffusion strategy enables the model to explore diverse paths during the optimization process, ensuring the quality of the final optimization result.
[0040] Step S4: Input the current multimodal conditional data into the trained diffusion model to obtain the OPC-corrected mask pattern.
[0041] After training, the diffusion model can be fed with the current multimodal conditional data, and after one iteration, the OPC-corrected mask pattern can be obtained.
[0042] Step S5: Post-processing and Verification. After diffusion optimization, post-processing is performed on the generated mask pattern to ensure that the final pattern meets actual production requirements and conforms to industrial application standards. The specific process includes:
[0043] The OPC mask pattern generated by the diffusion model is corrected to correct boundaries that do not conform to manufacturing rules;
[0044] The imaging effect and design rule compliance of the modified OPC mask pattern were verified using industry-standard photolithography simulation tools.
[0045] Furthermore, in this embodiment of the invention, to further improve the global robustness and detail fidelity of the correction results, the pattern processing is optimized by establishing a multi-scale joint optimization mechanism. Specifically, as... Figure 7As shown, the visual encoder constructs a multi-scale pyramid representation of the pattern, fusing pattern feature encoding information at different scales to generate word vectors that integrate pattern feature encoding information from different scales. These word vectors are then used as input to the model for more accurate control point updates. This mechanism can simultaneously consider the global structure and local details of the pattern, improving optimization accuracy and effectively balancing optimization needs across different scales. Multi-scale optimization not only enhances the model's adaptability to features of different sizes but also improves the consistency between global and local optimization.
[0046] Furthermore, in this embodiment of the invention, an optical proximity effect correction system based on a diffusion strategy is also provided, which automatically generates an OPC-corrected mask pattern using the aforementioned optical proximity effect correction method based on a diffusion strategy.
[0047] In summary, the optical proximity effect correction method and system based on diffusion strategy of this invention, by introducing a diffusion model, allows the diffusion-generated patterns to naturally possess diversity, effectively solving the multi-mode problem of solutions, thereby accelerating model convergence and reducing learning difficulty. It enables efficient optimization across multiple objectives, including the imaging accuracy of lithographic patterns, mask complexity, and robustness of the process window. Through the stepwise generation mechanism of the diffusion strategy, it can learn from large amounts of data and optimize different indicators. The introduction of the diffusion strategy generation process significantly improves computational efficiency, especially when facing large-scale data and complex tasks, exhibiting high processing power and lower computational cost. The model inference process based on the diffusion strategy can be parallelized, adapting to high-performance hardware, and significantly improving overall robustness and efficiency.
[0048] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for correcting the optical proximity effect based on a diffusion strategy, characterized in that, include: The pre-collected multimodal sample data is preprocessed and converted into word vectors that can be recognized by the large model. The sample data includes multimodal conditional data and OPC-corrected mask patterns. The multimodal conditional data includes an initial mask pattern in graphic format, an initial imaging error distribution map obtained from optical simulation, and process parameters and target design rule constraints in text format. The preprocessed multimodal sample data is sequentially input into the diffusion model for iterative training; The current multimodal conditional data is input into the trained diffusion model to obtain the OPC-corrected mask pattern.
2. The optical proximity effect correction method based on diffusion strategy as described in claim 1, characterized in that, The mask pattern is represented by dividing points clockwise from the top left corner, i.e., [(X1,Y1,L1),(X2,Y2,L2),(X3,Y3,L3),...,(X n ,Y n ,L n ]], where n is the number of segments, X and Y are the center points of each segment, and L represents the length of each segment.
3. The optical proximity effect correction method based on diffusion strategy as described in claim 1, characterized in that, The diffusion model used is a Transformer-based diffusion model.
4. The optical proximity effect correction method based on diffusion strategy as described in claim 3, characterized in that, The diffusion model includes multiple diffusion Transformer units, which are stacked together by an attention mechanism module, a cross-attention module, and a forward propagation module.
5. The optical proximity effect correction method based on diffusion strategy as described in claim 1, characterized in that, A visual encoder is used to convert the initial mask pattern, the OPC-corrected mask pattern, and the initial imaging error distribution map into word vectors.
6. The optical proximity effect correction method based on diffusion strategy as described in claim 1, characterized in that, An embedding layer is used to convert process parameters and target design rule constraints into word vectors.
7. The optical proximity effect correction method based on diffusion strategy as described in claim 1, characterized in that, Also includes: For image-formatted data, a multi-scale pyramid representation is constructed. The visual encoder encodes the features of the images at different scales and then fuses the encoded information from each scale to output word vectors.
8. The optical proximity effect correction method based on diffusion strategy as described in claim 1, characterized in that, Also includes: The OPC mask pattern generated by the diffusion model is corrected to correct boundaries that do not conform to manufacturing rules; The imaging effect and design rule compliance of the modified OPC mask pattern were verified using industry-standard photolithography simulation tools.
9. An optical proximity effect correction system based on a diffusion strategy, characterized in that, It automatically generates OPC-corrected mask patterns using the optical proximity effect correction method based on diffusion strategy as described in any one of claims 1-8.