Integrated Circuit Layout Verification Method, Apparatus, and Equipment Based on Graph-Text Comparison Learning

By employing a dual-tower neural network architecture based on image-text comparison learning, natural language-driven verification of integrated circuit layouts is achieved, solving the problems of high cost and high adaptation cost in traditional methods, and realizing fast and low-threshold layout verification.

CN123070044APending Publication Date: 2026-11-10WUHAN XINSAIWEISI TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611270267.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-20
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Traditional integrated circuit layout verification methods incur high costs for rule writing and maintenance, as well as high adaptation costs during process node migration. Furthermore, machine learning-based models require a large amount of labeled data for retraining, making it impossible to achieve design rule-driven verification in the form of natural language.

Method used

A graph-text comparison learning approach is adopted to construct a dual-tower neural network architecture. The visual tower and the linguistic tower convert the natural language description of the integrated circuit layout and design rules into a shared high-dimensional feature vector. Cosine similarity calculation is used to identify high-risk areas, thereby realizing zero-sample layout verification driven by natural language.

Benefits of technology

It significantly reduces verification costs and technical barriers, enables rapid rule adaptation and efficient layout verification, and allows for seamless migration between different process nodes.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123070044A_ABST
    Figure CN123070044A_ABST
Patent Text Reader

Abstract

This invention discloses an integrated circuit layout verification method, apparatus, and device based on image-text comparison learning, relating to the field of integrated circuit technology. The method includes constructing a dual-tower neural network architecture comprising a visual tower and a linguistic tower; acquiring local integrated circuit layout slices and corresponding textual diagnostic reports from historical tape-out data to form a training set for training the dual-tower neural network architecture; inputting natural language descriptions of design rules into the linguistic tower to obtain text feature vectors; simultaneously, the visual tower encodes and extracts the target integrated circuit layout through a sliding window, obtaining the target integrated circuit layout feature vector corresponding to each window and its coordinate position on the target integrated circuit layout; calculating the cosine similarity between the text feature vectors and the feature vectors of each target integrated circuit layout; and determining high-risk areas based on the cosine similarity calculation results to achieve integrated circuit layout verification. This application enables effective verification of integrated circuit layouts.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of integrated circuit technology, specifically to an integrated circuit layout verification method, apparatus, and equipment based on image-text comparison learning. Background Technology

[0002] Tape-out is the final node in the chip design cycle. Before the actual physical tape-out, physical verification is a crucial step in the wafer fabrication process. As integrated circuit process nodes evolve towards 5nm and below, the complexity of Design Rule Check (DRC) increases exponentially. Traditional DRC physical verification mainly suffers from the following problems: (1) High cost of rule writing and maintenance: Every time a new process node or a new design rule is introduced, it is necessary to rely on domain experts to manually rewrite extremely complex geometric calculation code (such as SVRF, TCL language). (2) Limitations of machine learning models: In recent years, hot spot detection models based on convolutional neural networks (CNN) or graph neural networks (GNN) introduced into the field of EDA (Electronic Design Automation) are discriminative models that rely heavily on a large amount of specific type of label data. When the design team migrates to a new process node, a large amount of data must be collected again and the model must be trained from scratch, resulting in extremely high adaptation costs for AI-based EDA tools during process migration.

[0003] It is evident that neither rule-hard-coding schemes based on dedicated verification languages ​​nor discriminative AI hotspot detection schemes based on CNNs and GNNs can directly drive integrated circuit layout verification with design rules in natural language form. Summary of the Invention

[0004] This application provides a method, apparatus, and device for verifying integrated circuit layouts based on image-text comparison learning, which can achieve effective verification of integrated circuit layouts.

[0005] In a first aspect, embodiments of this application provide an integrated circuit layout verification method based on image-text comparison learning, the integrated circuit layout verification method based on image-text comparison learning including: A dual-tower neural network architecture comprising a visual tower and a linguistic tower is constructed, and local integrated circuit layout slices and corresponding text diagnostic reports from historical tape-out data are obtained to form a training set for training the dual-tower neural network architecture. The input design rules are described in natural language and converted into a language tower to obtain text feature vectors. At the same time, the visual tower encodes and extracts the target integrated circuit layout through a sliding window to obtain the target integrated circuit layout feature vector and the coordinate position on the target integrated circuit layout corresponding to each window. Calculate the cosine similarity between the text feature vector and the feature vector of each target integrated circuit layout, and determine high-risk areas based on the cosine similarity calculation results to achieve integrated circuit layout verification; The visual tower is used to convert the segmented physical map image or polygon topology into a high-dimensional map feature vector through a visual Transformer, and the language tower is used to convert the design rule intent described in natural language into a text feature vector through a large language model.

[0006] In conjunction with the first aspect, in one implementation method, The visual tower includes an image segmentation and linear projection module, a position encoding module, a multi-head self-attention encoding layer, a classification label aggregation module, and a visual projection head; The image segmentation and linear projection module is used to segment the input integrated circuit layout into image blocks of fixed size and flatten them into a vector sequence, which is then mapped into a multidimensional embedded vector through linear projection. The location encoding module is used to embed overlay location information for each image block; The multi-head self-attention coding layer is composed of multiple stacked Transformer encoders and is used to capture the global dependencies between image patches through a multi-head self-attention mechanism. The classification tag aggregation module is used to add CLS tags to the head of the vector sequence, and after encoding, the output of the CLS tags is used as the aggregated representation of the entire integrated circuit layout slice; The visual projection head is composed of a multilayer perceptron (MLP) and is used to map the aggregated representation to a shared high-dimensional latent space and perform L2 normalization, thereby outputting an integrated circuit layout feature vector.

[0007] In conjunction with the first aspect, in one implementation method, The language tower includes a word segmenter, a word embedding layer, a Transformer encoding layer, a sentence vector extraction module, and a text projection head; The word segmenter is used to divide the text describing the design rules in natural language into a sequence of sub-words. The word embedding layer is used to map subwords into dense vectors and superimpose positional encodings; The Transformer encoding layer is used to perform semantic encoding on the text describing the natural language of the design rules based on a multi-layer self-attention structure of a large language model. The sentence vector extraction module is used to extract the output of the CLS marker position or to perform mean pooling on the word sequence to obtain a sentence-level semantic representation. The text projection head is composed of a multilayer perceptron (MLP) and is used to map sentence-level semantic representations to the same shared high-dimensional latent space as the visual tower and perform L2 normalization to output text feature vectors.

[0008] In conjunction with the first aspect, in one implementation method, The training set consists of positive sample pairs composed of local integrated circuit layout slices from historical tape-out data and corresponding text diagnostic reports; The text diagnostic report refers to the inspection result file output after performing design rule checks on the integrated circuit layout using physical verification tools. It includes the violation rule number, violation type, integrated circuit layout coordinates of the violation location, description of the violation geometry, and repair suggestions.

[0009] In conjunction with the first aspect, in one implementation method, During the training phase of the dual-tower neural network architecture, the InfoNCE contrastive loss function is applied. The InfoNCE contrastive loss function is specifically as follows:

[0010] in, Indicates the total loss. This represents the contrast loss from the visual to the language direction. This represents the contrast loss from language to visual direction.

[0011] In conjunction with the first aspect, in one implementation, the contrast loss from visual to language direction is specifically as follows:

[0012] in, This indicates the number of local integrated circuit layout slices and text diagnostic report sample pairs in a single training batch. Indicates the first Feature vectors of local integrated circuit layout slices Indicates and The first match The text feature vector of a text diagnostic report, Indicates the first The text feature vector of a text diagnostic report, Represents the cosine similarity function. This represents the adjustment coefficient.

[0013] In conjunction with the first aspect, in one implementation, the input design rules are described in natural language and converted into a language tower to obtain a text feature vector. Simultaneously, the visual tower encodes and extracts the target integrated circuit layout through a sliding window, obtaining the target integrated circuit layout feature vector corresponding to each window and its coordinate position on the target integrated circuit layout. Specifically, this includes: When performing integrated circuit layout verification, the design rules are input into a language pyramid in natural language description, which is then converted to obtain a text feature vector. A fixed-size window slides across the target integrated circuit layout of the entire chip at a set step size. A local integrated circuit layout image is captured for each window, normalized, and then sent to the vision tower. For each captured local integrated circuit layout image, image block segmentation, position encoding overlay, multi-head self-attention encoding, and projection mapping are performed sequentially to obtain the target integrated circuit layout feature vector corresponding to each window, as well as the coordinate position of each window on the target integrated circuit layout.

[0014] In conjunction with the first aspect, in one implementation, the calculation of the cosine similarity between the text feature vector and the feature vectors of each target integrated circuit layout, and the determination of high-risk areas based on the cosine similarity calculation results, to achieve integrated circuit layout verification, specifically includes: The cosine similarity between the text feature vector and the feature vector of each target integrated circuit layout is calculated sequentially, specifically as follows:

[0015] in, Represents the cosine similarity function. express and Cosine similarity between them Represents the text feature vector. This represents the feature vector of the current target integrated circuit layout. This represents the vector dot product operation. The L2 norm of a vector; Arrange the cosine similarity calculation results corresponding to all windows according to the coordinates of the target integrated circuit layout to generate a similarity heatmap; The dynamic threshold is determined based on the statistical characteristics of the similarity distribution, and the window corresponding to the cosine similarity calculation result exceeding the dynamic threshold is selected. Based on the coordinates of the selected windows on the target integrated circuit layout, high-risk areas of the target integrated circuit layout are identified, thereby enabling integrated circuit layout verification.

[0016] Secondly, embodiments of this application provide an integrated circuit layout verification device based on image-text comparison learning, the integrated circuit layout verification device based on image-text comparison learning comprising: The module is used to construct a dual-tower neural network architecture that includes a visual tower and a linguistic tower, and to obtain local integrated circuit layout slices and corresponding text diagnostic reports from historical tape-out data to form a training set for training the dual-tower neural network architecture. The execution module is used to input the natural language description of the design rules into the language tower to convert it into a text feature vector. At the same time, the visual tower encodes and extracts the target integrated circuit layout through a sliding window to obtain the target integrated circuit layout feature vector and the coordinate position on the target integrated circuit layout corresponding to each window. The verification module is used to calculate the cosine similarity between the text feature vector and the feature vector of each target integrated circuit layout, and to determine the high-risk area based on the cosine similarity calculation result, thereby realizing the verification of the integrated circuit layout. The visual tower is used to convert the segmented physical map image or polygon topology into a high-dimensional map feature vector through a visual Transformer, and the language tower is used to convert the design rule intent described in natural language into a text feature vector through a large language model.

[0017] Thirdly, embodiments of this application provide an integrated circuit layout verification device based on image-text comparison learning. The integrated circuit layout verification device based on image-text comparison learning includes a processor, a memory, and an integrated circuit layout verification program based on image-text comparison learning stored in the memory and executable by the processor. When the integrated circuit layout verification program based on image-text comparison learning is executed by the processor, it implements the steps of the integrated circuit layout verification method based on image-text comparison learning described above.

[0018] The beneficial effects of the technical solutions provided in this application include: (1) By using contrastive learning to align visual and semantic features, natural language can be used to find layout topologies and hotspots that have never been seen in the training set, which greatly accelerates the speed of rule adaptation when migrating from different processes. (2) Interact directly with the map through natural language prompts, replacing the expensive and error-prone hard-coded proprietary verification language, which greatly reduces verification costs and technical barriers; (3) By constructing sample pairs of cut map slices and diagnostic texts according to the violation coordinates, a learnable mapping between map geometric topology semantics and design rule semantics is established, making zero-sample map verification driven by natural language rules possible. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the integrated circuit layout verification method based on image-text comparison learning proposed in this application. Figure 2 A schematic diagram illustrating the process by which visual towers and linguistic towers map multimodal data to a shared latent space; Figure 3 A schematic diagram showing the identification of high-risk areas; Figure 4This is a schematic diagram of the functional modules of the integrated circuit layout verification device based on image-text comparison learning in this application; Figure 5 This is a schematic diagram of the hardware structure of the integrated circuit layout verification device based on image-text comparison learning in this application. Detailed Implementation

[0020] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0022] Firstly, embodiments of this application provide an integrated circuit layout verification method based on image-text contrastive learning. This method introduces a cross-modal vision-language contrastive learning mechanism into the underlying layout processing of EDA, employs a "layout slice-text diagnostic report" sample pair construction approach, and utilizes a zero-sample verification process oriented towards natural language for design rules. In practical applications, the integrated circuit layout verification method of this application is executed by a computer program module deployed on an EDA verification platform, can be integrated into existing physical verification processes, and carries the corresponding program instructions on a computer-readable storage medium.

[0023] In one embodiment, reference is made to Figure 1 , Figure 1 This is a flowchart illustrating the integrated circuit layout verification method based on image-text comparison learning proposed in this application. Figure 1 As shown, the integrated circuit layout verification method based on image-text comparison learning includes: S1: Construct a dual-tower neural network architecture that includes a vision tower (Vision Encoder) and a language tower (Language Encoder), and obtain local integrated circuit layout slices and corresponding text diagnostic reports from historical tape-out data to form a training set for training the dual-tower neural network architecture; The visual tower is used to convert the segmented physical layout image or polygon topology into a high-dimensional layout feature vector through a visual Transformer (ViT), and the language tower is used to convert the design rules intention described in natural language into a text feature vector through a large language model (LLM).

[0024] S2: Input the natural language description of the design rules into the language tower to convert it into a text feature vector. At the same time, the visual tower encodes and extracts the target integrated circuit layout through a sliding window to obtain the target integrated circuit layout feature vector and the coordinate position on the target integrated circuit layout corresponding to each window. S3: Calculate the cosine similarity between the text feature vector and the feature vector of each target integrated circuit layout, and determine the high-risk area based on the cosine similarity calculation result to realize the integrated circuit layout verification.

[0025] Furthermore, in one embodiment, the visual tower includes an image segmentation and linear projection module, a position encoding module, a multi-head self-attention encoding layer, a classification label aggregation module, and a visual projection head.

[0026] The image segmentation and linear projection module is used to segment the input integrated circuit layout into fixed-size image blocks and flatten them into a vector sequence, which is then mapped into a multi-dimensional embedded vector through linear projection. Its function is to discretize continuous geometric figures into computable serialized inputs.

[0027] The position encoding module is used to embed superimposed position information for each image block; its function is to preserve the spatial positional relationship of each geometric shape in the layout.

[0028] The multi-head self-attention coding layer consists of multiple stacked Transformer encoders and is used to capture global dependencies between image patches through a multi-head self-attention mechanism; its role is to model the contextual features of the map topology. The classification token aggregation module adds a CLS token to the beginning of the vector sequence, and after encoding, uses the output of the CLS token as the aggregated representation of the entire integrated circuit layout slice. CLS token, or Classification Token, is a classification marker.

[0029] The visual projection head consists of a multilayer perceptron (MLP) used to map the aggregated representation to a shared high-dimensional latent space and perform L2 normalization, thereby outputting an integrated circuit layout feature vector.

[0030] Furthermore, in one embodiment, the language tower includes a word segmenter, a word embedding layer, a Transformer encoding layer, a sentence vector extraction module, and a text projection head.

[0031] A word segmenter is used to divide the text describing the design rules in natural language into a sequence of sub-words.

[0032] The word embedding layer is used to map subwords into dense vectors and superimpose positional encodings.

[0033] The Transformer encoding layer is used in a multi-layer self-attention structure based on a large language model to semantically encode the text describing the natural language of design rules; its function is to extract the contextual semantics of elements such as process layer, constraint type, and constraint value in the design rule text.

[0034] The sentence vector extraction module is used to extract the output of CLS marker positions or to perform mean pooling on the word sequence to obtain a sentence-level semantic representation; The text projection head, composed of a multilayer perceptron (MLP), maps sentence-level semantic representations to a shared high-dimensional latent space identical to the visual tower, and then performs L2 normalization to output text feature vectors. See also Figure 2 The diagram illustrates the process by which visual towers and linguistic towers map multimodal data to a shared latent space.

[0035] It should be noted that the network structure of the visual tower and the language tower can be implemented using the well-known ViT and large language model architecture. The input of the visual tower is the physical integrated circuit layout image constructed by slicing the integrated circuit layout, and the input of the language tower is the design rule diagnostic text that is paired with the integrated circuit layout slices one by one according to the violation coordinates. The two towers are mapped to the same shared latent space through the projection head, thereby realizing the alignment between the geometric topology semantics of the integrated circuit layout and the semantics of the design rules.

[0036] Furthermore, in one embodiment, the training set consists of positive sample pairs composed of local integrated circuit layout slices and corresponding text diagnostic reports from historical tape-out data. Specifically, local integrated circuit layout slices and their corresponding text diagnostic reports are extracted from historical tape-out data to form positive samples, thereby training the dual-tower neural network architecture. The training scope includes all trainable parameters of the visual tower, the linguistic tower, and their respective projectors.

[0037] A text-based diagnostic report is an inspection result file output after performing design rule checks on an integrated circuit layout using physical verification tools (such as Calibre). It includes the violation rule number, violation type, the integrated circuit layout coordinates of the violation location, a description of the violation's geometry, and remediation suggestions. For example, a text-based diagnostic report might read, "Metal1 spacing is less than 0.09μm, location coordinates (x=123.5, y=456.2), it is recommended to increase the spacing between adjacent traces." Pairing a local integrated circuit layout slice with the text-based diagnostic report describing the violation in that slice creates a positive sample pair; within the same batch, mismatched local integrated circuit layout slices and text-based diagnostic reports are considered negative samples.

[0038] Furthermore, during the training phase of the dual-tower neural network architecture, the InfoNCE contrastive loss function (i.e., information-noise contrastive estimation loss) is applied. The InfoNCE contrastive loss function is specifically as follows:

[0039] in, Indicates the total loss. This represents the contrast loss from the visual to the language direction. This represents the contrast loss from language to visual direction. , Both structures are symmetrical, and the total loss is the average of the losses in the two directions to ensure bidirectional alignment of visual and semantic features.

[0040] Furthermore, regarding the contrast loss in the visual-to-language direction, specifically:

[0041] in, This indicates the number of local integrated circuit layout slice and text diagnostic report sample pairs in a single training batch, i.e., the number of local integrated circuit layout slice and text diagnostic report sample pairs in a training batch. Indicates the first Feature vectors of local integrated circuit layout slices Indicates and The first match The text feature vector of a text diagnostic report, Indicates the first The text feature vector of a text diagnostic report, Represents the cosine similarity function. This represents the adjustment coefficient, specifically an adjustable hyperparameter greater than zero, used to adjust the sharpness of the similarity distribution. The smaller the value, the stronger the ability of the dual-tower neural network architecture to distinguish negative samples that are difficult to differentiate.

[0042] The InfoNCE contrastive loss function drives a dual-tower neural network architecture to maximize the cosine similarity between the feature vectors of matching local integrated circuit layout slices and text diagnostic report sample pairs (bringing positive samples closer together) in a shared high-dimensional latent space, while minimizing the similarity between mismatched negative sample pairs.

[0043] Furthermore, in one embodiment, the input design rules are described in natural language and converted into a language tower to obtain a text feature vector. Simultaneously, the visual tower extracts the target integrated circuit layout through a sliding window, obtaining the target integrated circuit layout feature vector corresponding to each window and its coordinate position on the target integrated circuit layout. Specifically, this includes: S201: When performing integrated circuit layout verification, input the natural language description of the design rules into the language tower and convert it into a text feature vector; Specifically, in actual integrated circuit layout verification, users input new natural language descriptions of design rules that do not appear in the training set into the language tower. These descriptions generally include elements such as the process layer targeted by the design rule, the type of geometric constraint, and the constraint value. For example, "the minimum spacing of the traces in the metal layer Metal2 shall not be less than 0.10μm", or "the envelope spacing between the via Via1 and the adjacent metal line must be greater than 0.05μm". S202: Slide and traverse the target integrated circuit layout of the whole chip with a fixed-size window (such as W×W pixels) according to the set step size, capture a local integrated circuit layout image of each window, perform normalization processing, and then send it to the vision tower. S203: For each captured local integrated circuit layout image, perform image block segmentation, positional encoding overlay, multi-head self-attention encoding, and projection mapping sequentially to obtain the target integrated circuit layout feature vector corresponding to each window, as well as the coordinate position of each window on the target integrated circuit layout. That is, obtain the integrated circuit layout feature vector corresponding to each window and its coordinate position on the integrated circuit layout, and the integrated circuit layout feature vectors of all windows constitute the window vector set.

[0044] Furthermore, in one embodiment, the cosine similarity between the text feature vector and the feature vectors of each target integrated circuit layout is calculated, and high-risk areas are determined based on the cosine similarity calculation results to achieve integrated circuit layout verification. Specifically, this includes: S301: Calculate the cosine similarity between the text feature vector and the feature vector of each target integrated circuit layout in sequence, specifically:

[0045] in, Represents the cosine similarity function. express and Cosine similarity between them Represents the text feature vector. This represents the feature vector of the current target integrated circuit layout. This represents the vector dot product operation. The L2 norm of a vector; Specifically, the language tower encodes the natural language description of the input design rules into text feature vectors, and calculates the cosine similarity between each text feature vector and the feature vector of each target integrated circuit layout in the window vector set. S302: Arrange the cosine similarity calculation results corresponding to all windows according to the coordinates of the target integrated circuit layout to generate a similarity heatmap; S303: Determine a dynamic threshold based on the statistical characteristics of the similarity distribution (e.g., take the mean of the cosine similarity calculation results corresponding to all windows plus k times the standard deviation, or take the top p% quantile of the cosine similarity calculation results as the dynamic threshold), and filter out windows whose cosine similarity calculation results exceed the dynamic threshold; for example, if the cosine similarity calculation corresponding to the current window exceeds the dynamic threshold, then filter out the current window. S304: Based on the coordinates of the selected windows on the target integrated circuit layout, high-risk areas of the target integrated circuit layout are determined, enabling integrated circuit layout verification. See also... Figure 3 The diagram shown illustrates the identification of high-risk areas.

[0046] It should be further explained that, in one possible implementation, during the sample pair construction process, according to the coordinates of the violation layout recorded in the text diagnostic report, a local integrated circuit layout slice of a preset size is automatically cropped with the coordinates as the center, so that the integrated circuit layout slice is precisely aligned with the text diagnostic text in geometric position; the dynamic threshold in the inference stage can adopt a two-stage adaptive determination method: first, the distribution of the cosine similarity calculation results of all windows of the whole chip is statistically analyzed, and the initial threshold is determined by the mean plus k times the standard deviation (k is 2 to 3); then, windows exceeding the initial threshold are sorted in descending order according to the cosine similarity calculation results and adjacent windows are clustered and merged, and the coordinates of the merged region are output; the sliding window adopts a multi-scale hierarchical scanning strategy, firstly, a larger window is used to quickly and coarsely scan the whole chip to lock the candidate region, and then a smaller window is used to finely scan the candidate region, so as to balance the verification accuracy and computational efficiency.

[0047] Furthermore, the visual tower can be replaced by a convolutional neural network (such as ResNet) instead of the visual Transformer, the language tower can be replaced by a pre-trained language model such as BERT, and the contrastive loss function can be replaced by a pairwise contrastive loss based on Sigmoid instead of the InfoNCE loss.

[0048] Secondly, embodiments of this application also provide an integrated circuit layout verification device based on image-text comparison learning.

[0049] In one embodiment, reference is made to Figure 4 , Figure 4 This is a schematic diagram of the functional modules of the integrated circuit layout verification device based on image-text comparison learning according to this application. Figure 4 As shown, the integrated circuit layout verification device based on image-text comparison learning includes: a construction module, an execution module, and a verification module.

[0050] The construction module is used to construct a dual-tower neural network architecture including a visual tower and a linguistic tower, and to obtain local integrated circuit layout slices and corresponding text diagnostic reports from historical tape-out data to form a training set for training the dual-tower neural network architecture; the execution module is used to input the natural language description of the design rules into the linguistic tower to convert it into text feature vectors, while the visual tower encodes and extracts the target integrated circuit layout through a sliding window to obtain the target integrated circuit layout feature vector corresponding to each window and its coordinate position on the target integrated circuit layout; the verification module is used to calculate the cosine similarity between the text feature vector and the feature vectors of each target integrated circuit layout, and to determine high-risk areas based on the cosine similarity calculation results to achieve integrated circuit layout verification; The visual tower is used to convert the segmented physical map image or polygon topology into a high-dimensional map feature vector through a visual Transformer, and the language tower is used to convert the design rule intent described in natural language into a text feature vector through a large language model.

[0051] Thirdly, embodiments of this application provide an integrated circuit layout verification device based on image-text comparison learning. The integrated circuit layout verification device based on image-text comparison learning can be a personal computer (PC), laptop computer, server, or other device with data processing capabilities.

[0052] Reference Figure 5 , Figure 5 This is a schematic diagram of the hardware structure of an integrated circuit layout verification device based on image-text comparison learning, as described in an embodiment of this application. In this embodiment, the integrated circuit layout verification device based on image-text comparison learning may include a processor, a memory, a communication interface, and a communication bus.

[0053] The communication bus can be of any type and is used to interconnect the processor, memory, and communication interface.

[0054] The communication interface includes input / output (I / O) interfaces, physical interfaces, and logical interfaces used for interconnecting devices within the integrated circuit layout verification equipment based on image-text comparison learning, as well as interfaces used for interconnecting the integrated circuit layout verification equipment based on image-text comparison learning with other devices (such as other computing devices or user equipment). Physical interfaces can be Ethernet interfaces, fiber optic interfaces, ATM interfaces, etc.; user equipment can be displays, keyboards, etc.

[0055] Memory can be various types of storage media, such as random access memory (RAM), read-only memory (ROM), non-volatile RAM (NVRAM), flash memory, optical storage, hard disk, programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), etc.

[0056] The processor can be a general-purpose processor, which can call the integrated circuit layout verification program based on image-text comparison learning stored in memory and execute the integrated circuit layout verification method based on image-text comparison learning provided in the embodiments of this application. For example, the general-purpose processor can be a central processing unit (CPU). The method executed when the integrated circuit layout verification program based on image-text comparison learning is called can be referred to the various embodiments of the integrated circuit layout verification method based on image-text comparison learning in this application, and will not be repeated here.

[0057] Those skilled in the art will understand that Figure 5 The hardware structure shown does not constitute a limitation of this application and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0058] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus. The terms "first," "second," and "third," etc., are used to distinguish different objects, etc., and do not indicate a sequence, nor do they limit "first," "second," and "third" to different types.

[0059] In the description of the embodiments of this application, terms such as "exemplary," "for example," or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a concrete manner.

[0060] In the description of the embodiments of this application, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The "and / or" in the text is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in the description of the embodiments of this application, "multiple" means two or more.

[0061] In some processes described in the embodiments of this application, multiple operations or steps are included in a specific order. However, it should be understood that these operations or steps may not be executed in the order they appear in the embodiments of this application, or they may be executed in parallel. The sequence number of the operation is only used to distinguish different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed sequentially or in parallel, and these operations or steps may be combined.

[0062] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, 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 is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device to execute the methods described in the various embodiments of this application.

[0063] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for verifying integrated circuit layouts based on image-text comparison learning, characterized in that, The integrated circuit layout verification method based on image-text comparison learning includes: A dual-tower neural network architecture comprising a visual tower and a linguistic tower is constructed, and local integrated circuit layout slices and corresponding text diagnostic reports from historical tape-out data are obtained to form a training set for training the dual-tower neural network architecture. The input design rules are described in natural language and converted into a language tower to obtain text feature vectors. At the same time, the visual tower encodes and extracts the target integrated circuit layout through a sliding window to obtain the target integrated circuit layout feature vector and the coordinate position on the target integrated circuit layout corresponding to each window. Calculate the cosine similarity between the text feature vector and the feature vector of each target integrated circuit layout, and determine high-risk areas based on the cosine similarity calculation results to achieve integrated circuit layout verification; The visual tower is used to convert the segmented physical map image or polygon topology into a high-dimensional map feature vector through a visual Transformer, and the language tower is used to convert the design rule intent described in natural language into a text feature vector through a large language model.

2. The integrated circuit layout verification method based on image-text comparison learning as described in claim 1, characterized in that: The visual tower includes an image segmentation and linear projection module, a position encoding module, a multi-head self-attention encoding layer, a classification label aggregation module, and a visual projection head; The image segmentation and linear projection module is used to segment the input integrated circuit layout into image blocks of fixed size and flatten them into a vector sequence, which is then mapped into a multidimensional embedded vector through linear projection. The location encoding module is used to embed overlay location information for each image block; The multi-head self-attention coding layer is composed of multiple stacked Transformer encoders and is used to capture the global dependencies between image patches through a multi-head self-attention mechanism. The classification tag aggregation module is used to add CLS tags to the head of the vector sequence, and after encoding, the output of the CLS tags is used as the aggregated representation of the entire integrated circuit layout slice; The visual projection head is composed of a multilayer perceptron (MLP) and is used to map the aggregated representation to a shared high-dimensional latent space and perform L2 normalization, thereby outputting an integrated circuit layout feature vector.

3. The integrated circuit layout verification method based on image-text comparison learning as described in claim 1, characterized in that: The language tower includes a word segmenter, a word embedding layer, a Transformer encoding layer, a sentence vector extraction module, and a text projection head; The word segmenter is used to divide the text describing the design rules in natural language into a sequence of sub-words. The word embedding layer is used to map subwords into dense vectors and superimpose positional encodings; The Transformer encoding layer is used to perform semantic encoding on the text describing the natural language of the design rules based on a multi-layer self-attention structure of a large language model. The sentence vector extraction module is used to extract the output of the CLS marker position or to perform mean pooling on the word sequence to obtain a sentence-level semantic representation. The text projection head is composed of a multilayer perceptron (MLP) and is used to map sentence-level semantic representations to the same shared high-dimensional latent space as the visual tower and perform L2 normalization to output text feature vectors.

4. The integrated circuit layout verification method based on image-text comparison learning as described in claim 1, characterized in that: The training set consists of positive sample pairs composed of local integrated circuit layout slices from historical tape-out data and corresponding text diagnostic reports; The text diagnostic report refers to the inspection result file output after performing design rule checks on the integrated circuit layout using physical verification tools. It includes the violation rule number, violation type, integrated circuit layout coordinates of the violation location, description of the violation geometry, and repair suggestions.

5. The integrated circuit layout verification method based on image-text comparison learning as described in claim 1, characterized in that: During the training phase of the dual-tower neural network architecture, the InfoNCE contrastive loss function is applied. The InfoNCE contrastive loss function is specifically as follows: in, Indicates the total loss. This represents the contrast loss from the visual to the language direction. This represents the contrast loss from language to visual direction.

6. The integrated circuit layout verification method based on image-text comparison learning as described in claim 5, characterized in that, The contrast loss from visual to language direction is specifically as follows: in, This indicates the number of local integrated circuit layout slices and text diagnostic report sample pairs in a single training batch. Indicates the first Feature vectors of local integrated circuit layout slices Indicates and The first match The text feature vector of a text diagnostic report, Indicates the first The text feature vector of a text diagnostic report, Represents the cosine similarity function. This represents the adjustment coefficient.

7. The integrated circuit layout verification method based on image-text comparison learning as described in claim 1, characterized in that, The input design rules are described in natural language and converted into a language tower to obtain text feature vectors. Simultaneously, the visual tower extracts the target integrated circuit layout through a sliding window, obtaining the target integrated circuit layout feature vector and its coordinate position on the target integrated circuit layout corresponding to each window. Specifically, this includes: When performing integrated circuit layout verification, the design rules are input into a language pyramid in natural language description, which is then converted to obtain a text feature vector. A fixed-size window slides across the target integrated circuit layout of the entire chip at a set step size. A local integrated circuit layout image is captured for each window, normalized, and then sent to the vision tower. For each captured local integrated circuit layout image, image block segmentation, position encoding overlay, multi-head self-attention encoding, and projection mapping are performed sequentially to obtain the target integrated circuit layout feature vector corresponding to each window, as well as the coordinate position of each window on the target integrated circuit layout.

8. The integrated circuit layout verification method based on image-text comparison learning as described in claim 1, characterized in that, The calculation of the cosine similarity between the text feature vector and the feature vectors of each target integrated circuit layout, and the determination of high-risk areas based on the cosine similarity calculation results, to achieve integrated circuit layout verification, specifically includes: The cosine similarity between the text feature vector and the feature vector of each target integrated circuit layout is calculated sequentially, specifically as follows: in, Represents the cosine similarity function. express and Cosine similarity between them Represents the text feature vector. This represents the feature vector of the current target integrated circuit layout. This represents the vector dot product operation. The L2 norm of a vector; Arrange the cosine similarity calculation results corresponding to all windows according to the coordinates of the target integrated circuit layout to generate a similarity heatmap; The dynamic threshold is determined based on the statistical characteristics of the similarity distribution, and the window corresponding to the cosine similarity calculation result exceeding the dynamic threshold is selected. Based on the coordinates of the selected windows on the target integrated circuit layout, high-risk areas of the target integrated circuit layout are identified, thereby enabling integrated circuit layout verification.

9. An integrated circuit layout verification device based on image-text comparison learning, characterized in that, The integrated circuit layout verification device based on image-text comparison learning includes: The module is used to construct a dual-tower neural network architecture that includes a visual tower and a linguistic tower, and to obtain local integrated circuit layout slices and corresponding text diagnostic reports from historical tape-out data to form a training set for training the dual-tower neural network architecture. The execution module is used to input the natural language description of the design rules into the language tower to convert it into a text feature vector. At the same time, the visual tower encodes and extracts the target integrated circuit layout through a sliding window to obtain the target integrated circuit layout feature vector and the coordinate position on the target integrated circuit layout corresponding to each window. The verification module is used to calculate the cosine similarity between the text feature vector and the feature vector of each target integrated circuit layout, and to determine the high-risk area based on the cosine similarity calculation result, thereby realizing the verification of the integrated circuit layout. The visual tower is used to convert the segmented physical map image or polygon topology into a high-dimensional map feature vector through a visual Transformer, and the language tower is used to convert the design rule intent described in natural language into a text feature vector through a large language model.

10. An integrated circuit layout verification device based on image-text comparison learning, characterized in that, The integrated circuit layout verification device based on image-text comparison learning includes a processor, a memory, and an integrated circuit layout verification program based on image-text comparison learning stored in the memory and executable by the processor. When the integrated circuit layout verification program based on image-text comparison learning is executed by the processor, it implements the steps of the integrated circuit layout verification method based on image-text comparison learning as described in any one of claims 1 to 8.