Capacitance value extraction method and device based on sparse data machine learning technology

Through sparse data machine learning technology, the capacitance value of interconnected conductors in chip design is extracted, which solves the problems of low efficiency and accuracy in the existing technology, and realizes accurate modeling of complex conductor structures and efficient capacitance value extraction.

CN119047317AActive Publication Date: 2024-11-29PENG CHENG LAB

Patent Information

Application Number
CN202411169789.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-23
Publication Date
2024-11-29
Estimated Expiration
2044-08-23

AI Technical Summary

Technical Problem

Prior Art In chip design, the complex structure of interconnect conductors leads to a decrease in capacitance value extraction efficiency and accuracy, and computing resources consume a lot of time and cannot be accurately modeled.

Method used

Using a capacitance value extraction method based on sparse data machine learning technology, the conductor morphological features are extracted through the feature learning layer of the potential prediction model, and the electric field equation is encoded by the physical information layer to directly extract the capacitance value from the point cloud data to avoid parameterized conductor structure .

Benefits of technology

It improves the accuracy and efficiency of capacitance value extraction, saves computing resources, is suitable for complex interconnected conductor structures, and realizes accurate modeling of complex conductor structures and efficient capacitance value extraction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119047317A_ABST
    Figure CN119047317A_ABST
Patent Text Reader

Abstract

The invention provides a capacitance value extraction method and device based on a sparse data machine learning technology, and belongs to the technical field of artificial intelligence and the technical field of electronic circuits. Performing feature extraction on coordinate information of all sampling points of the target point cloud data through a feature learning layer of the potential prediction model to obtain conductor morphological features, and performing feature splicing on attribute information of each sampling point and the conductor morphological features to obtain target point cloud features of each sampling point; and performing physical information coding on the target point cloud features through a physical information layer of the potential prediction model to obtain the potential of each sampling point. The potentials of all sampling points on the Gaussian surface of the conductor serve as the target potential according to the type of the target sampling point, the capacitance value of the conductor is extracted according to the target potential, a field solver does not need to be used for calculating a large amount of training data in advance, calculation resources are saved, and meanwhile calculation efficiency is improved. In addition, the accuracy of capacitance value extraction can be improved by using the point cloud to depict the conductor structure.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the fields of artificial intelligence technology and electronic circuit technology, and in particular to a capacitance value extraction method and device based on sparse data machine learning technology. Background Art

[0002] In the related art, artificial intelligence technology is used to optimize the chip design (Electronic Design Automation, EDA) process to improve chip design efficiency. In the process of artificial intelligence-assisted chip design, parasitic parameters such as capacitance value extraction are involved. When performing the capacitance value extraction process, a series of parameters are usually introduced to engrave the interconnect conductor structure to obtain structural description data, and the capacitance value of the interconnect conductor structure is solved by a field solver based on the structural description data. However, in actual integrated circuits, the interconnects are densely distributed and the number of conductors is large, and the use of a field solver requires a lot of computing resources and time. In addition, the interconnect conductor structure is relatively complex, making it impossible to accurately model and pattern match the interconnect conductor structure, resulting in a decrease in the prediction accuracy of the capacitance value. Summary of the invention

[0003] The main purpose of the embodiments of the present application is to propose a capacitance value extraction method and device based on sparse data machine learning technology, aiming to improve the computational efficiency and accuracy of capacitance value extraction.

[0004] To achieve the above-mentioned purpose, a first aspect of an embodiment of the present application proposes a capacitance value extraction method based on sparse data machine learning technology, the method comprising:

[0005] Acquire target point cloud data of a target interconnect conductor structure, wherein the target interconnect conductor structure includes a conductor, a plurality of sampling points are arranged around the conductor, and the target point cloud data includes attribute information of each sampling point; the attribute information includes coordinate information of the sampling point and a target sampling point type;

[0006] Extracting features from the coordinate information of all sampling points in the target point cloud data through a feature learning layer of the electric potential prediction model to obtain conductor morphological features of the target point cloud data;

[0007] The attribute information of each sampling point in the target point cloud data is respectively spliced ​​with the conductor morphological feature to obtain the target point cloud feature of each sampling point;

[0008] Performing physical information encoding on the target point cloud features of all the sampling points through the physical information layer of the potential prediction model to obtain the potential of each of the sampling points;

[0009] Determine the sampling point located on the Gaussian surface of the conductor according to the target sampling point type of each sampling point in the target point cloud data, and use the potential of all sampling points located on the Gaussian surface of the conductor as the target potential;

[0010] A capacitance value of the conductor is extracted according to the target potential.

[0011] In some embodiments, the potential prediction model is trained according to the following steps:

[0012] Acquire sample interconnect conductor point cloud data; the sample interconnect conductor point cloud data includes sample sampling points and sample attributes of the sample sampling points; the sample attributes include sample sampling point types;

[0013] Performing potential prediction on the sample attribute by using a preset model to obtain a predicted potential of the sample sampling point under the sample sampling point type;

[0014] Acquire a reference potential of the sample sampling point according to the type of the sample sampling point;

[0015] Perform loss calculation according to the predicted potential and the reference potential to obtain a target loss;

[0016] The model parameters of the preset model are adjusted according to the target loss to obtain the electric potential prediction model.

[0017] In some embodiments, the performing loss calculation according to the predicted potential and the reference potential to obtain a target loss includes:

[0018] Perform a first loss calculation according to the predicted potential to obtain a first sub-loss;

[0019] Perform a second loss calculation based on the predicted potential and the reference potential to obtain a second sub-loss;

[0020] The first sub-loss and the second sub-loss are summed to obtain the target loss.

[0021] In some embodiments, the sample attributes further include point cloud coordinates of the sample sampling points, the point cloud coordinates include point cloud abscissas and point cloud ordinates, and the first loss calculation is performed according to the predicted potential to obtain the first sub-loss, including:

[0022] Obtaining a second-order derivative of the predicted potential with respect to the horizontal coordinate of the point cloud to obtain a first loss value;

[0023] Obtaining a second-order derivative of the predicted potential with respect to the ordinate of the point cloud to obtain a second loss value;

[0024] The first sub-loss is determined according to the first loss value and the second loss value.

[0025] In some embodiments, the sample sampling point type includes a first boundary sampling point subtype, a second boundary sampling point subtype, a third boundary sampling point subtype, and a fourth boundary sampling point subtype, and performing a second loss calculation according to the predicted potential and the reference potential to obtain a second sub-loss includes:

[0026] Perform loss calculation according to the predicted potential and the reference potential of the sample sampling point of the first boundary sampling point subtype to obtain a third loss value;

[0027] Perform loss calculation according to the predicted potential and the reference potential of the sample sampling point of the second boundary sampling point subtype to obtain a fourth loss value;

[0028] Perform loss calculation according to the predicted potential and the reference potential of the sample sampling point of the third boundary sampling point subtype to obtain a fifth loss value;

[0029] determining a predicted electric field strength according to the predicted electric potential of the sample sampling point of the fourth boundary sampling point subtype, determining a reference electric field strength according to the reference electric potential, and performing a loss calculation according to the predicted electric field strength and the reference electric field strength to obtain a sixth loss value;

[0030] The third loss value, the fourth loss value, the fifth loss value and the sixth loss value are summed to obtain the second sub-loss.

[0031] In some embodiments, the feature learning layer includes a plurality of network sublayers connected in series, and the feature learning layer of the potential prediction model extracts features of the coordinate information of all sampling points in the target point cloud data to obtain the conductor morphological features of the target point cloud data, including:

[0032] Performing feature extraction on the coordinate information of all sampling points in the target point cloud data through the feature learning layer to obtain the first point cloud feature output by each network sublayer;

[0033] The first point cloud features output by each network sub-layer are spliced ​​to obtain the conductor morphological features.

[0034] In some embodiments, the physical information layer includes a first convolution layer, a residual layer, and a second convolution layer, and the physical information layer of the potential prediction model is used to encode the target point cloud features of all the sampling points to obtain the potential of each of the sampling points, including:

[0035] Extracting the target point cloud features of all the sampling points through the first convolutional layer to obtain second point cloud features;

[0036] Performing feature extraction on the second point cloud feature through the residual layer to obtain a third point cloud feature;

[0037] The third point cloud feature is extracted through the second convolution layer to obtain the electric potential of each sampling point.

[0038] To achieve the above-mentioned purpose, a second aspect of an embodiment of the present application proposes a capacitance value extraction device based on sparse data machine learning technology, the device comprising:

[0039] an acquisition module, configured to acquire target point cloud data of a target interconnect conductor structure, wherein the target interconnect conductor structure comprises a conductor, a plurality of sampling points are arranged around the conductor, and the target point cloud data comprises attribute information of each sampling point; the attribute information comprises coordinate information of the sampling point and a target sampling point type;

[0040] A feature extraction module is used to extract features from the coordinate information of all sampling points in the target point cloud data through a feature learning layer of the electric potential prediction model to obtain conductor morphological features of the target point cloud data;

[0041] A feature splicing module is used to perform feature splicing on the attribute information of each sampling point in the target point cloud data and the conductor morphological feature to obtain the target point cloud feature of each sampling point;

[0042] A feature encoding module, used to perform physical information encoding on the target point cloud features of each sampling point through the physical information layer of the potential prediction model to obtain the potential of each sampling point;

[0043] A determination module, used for determining the sampling points located on the Gaussian surface of the conductor according to the target sampling point type of each sampling point in the target point cloud data, and taking the potentials of all sampling points located on the Gaussian surface of the conductor as the target potential;

[0044] A capacitance value extraction module is used to extract the capacitance value of the conductor according to the target potential.

[0045] To achieve the above-mentioned purpose, the third aspect of an embodiment of the present application proposes an electronic device, the electronic device includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the capacitance value extraction method based on sparse data machine learning technology of the above-mentioned first aspect is implemented.

[0046] To achieve the above-mentioned purpose, the fourth aspect of an embodiment of the present application proposes a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the capacitance value extraction method based on sparse data machine learning technology of the above-mentioned first aspect is implemented.

[0047] The capacitance value extraction method based on sparse data machine learning technology, the capacitance value extraction device based on sparse data machine learning technology, the electronic device and the computer-readable storage medium proposed in the present application, by acquiring the target point cloud data of the target interconnect conductor structure, extracting the capacitance value of the target interconnect conductor structure based on the target point cloud data, and at the same time, the accuracy of the capacitance value extraction can be improved by characterizing the conductor structure through the point cloud. The feature learning layer of the potential prediction model is used to extract the coordinate information of all sampling points in the target point cloud data, so as to directly model the target interconnect conductor structure using the feature learning layer, obtain the geometric features of the point cloud data, and obtain the conductor morphological features of the target point cloud data without introducing additional parameters to characterize the interconnect conductor structure. Because there is no need to parameterize the conductor structure, the present application has better adaptability and versatility for complex interconnect conductor structures, and realizes accurate modeling of complex interconnect conductor structures. The potential of the sampling point is not only related to the attribute information of the sampling point, but also to the shape and other features of the point cloud where the sampling point is located. In order to improve the accuracy of the potential prediction, the attribute information of each sampling point is feature spliced ​​with the conductor morphological features of the target point cloud data to obtain the target point cloud features of each sampling point. Solving the capacitance value through the field solver requires a lot of computing resources and time resources. In order to improve the efficiency of capacitance value extraction, the present application sets a physical information layer in the potential prediction model. The potential satisfies the electric field equation on the target point cloud data. The physical information layer of the potential prediction model encodes the physical information of the target point cloud features of each sampling point, so as to encode the physical information of the electric field equation using the physical information layer to obtain the potential of each sampling point. In order to extract the capacitance of the target interconnect conductor structure, the sampling point located on the Gaussian surface of the conductor is determined according to the target sampling point type of each sampling point, and the potential of all sampling points located on the Gaussian surface of the conductor is used as the target potential to solve the capacitance using the potential of the Gaussian surface sampling point, without having to use the field solver to calculate a large amount of training data in advance, thereby saving computing resources and improving the efficiency of capacitance extraction. Extracting the capacitance value of the conductor according to the target potential, rather than extracting the capacitance using the field solver and the structural description data obtained by the parameterized conductor structure, avoids the decrease in the capacitance value prediction accuracy caused by the inability of the parameters to accurately describe the conductor structure, and improves the accuracy of capacitance value extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 is a flow chart of a capacitance value extraction method based on sparse data machine learning technology provided in an embodiment of the present application;

[0049] Figure 2 is a schematic diagram of target point cloud data provided by an embodiment of the present application;

[0050] Figure 3 yes Figure 1 Flow chart of step S120 in FIG.

[0051] Figure 4 yes Figure 1 Flow chart of step S140 in FIG.

[0052] Figure 5a is another schematic diagram of target point cloud data provided by an embodiment of the present application;

[0053] Figure 5b is another schematic diagram of target point cloud data provided by an embodiment of the present application;

[0054] Figure 5c is another schematic diagram of target point cloud data provided by an embodiment of the present application;

[0055] Figure 5d yes Figure 5a Schematic diagram of the electric field of the target point cloud data;

[0056] Figure 5e yes Figure 5b Schematic diagram of the electric field of the target point cloud data;

[0057] Figure 5f yes Figure 5c Schematic diagram of the electric field of the target point cloud data;

[0058] Figure 6 is a flow chart of the training process of the electric potential prediction model provided in the embodiment of the present application;

[0059] Figure 7 yes Figure 6 Flow chart of step S640 in FIG.

[0060] Figure 8 yes Figure 7 Flowchart of step S710 in FIG.

[0061] Fig. 9 yes Figure 7 Flowchart of step S720 in FIG.

[0062] Fig.10 It is a structural schematic diagram of a capacitance value extraction device based on sparse data machine learning technology provided in an embodiment of the present application;

[0063] Fig.11 It is a schematic diagram of the hardware structure of the electronic device provided in the embodiment of the present application. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0065] It should be noted that, although the functional modules are divided in the device schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first", "second", etc. in the specification, claims and the above drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0066] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0067] In the related technology, artificial intelligence technology is used to assist chip design, such as using machine learning methods to determine the optimal circuit layout, interconnect parasitic parameter extraction, etc., to significantly shorten the time required for chip design and improve chip design efficiency. The extraction of parasitic parameters such as capacitance is involved in the timing analysis process of chip design. How to extract the parasitic parameters of interconnects efficiently and accurately is a basic problem in chip design. In actual integrated circuits, the distribution of interconnects is very dense and the number of conductors is large, making the circuit layout very complex, and the parasitic parameter extraction of large-scale layouts usually involves thousands of interconnect conductor data. In addition, circuit design is an iterative process. As long as the line changes, the parasitic parameters need to be re-extracted, which makes parasitic parameter extraction a very computationally intensive task. In addition, with the reduction in scale and increasing complexity of semiconductor process technology, the accuracy requirements for parasitic parameter extraction have been significantly improved. At the same time, advanced process structures such as low-dielectric constant dielectrics, non-vertical interface conductors, and bubble dielectrics have increased the complexity of interconnect modeling, making it difficult to parameterize complex conductor structures and layouts. At present, most of the methods used for parasitic parameter extraction use data-driven neural networks, machine learning methods, etc. The main idea is to characterize the interconnect conductor structure through some parameters, set a series of sampling values ​​for these parameters to obtain a batch of interconnect conductor structures, and solve their capacitance values ​​through a field solver. However, these parameters cannot accurately model complex interconnect conductor structures, and the field solver consumes more time and computing resources to solve the capacitance, resulting in a decrease in the prediction accuracy and efficiency of the capacitance.

[0068] Based on this, the embodiments of the present application provide a capacitance value extraction method based on sparse data machine learning technology, a capacitance value extraction device based on sparse data machine learning technology, an electronic device and a computer-readable storage medium, aiming to improve the accuracy and efficiency of capacitance extraction.

[0069] The capacitance value extraction method based on sparse data machine learning technology, the capacitance value extraction device based on sparse data machine learning technology, the electronic device and the computer-readable storage medium provided in the embodiments of the present application are specifically illustrated by the following embodiments. First, the capacitance value extraction method based on sparse data machine learning technology in the embodiments of the present application is described.

[0070] The capacitance value extraction method based on sparse data machine learning technology provided in the embodiment of the present application relates to the field of artificial intelligence technology and the field of electronic circuit technology. The capacitance value extraction method based on sparse data machine learning technology provided in the embodiment of the present application can be applied to the terminal, can also be applied to the server side, and can also be software running in the terminal or the server side. In some embodiments, the terminal can be a smart phone, a tablet computer, a laptop computer, a desktop computer, etc.; the server side can be configured as an independent physical server, or it can be configured as a server cluster or distributed system composed of multiple physical servers, and can also be configured to provide cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms and other basic cloud computing services. Cloud servers; software can be applications that implement capacitance value extraction methods, etc., but are not limited to the above forms.

[0071] The present application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present application can also be practiced in distributed computing environments, in which tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media including storage devices.

[0072] Figure 1 is an optional flow chart of a capacitance value extraction method based on sparse data machine learning technology provided in an embodiment of the present application, Figure 1 The method may include but is not limited to steps S110 to S160.

[0073] Step S110, obtaining target point cloud data of a target interconnect conductor structure, wherein the target interconnect conductor structure includes a conductor, a plurality of sampling points are arranged around the conductor, and the target point cloud data includes attribute information of each sampling point; the attribute information includes coordinate information of the sampling point and a target sampling point type;

[0074] Step S120, extracting features from the coordinate information of all sampling points in the target point cloud data through the feature learning layer of the electric potential prediction model to obtain the conductor morphological features of the target point cloud data;

[0075] Step S130, feature splicing is performed on the attribute information of each sampling point in the target point cloud data and the conductor morphological feature to obtain the target point cloud feature of each sampling point;

[0076] Step S140, performing physical information encoding on the target point cloud features of each sampling point through the physical information layer of the potential prediction model to obtain the potential of each sampling point;

[0077] Step S150, determining the sampling points located on the Gaussian surface of the conductor according to the target sampling point type of each sampling point in the target point cloud data, and taking the potentials of all sampling points located on the Gaussian surface of the conductor as the target potential;

[0078] Step S160: extracting the capacitance value of the conductor according to the target potential.

[0079] In step S110 of some embodiments, the target interconnect conductor structure is an interconnect conductor structure for extracting parasitic parameters, and the interconnect conductor structure refers to a conductive path for connecting different circuit elements (such as capacitors, resistors, transistors, etc.) in integrated circuit design. These conductive paths are usually made of metal or other conductive materials, and they form a complex network at multiple levels of the chip to achieve electrical connection between circuit elements. The point cloud of the target interconnect conductor structure can be generated by laser scanning, 3D modeling software, computer vision SLAM algorithm, etc., to obtain target point cloud data. The target interconnect conductor structure includes at least one conductor, and a plurality of sampling points are arranged around each conductor. The target point cloud data includes attribute information of each sampling point, and the attribute information is multi-dimensional data containing several attributes, including the point cloud coordinates of the sampling point, the target sampling point type, the number of conductors in the target interconnect conductor structure (data window) to which the sampling point belongs, the external normal vector, etc. The sampling points are divided into sampling points at the boundary of the data window, sampling points inside the calculation area, sampling points on the boundary of the conductor, and sampling points on the Gaussian surface of the conductor. The Gaussian surface is a pre-set closed surface, which can be of any shape, such as a plane, a sphere, a cylinder, etc. Among them, the sampling points of the data window boundary include the sampling points of the Dirichlet boundary condition and the sampling points of the Neumann boundary condition. The inside of the calculation area refers to the other areas in the data window where the target point cloud data is located except for all conductor areas. The sampling points on the conductor boundary include the sampling points on the main conductor boundary and the sampling points on the accompanying conductor boundary. The main conductor refers to the conductor whose capacitance is to be solved, and the accompanying conductor refers to other conductors that affect the capacitance of the main conductor. The target sampling point type is used to indicate the type of sampling point, which is one of the main conductor boundary sampling points, the accompanying conductor boundary sampling points, the Dirichlet boundary sampling points, the Neumann boundary sampling points, the sampling points inside the calculation area, and the Gaussian surface sampling points of the conductor.

[0080] The attribute information of the jth sampling point is represented as v j =(x j ,l j ,…,nnet j ), where x j represents the coordinate information (point cloud coordinates) of the jth sampling point, l j Indicates the boundary condition annotation, which is used to indicate the target sampling point type of the sampling point, nnet j Indicates the number of conductors in the data window. The target point cloud data of the target interconnect conductor structure is as follows: Figure 2As shown, the boundary conditions of the main conductor boundary sampling points are marked as 1, the boundary conditions of the accompanying conductor boundary sampling points are marked as 2, the boundary conditions of the Dirichlet boundary sampling points are marked as 3, the boundary conditions of the Newman boundary sampling points are marked as 4, the boundary conditions of the conductor Gaussian surface sampling points are marked as 5, and the boundary conditions inside the calculation area are marked as 6.

[0081] The electric potential prediction model is used to map the target point cloud data to the electric potential. The electric potential prediction model includes a feature learning layer and a physical information layer. The feature learning layer is located in the upper part of the electric potential prediction model and is used to geometrically encode the target point cloud data to extract the geometric features of the point cloud data. A neural network architecture similar to PointNet is used to construct the feature learning layer. The main idea of ​​this type of network is to use a multi-layer perceptron neural network with shared weights and a global symmetric pooling function to construct a low-dimensional representation from a set of point clouds. The physical information layer is located in the lower part of the electric potential prediction model and is used to encode the electric field equation to obtain the physical information of the electric field equation.

[0082] See also Figure 3 In some embodiments, the feature learning layer includes a plurality of network sublayers connected in series, and step S120 may include but is not limited to steps S310 to S320:

[0083] Step S310, extracting features from the coordinate information of all sampling points in the target point cloud data through the feature learning layer to obtain the first point cloud feature output by each network sublayer;

[0084] Step S320, performing feature splicing on the first point cloud features output by each network sub-layer to obtain conductor morphological features.

[0085] In step S310 of some embodiments, the feature learning layer includes a main network and a permutation invariant network layer. The main network uses a two-dimensional convolution operator, and the permutation invariant network layer can use a maximum pooling operator, an average pooling operator, a minimum operator, or a summation operator. Among them, the maximum pooling operator refers to applying a sliding window on the feature map, and selecting the maximum value from each window to generate a new feature map. The average pooling operator refers to calculating the average value of all values ​​in each window to generate a new feature map. The minimum operator refers to selecting the minimum value from each window to generate a new feature map. The summation operator refers to summing all values ​​in each window to generate a new feature map. The network sublayer includes a two-dimensional convolution layer or a permutation invariant network layer, and the feature learning layer includes five two-dimensional convolution layers connected in series and a permutation invariant network layer. The five two-dimensional convolution layers are represented as conv1, conv2, conv3, conv4 and conv5 from top to bottom. The parameters of the five 2D convolutional layers are as follows: the convolution kernel size of conv1 is 1*2, and the number of output channels is 64; the convolution kernel size of conv2 and conv3 is 1*1, and the number of output channels is 128; the convolution kernel size of conv4 is 1*1, and the number of output channels is 512; the convolution kernel size of conv5 is 1*1, and the number of output channels is 2048. The stride of all 2D convolution operations is [1,1]. Each 2D convolutional layer uses an activation function, and batch normalization (BN) is used before the activation function of each convolutional layer to solve the problem of internal covariate shift.

[0086] The attribute information includes the coordinate information of the sampling points. The coordinate information is used to indicate the point cloud coordinates of the sampling points and can be expressed in two-dimensional space coordinates. If the target point cloud data includes N sampling points, the point cloud coordinate set composed of the point cloud coordinates of the N sampling points is expressed as X N , X N The feature dimension of is N×2. The coordinate information of all sampling points in the target point cloud data, that is, the point cloud coordinate set, is input into the feature learning layer, and the point cloud coordinates of each sampling point are subjected to two-dimensional convolution processing through five two-dimensional convolutional layers connected in series to obtain the first point cloud feature output by each two-dimensional convolutional layer. The channel dimension is increased from 2 dimensions to 2048 dimensions through these five two-dimensional convolutional layers. In the spatial dimension, the first point cloud feature output by the last two-dimensional convolutional layer is subjected to permutation-invariant processing through a permutation-invariant network layer to aggregate the information shared by all pixels in the first point cloud feature, and obtain the first point cloud feature output by the permutation-invariant network layer. The number of feature channels of the first point cloud feature output by the permutation-invariant network layer is the same as the number of feature channels of the first point cloud feature output by the last two-dimensional convolutional layer, both of which are 2048. The first point cloud feature is the point cloud geometric feature of all sampling points in the target point cloud data for the coordinate information, including normals, curvatures, distances between sampling points, etc.

[0087] In step S320 of some embodiments, the first point cloud features output by the five two-dimensional convolutional layers and the first point cloud features output by the permutation invariant network layer are feature spliced ​​from the channel dimension to obtain the conductor morphological features of the target point cloud data. The conductor morphological features are global features used to characterize the point cloud geometric features of the sampling points. If the feature dimension of the conductor morphological features of a sampling point is expressed as: 64+128+128+512+2048+2048=4928, 64, 128, 128, 512, 2048 are the dimensions of the feature vector output by the convolutional layer, respectively, and 2048 is the dimension of the feature vector after the permutation invariant operation. The feature dimension of the conductor morphological features of N sampling points, i.e., the conductor morphological features of the target point cloud data, is N×4928.

[0088] In the above steps S310 to S320, the interconnect conductor structure is modeled directly using the feature extraction layer to extract the geometric features of the point cloud data without introducing an additional series of parameters to parameterize the interconnect conductor structure, thereby achieving accurate modeling of various types of conductor structures including complex conductor structures.

[0089] In step S130 of some embodiments, the potential u of the jth sampling point j It depends not only on the attribute information of the jth sampling point, but also on the shape characteristics of the point cloud where the jth sampling point is located. The attribute information of the jth sampling point is represented as v j , the conductor morphology characteristics of all sampling points are expressed as V i is the target point cloud data of the i-th sample. In order to ensure the accuracy of the subsequent potential prediction, the attribute information of the j-th sampling point is spliced ​​with the conductor morphology features of all sampling points in the target point cloud data, and the target point cloud features of the j-th sampling point are obtained as

[0090] See also Figure 4 In some embodiments, the physical information layer includes a first convolutional layer, a residual layer, and a second convolutional layer, and step S140 may include but is not limited to steps S410 to S430:

[0091] Step S410, extracting the target point cloud features of all sampling points through the first convolution layer to obtain second point cloud features;

[0092] Step S420, extracting features from the second point cloud features through a residual layer to obtain third point cloud features;

[0093] Step S430: extract the third point cloud features through the second convolution layer to obtain the potential of each sampling point.

[0094] In step S410 of some embodiments, the input of the physical information layer includes the first point cloud features output by each network sublayer in the feature learning layer and the conductor morphology features of all sampling points. The physical information layer is also composed of convolutional layers. Like the feature learning layer, each convolutional layer includes an activation function and a batch normalization operation. The first convolutional layer includes four two-dimensional convolutional layers connected in series, and the four two-dimensional convolutional layers are conv6, conv7, conv8 and conv9, respectively. The convolution kernel size is 1*1, the output channel number of conv1 is 256, the output channel number of conv7 and conv8 is 128, and the output channel number of con9 is 1. The target point cloud features of all sampling points are two-dimensionally convolved by four two-dimensional convolutional layers connected in series to obtain the second point cloud features composed of all sampling points. The number of channels of the feature can be reduced from 1928 dimensions to 1 dimension through these four two-dimensional convolutional layers.

[0095] In step S420 of some embodiments, in order to improve the approximation accuracy of the physical information layer to the physical information of the electric field equation, a residual layer may be added to the model. The residual layer includes a plurality of identical residual structures. The residual structure may be described by the following formula:

[0096] M1=φ(ConvR1(W 1 ,x1)+b 1 ),

[0097] N1=φ(ConvR2(W 2 ,M1)+b 2 ),

[0098] y1=N1+x1=F(x1)+x1,

[0099] Among them, ConvR1 and ConvR2 represent the first convolutional layer and the second convolutional layer of the residual structure, W 1 and b 1 Represents the weight parameters and bias parameters of the first convolutional layer; W 2 and b 2 represents the weight parameters and bias parameters of the second convolutional layer; x1 represents the feature vector input to the first convolutional layer; φ represents the activation function; M1 represents the output of the first convolutional layer or the input of the second convolutional layer; N1 is the output of the second convolutional layer; F represents the mapping function that maps x1 to N1; y1 represents the output of the residual structure.

[0100] Regarding the selection of the activation function φ, since the subsequent coding of the physical information of the electric field equation, namely the partial differential equation, requires that the electric potential prediction model is differentiable, a differentiable activation function is selected, such as sin, GeLU, tanh, etc.

[0101] The second point cloud features of each sampling point are extracted through the residual structure in the residual layer to capture more abstract and complex point cloud feature information, thereby obtaining the third point cloud features of each sampling point.

[0102] In step S430 of some embodiments, the second convolution layer includes a convolution layer conv10, and conv10 extracts the third point cloud features of all sampling points to obtain the electric potential of each sampling point, where the electric potential of the sampling point is a one-dimensional feature. The electric field is the first-order gradient of the electric potential, that is, Where E represents the electric field, represents the first-order gradient, and U represents the electric potential. The point cloud coordinates include the point cloud abscissa and the point cloud ordinate. The first-order gradient is a vector composed of the first-order partial derivative of the electric potential with respect to the point cloud abscissa and the first-order partial derivative of the electric potential with respect to the point cloud ordinate.

[0103] The activation function of the second convolutional layer can be selected based on prior knowledge. For example, if you want to ensure that the potential is in the range of [0,1], you can define the activation function as a sigmoid function. The definition of the sigmoid function is as follows:

[0104]

[0105] Where γ represents the feature vector input to the activation function.

[0106] If there is no prior knowledge, the second convolutional layer conv10 can have no activation function.

[0107] It should be noted that the selection of the number of output channels of the convolutional layer in the network structure is not fixed here, but can be adjusted according to the specific actual situation. The total number of sampling points, the complexity of the conductor structure, the number of samples of the conductor point cloud data, and the memory consumed by the neural network calculation, etc., all these factors will affect the fitting accuracy and training efficiency of the neural network, and need to be comprehensively considered after testing according to the actual situation.

[0108] Through the above steps S410 to S430, the electric potential of each sampling point can be obtained, so as to extract the capacitance of the conductor based on the electric potential without using a field solver, thereby improving the efficiency of capacitance extraction.

[0109] The target point cloud data of the first sample, the second sample, and the third sample are as follows: Figure 5a , Figure 5b , Figure 5cAs shown in the figure, the feature learning layer of the potential prediction model extracts features from the target point cloud data of the three samples respectively to obtain the geometric coding features of the three samples. For each sample, the geometric coding features are spliced ​​with the target point cloud data, and the potential of each sample is predicted by the physical information layer. The electric field of each sample is determined according to the potential of each sample. The electric field of the first sample is shown in Figure 5d As shown in Figure 2, the electric field of the second sample is Figure 5e The electric field of the third sample is shown in 5f.

[0110] In the related art, the structural description data is obtained by parameterizing the interconnect conductor structure, and the capacitance of the interconnect conductor structure is calculated in advance by a field solver, and the neural network model is trained using the structural description data and the capacitance. However, solving the capacitance by a field solver consumes a lot of computing resources and time resources, and the parametric modeling method cannot accurately characterize the complex interconnect conductor structure, resulting in a decrease in the efficiency and accuracy of capacitance extraction. In order to improve the capacitance extraction performance of the model, the present application adopts an unsupervised method for model training, which does not require the use of a field solver to calculate the capacitance in advance and does not require a parameterized conductor structure, which can save a lot of computing resources and computing time, and can also improve the accuracy of capacitance extraction.

[0111] See also Figure 6 In some embodiments, the training process of the electric potential prediction model may include but is not limited to steps S610 to S650:

[0112] Step S610, obtaining sample interconnect conductor point cloud data; the sample interconnect conductor point cloud data includes sample sampling points and sample attributes of the sample sampling points; the sample attributes include sample sampling point types;

[0113] Step S620, predicting the potential of the sample attribute by using a preset model to obtain the predicted potential of the sample sampling point under the sample sampling point type;

[0114] Step S630, obtaining a reference potential of the sample sampling point according to the type of the sample sampling point;

[0115] Step S640, performing loss calculation according to the predicted potential and the reference potential to obtain a target loss;

[0116] Step S650, adjusting the model parameters of the preset model according to the target loss to obtain the electric potential prediction model.

[0117] In step S610 of some embodiments, a point cloud data set is generated according to a batch of interconnect conductor structures, the batch of interconnect conductor structures includes M interconnect conductor structure samples, each interconnect conductor structure sample corresponds to a point cloud subset, and the point cloud data set is a set consisting of multiple point cloud subsets, each point cloud subset consists of a fixed number (such as N) of sampling points. Each point cloud subset is composed of multidimensional data that describes the characteristics of the point cloud subset, including the point cloud coordinates of all sampling points in the point cloud subset, the boundary condition annotations of all sampling points, the number of conductors contained in the conductor sample window, the external normal vectors of all sampling points, etc.

[0118] Each point cloud subset contains N sampling points. All sampling points in the point cloud subset are labeled, so each point cloud subset contains N×1 boundary condition annotations. This application writes the above-mentioned multidimensional point cloud data into a fixed HDF5 file by configuring a standard channel to achieve unified data storage. The specific configuration method of the standard channel is: the first data channel is named data, containing M×N×2 point cloud coordinate data, and the data type is float32 (32-bit floating point number); the second data channel is named label, containing M×N×1 annotation data (boundary condition annotation), and the data type is uint8 (8-bit unsigned integer); the third data channel is named nnet, containing M×1 data of type uint8, indicating the number of conductors contained in each point cloud subset; the fourth data channel is named nnett_gp, containing M×15 data of type uint8, indicating the number of Gaussian surface sampling points contained in each conductor; the fifth data channel is named normal, containing M×N b ×2 float32 data types, storing the coordinates of the external normal vectors of all boundary sampling points, N b Indicates the number of boundary sampling points. It can be assumed that the number of conductors contained in each conductor data window does not exceed 15. When the number of conductors n c When it is less than 15, the nth c The number of gauss surface sampling points after is recorded as 0.

[0119] 80% of the point cloud data can be randomly selected from the point cloud data set as the training set, denoted as m is the number of samples in the training set, that is, the number of point cloud subsets. 10% of the remaining 20% ​​is taken as the test set and validation set respectively. The test set is recorded as The validation set is denoted as l and k are the number of samples in the test set and validation set, respectively.

[0120] A point cloud subset is obtained from the training set, and the point cloud subset is characterized as a multi-dimensional data set containing several attributes to obtain the sample interconnect conductor point cloud data. The sample interconnect conductor point cloud data includes sample sampling points and sample attributes of the sample sampling points. The sample sampling points are sampling points in the point cloud subset. The sample attributes include the point cloud coordinates of the sample sampling points, the sample sampling point type, the number of conductors contained in the data window, the external normal vector, etc. The sample interconnect conductor point cloud data of the i-th sample can be expressed as:

[0121]

[0122] Among them, x j is the point cloud coordinate of the jth sample sampling point, l j Label the boundary conditions of the jth sample point, nnet j is the number of conductors in the data window, and N is the number of sampling points.

[0123] The point cloud coordinates include the point cloud horizontal coordinates and the point cloud vertical coordinates. The point cloud subset contains 4 different boundary conditions. The sample sampling point types include the conductor Gaussian surface sampling points, the sampling points in the calculation area and the boundary sampling points. The boundary sampling points include four types, namely the first boundary sampling point subtype, the second boundary sampling point subtype, the third boundary sampling point subtype and the fourth boundary sampling point subtype. The first boundary sampling point subtype is the main conductor boundary sampling point, the second boundary sampling point subtype is the accompanying conductor boundary sampling point, the third boundary sampling point subtype is the Dirichlet boundary sampling point (the sampling point of the Dirichlet boundary condition), and the fourth boundary sampling point subtype is the Newman boundary sampling point (the sampling point of the Newman boundary condition). The sample sampling point type is the same as the target sampling point type, and can be annotated with reference to the target sampling point type annotation method, which will not be repeated here.

[0124] The boundary conditions of the conductor will change with the change of the conductor structure of the interconnection line. The number and coordinates of the boundary sampling points corresponding to different boundary conditions are also changing. In order to facilitate the representation and calculation of boundary conditions, one-hot encoding, also known as one-bit effective encoding, is used to encode these boundary conditions to obtain the discrete variable l used to represent the point cloud data annotation. j , l jLabel the boundary conditions of the jth sample sampling point to ensure that the stability and accuracy of the training can be maintained when using low-precision calculations. One-hot encoding is a method used to represent discrete variables. This method uses a multi-bit state register to encode multiple states. Each state has its own independent register bit, and only one of them is valid at any time. Discrete variables can be converted into multi-dimensional vectors through one-hot encoding, which facilitates encoding boundary conditions into the capacitance solution model. Since the point cloud subset of this application includes 6 boundary conditions, a 6-bit state register is used to encode the labeled status of all sampling points in the point cloud subset. That is, using a 6-bit state register l j,one-hoo (dimension is M×N×6) The annotation status l of all sampling points in the point cloud subset j (dimension is M×N×1) for encoding.

[0125] In step S620 of some embodiments, the sample interconnect conductor point cloud data is input into a preset model for training an unsupervised capacitance solution model. The model structure of the preset model is the same as that of the potential prediction model, both of which include a feature learning layer and a physical information layer. The preset model is used to establish a point cloud subset V i To potential U i 's mapping. Represents point cloud subset V i The potential of (conductor data window). The feature learning layer extracts the point cloud coordinates in the sample interconnect conductor point cloud data to obtain the conductor morphological features of all sampling points in the point cloud data. The sample attributes of each sample sampling point are spliced ​​with the conductor morphological features of all sampling points to obtain the splicing features of each sample sampling point. The physical information layer predicts the potential of the splicing features of all sample sampling points in the point cloud data to obtain the predicted potential of each sample sampling point. The predicted potential u of the jth sample sampling point j The value of depends not only on the point cloud coordinate x of the sample sampling point j j , boundary condition annotation j The sample properties also depend on the shape of the point cloud subset where point j is located. The mathematical expression of the predicted potential is defined as:

[0126]

[0127] in, Indicates V i The geometric feature encoding, that is, the conductor morphological features of all sample sampling points; is the preset model, and Θ is the model parameter of the preset model.

[0128] This application uses point cloud data to characterize the interconnect conductor window data to reduce memory usage. The shape and distribution of interconnect conductors vary. It is very difficult to extract features of these geometric shapes and distributions using traditional parameterization methods. It is impossible to describe the local details of the conductor structure, resulting in low accuracy of capacitance prediction. The precision and accuracy of structural feature extraction determine the upper limit of capacitance extraction accuracy. For higher-precision capacitance extraction, an accurate structural feature extraction method is required. The geometric features of point cloud data are extracted using the feature learning layer, which can be used for feature processing of conductor data of arbitrary shapes and arrangements. Compared with the existing capacitance value extraction method, no additional parameterization of the conductor structure and position is required, which is more convenient and efficient, and can be widely used in conductor interconnect data in various situations. The complexity of the conductor medium will lead to complex electric field distribution around the conductor. The potential and derivative that satisfy the electric field equation are predicted through the physical information layer, so as to calculate the capacitance of the conductor. While maintaining high accuracy, no training data is required, saving a lot of computing time and computing resources. After the model is trained, the potential of the conductor data window can be predicted according to the interconnect conductor structure, and then the capacitance of the conductor can be solved according to the potential. In addition to predicting the electric field and potential values ​​of existing conductor windows in the training set, the trained model can also predict the electric field and potential values ​​of conductor structures that have not appeared in the training set. The preset model can predict the potential of the conductor structure with high accuracy, and then extract the conductor capacitance with high accuracy.

[0129] In step S630 of some embodiments, the preset model inputs V i Mapping to U i , U i Meet V i The electric field equation on the surface, the constraints of the electric field equation are expressed as: The predicted potentials of all sample points satisfy the constraints of the electric field equation. The constraints of the boundary conditions are expressed as:

[0130] U=u0, at the main body boundary γ p ,

[0131] U=u1, at the boundary of the accompanying conductor Υ a ,

[0132] U=u D , at the boundary Υ corresponding to the Dirichlet boundary condition D ,

[0133] At the boundary corresponding to the Newman boundary condition N ,

[0134] in, represents the first-order gradient; U represents the electric potential; q represents the normal electric field strength; k represents the dielectric constant, which can be a constant or a function that varies with the solution area; is the first-order gradient of the electric potential, expressed as Represents the electric potential U i Point cloud horizontal coordinate x j The first-order partial derivative of Represents the electric potential U i The y coordinate of the point cloud j The first-order partial derivative of ; n is the external normal vector.

[0135] Based on the constraints of the above boundary conditions, if the sample sampling point type is the first boundary sampling point subtype (main conductor boundary sampling point), the reference potential is the first subpotential u0; if the sample sampling point type is the second boundary sampling point subtype (accompanying conductor boundary sampling point), the reference potential is the second subpotential u1; if the sample sampling point type is the third boundary sampling point subtype (Dirichlet boundary sampling point), the reference potential is the third subpotential u D ; If the sample sampling point type is the fourth boundary sampling point subtype (Newman boundary sampling point), the reference potential is u N , and calculate the first-order gradient of the reference potential, and take the dot product of the first-order gradient and the external normal vector as the reference normal electric field strength If the sample sampling point type is a conductor Gaussian surface sampling point or a sampling point within the calculation area, there is no need to obtain the reference potential. It should be noted that the boundary sample sampling points need to calculate the reference potential, while the conductor Gaussian surface sampling points or the sampling points within the calculation area only need to ensure that the predicted potential meets the constraints of the electric field equation, and there is no need to calculate the reference potential, making the sample interconnect conductor point cloud data sparse data, further improving the efficiency of capacitance value extraction.

[0136] In step S640 of some embodiments, in order to measure the model training performance of the preset model, loss data needs to be calculated. The present application performs loss calculation based on the predicted potential output by the preset model, the reference potential obtained by using the constraint conditions of the boundary conditions, and the residual of the electric field equation to obtain the target loss, so as to adjust the model parameters of the preset model according to the target loss.

[0137] In step S650 of some embodiments, the Adam optimizer is used to train the preset model with a fixed learning rate, such as keeping the learning rate at 0.0005 and iterating for several epochs. When the target loss drops to a smaller value, the training of the preset model is stopped and the preset model at this time is used as the potential prediction model.

[0138] Through the above steps S610 to S650, there is no need to calculate a large amount of capacitance value data required for training through the field solver, the calculation process is more concise and efficient, a large amount of computing resources and computing time can be saved, and there is no need to parameterize the conductor structure, and it has better adaptability and versatility for complex conductor structures and complex media. The processing on the conductor structure is relatively flexible, and can be used for the capacitance parameter prediction of irregular (non-Manhattan) conductors without increasing the calculation complexity. It is suitable for the parameter extraction of chip conductor interconnects of various processes, including the capacitance prediction of interconnect conductors of arbitrary shapes, distributions and complex materials. While ensuring the accuracy of capacitance prediction, the speed of capacitance prediction is greatly improved.

[0139] See also Figure 7 In some embodiments, step S640 may include but is not limited to steps S710 to S730:

[0140] Step S710, performing a first loss calculation according to the predicted potential to obtain a first sub-loss;

[0141] Step S720, performing a second loss calculation according to the predicted potential and the reference potential to obtain a second sub-loss;

[0142] Step S730, summing the first sub-loss and the second sub-loss to obtain a target loss.

[0143] In step S710 of some embodiments, the target loss includes two parts: the residual part of the electric field equation and the boundary prediction error part. The target loss is the mean square sum of the residual part of the equation of all sample sampling points in the training set and the error of the boundary condition prediction value at the sampling point. The residual part of the electric field equation is used as the first sub-loss, and the boundary prediction error part is used as the second sub-loss. i In V i The electric field equation is satisfied, so the residual part of the electric field equation is added to the loss function to encode the physical information of the electric field equation in the preset model. In order to encode the residual of the electric field equation into the preset model, it is necessary to calculate the electric potential U i The second derivative of x j and j They represent the horizontal and vertical coordinates of the point cloud of the jth sample sampling point. In order to facilitate the calculation of capacitance, the first-order derivative of the potential also needs to be calculated Based on the first and second order derivatives of the electric potential, the network parameters of the neural network can be solved by using the automatic differentiation technique in the back propagation algorithm to obtain the capacitance of the conductor.

[0144] This application uses U θ To approximate the potential in the computational region in the data window And define the residual formula based on the electric field equation, which is expressed as:

[0145]

[0146] Among them, r θ is the residual of the electric field equation; θ is the model parameter of the preset model; U θ is the predicted potential of the preset model output when the model parameter is θ; represents the second-order derivative of the electric potential; x j and j They represent the horizontal and vertical coordinates of the point cloud of the sample sampling point respectively.

[0147] The loss is calculated based on the predicted potential of each sample point to obtain the first sub-loss. The first sub-loss is the mean square of the residuals of the equations of all sample points in the training set. The calculation method of the first sub-loss is as follows:

[0148]

[0149] Where M represents the number of samples in the training set, that is, the number of point cloud subsets; N is the number of sample sampling points in each point cloud subset; i represents the i-th sample; j represents the j-th sample sampling point; x i,j and i,j They respectively represent the point cloud horizontal coordinate and point cloud vertical coordinate of the jth sample sampling point in the i-th point cloud subset.

[0150] In step S720 of some embodiments, a second loss is calculated based on the predicted potential and the reference potential of each sample sampling point to obtain a second sub-loss, which is the sum of the mean square errors of the predicted potential and the reference potential of the sample sampling points under the four boundary conditions.

[0151] In step S730 of some embodiments, the first sub-loss and the second child loss Add them together to get the target loss. The target loss is expressed as:

[0152]

[0153] in, Target loss.

[0154] In the above steps S710 to S730, the physical information of the electric field equation can be encoded in the physical information layer through the first sub-loss, so that the predicted potential output by the preset model satisfies the electric field equation. The predicted potential of the boundary condition can be made close to the reference potential through the second sub-loss. The target loss is determined through the first sub-loss and the second sub-loss, so as to optimize the preset model based on the target loss so that the preset model outputs an accurate potential.

[0155] See also Figure 8In some embodiments, step S710 may include but is not limited to steps S810 to S830:

[0156] Step S810, obtaining the second-order derivative of the predicted potential with respect to the horizontal coordinate of the point cloud to obtain a first loss value;

[0157] Step S820, obtaining the second-order derivative of the predicted potential with respect to the vertical coordinate of the point cloud to obtain a second loss value;

[0158] Step S830: determine a first sub-loss according to the first loss value and the second loss value.

[0159] In step S810 of some embodiments, in the i-th point cloud subset, if the predicted potential of the j-th sample sampling point is represented by U θ , calculate the predicted potential for the point cloud horizontal coordinate x i,j The second-order derivative of , we get the first loss value. The first loss value is expressed as

[0160] In step S820 of some embodiments, the predicted potential U is calculated. θ The y coordinate of the point cloud i,j The second-order derivative of is used to obtain the second loss value. The second loss value is expressed as

[0161] In step S830 of some embodiments, the first loss value and the second loss value are summed to obtain the electric field equation residual r of the jth sample sampling point in the i-th point cloud subset. θ , calculate the residual r of the electric field equation at the jth sample point θ The square of the electric field equation residual r for all sample points in all point cloud subsets θ The square of is summed, and the mean of the sum is calculated, and the mean is used as the first sub-loss. The mean can be obtained by dividing the sum by (M×N), where M is the number of point cloud subsets and N is the number of sample points in the point cloud subset.

[0162] Through the above steps S810 to S830, the first sub-loss can be obtained, so as to adjust the predicted potential output by the preset model based on the first sub-loss, so that the predicted potential can satisfy the electric field equation, thereby improving the accuracy of the predicted potential prediction.

[0163] See also Fig. 9 In some embodiments, step S720 may include but is not limited to steps S910 to S950:

[0164] Step S910, performing loss calculation according to the predicted potential and the reference potential of the sample sampling points of the first boundary sampling point subtype to obtain a third loss value;

[0165] Step S920, performing loss calculation according to the predicted potential and the reference potential of the sample sampling points of the second boundary sampling point subtype to obtain a fourth loss value;

[0166] Step S930, performing loss calculation according to the predicted potential and the reference potential of the sample sampling points of the third boundary sampling point subtype to obtain a fifth loss value;

[0167] Step S940, determining a predicted electric field strength according to the predicted electric potential of the sample sampling points of the fourth boundary sampling point subtype, determining a reference electric field strength according to the reference electric potential, and performing loss calculation according to the predicted electric field strength and the reference electric field strength to obtain a sixth loss value;

[0168] Step S950, summing the third loss value, the fourth loss value, the fifth loss value and the sixth loss value to obtain a second sub-loss.

[0169] In step S910 of some embodiments, the first boundary sampling point subtype is a main conductor boundary sampling point, and the predicted potential of the sample sampling point of the first boundary sampling point subtype is expressed as The reference potential of the sample sampling point of the first boundary sampling point subtype is expressed as Calculate the square of the error between the predicted potential and the reference potential at the sample sampling point. The square of the error is expressed as:

[0170]

[0171] Among them, l j,one-hot represents the unique hot encoding of the boundary condition annotation of the jth sample sampling point; k c is the nominal value for the boundary condition of the main conductor, k c =0;·represents multiplication operation.

[0172] The sum of the square errors of all sample sampling points of the first boundary sampling point subtype in all point cloud subsets is calculated, and the mean of the sum is calculated to obtain the third loss value. The third loss value is the mean square error between the predicted potential and the reference potential of the main conductor boundary sampling point. The mean can be obtained by dividing the sum by the number of main conductor boundary sampling points in the training set (M×N c,i ) is obtained. The third loss value is expressed as:

[0173]

[0174] Where M is the number of point cloud subsets in the training set; N c,i is the number of sampling points of the dominant boundary in the i-th point cloud subset.

[0175] In step S920 of some embodiments, the second boundary sampling point subtype is an accompanying conductor boundary sampling point, the square of the error between the predicted potential U of the accompanying conductor boundary sampling point and the reference potential u1 is calculated, and the square of the error of all sample sampling points of the second boundary sampling point subtype in all point cloud subsets is summed, and the sum is divided by the number of all accompanying conductor boundary sampling points in the training set (M×N a,i ), and obtain a fourth loss value. The fourth loss value is the mean square error between the predicted potential and the reference potential of the sampling point at the conductor boundary. The fourth loss value is expressed as:

[0176]

[0177] Among them, N a,i is the number of sampling points of the accompanying conductor boundary in the i-th point cloud subset; k a is the annotated value of the conductor boundary condition, k a =1.

[0178] In step S930 of some embodiments, the third boundary sampling point subtype is a Dirichlet boundary sampling point, and the predicted potential U and the reference potential u of the Dirichlet boundary sampling point are calculated. D The square of the error between the two points is calculated, and the square of the error of the Dirichlet boundary sampling points in all point cloud subsets is summed, and the sum is divided by the number of all Dirichlet boundary sampling points in the training set. The fifth loss value is obtained. The fifth loss value is the mean square error between the predicted potential and the reference potential of the Dirichlet boundary sampling point. The fifth loss value is expressed as:

[0179]

[0180] in, is the number of Dirichlet frontier sampling points in the i-th point cloud subset; k d is the notated value of the Dirichlet boundary condition, k d =2.

[0181] In step S940 of some embodiments, the fourth boundary sampling point subtype is a Newman boundary sampling point, based on the electric field strength calculation formula Calculate the predicted electric field strength based on the predicted electric potential U and the external normal vector n of the Newman boundary sampling point, and calculate the reference electric field strength based on the reference electric potential and the external normal vector. Calculate the square of the error between the predicted electric field strength and the reference electric field strength, and sum the squares of the errors of all Newman boundary sampling points in all point cloud subsets. Divide the sum by the number of all Newman boundary sampling points in the training set. And divided by the square of the Newman boundary reference length, the sixth loss value is obtained. The sixth loss value is the mean square error between the predicted electric field strength at the Newman boundary sampling point and the reference electric field strength. The sixth loss value is expressed as:

[0182]

[0183] in, is the number of Newman boundary sampling points in the i-th point cloud subset; k n is the nominal value of the Newman boundary condition, k n =3;q θ To predict the electric field strength, θ is the model parameter of the preset model; Indicates the reference electric field strength; h n is the Newman boundary reference length.

[0184] Since the characteristic scales of different boundary conditions are different, in order to ensure the calculation accuracy, the error term of the Newman boundary condition needs to be multiplied by h n -2 .

[0185] The predicted electric field strength can also be determined by a preset model based on the input vector The reference electric field strength can also be set directly without having to be derived from the reference potential.

[0186] N c,i 、N a,i , are the number of main conductor boundary sampling points, adjoint conductor boundary sampling points, Dirichlet boundary sampling points and Newman boundary sampling points, respectively, and:

[0187] N c,i =∑l j,one-hot [:,:,k c ],

[0188] N a,i =∑l j,one-hot [:,:,k a ],

[0189]

[0190] In step S950 of some embodiments, the loss values ​​of the four boundary conditions, namely the third loss value, the fourth loss value, the fifth loss value and the sixth loss value, are summed to obtain a second sub-loss.

[0191] Through the above steps S910 to S950, the second sub-loss can be obtained, so as to adjust the predicted potentials of different boundary sampling points based on the second sub-loss, thereby improving the accuracy of the predicted potential.

[0192] In step S150 of some embodiments, sampling points whose target sampling point type is a conductor Gaussian surface sampling point are determined as sampling points located on the Gaussian surface of the conductor, and the potentials of all sampling points located on the Gaussian surface of the conductor are taken as target potentials.

[0193] In step S160 of some embodiments, the capacitance calculation formula is expressed as:

[0194]

[0195] Among them, C represents capacitance; Q represents charge; and U represents voltage or potential.

[0196] The interconnect capacitance is equivalent to a multi-conductor capacitor, and the conductors that make up the capacitor are the electrodes of the capacitor. To calculate the capacitance between any two electrodes (conductors) i and j, a bias voltage U can be set between the two electrodes. ij is 1, and the charge on the electrode surface is obtained:

[0197]

[0198] According to Gauss's theorem, we can get:

[0199] Q i =∫kE·dS,

[0200] Among them, Q i represents the charge of the interconnect conductor structure i (the i-th sample); k represents the dielectric constant; E represents the electric field; dS represents the closed surface S i The tiny surface element vector on S i is the Gaussian plane of the interconnect conductor structure i; · represents a multiplication operation.

[0201] By summing the above formulas in discrete form, the approximate charge value Q of the interconnect conductor structure i can be derived: i , expressed as:

[0202]

[0203] Among them, E i,j is the normal electric field at the jth sampling point on the Gaussian surface of conductor i; k i is the dielectric constant of the medium surrounding conductor i; is the change area of ​​the Gaussian surface of conductor i; U θ,i,j is the target potential of the jth sampling point on the Gaussian surface of conductor i; N g,i is the number of sampling points on the Gaussian surface of conductor i.

[0204] Interconnect conductor structure i includes nnet i After predicting the target potential of the sampling point on the Gaussian surface of the conductor, calculate the first-order derivative of the target potential of each Gaussian surface sampling point with respect to the point cloud coordinates of the sampling point, and obtain the target normal electric field of the Gaussian surface sampling point. Substitute the target normal electric field into the above discrete charge equation to obtain the capacitance value of the conductor. The capacitance calculation formula is expressed as:

[0205]

[0206] Where n represents the external normal vector of the sampling point on the Gaussian surface; represents the target electric field at the jth sampling point on the Gaussian surface of conductor i; h g is the sampling interval length.

[0207] It should be noted that the sampling points of the conductor Gaussian surface are based on h g is the sampling point uniformly sampled on the Gaussian surface by the reference length, and n is the external normal vector of the corresponding sampling point.

[0208] Given five interconnect conductor structures, Test case 1, Test case 2, Test case 3, Test case 4 and Test case 5, the capacitance extraction of the given interconnect conductor structures is performed using the capacitance value extraction method of the present application, the BEM boundary element method and the PINN physical neural network, and the capacitance extraction results are shown in Table 1.

[0209] Table 1

[0210]

[0211] As can be seen from Table 1, the capacitance prediction method of the present application has achieved good results in various capacitance extraction tasks. In terms of capacitance prediction accuracy, the present application has achieved results comparable to the boundary element method (BEM), with an error of less than 10%. In terms of prediction speed, the present application has significantly accelerated the prediction speed compared to methods such as the BEM boundary element method (traditional method) and the PINN physical neural network.

[0212] See also Fig.10 The embodiment of the present application further provides a capacitance value extraction device based on sparse data machine learning technology, which can implement the above-mentioned capacitance value extraction method based on sparse data machine learning technology, and the capacitance value extraction device includes:

[0213] An acquisition module 1010 is used to acquire target point cloud data of a target interconnect conductor structure, wherein the target interconnect conductor structure includes a conductor, and a plurality of sampling points are arranged around the conductor, and the target point cloud data includes attribute information of each sampling point; the attribute information includes coordinate information of the sampling point and a target sampling point type;

[0214] A feature extraction module 1020 is used to extract the attribute information of all sampling points in the target point cloud data through the feature learning layer of the electric potential prediction model to obtain the conductor morphological features of the target point cloud data;

[0215] The feature stitching module 1030 is used to stitch the attribute information of each sampling point in the target point cloud data with the conductor morphological feature to obtain the target point cloud feature of each sampling point;

[0216] A feature encoding module 1040 is used to perform physical information encoding on the target point cloud features of all sampling points through the physical information layer of the potential prediction model to obtain the potential of each sampling point;

[0217] A determination module 1050 is used to determine the sampling points located on the Gaussian surface of the conductor according to the target sampling point type of each sampling point in the target point cloud data, and use the potentials of all sampling points located on the Gaussian surface of the conductor as the target potential;

[0218] The capacitance value extraction module 1060 is used to extract the capacitance value of the conductor according to the target potential.

[0219] The specific implementation of the capacitance value extraction device is substantially the same as the specific implementation of the capacitance value extraction method described above, and will not be described in detail herein.

[0220] The embodiment of the present application also provides an electronic device, the electronic device includes a memory and a processor, the memory stores a computer program, and the processor implements the above capacitance value extraction method when executing the computer program. The electronic device can be any smart terminal including a tablet computer, a car computer, etc.

[0221] See also Fig.11 , Fig.11 The hardware structure of an electronic device of another embodiment is illustrated, and the electronic device includes:

[0222] The processor 1110 may be implemented by a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (Application Specific Integrated Circuit, ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present application;

[0223] The memory 1120 can be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 1120 can store an operating system and other applications. When the technical solution provided in the embodiment of this specification is implemented by software or firmware, the relevant program code is stored in the memory 1120, and the processor 1110 calls and executes the capacitance value extraction method of the embodiment of the present application;

[0224] Input / output interface 1130, used to implement information input and output;

[0225] The communication interface 1140 is used to realize the communication interaction between the device and other devices. The communication can be realized through a wired manner (such as USB, network cable, etc.) or a wireless manner (such as mobile network, WIFI, Bluetooth, etc.);

[0226] bus 1150 , which transmits information between the various components of the device (e.g., processor 1110 , memory 1120 , input / output interface 1130 , and communication interface 1140 );

[0227] The processor 1110 , the memory 1120 , the input / output interface 1130 , and the communication interface 1140 are connected to each other in communication within the device via the bus 1150 .

[0228] An embodiment of the present application further provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the above-mentioned capacitance value extraction method is implemented.

[0229] The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory may optionally include a memory remotely disposed relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0230] The embodiments described in the embodiments of the present application are intended to more clearly illustrate the technical solutions of the embodiments of the present application and do not constitute a limitation on the technical solutions provided in the embodiments of the present application. Those skilled in the art will appreciate that with the evolution of technology and the emergence of new application scenarios, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

[0231] Those skilled in the art will appreciate that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present application, and may include more or fewer steps than shown in the figures, or a combination of certain steps, or different steps.

[0232] The device embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separated, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0233] Those skilled in the art will appreciate that all or some of the steps in the methods disclosed above, and the functional modules / units in the systems and devices may be implemented as software, firmware, hardware, or a suitable combination thereof.

[0234] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0235] It should be understood that in the present application, "at least one (item)" means one or more, and "plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the objects associated before and after are in an "or" relationship. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0236] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the above units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. The mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0237] The units described above as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0238] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0239] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including multiple instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, referred to as ROM), random access memory (Random Access Memory, referred to as RAM), disk or optical disk and other media that can store programs.

[0240] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but the scope of the rights of the present invention is not limited thereto. Any modification, equivalent substitution and improvement made by a person skilled in the art without departing from the scope and essence of the present invention should be within the scope of the rights of the present invention.

Claims

1. A capacitance value extraction method based on sparse data machine learning technology, characterized in that: The method comprises: Acquire target point cloud data of a target interconnect conductor structure, wherein the target interconnect conductor structure includes a conductor, a plurality of sampling points are arranged around the conductor, and the target point cloud data includes attribute information of each sampling point; the attribute information includes coordinate information of the sampling point and a target sampling point type; Extracting features from the coordinate information of all sampling points in the target point cloud data through a feature learning layer of the electric potential prediction model to obtain conductor morphological features of the target point cloud data; The attribute information of each sampling point in the target point cloud data is respectively spliced ​​with the conductor morphological feature to obtain the target point cloud feature of each sampling point; Performing physical information encoding on the target point cloud features of all the sampling points through the physical information layer of the potential prediction model to obtain the potential of each of the sampling points; Determine the sampling point located on the Gaussian surface of the conductor according to the target sampling point type of each sampling point in the target point cloud data, and use the potential of all sampling points located on the Gaussian surface of the conductor as the target potential; A capacitance value of the conductor is extracted according to the target potential.

2. The capacitance value extraction method according to claim 1, characterized in that: The electric potential prediction model is trained according to the following steps: Acquire sample interconnect conductor point cloud data; the sample interconnect conductor point cloud data includes sample sampling points and sample attributes of the sample sampling points; the sample attributes include sample sampling point types; Performing potential prediction on the sample attribute by using a preset model to obtain a predicted potential of the sample sampling point under the sample sampling point type; Acquire a reference potential of the sample sampling point according to the type of the sample sampling point; Perform loss calculation according to the predicted potential and the reference potential to obtain a target loss; The model parameters of the preset model are adjusted according to the target loss to obtain the electric potential prediction model.

3. The capacitance value extraction method according to claim 2, characterized in that: The step of calculating the loss according to the predicted potential and the reference potential to obtain the target loss includes: Perform a first loss calculation according to the predicted potential to obtain a first sub-loss; Perform a second loss calculation according to the predicted potential and the reference potential to obtain a second sub-loss; The first sub-loss and the second sub-loss are summed to obtain the target loss.

4. The capacitance value extraction method according to claim 3, characterized in that: The sample attribute further includes the point cloud coordinates of the sample sampling point, the point cloud coordinates include the point cloud abscissa and the point cloud ordinate, and the first loss calculation is performed according to the predicted potential to obtain the first sub-loss, including: Obtaining a second-order derivative of the predicted potential with respect to the horizontal coordinate of the point cloud to obtain a first loss value; Obtaining a second-order derivative of the predicted potential with respect to the ordinate of the point cloud to obtain a second loss value; The first sub-loss is determined according to the first loss value and the second loss value.

5. The capacitance value extraction method according to claim 3, characterized in that: The sample sampling point type includes a first boundary sampling point subtype, a second boundary sampling point subtype, a third boundary sampling point subtype and a fourth boundary sampling point subtype, and performing a second loss calculation according to the predicted potential and the reference potential to obtain a second sub-loss includes: Perform loss calculation according to the predicted potential and the reference potential of the sample sampling point of the first boundary sampling point subtype to obtain a third loss value; Perform loss calculation according to the predicted potential and the reference potential of the sample sampling point of the second boundary sampling point subtype to obtain a fourth loss value; Perform loss calculation according to the predicted potential and the reference potential of the sample sampling point of the third boundary sampling point subtype to obtain a fifth loss value; determining a predicted electric field strength according to the predicted electric potential of the sample sampling point of the fourth boundary sampling point subtype, determining a reference electric field strength according to the reference electric potential, and performing a loss calculation according to the predicted electric field strength and the reference electric field strength to obtain a sixth loss value; The third loss value, the fourth loss value, the fifth loss value and the sixth loss value are summed to obtain the second sub-loss.

6. The capacitance value extraction method according to any one of claims 1 to 5, characterized in that: The feature learning layer includes a plurality of network sub-layers connected in series in sequence. The feature learning layer of the potential prediction model extracts features from the coordinate information of all sampling points in the target point cloud data to obtain the conductor morphological features of the target point cloud data, including: Performing feature extraction on the coordinate information of all sampling points in the target point cloud data through the feature learning layer to obtain the first point cloud feature output by each network sublayer; The first point cloud features output by each network sub-layer are spliced ​​to obtain the conductor morphological features.

7. The capacitance value extraction method according to any one of claims 1 to 5, characterized in that: The physical information layer includes a first convolution layer, a residual layer, and a second convolution layer. The physical information layer of the potential prediction model encodes the target point cloud features of all the sampling points by physical information encoding to obtain the potential of each sampling point, including: Extracting the target point cloud features of all the sampling points through the first convolutional layer to obtain second point cloud features; Performing feature extraction on the second point cloud feature through the residual layer to obtain a third point cloud feature; The third point cloud feature is extracted through the second convolution layer to obtain the electric potential of each sampling point.

8. A capacitance value extraction device based on sparse data machine learning technology, characterized in that: The device comprises: an acquisition module, configured to acquire target point cloud data of a target interconnect conductor structure, wherein the target interconnect conductor structure comprises a conductor, a plurality of sampling points are arranged around the conductor, and the target point cloud data comprises attribute information of each sampling point; the attribute information comprises coordinate information of the sampling point and a target sampling point type; A feature extraction module is used to extract features from the coordinate information of all sampling points in the target point cloud data through a feature learning layer of the electric potential prediction model to obtain conductor morphological features of the target point cloud data; A feature splicing module is used to perform feature splicing on the attribute information of each sampling point in the target point cloud data and the conductor morphological feature to obtain the target point cloud feature of each sampling point; A feature encoding module, used to perform physical information encoding on the target point cloud features of all the sampling points through the physical information layer of the potential prediction model to obtain the potential of each of the sampling points; A determination module, used for determining the sampling points located on the Gaussian surface of the conductor according to the target sampling point type of each sampling point in the target point cloud data, and taking the potentials of all sampling points located on the Gaussian surface of the conductor as the target potential; A capacitance value extraction module is used to extract the capacitance value of the conductor according to the target potential.

9. An electronic device, characterized in that: The electronic device includes a memory and a processor, the memory stores a computer program, and the processor implements the capacitance value extraction method based on sparse data machine learning technology as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the capacitance value extraction method based on sparse data machine learning technology described in any one of claims 1 to 7 is implemented.

Citation Information

Patent Citations

  • Intermittent finite element method-based integrated circuit interconnection line parasitic capacitance extraction method

    CN114357942A

  • Two-dimensional capacitance extraction method based on improved finite difference method

    CN114611455A

  • Parasitic capacitance extraction method, electronic equipment and readable storage medium

    CN117892679A

  • Part detection method, device and equipment based on visual identification and storage medium

    CN117953312A

Cited By

  • Industrial internet of things distributed data processing method and middle station

    CN120321239A

  • Industrial Internet of Things distributed data processing method and middle platform

    CN120321239B

  • Urban wireless signal field intensity prediction method and system based on three-dimensional scene modeling

    CN122433353A