Electronic device parasitic parameter point cloud deep learning prediction method and related device

Abstract

The invention discloses an electronic device parasitic parameter point cloud deep learning prediction method and a related device. The method comprises the following steps: generating a plurality of electronic device geometric files through an electronic device packaging geometric structure; different geometries in the electronic device geometric file are combined into a connected domain, grids are divided on the surfaces of all connected domains in the electronic device geometric file according to the set maximum side length, and a graph model containing grid node coordinates and the grid connection relation between nodes is established; searching a vertex closest to the center of the convergence plane from corresponding vertexes of the connected domain where the convergence plane is located, namely a convergence vertex, searching the shortest distances from all other vertexes of the connected domain where the convergence vertex is located to the convergence vertex, and marking the shortest distances in the vertex information of the graph model; creating a directed acyclic graph based on the shortest distance mark in the vertex information of the graph model; and establishing a point cloud deep learning model, and after training is completed, inputting point clouds of the directed graph and the undirected graph corresponding to the electronic device so as to calculate and obtain a parasitic parameter network of the electronic device.

Description

Technical Field

[0001] The present invention belongs to the field of electronic device design, and specifically relates to a method for predicting electronic device parasitic parameter point clouds through deep learning and related devices. Background Art

[0002] When designing electronic devices such as printed circuit boards (PCBs), power electronics and converters, and integrated circuits, parasitic parameter extraction is often necessary to ensure signal integrity, prevent overshoot, undershoot, ringing, and crosstalk. Existing methods often use numerical calculation software for this extraction. However, when electronic devices have complex structures or a large number of network ports, numerical methods often require significant computational resources. The design of electronic devices often requires continuous adjustments to device layout and structure before parasitic parameter extraction, a cumbersome and inefficient process. Summary of the Invention

[0003] The purpose of the present invention is to provide a deep learning prediction method and related devices for electronic device parasitic parameter point clouds to solve the problem of low design efficiency caused by high computing resource requirements in the existing technology. The present invention can be trained based on the numerical calculation results of different electronic devices and has strong generalization. Based on the present invention, efficient and accurate calculation of parasitic parameter networks under different layouts and structures of electronic devices can be achieved during the design stage.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions: A method for predicting electronic device parasitic parameters through point cloud deep learning includes the following steps: S1: Generate several electronic device geometry files based on the electronic device package geometry structure. After completing the source surface, sink surface and material property marking, extract the electronic device package parasitic parameter network. Based on the package parasitic parameter network, obtain the electronic device's parasitic inductance, parasitic resistance, parasitic capacitance and parasitic conductance matrix; S2: Merge different geometric bodies in the electronic device geometry file into a connected domain, mesh all connected domain surfaces in the electronic device geometry file according to the set maximum side length, and establish a graph model containing the mesh node coordinates and the mesh connection relationship between nodes, called a structured undirected graph. The vertex of the graph model stores the mesh node coordinate information corresponding to the vertex, the normal vector direction, and the material information of the geometric body to which the vertex belongs; S3: Find the vertex closest to the center of the sink surface among the corresponding vertices in the connected domain where the sink surface is located, called the sink vertex. Find the shortest distance from all other vertices in the connected domain where the sink vertex is located to the sink vertex, and mark the shortest distance in the vertex information of the graph model. S4: Based on the shortest distance mark in the vertex information of the graph model, a directed acyclic graph is created, called a source-sink directed graph. The vertex coordinate information, vertex normal vector direction, inter-vertex connection information, and vertex material properties in the source-sink directed graph are extracted to form a directed graph point cloud. The vertex coordinate information, vertex normal vector direction, inter-vertex connection information, and vertex material properties in the structured undirected graph are extracted to form an undirected graph point cloud. S5: Establish a point cloud deep learning model. The point cloud deep learning model takes a directed graph point cloud or an undirected graph point cloud as input and is trained with the parasitic inductance and parasitic resistance matrix or the parasitic capacitance and parasitic conductance matrix of the electronic device as output. After the training is completed, the directed graph and undirected graph point cloud corresponding to the electronic device are input to calculate the parasitic parameter network of the electronic device.

[0005] Furthermore, S1 uses an automatic layout generation method or a manual geometry modification method for electronic devices to adjust the chip position, package layout, and package material parameters in the electronic device, generate several electronic device geometry files, complete the source surface, sink surface, and material property marking, and use ANSYS Q3D to extract the electronic device's package parasitic parameter network.

[0006] Furthermore, in S2, different geometric bodies in the electronic device geometry file are merged into a connected domain when the different geometric bodies meet one of the following conditions; Condition 1: Different geometric bodies are adjacent and made of the same material; Condition 2: Different geometric bodies are adjacent and are both conductors.

[0007] Furthermore, in S2, the maximum side length is set according to the size of the electronic device, and the maximum side length is less than 1 mm; The grid division adopts a triangular grid, a quadrilateral grid, a pentagonal grid or a hexagonal grid; The material information of the geometric body to which the vertex belongs is one or more of the relative magnetic permeability, relative dielectric constant, electrical conductivity, and dielectric loss tangent of the material.

[0008] Furthermore, in S2, the structured undirected graph is an unweighted graph or a weighted graph whose weights are the Euclidean distances between vertices connected by the edges of the graph; In S3, the shortest distance from all other vertices in the connected domain where the sink vertex is located to the sink vertex is the shortest distance of an unweighted graph or the shortest path of a weighted graph.

[0009] Furthermore, in S4, the shortest distance mark in the vertex information of the graph model is used to determine the retention and direction of the original undirected graph edge. The specific determination method is as follows:

[0010] in, di and d j Represents a vertex in a structured undirected graph i and j The shortest distance to the sink vertex, ⊙ represents an operator, which is one of >, ≥, <, ≤. The above formula means that if the vertex in the undirected graph is i and j The distance to the sink vertex satisfies d i ⊙ d j ,and i and j There is an edge between v i , v j ) are connected, then the edge set ε in the source-sink directed graph D There is a line from the vertex i arrive j The directed edge ( v i , v j ), the vertex set of the source-sink directed graph is the same as the vertex set of the structural undirected graph, and the coordinate information, normal vector direction and material information of the geometric body to which the vertex belongs are the same. According to this judgment method, the source-sink directed graph is determined by the structural undirected graph.

[0011] Furthermore, the point cloud deep learning model described in S5 is formed by combining the following modules: a point cloud adjacency aggregation layer, a first PointNet network, a point group interaction calculation module, a path adjacency aggregation layer, a second PointNet network, a Transformer encoder layer, a point group similarity calculation module, a multi-layer perceptron layer, and a normalization layer: The first layer of the point cloud deep learning model is a point cloud adjacency aggregation layer, the second layer is the first PointNet network, the third layer is a point group interaction calculation module and several normalization layers and multi-layer perceptron layers connected thereto, the fourth layer is a path adjacency aggregation layer, the fifth layer is the second PointNet network, followed by several Transformer encoder layers forming the sixth layer, and the last layer is a point group similarity calculation module that outputs a parasitic parameter matrix; wherein: Point cloud adjacency aggregation layer: The point cloud adjacency aggregation layer takes a directed graph point cloud or an undirected graph point cloud as input and outputs a point cloud group. Specifically, first, a number of points in the point cloud are selected as the center points of the point set in the point cloud adjacency aggregation layer. Then, if the network takes a directed point cloud as input, all k points of the sampled point set center points are sampled. i Hop Neighbor, all k i Hop neighbors represent the jump from 1 to k iAll neighbors of the jump, the center point of the point set and each k corresponding to it i The jump neighbor is determined as a point set, and all point sets are output as a point cloud group, where i = 1~n, n represents the number of sampling times; if the network takes an undirected point cloud as input, the sampling distance from the center point of the point set is r i The other points in the array are regarded as a point set, and all point sets are output as a point cloud group, where i = 1~n, and n represents the number of sampling times; Path adjacency aggregation layer: The path adjacency aggregation layer takes the connection information between vertices in the point cloud group and the directed point cloud graph as input, and takes the RL path point cloud domain or the GC path point cloud domain as output; specifically, when calculating the parasitic inductance matrix and the parasitic resistance matrix, the vertex connection information in the directed point cloud graph is input, and the path adjacency aggregation layer first finds the point closest to the center of the source surface in each connected domain, and calls it the source vertex; then finds the shortest path from all source vertices to the corresponding sink vertex, called the RL path, selects all points or several points on the RL path, called the RL path key points, and for each RL path RL i , obtain the connected domain where the RL path is located and in the RL i The point set on the sphere with a radius of r on the key points of the path or the input point cloud group in the cylinder with a radius of r in the direction of the outgoing edge of the vertex in the path is called the RL path point cloud group RLGroup i , r takes a positive real number or multiple positive real numbers. If r takes multiple positive real numbers r1, r2, ..., r n , which means that the key points of the RL path are obtained on the sphere or cylinder with diameters of r1, r2, ..., r n When calculating the parasitic capacitance matrix and the parasitic conductance matrix, the path adjacency aggregation layer searches for all point cloud groups belonging to the same conductor connected domain to form a GC path point cloud group. The GC path point cloud group corresponding to all geometric bodies is called a GC path point cloud domain. Point group similarity calculation module: The point group similarity calculation module has three forms, one is the distance metric inverse similarity calculation module, the second is the dot product similarity calculation module, and the third is the hybrid similarity calculation module; if the point group similarity calculation module is a distance metric inverse similarity calculation module or a hybrid similarity calculation module, the point group similarity calculation module takes the point cloud group as input, otherwise, the point group similarity calculation module takes the point cloud group or the point cloud domain as input, and the point group similarity calculation module takes the similarity matrix between the point cloud group or the point cloud domain as output; the feature vectors of each point set in the point cloud group or the feature vectors of each point cloud group in the point cloud domain are first output as an A sequence feature vector matrix and a B sequence feature vector matrix after passing through two different multi-layer perceptron layers respectively. The eigenvector dimensions in the two sequence feature vector matrices are the same, and the number of heads of the point group similarity calculation module is defined. The number of heads is a positive integer, which is a factor of the eigenvector dimensions in the A sequence feature vector matrix and the B sequence feature vector matrix. The distance metric inverse similarity calculation module calculates the distance Odist between the center points of each point set in the input point cloud group. ij , i, j represent different or the same point sets, and the distance between the point sets themselves is defined as a positive real number ds Or 0, when defined as a positive real number, the similarity matrix is ​​obtained by filling the inverse of the distance between the center points of the point set. When defined as 0, the inverse of the distance between the center points of all point sets is added with a positive constant es to obtain the similarity matrix. When the distance metric inverse similarity is used, the number of module heads is fixed to 1. If the point group similarity calculation module is a dot product metric similarity calculation module, the sequence feature vector matrix is ​​divided into m sequence feature vector matrices in the feature vector dimension according to the number of module heads m, which are counted as (A1, A2, A3, ... A m ) and (B1,B2,B3,…B m ), calculate A respectively i B i T Obtain a sub-similarity matrix, i = 1 ~ m, and horizontally splice the sub-similarity matrix and output it as a similarity matrix; if the point group similarity calculation module is a hybrid similarity calculation module, multiply the similarity matrix obtained by the distance metric inverse similarity calculation module of the point cloud group and each sub-similarity matrix obtained by the dot product metric similarity calculation module element by element, and then horizontally splice the resulting matrix as the similarity matrix output; Point group interaction calculation module: The point group interaction calculation module is composed of a point group similarity calculation module and other modules. If the point group similarity calculation module adopts a distance metric inverse similarity calculation module or a hybrid similarity calculation module, the point group interaction calculation module takes the point cloud group as input and the sequence feature vector matrix corresponding to the point cloud group or the point cloud domain as output. Specifically, first, the feature vectors of each point set of the point cloud group or the feature vectors of each point cloud group in the point cloud domain are output as a C feature vector sequence after passing through a multi-layer perceptron layer. The feature vector dimension is the same as the feature vector dimension of the A feature vector sequence in the point group similarity calculation module. According to the number of module heads m, the sequence feature vector matrix is ​​equally divided into m sequence feature vector matrices in the feature vector dimension, which are counted as (C1, C2, C3, ..., C m ), and then each sub-similarity matrix A i B i T and C i After matrix multiplication, horizontal splicing is performed to obtain the sequence feature vector matrix.

[0012] A point cloud deep learning prediction system for electronic device parasitic parameters, comprising: Package parasitic parameter network extraction module: used to generate several electronic device geometry files based on the electronic device package geometry structure. After completing the source surface, sink surface and material attribute marking, it extracts the electronic device package parasitic parameter network. Based on the package parasitic parameter network, it obtains the electronic device's parasitic inductance, parasitic resistance, parasitic capacitance and parasitic conductance matrix. The graphical model construction module is used to merge different geometric bodies in the electronic device geometry file into a connected domain, divide the surfaces of all connected domains in the electronic device geometry file into meshes according to the set maximum side length, and establish a graphical model containing mesh node coordinates and mesh connection relationships between nodes, called a structural undirected graph. The vertex of the graphical model stores the corresponding mesh node coordinate information, normal vector direction, and material information of the geometric body to which the vertex belongs; Marking module: It is used to find the vertex closest to the center of the sink surface among the corresponding vertices in the connected domain where the sink surface is located, called the sink vertex, find the shortest distance from all other vertices in the connected domain where the sink vertex is located to the sink vertex, and mark the shortest distance in the vertex information of the graph model; Extraction module: used to create a directed acyclic graph based on the shortest distance mark in the vertex information of the graph model, called a source-sink directed graph, extract the vertex coordinate information, vertex normal vector direction, inter-vertex connection information and vertex material properties in the source-sink directed graph to form a directed graph point cloud, and extract the vertex coordinate information, vertex normal vector direction, inter-vertex connection information and vertex material properties in the structured undirected graph to form an undirected graph point cloud; Computing module: used to establish a point cloud deep learning model. The point cloud deep learning model takes a directed graph point cloud or an undirected graph point cloud as input and is trained with the parasitic inductance and parasitic resistance matrix or the parasitic capacitance and parasitic conductance matrix of the electronic device as output. After the training is completed, the directed graph and undirected graph point cloud corresponding to the electronic device are input to calculate the parasitic parameter network of the electronic device.

[0013] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the method for deep learning point cloud prediction of parasitic parameters of electronic devices are implemented.

[0014] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a method for deep learning point cloud prediction of parasitic parameters of electronic devices.

[0015] Compared with the prior art, the present invention has the following beneficial technical effects: The structural layout of electronic devices forms complex connected loops. Existing methods need to strike a balance between accuracy, efficiency, and generalization. For example, numerical calculation methods have higher accuracy and generalization, but their efficiency is low, which will greatly extend the optimization design cycle of electronic devices; simplified modeling methods often have higher computational efficiency, but their computational accuracy is low; existing neural network methods can achieve higher accuracy and efficiency for a certain layout or a certain type of layout, but the model generalization is insufficient. The method of the present invention is based on the point cloud deep learning method to achieve a complete characterization of the geometric structure of electronic devices, and based on the proposed point cloud adjacency aggregation layer and path adjacency aggregation layer and other structures, it can realize the automatic identification of connected loops, which greatly improves the generalization of the neural network while ensuring higher computational accuracy and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The drawings in the specification are used to provide further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0017] Figure 1 Schematic diagram of the structure of SiC power devices, where (a) is geometry 1 and (b) is geometry 2; Figure 2 Examples of the generated layouts, where (a) is the random generation of layout 1 for geometric structure 1, (b) is the random generation of layout 2 for geometric structure 1, and (c) is the random generation of layout 3 for geometric structure 1; (d) is the random generation of layout 1 for geometric structure 2; (e) is the random generation of layout 2 for geometric structure 2; and (f) is the random generation of layout 3 for geometric structure 2. Figure 3 Merge and mesh the surfaces for the geometry; Figure 4 Schematic diagram of the sink vertex and the connected domain where the sink vertex is located; Figure 5 It is a schematic diagram of the shortest distance from a local vertex to a sink vertex in a connected domain; Figure 6 is a local source-sink directed graph; Figure 7 It is an example of a point cloud deep learning model architecture; Figure 8 Schematic diagram of the parasitic inductance calculation error of the point cloud deep learning model, where (a) is the error diagram of geometry 1, and (b) is the error diagram of geometry 2; Figure 9 Schematic diagram of the parasitic resistance calculation error of the point cloud deep learning model, where (a) is the error diagram of geometry 1, and (b) is the error diagram of geometry 2; Figure 10 Schematic diagram of parasitic capacitance calculation error of point cloud deep learning model, where (a) is the error diagram of geometry 1 and (b) is the error diagram of geometry 2. DETAILED DESCRIPTION

[0018] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0019] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention 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 numbers used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes 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.

[0020] Example 1 The present invention proposes a method for predicting electronic device parasitic parameters using point cloud deep learning, which specifically includes the following steps: S1: Generate a large number of electronic device geometry files through the automatic generation method of electronic device package geometry structure or manual geometry structure modification method. After completing the source surface, sink surface and material property marking, use electromagnetic field numerical analysis software to extract the high-frequency package parasitic parameter network of the electronic device.

[0021] S2: Different geometric bodies in the generated electronic device geometry file can be merged into a connected domain when the different geometric bodies meet one of the following conditions: 1) Different geometric bodies are adjacent and made of the same material; 2) Different geometric bodies are adjacent and both are conductors; All connected domain surfaces in the geometry file are meshed according to the set maximum edge length, and a graph model containing the mesh node coordinates and the mesh connection relationship between nodes is established, which is called a structural undirected graph. The vertices of the graph model store the mesh node coordinate information corresponding to the vertex, the normal vector direction, and the material information of the geometric body to which the vertex belongs.

[0022] S3: Find the vertex closest to the center of the sink surface among the mesh vertices corresponding to the connected domain where the sink surface is located, which is called the sink vertex. Use the single-source shortest path algorithm to find the shortest distance from all other vertices in the connected domain where the sink vertex is located to the sink vertex, and mark the shortest distance in the vertex information of the graph model.

[0023] S4: Based on the shortest distance labels in the graph model's vertex information, a new directed acyclic graph is created, called a source-sink directed graph. The vertex coordinate information, vertex normal vector directions, inter-vertex connectivity information, and vertex material properties in the source-sink directed graph are extracted to form a directed graph point cloud. The vertex coordinate information, vertex normal vector directions, inter-vertex connectivity information, and vertex material properties in the structured undirected graph are extracted to form an undirected graph point cloud.

[0024] S5: Establish a point cloud deep learning model. The point cloud deep learning model takes a directed graph point cloud or an undirected graph point cloud as input and the parasitic inductance and parasitic resistance matrix or the parasitic capacitance and parasitic conductance matrix of the electronic device as output. The point cloud deep learning model is composed of the following modules used individually or in combination: a point cloud adjacency aggregation layer, a first PointNet network, a first point group interaction calculation module, a path adjacency aggregation layer, a second PointNet network, a Transformer encoder layer, a second point group similarity calculation module, a multilayer perceptron (MLP) layer, and a normalization layer. The point cloud deep learning model is trained. After training, the directed graph and undirected graph point cloud corresponding to the electronic device are input to calculate the parasitic parameter network of the electronic device.

[0025] In step S1, the electronic device layout geometry file is obtained by the automatic layout generation method, and two geometric structures are selected to be displayed in Figure 1 , several generated layouts of two geometric structures are selected and shown in Figure 2 ANSYS Q3D is used to extract the package parasitic parameter network matrix of SiC power devices. Since the insulating elements in SiC power devices are made of Al2O3 material, the dielectric loss tangent is not considered. Therefore, the calculation results only include the parasitic inductance matrix, parasitic resistance matrix, and parasitic capacitance matrix.

[0026] In step S2, first, different adjacent geometric bodies composed of the same material in the generated SiC power device geometry file are merged into a connected domain. Then, different adjacent geometric bodies that are both conductors are merged into the same connected domain. The surfaces of all geometric bodies in the geometry file are divided into triangular meshes according to the set maximum side length. In this example, the maximum side length of the mesh set for metal materials is 1mm, and the maximum side length of the mesh set for non-metallic materials is 5mm. The geometry after completing the geometry merging and surface meshing is displayed in Figure 3 .

[0027] In step S3, the vertex closest to the center of the sink is found among the mesh vertices corresponding to the connected domain where the sink is located. This vertex is called the sink vertex. Figure 4 The figure shows the sink vertex (the point with the largest vertex size) and the connected domain where the sink vertex is located. The single-source shortest path algorithm is used to find the shortest distance from all other vertices of the geometric body where the sink vertex is located to the sink vertex. The shortest distance calculated here is the shortest distance of the unweighted graph (that is, assuming that the weights of the edges are all 1), and the shortest distance is marked in the vertex information of the graph model. Figure 5 The shortest distance from the local vertex of the connected domain to the sink vertex is marked on the side of the vertex.

[0028] In step S4, based on the shortest distance mark in the vertex information of the graph model, ⊙ selects >, that is, when the vertex i and j If the distance to the sink vertex satisfies d i > d j And there is an edge between i and j ( v i , v j ) are connected, then there is an edge in the source-sink directed graph that starts from vertex i arrive j The directed edge ( v i , v j ). The local source-sink directed graph created is as follows Figure 6As shown in Figure 2, the vertex coordinate information, vertex normal vector direction, inter-vertex connection information, and vertex material properties in the source-sink directed graph are extracted to form a directed graph point cloud. The vertex coordinate information, vertex normal vector direction, inter-vertex connection information, and vertex material properties in the structured undirected graph are extracted to form an undirected graph point cloud.

[0029] In step S5, a point cloud deep learning model is established. This model takes directed and undirected graph point clouds as input and outputs the parasitic inductance, parasitic resistance, and parasitic capacitance matrices of the SiC power device. Two models are required: one that takes the directed point cloud graph as input and outputs the parasitic inductance and parasitic resistance matrix, known as the RL network model; and one that takes the undirected point cloud graph as input and outputs the parasitic capacitance matrix, known as the GC network model. In this example, both the RL network model and the GC network model consist of the following modules: a point cloud adjacency aggregation layer, a first PointNet network, a point group interaction calculation module, a path adjacency aggregation layer, a second PointNet network, a Transformer encoder layer, a point group similarity calculation module, a layer normalization layer (a type of normalization layer), and a multilayer perceptron (MLP) layer. The RL network model and the GC network model take the directed point cloud graph and the undirected point cloud graph as input respectively. After passing through the point cloud adjacency aggregation layer, they are transformed into the first point cloud group. The point set in each point cloud group passes through the first PointNet network and outputs a point set feature vector. The feature vectors of all point sets are concatenated into the first point cloud group feature vector matrix. The feature vector matrix information of the first point cloud group is integrated into the first point cloud group to obtain the second point cloud group. The second point cloud group inputs the first point group interaction calculation module to output the second point cloud group feature vector. The second point cloud group feature vector is input into the layer normalization layer and the MLP layer to output the third point cloud group feature vector. The third point cloud group feature vector is integrated into the second point cloud group to obtain the third point cloud group. The third point cloud group inputs the path adjacency aggregation layer and outputs the first point cloud domain. After passing through the second PointNet network, each point cloud group and their feature vector matrix in the first point cloud domain are transformed into the fourth point cloud group feature vector. The matrix composed of all the fourth point cloud group feature vectors is called the first point cloud domain feature matrix. After the first point cloud domain feature matrix passes through the Transformer encoder layer, it outputs the second point cloud domain feature matrix. After the second point cloud domain feature matrix passes through the second point group similarity calculation module, the output result is obtained, that is, the parasitic inductance matrix and the parasitic resistance matrix or the parasitic capacitance matrix.

[0030] In the RL network model, the point group similarity calculation module in the point group interaction calculation module adopts the hybrid similarity calculation module, and the number of module heads is 1. The point group similarity calculation module adopts the dot product similarity calculation module, and the number of module heads is 2. In the GC network model, the point group similarity calculation module in the point group interaction calculation module adopts the distance metric inverse similarity calculation module, and the number of module heads is 1. The point group similarity calculation module adopts the dot product similarity calculation module, and the number of module heads is 1. The network model structure is as follows Figure 7 shown.

[0031] The deep learning model is trained. After the training is completed, the directed graph and undirected graph point cloud corresponding to the electronic device are input to calculate the parasitic parameter network of the electrical device. The parasitic inductance matrix error, parasitic resistance matrix error and parasitic capacitance matrix error obtained are as follows: Figure 8 , as shown in 9 and 10.

[0032] As can be seen, the deep learning model, trained using both the Geometry 1 and Geometry 2 datasets, can accurately calculate the parasitic parameter networks for both geometries, validating the accuracy and generalizability of the proposed method. Furthermore, the model processes a single layout in an average of approximately 0.5 seconds, making it approximately 900 times more efficient than numerical methods.

[0033] Example 2 This embodiment provides a system for predicting electronic device parasitic parameters using point cloud deep learning, including: Package parasitic parameter network extraction module: used to generate several electronic device geometry files based on the electronic device package geometry structure. After completing the source surface, sink surface and material attribute marking, it extracts the electronic device package parasitic parameter network. Based on the package parasitic parameter network, it obtains the electronic device's parasitic inductance, parasitic resistance, parasitic capacitance and parasitic conductance matrix. The graphical model construction module is used to merge different geometric bodies in the electronic device geometry file into a connected domain, divide the surfaces of all connected domains in the electronic device geometry file into meshes according to the set maximum side length, and establish a graphical model containing mesh node coordinates and mesh connection relationships between nodes, called a structural undirected graph. The vertex of the graphical model stores the corresponding mesh node coordinate information, normal vector direction, and material information of the geometric body to which the vertex belongs; Marking module: It is used to find the vertex closest to the center of the sink surface among the corresponding vertices in the connected domain where the sink surface is located, called the sink vertex, find the shortest distance from all other vertices in the connected domain where the sink vertex is located to the sink vertex, and mark the shortest distance in the vertex information of the graph model; Extraction module: used to create a directed acyclic graph based on the shortest distance mark in the vertex information of the graph model, called a source-sink directed graph, extract the vertex coordinate information, vertex normal vector direction, inter-vertex connection information and vertex material properties in the source-sink directed graph to form a directed graph point cloud, and extract the vertex coordinate information, vertex normal vector direction, inter-vertex connection information and vertex material properties in the structured undirected graph to form an undirected graph point cloud; Computing module: used to establish a point cloud deep learning model. The point cloud deep learning model takes a directed graph point cloud or an undirected graph point cloud as input and is trained with the parasitic inductance and parasitic resistance matrix or the parasitic capacitance and parasitic conductance matrix of the electronic device as output. After the training is completed, the directed graph and undirected graph point cloud corresponding to the electronic device are input to calculate the parasitic parameter network of the electronic device.

[0034] Example 3 This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the method for deep learning point cloud prediction of parasitic parameters of electronic devices are implemented.

[0035] Example 4 This embodiment provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the steps of the method for deep learning point cloud prediction of parasitic parameters of electronic devices.

[0036] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0037] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1A device that provides the functions specified in a block or multiple blocks.

[0038] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0039] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0040] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit its scope of protection. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that after reading the present invention, those skilled in the art may still make various changes, modifications or equivalent substitutions to the specific implementation methods of the invention, but these changes, modifications or equivalent substitutions are all within the scope of protection of the pending claims of the invention.

Claims

1. A method for predicting electronic device parasitic parameters through point cloud deep learning, characterized in that: The steps include: S1: Generate several electronic device geometry files based on the electronic device package geometry structure. After completing the source surface, sink surface and material property marking, extract the electronic device package parasitic parameter network. Based on the package parasitic parameter network, obtain the electronic device's parasitic inductance, parasitic resistance, parasitic capacitance and parasitic conductance matrix; S2: Merge different geometric bodies in the electronic device geometry file into a connected domain, mesh all connected domain surfaces in the electronic device geometry file according to the set maximum side length, and establish a graph model containing the mesh node coordinates and the mesh connection relationship between nodes, called a structured undirected graph. The vertex of the graph model stores the mesh node coordinate information corresponding to the vertex, the normal vector direction, and the material information of the geometric body to which the vertex belongs; S3: Find the vertex closest to the center of the sink surface among the corresponding vertices in the connected domain where the sink surface is located, called the sink vertex. Find the shortest distance from all other vertices in the connected domain where the sink vertex is located to the sink vertex, and mark the shortest distance in the vertex information of the graph model. S4: Based on the shortest distance mark in the vertex information of the graph model, a directed acyclic graph is created, called a source-sink directed graph. The vertex coordinate information, vertex normal vector direction, inter-vertex connection information, and vertex material properties in the source-sink directed graph are extracted to form a directed graph point cloud. The vertex coordinate information, vertex normal vector direction, inter-vertex connection information, and vertex material properties in the structured undirected graph are extracted to form an undirected graph point cloud. S5: Establish a point cloud deep learning model. The point cloud deep learning model takes a directed graph point cloud or an undirected graph point cloud as input and is trained with the parasitic inductance and parasitic resistance matrix or the parasitic capacitance and parasitic conductance matrix of the electronic device as output. After the training is completed, the directed graph and undirected graph point cloud corresponding to the electronic device are input to calculate the parasitic parameter network of the electronic device.

2. The method for predicting electronic device parasitic parameters using point cloud deep learning according to claim 1, wherein: In S1, the automatic layout generation method of electronic devices or the manual geometry modification method is used to adjust the position of the chip in the electronic device, the packaging layout method and the packaging material parameters, and several electronic device geometry files are generated. After completing the source surface, sink surface and material property marking, ANSYS Q3D is used to extract the electronic device packaging parasitic parameter network.

3. The method for predicting electronic device parasitic parameters using point cloud deep learning according to claim 1, wherein: In S2, different geometric bodies in the electronic device geometry file are merged into a connected domain when the different geometric bodies meet one of the following conditions; Condition 1: Different geometric bodies are adjacent and made of the same material; Condition 2: Different geometric bodies are adjacent and are both conductors.

4. The method for predicting electronic device parasitic parameters using point cloud deep learning according to claim 1, wherein: In S2, the maximum side length is set according to the size of the electronic device, and the maximum side length is less than 1 mm; The grid division adopts a triangular grid, a quadrilateral grid, a pentagonal grid or a hexagonal grid; The material information of the geometric body to which the vertex belongs is one or more of the relative magnetic permeability, relative dielectric constant, electrical conductivity, and dielectric loss tangent of the material.

5. The method for predicting electronic device parasitic parameters using point cloud deep learning according to claim 1, wherein: In S2, the structured undirected graph is an unweighted graph or a weighted graph whose weights are the Euclidean distances between vertices connected by the edges of the graph; In S3, the shortest distance from all other vertices in the connected domain where the sink vertex is located to the sink vertex is the shortest distance of an unweighted graph or the shortest path of a weighted graph.

6. The method for predicting electronic device parasitic parameters using point cloud deep learning according to claim 1, wherein: In S4, the shortest distance mark in the vertex information of the graph model is used to determine the retention and direction of the original undirected graph edge. The specific determination method is as follows: in, d i and d j Represents a vertex in a structured undirected graph i and j The shortest distance to the sink vertex, ⊙ represents an operator, which is one of >, ≥, <, ≤. The above formula means that if the vertex in the undirected graph is i and j The distance to the sink vertex satisfies d i ⊙ d j ,and i and j There is an edge between v i , v j ) are connected, then the edge set ε in the source-sink directed graph D There is a line from the vertex i arrive j The directed edge ( v i , v j ), the vertex set of the source-sink directed graph is the same as the vertex set of the structural undirected graph, and the coordinate information, normal vector direction and material information of the geometric body to which the vertex belongs are the same. According to this judgment method, the source-sink directed graph is determined by the structural undirected graph.

7. The method for predicting electronic device parasitic parameters using point cloud deep learning according to claim 1, wherein: The point cloud deep learning model described in S5 is composed of the following modules: point cloud adjacency aggregation layer, first PointNet network, point group interaction calculation module, path adjacency aggregation layer, second PointNet network, Transformer encoder layer, point group similarity calculation module, multi-layer perceptron layer and normalization layer: The first layer of the point cloud deep learning model is a point cloud adjacency aggregation layer, the second layer is the first PointNet network, the third layer is a point group interaction calculation module and several normalization layers and multi-layer perceptron layers connected thereto, the fourth layer is a path adjacency aggregation layer, the fifth layer is the second PointNet network, followed by several Transformer encoder layers forming the sixth layer, and the last layer is a point group similarity calculation module that outputs a parasitic parameter matrix; wherein: Point cloud adjacency aggregation layer: The point cloud adjacency aggregation layer takes a directed graph point cloud or an undirected graph point cloud as input and outputs a point cloud group. Specifically, first, a number of points in the point cloud are selected as the center points of the point set in the point cloud adjacency aggregation layer. Then, if the network takes a directed point cloud as input, all k points of the sampled point set center points are sampled. i Hop Neighbor, all k i Hop neighbors represent the jump from 1 to k i All neighbors of the jump, the center point of the point set and each k corresponding to it i The jump neighbor is determined as a point set, and all point sets are output as a point cloud group, where i = 1~n, n represents the number of sampling times; if the network takes an undirected point cloud as input, the sampling distance from the center point of the point set is r i The other points in the array are regarded as a point set, and all point sets are output as a point cloud group, where i = 1~n, and n represents the number of sampling times; Path adjacency aggregation layer: The path adjacency aggregation layer takes the connection information between vertices in the point cloud group and the directed point cloud graph as input, and takes the RL path point cloud domain or the GC path point cloud domain as output; specifically, when calculating the parasitic inductance matrix and the parasitic resistance matrix, the vertex connection information in the directed point cloud graph is input, and the path adjacency aggregation layer first finds the point closest to the center of the source surface in each connected domain, and calls it the source vertex; then finds the shortest path from all source vertices to the corresponding sink vertex, called the RL path, selects all points or several points on the RL path, called the RL path key points, and for each RL path RL i , obtain the connected domain where the RL path is located and in the RL i The point set on the sphere with a radius of r on the key points of the path or the input point cloud group in the cylinder with a radius of r in the direction of the outgoing edge of the vertex in the path is called the RL path point cloud group RLGroup i , r takes a positive real number or multiple positive real numbers. If r takes multiple positive real numbers r1, r2, ..., r n , which means that the key points of the RL path are obtained on the sphere or cylinder with diameters of r1, r2, ..., r n When calculating the parasitic capacitance matrix and the parasitic conductance matrix, the path adjacency aggregation layer searches for all point cloud groups belonging to the same conductor connected domain to form a GC path point cloud group. The GC path point cloud group corresponding to all geometric bodies is called a GC path point cloud domain. Point group similarity calculation module: The point group similarity calculation module has three forms, one is the distance metric inverse similarity calculation module, the second is the dot product similarity calculation module, and the third is the hybrid similarity calculation module; if the point group similarity calculation module is a distance metric inverse similarity calculation module or a hybrid similarity calculation module, the point group similarity calculation module takes the point cloud group as input, otherwise, the point group similarity calculation module takes the point cloud group or the point cloud domain as input, and the point group similarity calculation module takes the similarity matrix between the point cloud group or the point cloud domain as output; the feature vectors of each point set in the point cloud group or the feature vectors of each point cloud group in the point cloud domain are first output as an A sequence feature vector matrix and a B sequence feature vector matrix after passing through two different multi-layer perceptron layers respectively. The eigenvector dimensions in the two sequence feature vector matrices are the same, and the number of heads of the point group similarity calculation module is defined. The number of heads is a positive integer, which is a factor of the eigenvector dimensions in the A sequence feature vector matrix and the B sequence feature vector matrix. The distance metric inverse similarity calculation module calculates the distance Odist between the center points of each point set in the input point cloud group. ij , i, j represent different or the same point sets, and the distance between the point sets themselves is defined as a positive real number ds Or 0, when defined as a positive real number, the similarity matrix is ​​obtained by filling the inverse of the distance between the center points of the point set. When defined as 0, the inverse of the distance between the center points of all point sets is added with a positive constant es to obtain the similarity matrix. When the distance metric inverse similarity is used, the number of module heads is fixed to 1. If the point group similarity calculation module is a dot product metric similarity calculation module, the sequence feature vector matrix is ​​divided into m sequence feature vector matrices in the feature vector dimension according to the number of module heads m, which are counted as (A1, A2, A3, ... A m ) and (B1,B2,B3,…B m ), calculate A respectively i B i T Obtain a sub-similarity matrix, i = 1 ~ m, and horizontally splice the sub-similarity matrix and output it as a similarity matrix; if the point group similarity calculation module is a hybrid similarity calculation module, multiply the similarity matrix obtained by the distance metric inverse similarity calculation module of the point cloud group and each sub-similarity matrix obtained by the dot product metric similarity calculation module element by element, and then horizontally splice the resulting matrix as the similarity matrix output; Point group interaction calculation module: The point group interaction calculation module is composed of a point group similarity calculation module and other modules. If the point group similarity calculation module adopts a distance metric inverse similarity calculation module or a hybrid similarity calculation module, the point group interaction calculation module takes the point cloud group as input and the sequence feature vector matrix corresponding to the point cloud group or the point cloud domain as output. Specifically, first, the feature vectors of each point set of the point cloud group or the feature vectors of each point cloud group in the point cloud domain are output as a C feature vector sequence after passing through a multi-layer perceptron layer. The feature vector dimension is the same as the feature vector dimension of the A feature vector sequence in the point group similarity calculation module. According to the number of module heads m, the sequence feature vector matrix is ​​equally divided into m sequence feature vector matrices in the feature vector dimension, which are counted as (C1, C2, C3, ..., C m ), and then each sub-similarity matrix A i B i T and C i After matrix multiplication, horizontal splicing is performed to obtain the sequence feature vector matrix.

8. A deep learning prediction system for electronic device parasitic parameter point cloud, characterized in that: include: Package parasitic parameter network extraction module: used to generate several electronic device geometry files based on the electronic device package geometry structure. After completing the source surface, sink surface and material attribute marking, it extracts the electronic device package parasitic parameter network. Based on the package parasitic parameter network, it obtains the electronic device's parasitic inductance, parasitic resistance, parasitic capacitance and parasitic conductance matrix. The graphical model construction module is used to merge different geometric bodies in the electronic device geometry file into a connected domain, divide the surfaces of all connected domains in the electronic device geometry file into meshes according to the set maximum side length, and establish a graphical model containing mesh node coordinates and mesh connection relationships between nodes, called a structural undirected graph. The vertex of the graphical model stores the corresponding mesh node coordinate information, normal vector direction, and material information of the geometric body to which the vertex belongs; Marking module: It is used to find the vertex closest to the center of the sink surface among the corresponding vertices in the connected domain where the sink surface is located, called the sink vertex, find the shortest distance from all other vertices in the connected domain where the sink vertex is located to the sink vertex, and mark the shortest distance in the vertex information of the graph model; Extraction module: used to create a directed acyclic graph based on the shortest distance mark in the vertex information of the graph model, called a source-sink directed graph, extract the vertex coordinate information, vertex normal vector direction, inter-vertex connection information and vertex material properties in the source-sink directed graph to form a directed graph point cloud, and extract the vertex coordinate information, vertex normal vector direction, inter-vertex connection information and vertex material properties in the structured undirected graph to form an undirected graph point cloud; Computing module: used to establish a point cloud deep learning model. The point cloud deep learning model takes a directed graph point cloud or an undirected graph point cloud as input and is trained with the parasitic inductance and parasitic resistance matrix or the parasitic capacitance and parasitic conductance matrix of the electronic device as output. After the training is completed, the directed graph and undirected graph point cloud corresponding to the electronic device are input to calculate the parasitic parameter network of the electronic device.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the electronic device parasitic parameter point cloud deep learning prediction method as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method for deep learning point cloud prediction of parasitic parameters of electronic devices as described in any one of claims 1 to 7 are implemented.

Images