A chip simulation design verification method based on finite element analysis
By combining adaptive mesh generation and conditional generative adversarial network correction, the problems of computational resource overload and solution distortion in chip simulation design verification are solved, and high-precision identification of thermal stress and electromagnetic coupling interference regions is achieved, thereby improving simulation efficiency and accuracy.
Patent Information
- Application Number
- CN202611334797.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-31
- Publication Date
- 2026-11-17
AI Technical Summary
In current chip simulation design verification, traditional finite element analysis struggles to balance computational accuracy and efficiency when dealing with advanced process chips. Mesh generation and solution methods cannot accurately identify thermal stress and electromagnetic coupling interference regions, leading to computational resource overload and solution distortion.
A method combining adaptive non-uniform mesh generation, reduced-order model solving, and conditional generative adversarial networks is adopted. The adaptive mesh generation generates a fine mesh set, and the conditional generative adversarial network is used to correct the field offset of the reduced-order solution, ensuring that the physical field distribution conforms to the physical conservation laws.
It achieves high-precision multiphysics simulation under limited computing resources, accurately identifies the thermal stress and electromagnetic coupling interference regions of the chip, and improves the confidence of the simulation results and computational efficiency.
Smart Images

Figure CN123117009A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of chip simulation and verification technology, specifically to a chip simulation design and verification method based on finite element analysis. Background Technology
[0002] In chip simulation design and verification, it is often necessary to obtain the physical field distributions such as temperature field, electric potential field, and stress field inside the chip through finite element analysis to evaluate thermal reliability and signal integrity. Traditional solutions typically perform uniform or simple geometric two-level meshing on the design layout, then construct a full-order finite element stiffness matrix and solve it directly to obtain initial physical field data. Extrema points are then extracted based on this initial field as hotspots or electromagnetic interference regions. This approach struggles to balance computational accuracy and efficiency in advanced process chips with highly drastically different geometric features and significant material anisotropy. On one hand, uniform mesh refinement leads to a sharp increase in the scale of unknowns, and the storage and computational overhead of full-order solutions exceeds the capacity of a single machine. On the other hand, insufficient mesh resolution in local areas can cause gradient distortion at critical microstructures such as channels and via arrays, resulting in shifts or omissions in thermal stress and electromagnetic coupling interference regions. The physical fields obtained through direct solutions often violate fundamental laws such as the second law of thermodynamics or charge conservation due to numerical truncation errors and the lack of physical constraints, making subsequent hotspot labeling and interference determination lack physically consistent support. At the mesh generation level, existing octree or hierarchical meshing methods lack a continuous quantization mechanism for microstructural features such as curvature and linewidth gradients, failing to drive adaptive localization of refined boundaries. At the solution level, while full-order projection reduction can compress system dimensions, uncorrected reduced-order solutions exhibit pseudo-oscillations or amplitude degradation in local boundary layers and strong gradient regions, affecting the precise spatial localization of temperature and potential gradient extrema. To address these issues, this method aims to solve how, with limited computational resources, it can reduce the degrees of freedom of the solution through non-uniform adaptive discretization and correct the field offset caused by the reduced-order solution using a data-driven approach, thereby accurately delineating thermal stress hotspots and electromagnetic coupling interference regions while ensuring the correctness of multiphysics. Summary of the Invention
[0003] This paper presents a chip simulation design verification method based on finite element analysis, which balances the computational efficiency of finite element solution with the accuracy of physical field prediction, thereby improving the analysis confidence of chip thermal stress and electromagnetic coupling interference regions.
[0004] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a chip simulation design verification method based on finite element analysis. This method combines adaptive non-uniform mesh generation, reduced-order model solving, and conditional generative adversarial network physics field correction to achieve efficient and high-precision multiphysics simulation that conforms to physical conservation laws, and automatically locates thermal stress and electromagnetic interference risk areas in the chip layout. Specifically, it includes: The design layout data of the target chip is acquired, and multi-level non-uniform meshing processing is performed on the design layout data to generate an adaptive mesh set. Based on the adaptive mesh set, a finite element stiffness matrix and load vector are constructed. The finite element stiffness matrix and load vector are then solved by projection reduction using a reduced-order model solver to generate initial field distribution data. A pre-trained conditional generative adversarial network is invoked to perform physical field prediction and correction processing on the initial field distribution data to generate complete physical field distribution data of the target chip within a preset computational domain. Based on the complete physical field distribution data, the thermal stress hotspot region and electromagnetic coupling interference region of the target chip are identified, and the verification and evaluation results of the target chip are generated.
[0005] As a preferred embodiment of the present invention, the step of performing multi-level non-uniform meshing processing on the design layout data to generate an adaptive mesh set includes: extracting geometric topological features from the design layout data to identify microstructure feature data of transistor channel regions, interconnect metal line boundary regions, and via array regions; calculating a set of mesh refinement thresholds for each region of the design layout data based on the curvature change rate and linewidth gradient values in the microstructure feature data; performing octree recursive partitioning on the design layout data based on the set of mesh refinement thresholds to generate an initial multi-level mesh structure; and performing dangling node constraint processing between adjacent levels on the initial multi-level mesh structure to generate the adaptive mesh set. This solution automatically refines the mesh in key microstructure regions according to the geometric complexity of the layout, while maintaining a coarser mesh in flat regions. This ensures high-gradient physics field capture accuracy while significantly reducing the mesh cell size and decreasing the computational load of subsequent solutions.
[0006] As a preferred embodiment of the present invention, the construction of the finite element stiffness matrix and load vector based on the adaptive mesh set includes: traversing each volume element in the adaptive mesh set and extracting the shape function and spatial derivative data of the shape function for each volume element; performing numerical integration processing on the spatial derivative data and a preset anisotropic material constitutive matrix to generate the element stiffness matrix of each volume element; assembling the element stiffness matrix of each volume element according to the global node number of the adaptive mesh set to generate the finite element stiffness matrix; and calculating and generating the load vector based on the volume load density data of the integration points within each volume element and the surface load density data of the boundary surface. The anisotropic material constitutive matrix is constructed according to the material properties of the interconnect metal layer, dielectric layer, and semiconductor layer of the target chip, respectively. The material properties include the thermal conductivity tensor, electrical conductivity tensor, and coefficient of thermal expansion tensor, so that the stiffness matrix accurately reflects the thermal-electric-mechanical coupling characteristics of the multilayer heterogeneous materials of the chip. By combining a reduced-order model solver, the high-dimensional finite element system is projected onto a low-dimensional subspace for rapid solution, obtaining initial field distribution data, significantly shortening simulation time, and providing excellent initial values for subsequent physical field correction.
[0007] As a preferred technical solution of the present invention, the step of calling a pre-trained conditional generative adversarial network to perform physical field prediction correction processing on the initial field distribution data to generate complete physical field distribution data includes: using the initial field distribution data as conditional constraint data of the generator input layer of the conditional generative adversarial network, inputting it to the generator encoder for latent space feature mapping processing to generate latent space feature vectors; the latent space feature vectors are multi-scale feature fusion vectors output after being downsampled stepwise through multiple convolutional layers and pooling layers of the generator encoder, used to characterize the global topological features and local mutation features of the initial field distribution data; inputting the latent space feature vectors to the generator decoder for spatial reconstruction processing to generate reconstructed field distribution data; inputting the reconstructed field distribution data to a discriminator for physical consistency discrimination processing to generate discriminator score data; performing iterative optimization processing on the reconstructed field distribution data based on the discriminator score data, and outputting the complete physical field distribution data when the discriminator score data exceeds a preset discrimination threshold. The physical consistency discrimination process includes: the discriminator performing thermodynamic second law constraint discrimination, charge conservation constraint discrimination, and momentum balance constraint discrimination on the reconstructed field distribution data, respectively. By conditionally correcting the reduced-order approximate solution through a generative adversarial network, detailed features can be recovered and physical conservation laws can be satisfied, significantly improving the accuracy and reliability of physical field prediction while maintaining computational efficiency.
[0008] As a preferred embodiment of the present invention, the step of identifying the thermal stress hotspot region and electromagnetic coupling interference region of the target chip based on complete physical field distribution data and generating verification and evaluation results includes: performing spatial gradient field calculation processing on the complete physical field distribution data to generate temperature gradient field data and electric potential gradient field data; performing local extremum search processing on the temperature gradient field data and the electric potential gradient field data respectively to identify the temperature gradient extremum point set and the electric potential gradient extremum point set, wherein the temperature gradient extremum point set corresponds to the thermal stress hotspot region and the electric potential gradient extremum point set corresponds to the electromagnetic coupling interference region; and performing superposition and comparison processing on the geometric coordinate data of the thermal stress hotspot region and the electromagnetic coupling interference region with the design layout data to generate the verification and evaluation results. The spatial gradient field calculation process includes: performing global Laplace smoothing on the temperature field component data in the complete physical field distribution data to generate smoothed temperature field data, and performing finite difference gradient calculation on the smoothed temperature field data to generate temperature gradient field data; performing spatial Lorentz normalization on the electric potential field component data in the complete physical field distribution data to generate normalized electric potential field data, and performing piecewise cubic Hermitian interpolation between the node values of the normalized electric potential field data to generate electric potential gradient field data. The local extremum search process includes: marking continuous grid node regions with gradient magnitudes greater than a preset thermal stress threshold as thermal stress hotspot regions in the temperature gradient field data; and marking continuous grid node regions with gradient magnitudes greater than a preset electromagnetic interference threshold as electromagnetic coupling interference regions in the electric potential gradient field data. By accurately capturing the locations of drastic temperature and potential changes through smoothing and interpolation gradient calculations, and locating risk areas based on adaptive thresholds, the results, when overlaid with the layout, directly point out design weaknesses, providing precise guidance for chip design iteration.
[0009] The technical effects and advantages provided by the present invention in the above technical solution are as follows: A multi-level non-uniform mesh generation approach combined with a reduced-order model solution reduces computational complexity at both the discrete scale and algebraic solution levels. When extracting geometric and topological features from the design layout data, the mesh refinement threshold is calculated region-by-region based on the rate of curvature change and linewidth gradient values in the transistor channel region, interconnect metal line boundary region, and via array region. This drives octree recursive partitioning and dangling node constraints, forming an adaptive mesh set that is refined in gradient-dense regions and coarser in gradually varying regions. This discretization method avoids redundancy in degrees of freedom caused by global uniform refinement, significantly reducing the number of volume elements while preserving the solution accuracy at boundary layers and sharp corners. After constructing the finite element stiffness matrix and load vector based on this adaptive mesh, a reduced-order model solver is introduced to project and reduce the order of the entire system, compressing high-dimensional matrix operations into a low-dimensional subspace. The resulting initial field distribution data can be generated quickly while maintaining the overall trend. The combination of these two approaches results in computation time and memory usage increasing approximately linearly with problem size, meeting the high-resolution simulation requirements of advanced process chip layouts. By using a pre-trained conditional generative adversarial network (GAN) to correct the physical field prediction of the initial field distribution, errors introduced by the reduced-order solution in local high-gradient regions can be compensated. The network uses the initial field distribution data as generator constraints, extracts multi-scale feature fusion vectors through the encoder, reconstructs the spatial field through the decoder, and then uses a discriminator to perform physical consistency judgment on the reconstructed field based on the second law of thermodynamics, charge conservation constraints, and momentum balance constraints, iterating and optimizing until the discrimination score exceeds a preset threshold. This process ensures that the output complete physical field distribution data strictly obeys the constraints of the governing equations, eliminates spurious oscillations and amplitude decay, and enables the temperature gradient extreme point set and the electric potential gradient extreme point set to accurately converge to the actual thermal stress hotspot region and electromagnetic coupling interference region. This allows for accurate verification and evaluation results when the design layout is overlaid and compared. Attached Figure Description
[0010] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.
[0011] Figure 1 This is a flowchart of a chip simulation design and verification method based on finite element analysis; Figure 2 It is a graph showing the rate of change of curvature of microstructure features in the design layout and the distribution of linewidth gradient values; Figure 3 It is a box plot of the diagonal elements of the stiffness matrix of different material layer elements; Figure 4 This is a comparison chart of the temperature field reconstructed by a conditional generative adversarial network and the temperature field solved by the full-order finite element method. Detailed Implementation
[0012] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0013] See Figure 1 This invention provides a chip simulation design verification method based on finite element analysis. The method acquires the design layout data of the target chip, performs multi-level non-uniform mesh generation on the design layout data to generate an adaptive mesh set, constructs a finite element stiffness matrix and load vector based on the adaptive mesh set, and performs projection-reduction-order solution processing on the finite element stiffness matrix and load vector using a reduced-order model solver to generate initial field distribution data, calls a pre-trained conditional generative adversarial network to perform physical field prediction correction processing on the initial field distribution data, and generates complete physical field distribution data of the target chip within a preset computational domain, based on the complete physical field distribution data, identifies the thermal stress hotspot region and electromagnetic coupling interference region of the target chip, and generates the verification evaluation result of the target chip.
[0014] Example 1: In practice, geometric topology feature extraction processing is performed on the design layout data to identify the microstructure features of transistor channel regions, interconnect metal line boundary regions, and via array regions. Geometric topology feature extraction processing obtains graphic information from multiple layout layers by parsing the layout file of the design layout data. In the overlapping region of the polysilicon layer and the active region layer, the transistor channel region is identified through Boolean AND operations; at the edges of the interconnect metal layer patterns, interconnect metal line boundary regions are identified by extracting the pattern boundaries and searching along the boundaries; in the via layer, via patterns are identified and the via density distribution is calculated, marking continuous regions with via density exceeding a preset density threshold as via array regions. The transistor channel region, interconnect metal line boundary region, and via array region constitute the microstructure feature data.
[0015] In calculating the set of mesh refinement thresholds for each region of the design layout data based on the rate of curvature change and linewidth gradient values in the microstructure feature data, the rate of curvature change and linewidth gradient values of the geometric contour are calculated for each local region covered by the microstructure feature data. The rate of curvature change is obtained by extracting the arc-length parameterized curve of the local region's geometric contour, calculating the second derivative of the curve and the curvature, and then calculating the rate of curvature change along the arc-length direction. The linewidth gradient value is obtained by calculating the derivative of the linewidth along the extension direction of the interconnect metal line boundary region. The mesh refinement index is calculated using the rate of curvature change and the linewidth gradient value. The formula for calculating the mesh refinement index is: in, The curvature change rate is taken as the reference value, which is the maximum value of the curvature change rate of all local areas in the design layout data; This is the reference value for linewidth gradient, taken as the maximum value of the linewidth gradient of all local areas in the design layout data. This indicates the grid refinement index for the current local area. This represents the rate of change of curvature of the geometric contour of the current local region. This represents the linewidth gradient value of the current local region. and These are the weighting coefficients. The value is 0.7. The value is set to 0.3; the rationale for this value is that in the transistor channel region and via array region, the influence of curvature variation on thermal stress concentration and electromagnetic singularity is greater than that of linewidth gradient. Therefore, it is assigned the value of 0.3. Higher weighting. Refinement index for each local region's grid. The set of mesh refinement thresholds is formed by comparing the results with a preset refinement threshold range. The refinement threshold range is divided into three intervals: when... When, the corresponding local area's mesh refinement level is level three; when When, the corresponding local area's mesh refinement level is level two; when At that time, the mesh refinement level for the corresponding local area is Level 1 refinement. The mesh refinement threshold set records the mesh refinement level corresponding to each local area in the design layout data.
[0016] Based on a set of mesh refinement thresholds, the design layout data is recursively partitioned into octrees to generate an initial multi-level mesh structure. The octree recursive partitioning process uses a cuboid enclosing the entire computational domain of the design layout data as the root node. Starting from the root node, the following process is recursively executed: The mesh refinement level corresponding to the area covered by the current octree node in the set of mesh refinement thresholds is checked. If the minimum mesh size required for the current mesh refinement level is smaller than the size of the current octree node, then the current octree node is uniformly partitioned into eight child nodes along three coordinate directions. This checking and partitioning process is repeated for each child node until all child nodes meet the size requirements specified by the corresponding mesh refinement level, or the preset maximum partitioning level is reached. After completing the recursive partitioning, an octree structure containing nodes of multiple size levels is obtained, forming the initial multi-level mesh structure.
[0017] The initial multi-level mesh structure is subjected to dangling node constraint processing between adjacent levels to generate an adaptive mesh set. The dangling node constraint processing identifies dangling nodes located at the interface between adjacent elements at different refinement levels that do not share corner points with adjacent elements. For each identified dangling node, a multi-point constraint equation is constructed, representing the physical field degrees of freedom (PDO) of the dangling node as a linear combination of the PDOs of all corner nodes on the larger element surface where the dangling node is located. The coefficients of the linear combination are calculated based on the normalized position coordinates of the dangling node on the larger element surface. By applying the multi-point constraint equation, the additional degrees of freedom introduced by the dangling nodes are eliminated, ensuring that adjacent elements at different levels satisfy the physical field continuity condition at the interface. After completing all dangling node constraint processing, an adaptive mesh set is obtained, jointly defined by the octree subdivision elements and the dangling node constraint relationships.
[0018] See Figure 2 The figure shows the linewidth gradient values of each local region in three microstructure feature regions: the transistor channel region, the interconnect metal line boundary region, and the via array region. With rate of change of curvature The distribution of the line width gradient is shown on the x-axis. The range is from 0 to approximately 1.3, and the vertical axis represents the rate of change of curvature. The range is from 0 to approximately 2.2. The legend distinguishes three types of scatter symbols: the transistor channel region is represented by blue dots, the interconnect metal line boundary region is represented by orange squares, and the via array region is represented by green triangles.
[0019] As shown in the figure, the linewidth gradient values of the transistor channel region are concentrated in the range of 0.2 to 0.6, while the curvature change rate is mainly concentrated between 0.5 and 1.3, indicating a moderately high level. The linewidth gradient values of the via array region are lower, mainly distributed in the range of 0 to 0.4, while the curvature change rate is distributed in the higher range of 1.0 to 2.2, showing the characteristic of drastic changes in the geometric contour of this region. The linewidth gradient values of the interconnect metal line boundary region are higher, mainly concentrated in the range of 0.6 to 1.3, while the curvature change rate is lower, concentrated between 0 and 0.6, reflecting the characteristic of significant linewidth changes but gentle curvature changes in this region.
[0020] Example 2: In practical implementation, the process of constructing the finite element stiffness matrix and load vector based on the adaptive mesh set begins by traversing each volume element in the adaptive mesh set. The adaptive mesh set contains multiple volume elements, each defined by a set of nodes, each node having a unique global node number within the adaptive mesh set. For the currently traversed volume element, the shape function and its spatial derivative data are extracted. The shape function is an interpolation function predetermined based on the geometry of the volume element. The input of the shape function is the local coordinates in the local coordinate system of the volume element, and the output is the interpolation coefficient of the physical field value at any point inside the volume element with respect to the node values of the volume element. The spatial derivative data of the shape function is calculated by taking the partial derivative of the shape function with respect to the local coordinates and multiplying it by the inverse of the Jacobian matrix of the coordinate transformation from local coordinates to global coordinates. The spatial derivative data of the shape function constitutes the shape function derivative matrix, where each element corresponds to the partial derivative of a shape function with respect to a global coordinate component.
[0021] The anisotropic material constitutive matrix is constructed based on the material properties of the interconnect metal layer, dielectric layer, and semiconductor layer of the target chip. For regions marked as interconnect metal layers in the target chip design layout data, the elements of the anisotropic material constitutive matrix are determined by the thermal conductivity tensor, electrical conductivity tensor, and coefficient of thermal expansion tensor of the interconnect metal layer; for regions marked as dielectric layers, the elements are determined by the thermal conductivity tensor, electrical conductivity tensor, and coefficient of thermal expansion tensor of the dielectric layer; and for regions marked as semiconductor layers, the elements are determined by the thermal conductivity tensor, electrical conductivity tensor, and coefficient of thermal expansion tensor of the semiconductor layer. Each component of the thermal conductivity tensor, electrical conductivity tensor, and coefficient of thermal expansion tensor is defined in the local material coordinate system and transformed to the global coordinate system using a coordinate transformation matrix before being filled into the corresponding positions in the anisotropic material constitutive matrix. The anisotropic material constitutive matrix is a symmetric positive definite matrix, and its dimensions match the number of degrees of freedom of the physical field being solved.
[0022] After extracting the spatial derivative data of the shape functions and obtaining the anisotropic material constitutive matrix corresponding to the material region where the bulk element is located, numerical integration is performed on the spatial derivative data and the anisotropic material constitutive matrix to generate the element stiffness matrix of the bulk element. The numerical integration is implemented using Gaussian integration. A predetermined number of Gaussian integration points are selected within the bulk element. At each Gaussian integration point, the product of the transpose of the shape function derivative matrix, the anisotropic material constitutive matrix, and the shape function derivative matrix is calculated. This product is then multiplied by the integration weight of the Gaussian integration point and the determinant of the coordinate transformation Jacobian matrix. The results from all Gaussian integration points are summed to obtain the element stiffness matrix. The expression for calculating the element stiffness matrix is: in, This represents the element stiffness matrix corresponding to the current volume element; This represents the total number of Gaussian integration points within a volume element. For an eight-node hexahedral element, The value is 8; The index of the Gaussian integration point; Indicates the first The shape function derivative matrix is calculated at each Gaussian integration point, and the matrix elements are the partial derivatives of the shape function with respect to the global coordinates; for The transpose of the matrix; This is the anisotropic material constitutive matrix corresponding to the material region where the bulk element is located; Jacobian matrix for coordinate transformation The determinant value reflects the volume scaling factor from the local coordinate element to the global coordinate element; For the first The integral weighting coefficients corresponding to each Gaussian integration point; symbol Indicates all Summing the results of Gaussian integrals at each point.
[0023] After obtaining the element stiffness matrix of each solid element, the element stiffness matrix of each solid element is assembled according to the global node numbers of the adaptive mesh set to generate the finite element stiffness matrix. The assembly process is as follows: an initial all-zero global stiffness matrix is established with a dimension equal to the total number of nodes in the adaptive mesh set multiplied by the degrees of freedom of each node; for the element stiffness matrix of each solid element, the local node number of the solid element corresponding to the row and column index of each element in the element stiffness matrix is extracted, and the local node number is mapped to the global node number in the adaptive mesh set; the target row and column position of the element stiffness matrix element in the global stiffness matrix is determined according to the global node number and the degree of freedom offset, and the element stiffness matrix element is accumulated to the target row and column position of the global stiffness matrix; after all solid elements are processed, the finite element stiffness matrix is obtained.
[0024] A load vector is calculated based on the volume load density data at the integration points within each volume element and the surface load density data at the boundary surfaces. The volume load density data includes the heat generation rate density distribution of heat sources and the current source density distribution within the target chip, obtained from the chip design parameters. For each volume element, at the Gaussian integration points within the volume element, shape function interpolation and coordinate transformation are used to map the volume load density to the nodal load contribution, which is then accumulated to the corresponding position in the load vector. The boundary surface load density data includes the heat flux density and current density on the chip boundary, obtained from the boundary conditions. For volume elements located at the boundary of the computational domain, the surfaces within the volume element that belong to the boundary and have applied surface loads are identified. Surface integration is performed on these surfaces, and the surface load density is converted into nodal load contributions through integration operations related to shape functions and coordinate transformations, also accumulated to the corresponding position in the load vector. The complete load vector is obtained by superimposing the volume load contributions within the volume element and the surface load contributions at the boundary surfaces.
[0025] See Figure 3 The figure, presented as a box plot, illustrates the distribution of the average diagonal elements of the element stiffness matrix for the three material regions—interconnect metal layer, dielectric layer, and semiconductor layer—within the adaptive mesh set. The horizontal axis represents the interconnect metal layer, dielectric layer, and semiconductor layer, respectively, while the vertical axis represents the average diagonal elements of the element stiffness matrix, ranging from 0 to approximately 350,000. Each scatter point in the figure represents the specific numerical distribution of the diagonal elements of the element stiffness matrix within the corresponding material region, and the box plots reflect the median, upper and lower quartiles, and outlier ranges.
[0026] Specifically, the average value of the diagonal elements of the stiffness matrix corresponding to the interconnect metal layer is generally high, with data points mainly distributed between approximately 80,000 and 330,000, and a median of approximately 195,000. The box width and whisker length are both relatively large, indicating significant differences and distribution ranges in the stiffness matrix elements of the interconnect metal layer material region. The average value of the diagonal elements of the stiffness matrix corresponding to the dielectric layer is significantly lower than that of the interconnect metal layer, mainly concentrated in the range of 30,000 to 80,000, with a median of approximately 50,000. The box width and whisker length are relatively narrow, indicating that the values of the stiffness matrix elements in this layer are relatively concentrated and stable. The average value of the diagonal elements of the stiffness matrix of the semiconductor layer is between that of the interconnect metal layer and the dielectric layer, with a median of approximately 90,000 and a value range from approximately 20,000 to approximately 170,000, showing a relatively dispersed distribution and a moderate box width.
[0027] Example 3: In practice, a pre-trained Conditional Generative Adversarial Network (CGAN) is invoked to perform physical field prediction and correction processing on the initial field distribution data. The process of generating complete physical field distribution data relies on a pre-trained CGAN. The CGAN consists of two sub-networks: a generator and a discriminator. The generator contains a generator encoder and a generator decoder. The generator encoder receives the initial field distribution data as conditional constraint data and is composed of multiple convolutional and pooling layers stacked sequentially for progressive downsampling and extraction of multi-scale features. The first convolutional layer of the generator encoder takes the temperature, electric potential, and stress components of the initial field distribution data as input channels and extracts low-level texture features through convolution. The subsequent three convolutional and pooling layers progressively reduce the spatial size of the feature map to half, a quarter, and an eighth of the original size, respectively, while progressively expanding the number of channels to two, four, and eight times the initial number of channels. The multi-scale feature map output by the generator encoder is flattened and fully connected to form a latent space feature vector. The latent space feature vector is a multi-scale feature fusion vector, which contains both the global topological features of the initial field distribution data preserved by downsampling and the local mutation features preserved by skip connections.
[0028] The generator decoder receives latent space feature vectors and transforms them into feature tensors of the same size as the deepest feature map of the generator encoder through fully connected layers and deformation reshaping operations. Then, it gradually restores spatial resolution through progressive upsampling and deconvolution operations. After each upsampling operation, the generator decoder concatenates the upsampling result with the corresponding scale feature map from the generator encoder along the channel dimension, forming multi-scale skip connections. This allows low-level detail information to be directly passed to the reconstruction process. Finally, the generator decoder outputs reconstructed field distribution data, which includes reconstructed values of the temperature, electric potential, and stress components within the same computational domain as the initial field distribution data.
[0029] The discriminator receives the reconstructed field distribution data, performs physical consistency discrimination processing on the data, and generates discriminator score data. The physical consistency discrimination processing includes constraints based on the second law of thermodynamics, charge conservation, and momentum balance. The discriminator consists of a physical constraint calculation module and a score aggregation module. The physical constraint calculation module calculates the second law of thermodynamics, charge conservation, and momentum balance constraints at the discrete grid nodes of the reconstructed field distribution data. The second law of thermodynamics constraint is obtained by calculating whether the local entropy generation rate at each grid node satisfies the non-negativity condition; the charge conservation constraint is obtained by calculating the modulus of the current density divergence at each grid node; and the momentum balance constraint is obtained by calculating the residual modulus of the stress tensor divergence and the volume load density at each grid node. The score aggregation module maps the three constraint criteria to discriminator score data through a weighted summation, as shown in the following formula: in, As a thermodynamic constraint reference value, the maximum value of the total integral of the negative local entropy generation rate of physically consistent samples in the training dataset is taken. The maximum value of the root mean square value of the current density divergence is taken as the reference value for the charge conservation constraint, and the maximum value of the root mean square value of the current density divergence of physically consistent samples in the training dataset is taken. As a reference value for momentum balance constraints, the maximum value of the root mean square difference between the stress tensor divergence and the volume load density of physically consistent samples in the training dataset is taken. This represents the discriminator scoring data; This represents the constraint criterion term of the second law of thermodynamics, specifically the total integral of the negative region of the local entropy generation rate of the temperature field in the reconstructed field distribution data. This represents the charge conservation constraint criterion, specifically the root mean square value of the current density divergence in the reconstructed field distribution data over the entire computational domain. This represents the momentum balance constraint criterion, specifically the root mean square value of the difference between the stress tensor divergence and the volume load density in the reconstructed field distribution data over the entire computational domain. , , These are weighting coefficients. The value is 0.4. The value is 0.3. The value is set to 0.3. The weighting coefficient is determined based on the following: In the joint simulation of thermal stress and electromagnetic coupling of the chip, violations of thermodynamic constraints can lead to misjudgments of hot spots. The severity of these violations is higher than that of local violations of charge conservation constraints and momentum balance constraints. Therefore, the second law of thermodynamics constraint discrimination term is given a slightly higher weight.
[0030] The training process of the Conditional Generative Adversarial Network (GAN) employs an alternating training approach for the generator and discriminator. The training dataset consists of initial field distribution data obtained from reduced-order finite element method (FEM) solutions corresponding to multiple chip design layouts, paired with complete physical field distribution data obtained from full-order FEM solutions. The generator's training loss function is a weighted sum of reconstruction error loss and adversarial loss. The reconstruction error loss is calculated using the L1 norm between the reconstructed field distribution data and the complete physical field distribution data in the training dataset. The adversarial loss is the negative logarithm of the probability that the reconstructed field distribution data generated by the generator is physically identical to the discriminator's. The discriminator's training loss function is the L2 loss between the discriminator's score on the complete physical field distribution data in the training dataset and a perfect score, plus the L2 loss between the discriminator's score on the generator's reconstructed field distribution data and a score of zero. During training, the Adam optimizer was used with an initial learning rate of 0.0002 and a batch size of 16. Training was conducted for 200 epochs. In each epoch, the discriminator was trained once, and then the generator was trained once, until the loss function converged, resulting in a pre-trained conditional generative adversarial network.
[0031] When using a pre-trained conditional generative adversarial network (GAN) for physics field prediction correction, the initial field distribution data is used as the conditional constraint data for the generator's input layer. This data is then input to the generator encoder for latent space feature mapping to generate latent space feature vectors. In some embodiments, the initial field distribution data is normalized to a data distribution with a mean of 0 and a standard deviation of 1 before input, with the normalization parameters remaining consistent with those used in the training phase. The generator encoder calculates the output latent space feature vectors through forward propagation. Optionally, the latent space feature vectors can also be used as variables to be optimized during the inference phase for iterative optimization. Specifically, the latent space feature vectors are input to the generator decoder for spatial reconstruction to generate reconstructed field distribution data. The reconstructed field distribution data is then input to the discriminator for physical consistency discrimination to generate discriminator score data. If the discriminator score data does not exceed a preset discrimination threshold, the generator decoder parameters and discriminator parameters are fixed. Using the discriminator score data as the objective function, gradient ascent optimization is performed on the latent space feature vectors to update them. The reconstructed field distribution data is then regenerated based on the updated latent space feature vectors, and the discriminator score data is recalculated. Repeat the iterative optimization process described above. When the discriminator score data exceeds the preset discrimination threshold, stop the iteration and output the current reconstructed field distribution data as the complete physical field distribution data. The preset discrimination threshold is set to 0.95. This value is based on the fact that on the physically consistent samples in the training set, the mean of the discriminator score data is approximately 0.98, and the standard deviation is approximately 0.02. Setting the threshold slightly lower than the mean of 0.95 can ensure physical rationality while avoiding excessively long iterative optimization time.
[0032] See Figure 4The figure shows a scatter plot comparing the reconstructed temperature field values obtained by using a pre-trained conditional generative adversarial network to predict and correct the temperature fields of different material layers in the target chip design layout in Example 3, with the temperature field values obtained through full-order finite element analysis. The horizontal axis represents the temperature field value obtained through full-order finite element analysis, and the vertical axis represents the reconstructed temperature field value generated by the conditional generative adversarial network. The temperature unit is Kelvin (K). The scatter points of different colors and shapes in the figure correspond to the interconnect metal layers (red dots), dielectric layers (blue squares), and semiconductor layers (green triangles) in the chip design layout, respectively. The black dashed line in the figure is the reference line where the reconstructed temperature field value is completely consistent with the full-order finite element temperature field value under ideal conditions. .
[0033] As shown in the figure, the reconstructed temperature field scatter points for all material layers are roughly distributed near the reference line, indicating that the physical field prediction correction processing using conditional generative adversarial networks can accurately reconstruct the complete temperature field distribution of the target chip within the preset computational domain. The temperature reconstruction scatter points for the interconnect metal layer are concentrated in the temperature range of approximately 355K to 380K, with relatively clustered points and distribution basically along the reference line, indicating high accuracy in temperature field reconstruction for this layer. The temperature reconstruction scatter points for the dielectric layer are mainly distributed in the lower temperature range of 320K to 355K, with the scatter points forming a narrow band close to the reference line, indicating that the prediction correction of the dielectric layer's temperature field is relatively stable and accurate. The temperature scatter points for the semiconductor layer cover a wider temperature range of approximately 335K to 395K, with the scatter points slightly more dispersed than those for the other two layers, but still generally distributed around the reference line, demonstrating the predictive capability of the method of this invention in the high-temperature range of the semiconductor layer.
[0034] Example 4: In practical implementation, spatial gradient field calculations are performed on the complete physical field distribution data to generate temperature gradient field data and electric potential gradient field data. The spatial gradient field calculation is executed at each node of the adaptive mesh set, using the spatial derivatives of the finite element shape functions to calculate the gradient vectors of the temperature and electric potential components in the complete physical field distribution data. For the temperature component, the gradient vectors are calculated at the node of the adaptive mesh set. At the nth node, use the surrounding node... The temperature field component is calculated by weighted averaging of the shape function space derivative data of all volume elements at each node at that node. , , The partial derivatives in the three spatial coordinate directions, the three partial derivatives constitute the first... The temperature gradient vectors at all nodes. The temperature gradient vectors of all nodes constitute the temperature gradient field data. For the electric potential field component, the same gradient calculation process as for the temperature field component is used, starting at the _th ... At each node, the partial derivatives of the electric potential field components in the three spatial coordinate directions are calculated using the spatial derivative data of the shape function. These three partial derivatives constitute the first... The potential gradient vectors at each node. The potential gradient vectors of all nodes constitute the potential gradient field data.
[0035] During the local extremum search process on the temperature gradient field data and the potential gradient field data respectively, the extreme point sets of the temperature gradient and the potential gradient were identified. In the temperature gradient field data, the local extremum search process calculates the magnitude of the temperature gradient vector at each node. The formula for calculating the magnitude of the temperature gradient is: in, Represents the first in the adaptive mesh set Temperature gradient magnitude at each node; This indicates that the temperature field component in the complete physical field distribution data is at the 1st... At each node Partial derivatives in the coordinate directions; Indicates the temperature field component in the th... At each node Partial derivatives in the coordinate directions; Indicates the temperature field component in the th... At each node Partial derivatives in the coordinate directions. The calculated temperature gradient magnitude for each node. Compare with a preset thermal stress threshold. The preset thermal stress threshold is set to... The basis for this value is: within the typical operating temperature range of the chip, 320K to 380K, based on the interfacial thermal resistance coefficient and thermal expansion mismatch parameters between the copper interconnect and the low dielectric constant medium, when the temperature gradient modulus exceeds... At this point, the Mises stress at the interface will exceed the yield strength threshold of copper; therefore, this value is set as the threshold for determining the hot spot region of thermal stress. The temperature gradient modulus is then... Greater than the preset thermal stress threshold The nodes are marked as extreme points of the temperature gradient. All marked nodes constitute the set of extreme points of the temperature gradient, which corresponds to the hot spot region of thermal stress.
[0036] In the potential gradient field data, a local extremum search process similar to that used for the temperature gradient field data is performed. The magnitude of the potential gradient vector at each node is calculated, and the calculated magnitude of the potential gradient at each node is compared with a preset electromagnetic interference threshold. The preset electromagnetic interference threshold is set to... The basis for this value is: based on the typical operating voltage swing of the low-voltage differential signal line within the chip and the parasitic coupling capacitance model of adjacent interconnects, when the potential gradient magnitude exceeds... When the crosstalk voltage on adjacent interconnects exceeds 30% of the receiver's noise margin, it reaches the critical value for electromagnetic coupling interference (EMI) criteria. Therefore, this value is set as the threshold for determining the EMI region. The potential gradient magnitude is set to be greater than a preset EMI threshold. The nodes are marked as potential gradient extrema points, and all marked nodes constitute the potential gradient extrema point set, which corresponds to the electromagnetic coupling interference region.
[0037] In some embodiments, for the set of temperature gradient extrema, an octree-based neighborhood-based region growing method is used to aggregate spatially adjacent nodes whose gradient magnitudes all exceed a preset thermal stress threshold into connected temperature extrema regions. These connected temperature extrema regions are then used as the final output thermal stress hotspot regions. Similarly, for the set of potential gradient extrema, the same region growing method is used to aggregate spatially adjacent nodes whose gradient magnitudes all exceed a preset electromagnetic interference threshold into connected potential extrema regions. These connected potential extrema regions are then used as the final output electromagnetic coupling interference regions.
[0038] Verification and evaluation results are generated by overlaying and comparing the geometric coordinate data of thermal stress hotspot regions and electromagnetic coupling interference regions with the design layout data. The overlay and comparison process involves: extracting the three-dimensional spatial geometric coordinate data of all nodes in the thermal stress hotspot regions; mapping this data to the two-dimensional layout coordinate system of the design layout data to determine the projection position of the thermal stress hotspot regions on the layout plane; and identifying the metal interconnect layer, dielectric region, or semiconductor device region corresponding to the projection position by querying the layer information of the design layout data. For each thermal stress hotspot region, the verification and evaluation results record the geometric coordinate range, peak temperature gradient magnitude, involved layout layer name, and circuit network name of the thermal stress hotspot region. Simultaneously, the three-dimensional spatial geometric coordinate data of all nodes in the electromagnetic coupling interference regions are extracted, and the same mapping and query process is performed. For each electromagnetic coupling interference region, the verification and evaluation results record the geometric coordinate range, peak potential gradient magnitude, involved interconnect number, and name of the interfered signal network. The verification and evaluation results are output in the form of a structured report, which includes a list of thermal stress hotspots, a list of electromagnetic coupling interference areas, and overlay annotations of each area with the design layout data.
[0039] Example 5: In practice, global Laplace smoothing is performed on the temperature field component data in the complete physical field distribution data to generate smoothed temperature field data. Global Laplace smoothing iteratively executes the Laplace smoothing operator on all internal nodes of the adaptive mesh set. In each iteration of the smoothing operator, for the th node in the adaptive mesh set... For each internal node, the temperature field component value at that internal node is updated to the arithmetic mean of the temperature field component values at all adjacent nodes directly connected to that internal node via element edges. This iterative process continues until a preset convergence condition is met. The preset convergence condition is that the maximum absolute value of the change in the temperature field component values of all internal nodes between two consecutive iterations is less than the convergence tolerance. The convergence tolerance is set to... The basis for this value is: the temperature resolution in steady-state thermal analysis of chips is typically 100°C. Setting the convergence tolerance to the order of one-tenth of the temperature resolution ensures that the numerical error introduced by Laplace smoothing is far less than the perceptible temperature differences in the analysis. The maximum number of iterations is set to 500. When 500 iterations have been completed and the convergence condition is still not met, iteration stops, and the current result is used as the smoothed temperature field data. Nodes located on the boundary of the computational domain in the adaptive mesh set do not participate in the above update process; the temperature field component values at the boundary nodes remain the original temperature field component values in the complete physical field distribution data throughout the global Laplace smoothing process.
[0040] After obtaining smoothed temperature field data through global Laplace smoothing, finite difference gradient calculation is performed on the smoothed temperature field data to generate temperature gradient field data. The finite difference gradient calculation is performed on each node of the adaptive mesh set. For the _th ... The node is constructed with the first node. A three-dimensional local difference template centered on a node, the difference template along... Coordinate direction Coordinate direction and Each coordinate direction includes forward difference points and backward difference points. When the... When a node is located inside the computation domain, along Coordinate direction search and the first The two neighboring nodes with the smallest distance between nodes, one of which is in The coordinates are in the positive direction, and another adjacent node is in Negative coordinate direction; if two adjacent nodes (positive and negative) cannot be found simultaneously in a certain coordinate direction, then one-sided difference is used along that coordinate direction. For coordinate directions where two adjacent nodes can be obtained, the central difference scheme is used to calculate the partial derivatives; for coordinate directions where only one adjacent node can be obtained, the forward or backward difference scheme is used to calculate the partial derivatives. Temperature field components along... Formulas for calculating partial derivatives along coordinate directions, along Formulas for calculating partial derivatives along coordinate directions and along The formulas for calculating partial derivatives in the coordinate directions are all executed according to the above difference scheme. The three partial derivative arrays at the nth node become the nth node. The temperature gradient vectors at each node, together with the temperature gradient vectors of all nodes, constitute the temperature gradient field data.
[0041] In some embodiments, spatial Lorentz normalization is performed on the electric potential field component data in the complete physical field distribution data to generate normalized electric potential field data. The spatial Lorentz normalization process is as follows: First, the global maximum and global minimum values of the electric potential field component data in the complete physical field distribution data are calculated within the entire preset computational domain. Then, linear minimum-maximum normalization is performed on the electric potential field component data using the global maximum and global minimum values, mapping the electric potential field component data to... The interval. The linear minimum-maximum normalization is calculated as follows: for the ... The potential field component values at each node are subtracted from the global minimum, divided by the difference between the global maximum and global minimum, multiplied by 2, and then subtracted by 1. Next, a Lorentz nonlinear transformation is performed on the linear minimum-maximum normalized result. The formula for the Lorentz nonlinear transformation is: in, In the normalized potential field data, the first... Lorentz normalized values of the electric potential field at each node; Indicates the first The intermediate values of the potential field component data at each node after linear min-max normalization. The range of values is limited to Within the range; The scaling parameter represents the Lorentz transform. Values . Values The basis is: the scaling parameter of the Lorentz transform. Controlling the intensity of nonlinear compression, when Pick At that time, normalized median The width of the linear preservation interval near zero matches the numerical range of the potential gradient feature mutation region, which can compress the abnormal mutation amplitude of the large gradient value region in the potential gradient field data to a stable range, while maintaining the local contrast of the small gradient value region.
[0042] After obtaining the normalized potential field data, piecewise cubic Hermitian interpolation is performed between the node values of the normalized potential field data to generate the potential gradient field data. The piecewise cubic Hermitian interpolation is performed on each element edge of the adaptive mesh set. For each element edge connected in the adaptive mesh set... The node and the first For each element edge of a given node, the normalized potential field data values at the two endpoints of the element edge, along with the tangential derivatives of the normalized potential field data along the element edge direction at those endpoints, are extracted. The tangential derivatives are calculated using finite difference methods, constructing a difference scheme at the endpoints using the normalized potential field data values of neighboring nodes along the element edge direction. Using the normalized potential field data values at the two endpoints of the element edge and the tangential derivatives at those endpoints—a total of four conditions—a piecewise cubic Hermitian polynomial is constructed on that element edge. The function value and first derivative of the piecewise cubic Hermitian polynomial are continuous at any parameter position on the element edge. Along each element edge, the first derivative of the piecewise cubic Hermitian polynomial is calculated, and the result is multiplied by the unit tangent vector along the element edge direction to obtain the potential gradient vector at each position on the element edge. The potential gradient vector at each sampling point on the element edge is mapped to the corresponding grid node position of the adaptive grid set according to the spatial position of the sampling point. After completing the interpolation processing of all element edges, the potential gradient field data is generated.
[0043] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.
Claims
1. A chip simulation design verification method based on finite element analysis, characterized by, The method includes: Obtain the design layout data of the target chip, perform multi-level non-uniform mesh partitioning processing on the design layout data, and generate an adaptive mesh set; Based on the adaptive mesh set, a finite element stiffness matrix and load vector are constructed. The finite element stiffness matrix and load vector are then subjected to projection reduction solution processing by a reduced-order model solver to generate initial field distribution data. The pre-trained conditional generative adversarial network is invoked to perform physical field prediction and correction processing on the initial field distribution data, thereby generating complete physical field distribution data of the target chip within a preset computing domain; Based on the complete physical field distribution data, identify the thermal stress hotspot region and electromagnetic coupling interference region of the target chip, and generate the verification and evaluation results of the target chip.
2. The chip simulation design verification method based on finite element analysis according to claim 1, wherein, The step of performing multi-level non-uniform grid partitioning processing on the design layout data to generate an adaptive grid set includes: Geometric topology feature extraction processing is performed on the design layout data to identify the microstructure features of transistor channel region, interconnect metal line boundary region and via array region in the design layout data; Based on the curvature change rate and linewidth gradient value in the microstructure feature data, calculate the set of mesh refinement thresholds for each region of the design layout data; Based on the set of mesh refinement thresholds, the design layout data is recursively partitioned into an octree to generate an initial multi-level mesh structure. The initial multi-level mesh structure is subjected to dangling node constraint processing between adjacent levels to generate the adaptive mesh set.
3. The chip simulation design verification method based on finite element analysis according to claim 1, characterized in that, The pre-trained conditional generative adversarial network is invoked to perform physical field prediction and correction processing on the initial field distribution data, generating complete physical field distribution data of the target chip within a preset computational domain, including: The initial field distribution data is used as the conditional constraint data of the generator input layer of the conditional generative adversarial network, and is input to the generator encoder of the conditional generative adversarial network for latent space feature mapping processing to generate latent space feature vectors. The latent space feature vector is input into the generator decoder of the conditional generative adversarial network for spatial reconstruction processing to generate reconstructed field distribution data. The reconstructed field distribution data is input into the discriminator of the conditional generative adversarial network for physical consistency discrimination processing, and discriminator score data is generated. The reconstructed field distribution data is iteratively optimized based on the discriminator score data. When the discriminator score data exceeds a preset discrimination threshold, the complete physical field distribution data is output.
4. The chip simulation design verification method based on finite element analysis of claim 1, wherein, The step of identifying the thermal stress hotspot region and electromagnetic coupling interference region of the target chip based on the complete physical field distribution data, and generating the verification and evaluation results of the target chip, includes: Spatial gradient field calculations are performed on the complete physical field distribution data to generate temperature gradient field data and electric potential gradient field data. Local extremum search processing is performed on the temperature gradient field data and the electric potential gradient field data respectively to identify the temperature gradient extremum point set and the electric potential gradient extremum point set. The temperature gradient extremum point set corresponds to the thermal stress hot spot region, and the electric potential gradient extremum point set corresponds to the electromagnetic coupling interference region. The geometric coordinate data of the thermal stress hotspot area and the electromagnetic coupling interference area are overlaid and compared with the design layout data to generate the verification and evaluation results.
5. The chip simulation design verification method based on finite element analysis according to claim 2, wherein, The construction of the finite element stiffness matrix and load vector based on the adaptive mesh set includes: Traverse each volume element in the adaptive mesh set and extract the shape function and spatial derivative data of each volume element; The spatial derivative data is numerically integrated with the preset anisotropic material constitutive matrix to generate the element stiffness matrix of each volume element. The element stiffness matrix of each solid element is assembled according to the global node number of the adaptive mesh set to generate the finite element stiffness matrix; The load vector is calculated and generated based on the volume load density data of the integration points within each volume element and the surface load density data of the boundary surface.
6. The chip simulation design verification method based on finite element analysis according to claim 4, wherein, The step of performing spatial gradient field calculations on the complete physical field distribution data to generate temperature gradient field data and electric potential gradient field data includes: Global Laplace smoothing is performed on the temperature field component data in the complete physical field distribution data to generate smoothed temperature field data. Perform finite difference gradient calculation on the smoothed temperature field data to generate the temperature gradient field data; Spatial Lorentz normalization is performed on the electric potential field component data in the complete physical field distribution data to generate normalized electric potential field data; Piecewise cubic Hermitian interpolation is performed between the node values of the normalized potential field data to generate the potential gradient field data.
7. The chip simulation design verification method based on finite element analysis according to claim 3, wherein, The physical consistency determination process includes: The discriminator performs thermodynamic second law constraint discrimination, charge conservation constraint discrimination and momentum balance constraint discrimination on the reconstructed field distribution data respectively, and generates the discriminator score data.
8. The chip simulation design verification method based on finite element analysis of claim 3, wherein, The latent space feature vector is a multi-scale feature fusion vector output by downsampling through multiple convolutional and pooling layers of the generator encoder. The multi-scale feature fusion vector is used to characterize the global topological features and local mutation features of the initial field distribution data.
9. The chip simulation design verification method based on finite element analysis of claim 4, wherein, The local extremum search process includes: In the temperature gradient field data, continuous grid node regions with gradient magnitude values greater than a preset thermal stress threshold are marked as thermal stress hotspot regions; In the potential gradient field data, continuous grid node regions with gradient magnitudes greater than a preset electromagnetic interference threshold are marked as electromagnetic coupling interference regions.
10. The chip simulation design verification method based on finite element analysis of claim 5, wherein, The anisotropic material constitutive matrix is constructed based on the material properties of the interconnect metal layer, dielectric layer and semiconductor layer of the target chip, respectively. The material properties include thermal conductivity tensor, electrical conductivity tensor and coefficient of thermal expansion tensor.
