AI copper-embedded plate-thermal structure coupling layout design method

Abstract

The application relates to the technical field of computer-aided design, and particularly discloses an AI copper-embedded plate-thermal structure coupling layout design method, multiple groups of copper-embedded plate layout parameters meeting manufacturing constraints are collected, and thermal-structure coupling simulation is performed to obtain junction temperature peak values and stress peak values; density clustering is performed according to connectivity measurement between layout parameter samples, and a design space is divided into multiple feasible layout regions separated by manufacturing constraint boundaries; a local proxy mapping is constructed for each region, optimal potential values and exploration demand values of each region are extracted based on posterior distribution thereof; a two-dimensional preference pair formed by potential difference degrees and exploration difference degrees of the regions is taken as a decision basis to divide potential preference areas or exploration preference areas, a region to be developed is selected, a candidate layout point is selected, simulation is performed, sampling data is updated, and the process is repeated until convergence; after non-dominated front solutions of the regions are combined, a final layout scheme is output; and the application realizes efficient global layout optimization under a discrete feasible region with process constraints.

Description

Technical Field

[0001] This invention relates to the field of computer-aided design technology, specifically to an AI-embedded copper plate-thermal structure coupled layout design method. Background Technology

[0002] In the field of electronic packaging substrate design, the layout of copper-embedded substrates directly affects the heat dissipation efficiency and structural reliability of power devices. Its design must simultaneously consider the optimization of heat conduction paths and the control of thermal expansion mismatch stress, which is a typical multi-objective coupled layout design problem. Existing computer-aided design methods usually adopt parametric modeling combined with finite element simulation. A global surrogate model is used to fit the response surface of the design space, and then evolutionary algorithms or active learning strategies are used to search for the optimal layout scheme in the global space.

[0003] In copper-embedded panel layout design, manufacturing constraints cause the design space to be discretized into multiple unconnected feasible layout regions. Existing auxiliary design methods based on global surrogate models treat all design variables as a continuous and unified space. Sample points across regions are incorrectly included in the same covariance structure for interpolation fitting due to their close spatial distance, resulting in severe distortion of the response surface of the surrogate model on both sides of the restricted area boundary. Active learning sampling strategies continue to oversample in fully sampled local regions based on global uncertainties, making it difficult to cross constraint boundaries to perceive the exploration value of other potential feasible regions. As a result, the optimization process cannot converge to the global optimal frontier within the limited simulation budget, and the reliability and design efficiency of global optimization are insufficient. Summary of the Invention

[0004] The purpose of this invention is to provide an AI copper-embedded board-thermal structure coupling layout design method to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions: AI-embedded copper board - thermal structure coupling layout design method, including the following steps: S1. Collect multiple sets of AI copper-embedded board layout parameters that meet manufacturing constraints, and perform thermal-structural coupling simulation on each set of layout parameters to obtain the corresponding junction temperature peak and stress peak. S2, based on the connectivity between layout parameters, performs density clustering to divide the design space into multiple feasible layout regions; S3. For each feasible layout region, construct a local proxy map using the responses of the sampled points, and extract the optimal potential value and exploration demand value of the feasible layout region based on the posterior distribution of the local proxy map. S4. Based on the optimal potential value and exploration requirement value of each feasible layout region, select the region to be developed. Within the region to be developed, select candidate layout points for simulation based on local proxy mapping, and update the sampling data of the region to be developed. S5, repeat S2 to S3 until convergence, merge the frontier solutions of each region and output the final layout scheme.

[0006] As a further aspect of the present invention: obtaining the corresponding junction temperature peak and stress peak specifically includes: Based on each set of layout parameters, establish the geometric configuration and apply thermal loads and constraints, then solve the heat conduction equation to obtain the initial temperature field. The thermal-mechanical sensitivity of each element is calculated by multiplying the temperature gradient of each region by the coefficient of thermal expansion of the material. Based on the thermal-mechanical sensitivity, the interpolation weights of high-gradient regions are refined, and the load mapping is reconstructed. The reconstructed thermal load is transferred step by step to the equilibrium equations to solve the stress field; After convergence, the maximum global temperature and the maximum equivalent stress are extracted as the junction temperature peak and stress peak, respectively.

[0007] As a further aspect of the present invention: the reconstructed load mapping specifically includes: Based on thermo-mechanical sensitivity, interpolation weights are added to interpolation nodes in high gradient regions where the temperature gradient exceeds a preset threshold, and the thermal load mapping from the temperature field to the stress field is reconstructed in a weighted manner.

[0008] As a further aspect of the present invention: the division of the design space into multiple feasible layout areas specifically includes: Calculate the shortest continuous path length for every two layout parameter samples to bypass the manufacturing constraint boundary within the feasible region, and use the shortest continuous path length as a connectivity measure. Using connectivity metrics as input, local density peaks are identified and sample clusters are divided, with each cluster defining an independent feasible layout region.

[0009] As a further aspect of the present invention: the identification of local density peaks and the division of sample clusters specifically includes: Based on the connectivity metric between each layout sample and other samples in its neighborhood, the local connectivity density of each sample is calculated. Based on the local connectivity density of each sample and the shortest connected path length between it and samples with higher density, the deviation judgment value of each sample is calculated, and samples with deviation judgment values ​​greater than a preset threshold are marked as peak points. Using each peak point as the center, the remaining samples are assigned to the cluster containing the peak point with the minimum connected path length, thus completing the sample clustering.

[0010] As a further aspect of the present invention: the extraction of the optimal potential value and exploration requirement value of the feasible layout region specifically includes: Based on the posterior distribution, generate predicted response values ​​and prediction variances for multiple sets of candidate layout points; For each candidate layout point, non-dominated sorting is performed with the predicted response value and predicted variance as two independent objectives, and the two extreme points on the non-dominated frontier are extracted. The value with the smallest predicted response between the two extreme points is taken as the optimal potential value, and the value with the largest predicted variance is taken as the exploration demand value.

[0011] As a further aspect of the present invention: the extraction of the two extreme points on the non-dominated front specifically includes: Compare the predicted response value and predicted variance of each candidate layout point one by one. Points whose predicted response value and predicted variance are worse than any other point are marked as dominated points. Undominated points form the undominated frontier. In the non-dominated frontier, the minimum predicted response value and the maximum predicted variance are used as screening objectives, respectively. The point with the minimum predicted response value is taken as the first extreme point, and the point with the maximum predicted variance is taken as the second extreme point, thus completing the extraction of the two extreme points.

[0012] As a further aspect of the present invention: the updating of the sampling data of the area to be developed specifically includes: Based on the difference between the optimal potential value and the exploration demand value of each feasible layout region and the corresponding global extreme value, a two-dimensional preference pair consisting of the potential difference degree and the exploration difference degree is generated for each feasible layout region. Based on the relative magnitude of the two differences in the two-dimensional preference pair, the feasible layout area is divided into a potential preference area or an exploratory preference area. The region with the smallest potential difference among the regions classified as potential preference zones will be selected as the region to be developed, or the region with the largest exploration difference among the regions classified as exploration preference zones will be selected as the region to be developed. Within the selected development area, the candidate layout point with the smallest predicted response value is selected based on its local proxy mapping to perform simulation, and the newly added sampled data is incorporated into the existing sampled data of the feasible layout area.

[0013] As a further aspect of the present invention: selecting the region with the largest exploration difference among the regions divided into exploration preference areas as the region to be developed specifically includes: The potential difference and exploration difference of each feasible layout region are normalized to obtain normalized potential value and normalized exploration value; Calculate the difference between the normalized potential value and the normalized exploration value; Then, regions with values ​​greater than zero are identified as potential preference regions, and regions with differences less than zero are identified as exploration preference regions.

[0014] As a further aspect of the present invention: the final output layout scheme specifically includes: Collect local frontier solutions for each feasible layout region, and summarize all local frontier solutions into a global candidate set; For all front solutions in the global candidate set, compare their junction temperature peak value and stress peak value one by one, eliminate the dominated solutions whose junction temperature and stress are inferior to any other front solution, and retain the undominated front solutions to form the global optimal front set. From the globally optimal frontier set, the one with the smallest junction temperature peak or the smallest stress peak is selected as the final layout scheme output.

[0015] The beneficial effects of this invention are: (1) This invention performs density clustering by calculating the connectivity measure between layout parameter samples based on manufacturing constraint boundaries, dividing the design space into multiple feasible layout regions that are separated from each other, so that the sample distribution in each region has clear topological boundaries, effectively avoiding the response surface distortion problem caused by cross-region interpolation in the discrete feasible domain of the traditional global proxy model.

[0016] (2) This invention extracts the optimal potential value and exploration demand value of each region to form a two-dimensional preference pair. Based on the normalized difference, the region is divided into a potential preference area or an exploration preference area. Based on this, the region to be developed is dynamically selected, so that the limited computing resources are allocated differently according to the actual development value and exploration value of each region. This avoids the problem of local regions being oversaturated and other regions being underexplored due to global equal weight sampling. Attached Figure Description

[0017] The invention will now be further described with reference to the accompanying drawings.

[0018] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a flowchart of the reconstructed load mapping in this invention; Figure 3 This is a flowchart illustrating how the design space is divided into multiple feasible layout areas in this invention. Detailed Implementation

[0019] 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, and 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.

[0020] Please see Figure 1 As shown, this invention is an AI copper-embedded board-thermal structure coupled layout design method, including the following steps: S1. Collect multiple sets of AI copper-embedded board layout parameters that meet manufacturing constraints, and perform thermal-structural coupling simulation on each set of layout parameters to obtain the corresponding junction temperature peak and stress peak. S2, based on the connectivity between layout parameters, performs density clustering to divide the design space into multiple feasible layout regions; S3. For each feasible layout region, construct a local proxy map using the responses of the sampled points, and extract the optimal potential value and exploration demand value of the feasible layout region based on the posterior distribution of the local proxy map. S4. Based on the optimal potential value and exploration requirement value of each feasible layout region, select the region to be developed. Within the region to be developed, select candidate layout points for simulation based on local proxy mapping, and update the sampling data of the region to be developed. S5, repeat S2 to S3 until convergence, merge the frontier solutions of each region and output the final layout scheme.

[0021] Please see Figure 2 As shown, in S1, multiple sets of AI copper-embedded board layout parameters that meet manufacturing constraints are collected, and thermal-structural coupling simulation is performed on each set of layout parameters to obtain the corresponding junction temperature peak and stress peak, specifically including: The specific steps for collecting multiple sets of copper-embedded board layout parameters that meet manufacturing constraints are as follows: In industrial design software, for the copper-embedded board structure to be optimized, the length, width, height of the copper blocks, as well as their lateral and longitudinal offsets within the substrate plane, are used as variable dimensions of the layout parameters. The minimum spacing between adjacent copper blocks is set to 0.3 mm, and the minimum distance from the edge of the copper block to the edge of the substrate is set to 0.5 mm as manufacturing constraint boundaries. The Latin hypercube sampling method is used to generate multiple sets of candidate layout parameters in the design space. Each set of parameters satisfies the above spacing and edge constraints, forming the initial sample set.

[0022] The specific steps for performing thermal-structural coupling simulation for each set of layout parameters are as follows: Based on the current group layout parameters, an assembly model containing the geometry of copper blocks and substrates is established in the finite element preprocessing environment. Heat flux density load and mechanical fixed constraint conditions consistent with actual working conditions are applied to the assembly model. Based on this, a finite element method (FEM) is used to discretize and solve the heat conduction equation. The solution process for the heat conduction equation is as follows: The domain to be solved is divided into multiple discrete elements. A balance relationship between temperature and heat flux density is established for each element. The temperature values ​​of each node are solved iteratively until the temperature residual of the entire domain converges to the preset standard, thus obtaining the initial temperature field distribution.

[0023] After obtaining the initial temperature field distribution, the specific steps for calculating the thermo-mechanical sensitivity of each element are as follows: Extract the temperature gradient value of each element in the initial temperature field. This temperature gradient value is the rate of temperature change of the element in three spatial directions. The temperature gradient value is multiplied by the coefficient of thermal expansion of the material at the location of the element, and the resulting product is defined as the thermo-mechanical sensitivity of the element.

[0024] The coefficient of thermal expansion of copper is set to be [value missing]. The coefficient of thermal expansion of the silicon nitride substrate material is set to per Kelvin. For each Kelvin, the material properties at each location are determined based on the material type of that location in the geometric configuration.

[0025] The specific operation for reconstructing the load mapping based on thermal-mechanical sensitivity is as follows: A temperature gradient threshold is set, which is determined based on the statistical upper quartile of the distribution of temperature gradient values ​​of all units. Regions with temperature gradient values ​​exceeding this threshold are marked as high gradient regions. For interpolation nodes in high gradient regions, the interpolation weight is increased based on their thermal-mechanical sensitivity values, with the interpolation weight increasing by 0.3 to 0.5 units for each unit increase in sensitivity value. For interpolation nodes that are not marked as high gradient regions, their original interpolation weights remain unchanged; The weights of all interpolation nodes are normalized to bring the total interpolation weights of the entire domain to the same order of magnitude, thus reconstructing the thermal load mapping from the temperature field to the stress field.

[0026] The specific operation of using the reconstructed thermal load mapping as input and progressively passing it to the stress balance equation for solution is as follows: In the process of solving the equilibrium equation of the stress field, the reconstructed temperature values ​​of each node are first distributed to the stress field integration point according to the normalized weight. The temperature value at each integration point is equal to the sum of the products of the temperature values ​​of the surrounding interpolated nodes and the normalized weight of that node. Then, the temperature value at each integration point is multiplied by the thermal expansion coefficient of the material at that point to obtain the free strain value at that point. The free strain value is introduced into the stress equilibrium equation as the initial strain component. Combined with the mechanical fixed constraint conditions, the stress distribution in each discrete element is solved. After multiple iterations, the displacement field residual and the stress field residual converge to the preset standard, and the stress field distribution is obtained.

[0027] The convergence criteria used in solving the stress balance equation are: the maximum displacement change over the entire domain between two consecutive iterations is less than 0.001 mm and the maximum stress change over the entire domain is less than 1 MPa.

[0028] After both the heat conduction equation and the stress balance equation converge, the maximum value in the global temperature field is extracted as the junction temperature peak value corresponding to the set of layout parameters, and the maximum value of the equivalent stress in the global stress field is extracted as the stress peak value corresponding to the set of layout parameters.

[0029] The equivalent stress is calculated as follows: the stress values ​​in the three principal stress directions of each element are combined according to the von Mises equivalence criterion, and the combined result is used as the equivalent stress value of the element.

[0030] The current group layout parameters are associated with and stored with the corresponding junction temperature peak and stress peak, serving as simulation response labels for the group parameters, thus completing the coupled simulation process of the current group layout parameters.

[0031] Repeat all the above operations for each set of layout parameters in the initial sample set to obtain the junction temperature peak and stress peak corresponding to each set of layout parameters.

[0032] Please see Figure 3 As shown, in S2, density clustering is performed based on the connectivity between layout parameters to divide the design space into multiple feasible layout regions, specifically including: The specific operation for calculating the shortest consecutive path length of every two layout parameter samples around the manufacturing constraint boundary within the feasible region is as follows: All layout parameter sample points that meet manufacturing constraints are calibrated in the design space. The geometric boundary enclosed by the minimum spacing between adjacent copper blocks of 0.3 mm and the minimum distance from the edge of the copper block to the edge of the substrate of 0.5 mm is used as the non-crossable constraint boundary. The area in the design space outside the constraint boundary is defined as the feasible region. For any two layout parameter sample points, the path search method is used within the feasible region to find the shortest continuous path connecting the two sample points without crossing any constraint boundaries. The length of the shortest continuous path is obtained by discretizing the path into multiple small straight line segments and accumulating the lengths of each segment. The accumulated value is used as the connectivity measure between the two sample points.

[0033] The specific operation for identifying local density peaks using connectivity metrics as input is as follows: Set a neighborhood radius, which is 0.3 times the arithmetic mean of the connectivity metrics among all sample pairs; Based on the connectivity metric between each sample and other samples falling within its neighborhood radius, the local connectivity density of each sample is calculated using the following formula: ; in, Indicates the first Local connectivity density of each sample Indicates falling into the first The set of all other samples within the neighborhood radius of a given sample. Indicates the first The sample and the first A measure of connectivity between samples; According to this formula, the shorter the connection path length between a sample and its surrounding samples, the greater the local connectivity density value it obtains.

[0034] After obtaining the local connectivity density of each sample, the specific operation for calculating the deviation judgment value is as follows: For each sample, among all other samples whose local connectivity density is greater than that sample, select the sample with the smallest connectivity metric with that sample, and record the smallest connectivity metric as the density-distance value of that sample. If there are no other samples with a local connectivity density greater than that of the sample, then the density-distance value of the sample is set to the maximum value among all sample pairs in the connectivity metric. The deviation judgment value is calculated using the following formula: ; in, Indicates the first The deviation judgment value of each sample, This represents the local connectivity density value of the sample. This is the density-distance value for the sample.

[0035] According to this formula, the greater the local connectivity density of a sample and the greater the distance between it and samples with higher density, the greater its deviation from the judgment value.

[0036] A deviation judgment threshold is set, and the threshold is determined as follows: Arrange all the deviation judgment values ​​of the samples in descending order, and take the deviation judgment values ​​in the top 5% of the arrangement as the threshold; Samples that deviate from the judgment value by more than the threshold are marked as peak points, and each peak point represents the cluster center of each cluster within the feasible region.

[0037] After completing the peak point labeling, the specific steps for sample cluster partitioning are as follows: For each non-peak point sample, calculate the connectivity metric between the sample and each peak point, select the peak point with the smallest connectivity metric with the sample, and assign the sample to the cluster of the peak point. The above-mentioned inclusion operation is performed on all non-peak point samples one by one until all samples are assigned to the cluster where their corresponding peak points are located. Thus, each cluster defines an independent feasible layout region. Each feasible layout region is separated by an inaccessible region formed by manufacturing constraint boundaries. Layout parameter samples in different clusters are not connected to each other.

[0038] The partitioning results are output, and the layout parameters of all samples within each feasible layout region, as well as the corresponding peak junction temperature and peak stress response data, are stored in association.

[0039] In S3, for each feasible layout region, a local proxy map is constructed using the responses of the sampled points. Based on the posterior distribution of the local proxy map, the optimal potential value and exploration requirement value of the feasible layout region are extracted, specifically including: The specific steps for generating predicted response values ​​and predicted variances for multiple candidate layout points based on the posterior distribution are as follows: For each feasible layout region, the layout parameters of the sampled points in the region are used as input, and the junction temperature peak and stress peak corresponding to the sampled points in the region are used as output. In the subspace spanned by the layout parameters of the sampled points, the Gaussian process regression method is used to construct the mapping relationship between the layout parameters and the junction temperature peak, as well as the mapping relationship between the layout parameters and the stress peak. This mapping relationship is characterized by a mean function and a covariance function, where the mean function reflects the central estimate of the predicted response value at the candidate layout point, and the covariance function reflects the degree of uncertainty of the predicted response value. In each feasible layout region, the range enclosed by the minimum and maximum values ​​of each layout parameter variable dimension in the region is selected as the candidate space. Multiple sets of candidate layout points are generated in the candidate space by sampling at equal intervals. The number of candidate layout points is set to twice the number of sampled points in the region. For each candidate layout point, the predicted values ​​of the junction temperature peak and stress peak are output according to the mean function of the posterior distribution of the Gaussian process regression, and the predicted variances of the junction temperature peak and stress peak are output according to the covariance function of the posterior distribution. The predicted values ​​of the junction temperature peak and stress peak for each candidate layout point are combined to form the predicted response value for that candidate layout point, and the predicted variances of the junction temperature peak and stress peak for each candidate layout point are combined to form the predicted variance for that candidate layout point. The calculation method for the predicted junction temperature peak value is as follows: based on the junction temperature peak data of the sampled points, according to the distance measurement between the candidate layout point and each sampled point in the layout parameter space, the weight of the junction temperature peak of each sampled point is determined by the covariance function and then the weighted sum is performed. The calculation method for the predicted peak stress value is the same as that for the predicted peak junction temperature value; the method for determining the predicted variance of the peak junction temperature value is as follows: The combination operation is performed based on the covariance function values ​​between the candidate layout point and each sampled point, as well as between each sampled point. The specific process of the combination operation is to subtract the covariance value of the candidate layout point itself from the weighted covariance value between each sampled point. The method for determining the variance of peak stress prediction is the same as that for determining the variance of peak junction temperature prediction.

[0040] The specific operation for non-dominated sorting of each candidate layout point is as follows: Each candidate layout point is considered as a point with two attributes, where the first attribute is the predicted response value of the candidate layout point and the second attribute is the predicted variance of the candidate layout point. Compare the first and second attributes of any two candidate layout points one by one. If the predicted response value of the first candidate point is greater than or equal to the predicted response value of the second candidate point and the prediction variance of the first candidate point is greater than or equal to the prediction variance of the second candidate point, then the first candidate point is determined to be inferior to the second candidate point, and the first candidate point is marked as the dominated point. Conversely, if the predicted response value of the first candidate point is less than the predicted response value of the second candidate point and the predicted variance of the first candidate point is less than the predicted variance of the second candidate point, then the first candidate point is determined to be superior to the second candidate point, and the second candidate point is marked as the dominated point. Perform the pairwise comparison operation described above on all candidate layout points, extract all candidate layout points that are not marked as dominated points, and form the non-dominated frontier of the feasible layout region.

[0041] The specific steps for extracting two extreme points from the non-dominated frontier are as follows: Obtain the predicted response values ​​of all candidate layout points in the non-dominated frontier, select the candidate layout point with the smallest predicted response value, and mark this point as the first extreme point; Obtain the predicted variance of all candidate layout points in the non-dominated frontier, and select the candidate layout point with the largest predicted variance, marking this point as the second extreme point. When the candidate layout point with the smallest predicted response value is also the candidate layout point with the largest predicted variance, then in the non-dominated frontier, the second smallest predicted response value is taken as the first extreme point and the second largest predicted variance is taken as the second extreme point, ensuring that the two extreme points are different candidate layout points.

[0042] After extracting the two extreme points, the predicted response value corresponding to the first extreme point is taken as the optimal potential value of the feasible layout region, and the predicted variance corresponding to the second extreme point is taken as the exploration demand value of the feasible layout region. The optimal potential value and exploration demand value of each feasible layout region are output.

[0043] In S4 and S5, the optimal potential value and exploration requirement value of each feasible layout region are used as the decision-making basis to select the regions to be developed. Within the regions to be developed, candidate layout points are selected based on local proxy mapping for simulation. The sampling data of the regions to be developed are updated, and S2 to S3 are repeated until convergence. After merging the frontier solutions of each region, the final layout scheme is output, which includes: The specific operation for generating a two-dimensional preference pair consisting of potential difference degree and exploration difference degree is as follows: Determine the global optimal potential value and the global maximum exploration requirement value. The global optimal potential value is the minimum value among the optimal potential values ​​in all feasible layout regions, and the global maximum exploration requirement value is the maximum value among the exploration requirement values ​​in all feasible layout regions. For each feasible layout region, its optimal potential value is subtracted from the global optimal potential value, and the difference is defined as the potential difference degree of that region. The absolute value of the difference between the exploration requirement value and the global maximum exploration requirement value is defined as the exploration difference degree of the region. Each feasible layout region constitutes a two-dimensional preference pair with its potential difference degree and exploration difference degree.

[0044] The specific steps for dividing the feasible layout area into potential preference zones or exploration preference zones are as follows: The potential difference and exploration difference of each feasible layout region are normalized. The normalization method of potential difference is to divide the potential difference of the region by the maximum potential difference of all regions, and the quotient is used as the normalized potential value. The normalization method for the exploration difference degree is to divide the exploration difference degree of the region by the maximum value of the exploration difference degree of all regions, and use the quotient as the normalized exploration value; calculate the difference between the normalized potential value and the normalized exploration value; When the difference is greater than zero, the region is identified as a potential preference region, indicating that the current optimal potential value of the region is relatively large from the global optimal potential value, and its development value is higher than its exploration value. When the difference is less than zero, the area is identified as an exploration preference area, indicating that the current exploration demand value of the area is relatively large compared with the global maximum exploration demand value, and its exploration value is higher than its development value. When the difference is zero, the area is identified as an exploration preference zone, so that resources are prioritized for areas that have not been fully explored.

[0045] The specific steps for selecting the area to be developed are as follows: If a potential preference area exists, then select the area with the smallest potential difference among all potential preference areas as the area to be developed. If there are no potential preference areas but only exploration preference areas, then the area with the largest exploration difference degree among all exploration preference areas is selected as the area to be developed. Within the selected development area, based on the current Gaussian process regression posterior distribution of the area, the candidate layout point with the smallest predicted response value is selected from all candidate layout points in the candidate space of the area. This candidate layout point is then submitted for finite element thermal-structural coupling simulation to obtain the junction temperature peak and stress peak corresponding to the candidate layout point. The newly added layout parameters and their corresponding junction temperature peak and stress peak are then incorporated into the existing sampling data of the feasible layout area to complete the update of the sampling data of the area.

[0046] The specific steps for repeating the aforementioned steps until the convergence condition is met are as follows: Set the maximum number of iterations. The maximum number of iterations is half the number of initial sampling points, rounded down to the integer part. In each iteration, the following steps are performed sequentially: clustering of feasible layout regions, construction of local proxy mappings for each region, extraction of optimal potential values ​​and exploration requirements, selection of regions to be developed, simulation of candidate layout points, and updating of sampled data. After each iteration, the minimum value among the optimal potential values ​​in all feasible layout regions is calculated and recorded. When the change in the minimum value is less than 1% of the value in three consecutive iterations, the convergence condition is considered met, and the iteration process is terminated. If the maximum number of iterations is reached but the convergence condition is still not met, the iteration process will also terminate.

[0047] The specific steps for merging the frontier solutions of each region to output the final layout scheme are as follows: Collect all candidate layout points and their corresponding junction temperature peaks and stress peaks generated by each feasible layout region during each iteration, and summarize all candidate layout points to form a global candidate set. Compare any two candidate layout points in the global candidate set. If the junction temperature peak of the first candidate layout point is greater than or equal to the junction temperature peak of the second candidate layout point and the stress peak of the first candidate layout point is greater than or equal to the stress peak of the second candidate layout point, then mark the first candidate layout point as the global dominated point. If the peak junction temperature of the first candidate layout point is less than the peak junction temperature of the second candidate layout point and the peak stress of the first candidate layout point is less than the peak stress of the second candidate layout point, then the second candidate layout point is marked as the globally dominated point. After performing pairwise comparisons on all candidate layout points, all candidate layout points that are not marked as globally dominated points are extracted to form a set of globally optimal frontiers; The layout parameters corresponding to the candidate layout point with the smallest junction temperature peak value are selected from the global optimal frontier set as the first candidate output scheme, and the layout parameters corresponding to the candidate layout point with the smallest stress peak value are selected as the second candidate output scheme. When thermal management is the primary objective of engineering requirements, the first candidate output scheme will be output as the final layout scheme. When structural reliability is the primary objective of engineering requirements, the second candidate output scheme will be output as the final layout scheme.

[0048] The working principle of this invention is as follows: Multiple sets of copper-embedded board layout parameters satisfying manufacturing constraints are collected and thermal-structural coupled simulations are performed to obtain the junction temperature peak and stress peak corresponding to each layout parameter. Density clustering is performed based on the connectivity measure between layout parameter samples, dividing the design space into multiple feasible layout regions separated by manufacturing constraint boundaries. For each feasible layout region, a local surrogate mapping of Gaussian process regression is constructed using the responses of the sampled points, and the optimal potential value and exploration demand value of each region are extracted based on its posterior distribution. Using the two-dimensional preference pair composed of the potential difference degree and exploration difference degree of each region as the decision basis, potential preference regions or exploration preference regions are divided and regions to be developed are selected. Within these regions, the candidate layout point with the smallest predicted response value is selected, simulation is performed, and the sampled data is updated. The above clustering and extraction steps are repeated until convergence. After merging the non-dominated front solutions of each region, the final layout scheme with the smallest junction temperature peak or stress peak is output.

[0049] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. An AI-embedded copper plate-thermal structure coupled layout design method, characterized in that, Includes the following steps: S1. Collect multiple sets of AI copper-embedded board layout parameters that meet manufacturing constraints, and perform thermal-structural coupling simulation on each set of layout parameters to obtain the corresponding junction temperature peak and stress peak. S2, based on the connectivity between layout parameters, performs density clustering to divide the design space into multiple feasible layout regions; S3. For each feasible layout region, construct a local proxy map using the responses of the sampled points, and extract the optimal potential value and exploration demand value of the feasible layout region based on the posterior distribution of the local proxy map. S4. Based on the optimal potential value and exploration requirement value of each feasible layout region, select the region to be developed. Within the region to be developed, select candidate layout points for simulation based on local proxy mapping, and update the sampling data of the region to be developed. S5, repeat S2 to S3 until convergence, merge the frontier solutions of each region and output the final layout scheme.

2. The AI ​​copper-embedded board-thermal structure coupling layout design method according to claim 1, characterized in that, Obtaining the corresponding junction temperature peak and stress peak specifically includes: Based on each set of layout parameters, establish the geometric configuration and apply thermal loads and constraints, then solve the heat conduction equation to obtain the initial temperature field; The thermal-mechanical sensitivity of each element is calculated by multiplying the temperature gradient of each region by the coefficient of thermal expansion of the material. Based on the thermal-mechanical sensitivity, the interpolation weights of high-gradient regions are refined, and the load mapping is reconstructed. The reconstructed thermal load is transferred step by step to the equilibrium equations to solve the stress field; After convergence, the maximum global temperature and the maximum equivalent stress are extracted as the junction temperature peak and stress peak, respectively.

3. The AI ​​copper-embedded board-thermal structure coupling layout design method according to claim 2, characterized in that, The reconstructed load mapping specifically includes: Based on thermo-mechanical sensitivity, interpolation weights are added to interpolation nodes in high gradient regions where the temperature gradient exceeds a preset threshold, and the thermal load mapping from the temperature field to the stress field is reconstructed in a weighted manner.

4. The AI ​​copper-embedded board-thermal structure coupling layout design method according to claim 1, characterized in that, The division of the design space into multiple feasible layout areas specifically includes: Calculate the shortest continuous path length for every two layout parameter samples to bypass the manufacturing constraint boundary within the feasible region, and use the shortest continuous path length as a connectivity measure. Using connectivity metrics as input, local density peaks are identified and sample clusters are divided, with each cluster defining an independent feasible layout region.

5. The AI ​​copper-embedded plate-thermal structure coupling layout design method according to claim 4, characterized in that, The process of identifying local density peaks and dividing sample clusters specifically includes: Based on the connectivity metric between each layout sample and other samples in its neighborhood, the local connectivity density of each sample is calculated. Based on the local connectivity density of each sample and the shortest connected path length between it and samples with higher density, the deviation judgment value of each sample is calculated, and samples with deviation judgment values ​​greater than a preset threshold are marked as peak points. Using each peak point as the center, the remaining samples are assigned to the cluster containing the peak point with the minimum connected path length, thus completing the sample clustering.

6. The AI ​​copper-embedded board-thermal structure coupling layout design method according to claim 1, characterized in that, The extraction of the optimal potential value and exploration requirement value of the feasible layout region specifically includes: Based on the posterior distribution, generate predicted response values ​​and prediction variances for multiple sets of candidate layout points; For each candidate layout point, non-dominated sorting is performed with the predicted response value and predicted variance as two independent objectives, and the two extreme points on the non-dominated frontier are extracted. The value with the smallest predicted response between the two extreme points is taken as the optimal potential value, and the value with the largest predicted variance is taken as the exploration demand value.

7. The AI ​​copper-embedded board-thermal structure coupling layout design method according to claim 6, characterized in that, The extraction of the two extreme points on the non-dominated front specifically includes: Compare the predicted response value and predicted variance of each candidate layout point one by one. Points whose predicted response value and predicted variance are worse than any other point are marked as dominated points. Undominated points form the undominated frontier. In the non-dominated frontier, the minimum predicted response value and the maximum predicted variance are used as screening objectives, respectively. The point with the minimum predicted response value is taken as the first extreme point, and the point with the maximum predicted variance is taken as the second extreme point, thus completing the extraction of the two extreme points.

8. The AI ​​copper-embedded board-thermal structure coupling layout design method according to claim 1, characterized in that, The updating of the sampled data for the area to be developed specifically includes: Based on the difference between the optimal potential value and the exploration demand value of each feasible layout region and the corresponding global extreme value, a two-dimensional preference pair consisting of the potential difference degree and the exploration difference degree is generated for each feasible layout region. Based on the relative magnitude of the two differences in the two-dimensional preference pair, the feasible layout area is divided into a potential preference area or an exploratory preference area. The region with the smallest potential difference among the regions classified as potential preference zones will be selected as the region to be developed, or the region with the largest exploration difference among the regions classified as exploration preference zones will be selected as the region to be developed. Within the selected development area, the candidate layout point with the smallest predicted response value is selected based on its local proxy mapping to perform simulation, and the newly added sampled data is incorporated into the existing sampled data of the feasible layout area.

9. The AI ​​copper-embedded plate-thermal structure coupling layout design method according to claim 8, characterized in that, The region with the largest exploration difference among the regions divided into exploration preference zones is selected as the region to be developed, specifically including: The potential difference and exploration difference of each feasible layout region are normalized to obtain normalized potential value and normalized exploration value; Calculate the difference between the normalized potential value and the normalized exploration value; Then, regions with values ​​greater than zero are identified as potential preference regions, and regions with differences less than zero are identified as exploration preference regions.

10. The AI ​​copper-embedded board-thermal structure coupling layout design method according to claim 1, characterized in that, The final output layout scheme specifically includes: Collect local frontier solutions for each feasible layout region, and summarize all local frontier solutions into a global candidate set; For all front solutions in the global candidate set, compare their junction temperature peak value and stress peak value one by one, eliminate the dominated solutions whose junction temperature and stress are inferior to any other front solution, and retain the undominated front solutions to form the global optimal front set. From the globally optimal frontier set, the one with the smallest junction temperature peak or the smallest stress peak is selected as the final layout scheme output.

Images