A Co-optimization Design Method for High-Frequency Electromagnetic Characteristics and Thermal Stability of Semiconductor Devices

Abstract

This invention provides a method for synergistic optimization of high-frequency electromagnetic properties and thermal stability of semiconductor devices, relating to the field of semiconductor device design technology. The method involves arranging electromagnetic field detection points at device grid nodes to establish a mapping relationship between electromagnetic field energy loss and temperature, analyzing heat distribution characteristics using a multi-scale decomposition network, predicting temperature change trends using a recurrent neural network, constructing material parameter mapping relationships, and optimizing the distribution of dielectric constant and conductivity. This achieves synergistic optimization of electromagnetic properties and thermal stability, effectively improving the performance stability and reliability of semiconductor devices.

Description

Technical Field

[0001] This invention relates to semiconductor device design technology, and more particularly to a method for synergistic optimization of high-frequency electromagnetic characteristics and thermal stability of semiconductor devices. Background Technology

[0002] Semiconductor devices generate a lot of heat when operating at high frequencies. Traditional device design methods often separate and optimize electromagnetic and thermal characteristics separately, which leads to a large deviation between the device's performance and design expectations during actual operation, affecting the device's reliability and lifespan.

[0003] Existing technologies typically employ simplified thermal models to predict device temperatures, which fail to accurately reflect the conversion of electromagnetic energy loss into heat. Furthermore, they lack precise identification of hotspot regions and analysis of temperature evolution patterns, resulting in a lack of effective theoretical guidance for device structure optimization. Simultaneously, the optimization of material parameters fails to fully consider the correlation between thermal conductivity characteristics and material parameters, making it difficult to achieve synergistic optimization of device electromagnetic properties and thermal stability.

[0004] Therefore, there is an urgent need for a design method that can accurately describe the relationship between electromagnetic field energy loss and temperature distribution, and optimize material parameters by combining thermal conduction characteristics, so as to improve the performance stability of semiconductor devices under high-frequency operating conditions. Summary of the Invention

[0005] This invention provides a method for synergistic optimization design of high-frequency electromagnetic characteristics and thermal stability of semiconductor devices, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides a method for synergistically optimizing the high-frequency electromagnetic characteristics and thermal stability of semiconductor devices, comprising:

[0007] Electromagnetic field detection points are arranged at the grid nodes of semiconductor devices to collect the electric field strength and magnetic field strength of the electromagnetic field detection points. The electromagnetic field energy loss of the grid nodes is calculated based on the electric field strength and magnetic field strength. A piecewise linear interpolation method is used to establish the mapping relationship between electromagnetic field energy loss and temperature. The electromagnetic field energy loss is converted into heat source power density based on the mapping relationship.

[0008] The heat source power density is input into a multi-scale decomposition network to obtain heat distribution characteristic data. Based on the heat distribution characteristic data, the location information of hotspot areas is determined, and time-series temperature data of hotspot areas is established. The time-series temperature data is input into a recurrent neural network to obtain the temperature change trend of hotspot areas. Based on the temperature change trend, a temperature gradient matrix is ​​generated.

[0009] The thermal conduction characteristics are obtained by performing eigenvalue decomposition on the temperature gradient matrix. Based on the thermal conduction characteristics, the device region is divided and the thermal conduction channels between regions are established. The thermal conduction channels are associated with the temperature gradient to construct the material parameter mapping relationship. The dielectric constant distribution and conductivity distribution are generated through parameter sensitivity analysis and boundary continuity optimization.

[0010] Using the dielectric constant distribution and conductivity distribution as initial parameters, material parameters are configured for each region of the semiconductor device. Based on the configuration results, a new temperature distribution is calculated. The new temperature distribution is compared with a preset temperature threshold. When the temperature distribution meets the preset threshold requirements, the current dielectric constant distribution and conductivity distribution are determined as the optimized structural parameters of the semiconductor device.

[0011] In one alternative embodiment,

[0012] Electromagnetic field detection points are arranged at the grid nodes of the semiconductor device to collect the electric and magnetic field strengths of the detection points. The electromagnetic field energy loss of the grid nodes is calculated based on the electric and magnetic field strengths, including:

[0013] Electromagnetic field detection points are arranged at the grid nodes. Each electromagnetic field detection point includes an electric field detector and a magnetic field detector. The output signals of the electric field detector and the magnetic field detector are sampled to obtain electromagnetic field sampling data.

[0014] The electromagnetic field sampling data is filtered to obtain the electric field strength and magnetic field strength at the electromagnetic field detection point.

[0015] The temperature data of the semiconductor device is obtained, the material parameters are determined based on the temperature data, the dielectric loss value is calculated based on the electric field strength, the magnetic loss value is calculated based on the magnetic field strength, and the conduction loss value is calculated based on the electric field strength and the material parameters.

[0016] The electromagnetic energy loss of the grid node is obtained by adding the dielectric loss value, magnetic loss value and conduction loss value.

[0017] In one alternative embodiment,

[0018] A piecewise linear interpolation method is used to establish the mapping relationship between electromagnetic field energy loss and temperature. Based on this mapping relationship, the electromagnetic field energy loss is converted into heat source power density, including:

[0019] Adaptive filtering is performed on the electromagnetic field energy loss data. The filtering parameters are dynamically adjusted according to the signal change amplitude to obtain the filtered electromagnetic field energy loss data. The distribution probability of the filtered electromagnetic field energy loss data is calculated, and the inflection point of the probability change is determined as the segmentation point. The range of electromagnetic field energy loss is divided into multiple sub-intervals.

[0020] Within each sub-interval, the correlation coefficient between electromagnetic field energy loss data and temperature data is calculated. The correlation coefficient is used as a weight value, and a piecewise linear interpolation method is used to establish a sub-interval mapping function for the weighted data.

[0021] Calculate the slope difference of the mapping function of adjacent sub-intervals at the segmentation point, adjust the mapping function coefficients according to the slope difference, so that the mapping function of adjacent sub-intervals is continuous at the segmentation point, and obtain the piecewise linear mapping relationship between electromagnetic field energy loss and temperature.

[0022] Real-time electromagnetic field energy loss data of semiconductor devices is acquired. The real-time electromagnetic field energy loss data is substituted into the piecewise linear mapping relationship to obtain the temperature prediction value. The corresponding material parameter correction factor is obtained based on the temperature prediction value. The real-time electromagnetic field energy loss data is multiplied by the material parameter correction factor to obtain the heat source power density.

[0023] In one alternative embodiment,

[0024] The heat source power density is input into a multi-scale decomposition network to obtain heat distribution characteristic data. Based on the heat distribution characteristic data, the location information of hotspot regions is determined, and time-series temperature data of the hotspot regions is established, including:

[0025] The heat source power density is input into a multi-scale decomposition network for signal decomposition to obtain a decomposition coefficient matrix group, which is then divided into multiple scale levels.

[0026] Local statistical features are calculated for the decomposition coefficient matrices at multiple scale levels to obtain the feature variance matrix and feature mean matrix. A dynamic decision function is constructed based on the feature variance matrix and feature mean matrix. The decomposition coefficient matrix set is compared with the dynamic decision function to obtain heat distribution feature data.

[0027] The location relationships of hotspot areas are calculated based on the heat distribution characteristic data. Density clustering is used to analyze and process the location relationships to obtain the location information of the hotspot areas. Temperature data is collected based on the location information of the hotspot areas, and the temperature data is organized in chronological order to establish time-series temperature data of the hotspot areas.

[0028] In one alternative embodiment,

[0029] Time-series temperature data is input into a recurrent neural network to obtain the temperature change trend of hotspot areas. A temperature gradient matrix is ​​then generated based on this trend, including:

[0030] Feature enhancement processing is performed on time-series temperature data to extract temperature values, temperature changes, temperature change rates, and hotspot volume, and to construct temperature feature vectors.

[0031] Based on the temperature feature vector, calculate the temperature difference between the hot spot area and the adjacent area, obtain the shape features and location information of the hot spot area, and generate spatial correlation features.

[0032] The temperature feature vector and spatial correlation features are input into a two-layer recurrent neural network. The first layer of the network extracts the short-term temperature change features and temperature evolution patterns. The short-term temperature change features and temperature evolution patterns are then input into the second layer of the network to establish a long-term temperature dependency relationship and obtain the temperature change trend of the hotspot area.

[0033] The magnitude and direction of the spatial temperature gradient are calculated based on the temperature change trend. The distribution information of gradient singular points is extracted, and the magnitude and direction of the spatial temperature gradient and the distribution information of singular points are combined to generate a temperature gradient matrix.

[0034] In one alternative embodiment,

[0035] Eigenvalue decomposition of the temperature gradient matrix yields thermal conductivity characteristics. Based on these characteristics, device regions are divided and inter-regional thermal conduction channels are established. The thermal conduction channels are correlated with the temperature gradient to construct a material parameter mapping relationship. Through parameter sensitivity analysis and boundary continuity optimization, the final dielectric constant and conductivity distributions are generated, including:

[0036] Eigenvalue decomposition and eigenvector decomposition are performed on the temperature gradient matrix to obtain heat conduction direction feature data. A gradient intensity distribution map is generated based on the heat conduction direction feature data. The device is divided into regions according to the gradient intensity distribution map to obtain material parameter priority region data. A heat conduction channel between adjacent regions is constructed based on the material parameter priority region data.

[0037] The heat conduction channel is correlated with the temperature gradient matrix to establish a mapping relationship between temperature gradient and dielectric constant and between temperature and conductivity. Adaptive weights of material parameters are calculated based on the mapping relationship.

[0038] The parameter sensitivity of each region is calculated using the adaptive weights, and the parameter sensitivity is compared with a preset threshold to obtain the parameter update interval. Local parameter adjustment data is then generated based on the parameter update interval.

[0039] Boundary continuity analysis is performed on the local parameter adjustment data, the parameter gradient and boundary features of adjacent regions are calculated, a boundary feature vector is constructed, the boundary type is classified based on the boundary feature vector, the comprehensive weight of the region is calculated according to the obtained boundary type, the parameters are processed in a hierarchical smoothing manner, and a parameter smoothing transition scheme is generated.

[0040] Material physical constraints are applied to the parameter smooth transition scheme to establish a mapping relationship between the temperature field and material parameters, and the dielectric constant distribution and conductivity distribution of each region of the device are calculated.

[0041] In one alternative embodiment,

[0042] Boundary continuity analysis is performed on the local parameter adjustment data to calculate the parameter gradient and boundary features of adjacent regions, construct boundary feature vectors, classify boundary types based on the boundary feature vectors, calculate the comprehensive weight of the regions according to the obtained boundary types, perform hierarchical smoothing processing on the parameters, and generate a parameter smoothing transition scheme, including:

[0043] A multi-directional difference template is used to analyze local parameter adjustment data to obtain parameter gradient information. Based on the parameter gradient information, the normal gradient, tangential gradient and boundary curvature are calculated, and boundary features are extracted.

[0044] The boundary features are normalized, and the normalized parameter gradient magnitude, gradient direction, boundary curvature, and parameter jump variables are combined to construct a boundary feature vector.

[0045] A boundary grading matrix is ​​established based on the boundary feature vector. The first boundary grading is determined according to the parameter gradient magnitude, the second boundary grading is determined by combining the gradient direction continuity, and the third boundary grading is determined by the boundary curvature characteristics.

[0046] A weighted correlation matrix is ​​constructed based on the boundary classification. The boundary classification weight, regional area weight, parameter importance weight, and location relevance weight are input into the weighted correlation matrix to calculate the comprehensive regional weight.

[0047] Based on the comprehensive weight of the region, a smoothing coefficient is selected. The first smoothing coefficient is used for the first boundary graded region, the second smoothing coefficient is used for the second boundary graded region, and the third smoothing coefficient is used for the third boundary graded region. The boundary continuity of the smoothing result is evaluated, and when the boundary continuity meets the preset conditions, the parameter smoothing transition scheme is output.

[0048] A second aspect of the present invention provides an electronic device, comprising:

[0049] processor;

[0050] Memory used to store processor-executable instructions;

[0051] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0052] A third aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0053] In this embodiment, by establishing a mapping relationship between electromagnetic field energy loss and temperature, a tight coupling between electromagnetic field analysis and thermal analysis is achieved. This overcomes the optimization inconsistency problem caused by the independent analysis of electromagnetic and thermal characteristics in traditional design methods, making the design results more accurate and reliable. The use of a combination of multi-scale decomposition networks and recurrent neural networks to process time-series temperature data in hotspot regions effectively captures the hotspot distribution and temperature evolution trends of semiconductor devices under high-frequency operating conditions, improving the accuracy of thermal stability prediction, making thermal management strategies more targeted, and significantly enhancing the thermal reliability of the device. Through parameter sensitivity analysis and boundary continuity optimization, a mapping relationship between material parameters and temperature gradients is established, enabling adaptive optimization of device structural parameters. This allows the optimized semiconductor device to maintain good electromagnetic characteristics while possessing superior thermal stability, extending device lifespan, and improving the overall performance and reliability of the system. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating the collaborative optimization design method for high-frequency electromagnetic characteristics and thermal stability of semiconductor devices according to an embodiment of the present invention.

[0055] Figure 2 This is a comparison chart of the efficiency of electromagnetic field energy loss data processing methods according to embodiments of the present invention;

[0056] Figure 3 This is a comparison chart of the effects of temperature gradient analysis and material parameter optimization in embodiments of the present invention. Detailed Implementation

[0057] 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, 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.

[0058] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0059] Figure 1 This is a flowchart illustrating the collaborative optimization design method for high-frequency electromagnetic characteristics and thermal stability of semiconductor devices according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0060] Electromagnetic field detection points are arranged at the grid nodes of semiconductor devices to collect the electric field strength and magnetic field strength of the electromagnetic field detection points. The electromagnetic field energy loss of the grid nodes is calculated based on the electric field strength and magnetic field strength. A piecewise linear interpolation method is used to establish the mapping relationship between electromagnetic field energy loss and temperature. The electromagnetic field energy loss is converted into heat source power density based on the mapping relationship.

[0061] The heat source power density is input into a multi-scale decomposition network to obtain heat distribution characteristic data. Based on the heat distribution characteristic data, the location information of hotspot areas is determined, and time-series temperature data of hotspot areas is established. The time-series temperature data is input into a recurrent neural network to obtain the temperature change trend of hotspot areas. Based on the temperature change trend, a temperature gradient matrix is ​​generated.

[0062] The thermal conduction characteristics are obtained by performing eigenvalue decomposition on the temperature gradient matrix. Based on the thermal conduction characteristics, the device region is divided and the thermal conduction channels between regions are established. The thermal conduction channels are associated with the temperature gradient to construct the material parameter mapping relationship. The dielectric constant distribution and conductivity distribution are generated through parameter sensitivity analysis and boundary continuity optimization.

[0063] Using the dielectric constant distribution and conductivity distribution as initial parameters, material parameters are configured for each region of the semiconductor device. Based on the configuration results, a new temperature distribution is calculated. The new temperature distribution is compared with a preset temperature threshold. When the temperature distribution meets the preset threshold requirements, the current dielectric constant distribution and conductivity distribution are determined as the optimized structural parameters of the semiconductor device.

[0064] In one optional implementation, electromagnetic field detection points are arranged at the grid nodes of the semiconductor device to collect the electric and magnetic field strengths of the detection points. The electromagnetic field energy loss of the grid nodes is calculated based on the electric and magnetic field strengths, including:

[0065] Electromagnetic field detection points are arranged at the grid nodes. Each electromagnetic field detection point includes an electric field detector and a magnetic field detector. The output signals of the electric field detector and the magnetic field detector are sampled to obtain electromagnetic field sampling data.

[0066] The electromagnetic field sampling data is filtered to obtain the electric field strength and magnetic field strength at the electromagnetic field detection point.

[0067] The temperature data of the semiconductor device is obtained, the material parameters are determined based on the temperature data, the dielectric loss value is calculated based on the electric field strength, the magnetic loss value is calculated based on the magnetic field strength, and the conduction loss value is calculated based on the electric field strength and the material parameters.

[0068] The electromagnetic energy loss of the grid node is obtained by adding the dielectric loss value, magnetic loss value and conduction loss value.

[0069] For example, electromagnetic field detection points are arranged at grid nodes of a semiconductor device. Each electromagnetic field detection point consists of a set of electric field detectors and magnetic field detectors. The electric field detectors acquire the amplitude changes of the electric field signal at the node, and the magnetic field detectors simultaneously acquire the amplitude changes of the spatial magnetic field. The output signals of the two types of sensors are sampled at a high frequency within the time domain, with a sampling frequency of no less than 100 mega-samples per second, to ensure that all fluctuation characteristics of the electromagnetic response of the high-frequency device under operating conditions are captured. The raw data obtained from sampling includes unstable noise and background disturbance signals. To improve data accuracy, a sliding window filtering method is used for processing. Each window contains 128 consecutive data points. A dynamic threshold range is set to filter out high-frequency noise and electromagnetic background interference, outputting the effective electromagnetic intensity values ​​of the electric field detection points and the magnetic field detection points. For the electric field intensity, its instantaneous peak value and time average value are recorded; for the magnetic field intensity, its maximum fluctuation amplitude and dominant direction are extracted.

[0070] To meet the requirements of thermal response coupling modeling, temperature sensors are deployed at the spatial locations corresponding to the aforementioned electromagnetic sampling nodes to collect temperature data during the operation of the semiconductor devices. The temperature acquisition interval is set to 0.5 seconds to ensure sufficient coverage of the heat accumulation process. The collected temperature data is used to identify the thermal state of the device during its operating phase, and combined with the material parameter mapping information at different temperatures in the material data sheet, the relative permittivity, permeability, and conductivity parameters at the current node are determined. These parameters serve as the basis for calculating electromagnetic field energy loss. Based on the combination of electric field strength and relative permittivity, the dielectric loss value is determined; based on the relationship between magnetic field strength and permeability, the magnetic loss value is determined; and simultaneously, based on the relationship between electric field strength and conductivity, the conduction loss value is calculated. After unified normalization, the three loss values ​​are weighted and superimposed to obtain the electromagnetic field energy loss per unit time for that grid node.

[0071] Taking a high-frequency RF switch device with a structural size of 8 mm × 8 mm as an example, a 64 × 64 grid node network is established on its surface, with each node deploying one composite electromagnetic field detection point and one temperature sensor. Under operating conditions of 5 GHz and channel power consumption of 0.5 W, a typical node detects a peak electric field strength of 280 V / m and a magnetic field strength of 3.2 A / m in the central region, corresponding to a temperature of 78°C. Based on data from its material library, the relative permittivity at this point is 4.5, the conductivity is 1.6 S / m, and the permeability is 1.2. Based on the above physical quantity relationships, the estimated dielectric loss, magnetic loss, and conduction loss are approximately 0.013 W, 0.005 W, and 0.007 W, respectively. The sum of these three values ​​yields an electromagnetic field energy loss of 0.025 W per unit time for this node.

[0072] After obtaining the energy loss of all grid nodes, a heat source power density distribution map is constructed in the form of a two-dimensional matrix. This map is then input into the multi-scale decomposition network model. The multi-scale decomposition network, based on a hierarchical residual structure and spatial attention mechanism, extracts the spatial features and gradient change boundaries of densely populated hotspot regions in the heat source distribution. The heat distribution feature data is dimensionality-reduced and compressed using convolution operations, and then reconstructed into an identifiable hotspot region distribution map through a cross-scale fusion module. Significant regions of heat distribution features are marked as potential high-temperature clusters, and their heat accumulation changes are further extracted to generate time-series temperature data corresponding to each hotspot region.

[0073] Time-series temperature data is input into a recurrent neural network for processing. The network structure employs bidirectional memory units, enabling it to simultaneously capture the rate of temperature rise and the trend of temperature fluctuations. For each hotspot region, a predicted temperature change sequence for the next 5 seconds is output. The difference between the current temperature value and the predicted value is then used to construct a local temperature gradient matrix. Each temperature gradient matrix contains information on the direction and rate of temperature change for 16 grid points within the hotspot region, serving as input for understanding the heat conduction trend.

[0074] After obtaining the temperature gradient matrix, it is transformed into dominant heat conduction eigenvectors using eigenvalue decomposition. These vectors represent the main direction and intensity of heat propagation in space. Through cluster analysis of the heat conduction eigenvectors, the entire device structure is divided into multiple regional units, each with similar heat conduction characteristics. The heat exchange relationship between regions is modeled as heat conduction channels. The length, cross-sectional area, and boundary thermal resistance of the heat conduction channels are comprehensively evaluated to establish a description of the heat conduction flux of each channel.

[0075] By establishing the correspondence between heat conduction channels and temperature gradient matrices, a material parameter mapping network is further constructed. This network aims to maximize heat conduction efficiency and automatically adjusts the dielectric constant and conductivity distribution within the region. Before configuring the material parameters, the sensitivity of each parameter is analyzed to identify key parameters affecting the uniform diffusion of heat, and the continuity of parameters at the boundaries is adjusted to avoid heat flow distortion caused by abrupt parameter changes. Finally, dielectric constant and conductivity distribution maps for each region are obtained.

[0076] Using the aforementioned parameter distribution as input, the material configuration of each functional region of the semiconductor device is updated. Based on the new parameters, the temperature distribution of the device under high-frequency operating conditions is recalculated, obtaining a steady-state thermal field map of the entire device. The highest temperature in the thermal field map is compared with a set temperature threshold. If all node temperatures are below the threshold, the parameter configuration scheme has met the optimization requirements. At this point, the current dielectric constant distribution map and conductivity distribution map are used as the final co-optimization design result of the semiconductor device to confirm the overall optimization process between high-frequency electromagnetic characteristics and thermal stability. If the calculation results do not meet the set threshold, the parameter sensitivity evaluation and regional material parameter update steps are restarted until optimization is complete.

[0077] In this embodiment, the electric and magnetic field strengths are accurately acquired through electromagnetic field detection points, and energy loss is calculated by combining this with temperature data, achieving an effective mapping of electromagnetic behavior to thermal effects. The introduction of multi-scale decomposition and a recursive network structure makes the identification of hotspot regions and temperature trend prediction more accurate, providing data support for subsequent material parameter adjustments. Through thermal conduction characteristic analysis and region division, a thermal management-oriented material distribution optimization mechanism is established, effectively reducing the risk of localized temperature rise. Ultimately, a device design scheme that combines high-frequency response capability and thermal stability is obtained, improving the device's operational reliability and service life.

[0078] In one optional implementation, a piecewise linear interpolation method is used to establish the mapping relationship between electromagnetic field energy loss and temperature. Based on this mapping relationship, the electromagnetic field energy loss is converted into heat source power density, including:

[0079] Adaptive filtering is performed on the electromagnetic field energy loss data. The filtering parameters are dynamically adjusted according to the signal change amplitude to obtain the filtered electromagnetic field energy loss data. The distribution probability of the filtered electromagnetic field energy loss data is calculated, and the inflection point of the probability change is determined as the segmentation point. The range of electromagnetic field energy loss is divided into multiple sub-intervals.

[0080] Within each sub-interval, the correlation coefficient between electromagnetic field energy loss data and temperature data is calculated. The correlation coefficient is used as a weight value, and a piecewise linear interpolation method is used to establish a sub-interval mapping function for the weighted data.

[0081] Calculate the slope difference of the mapping function of adjacent sub-intervals at the segmentation point, adjust the mapping function coefficients according to the slope difference, so that the mapping function of adjacent sub-intervals is continuous at the segmentation point, and obtain the piecewise linear mapping relationship between electromagnetic field energy loss and temperature.

[0082] Real-time electromagnetic field energy loss data of semiconductor devices is acquired. The real-time electromagnetic field energy loss data is substituted into the piecewise linear mapping relationship to obtain the temperature prediction value. The corresponding material parameter correction factor is obtained based on the temperature prediction value. The real-time electromagnetic field energy loss data is multiplied by the material parameter correction factor to obtain the heat source power density.

[0083] In one implementation, a piecewise linear interpolation method is used to establish the mapping relationship between electromagnetic field energy loss and temperature. The method of converting electromagnetic field energy loss into heat source power density based on the mapping relationship can be specifically implemented as follows:

[0084] When performing adaptive filtering on electromagnetic field energy loss data, the filtering parameters can be dynamically adjusted during runtime to adapt to changes in signal characteristics. Specifically, a bidirectional sliding window method is used to process the acquired raw electromagnetic field energy loss data. The initial window size is set to 5% of the total number of data points, and the standard deviation σ and mean μ are calculated within the window. When the data point value exceeds the range of μ±2σ, it is identified as a region with large fluctuations, and the window size is reduced to half of its original size to preserve signal details. When the data point values ​​are all within the range of μ±0.5σ, it is identified as a stable region, and the window size is increased to twice its original size, with a maximum not exceeding 10% of the total number of data points, to enhance the filtering effect. In this way, the characteristics of drastically changing regions are preserved, while more robust smoothing is applied to stable regions. For example, for the energy loss data of a semiconductor device, the window size is adjusted to 3 points in the interface region where data changes drastically, and more than 80% of the edge features are preserved after filtering; while in the homogeneous material region, the window is expanded to 15 points, and the noise is reduced by 92%.

[0085] After obtaining the filtered electromagnetic field energy loss data, its probability distribution is calculated to determine the segmentation points. The filtered loss data is sorted by size and divided into 100 equally spaced intervals. The number of data points in each interval is counted and divided by the total number of data points to obtain the probability distribution curve. By calculating the first-order difference of the probability curve, the locations where the difference value changes significantly are identified as segmentation points. In practical applications, analysis of a certain type of power device shows that when the difference value exceeds three times the average difference value, it can be identified as a valid segmentation point. For example, for a set of energy loss distributions containing 10,000 data points, this method identified four key segmentation points, located at 0.05 W / mm². 3 0.35W / mm 3 1.2W / mm 3 and 2.8W / mm 3 The entire loss range is divided into 5 sub-intervals.

[0086] Within each sub-interval, a mapping relationship between electromagnetic field energy loss and temperature needs to be established. First, the Pearson correlation coefficient *r* between the energy loss data and the corresponding temperature data within each sub-interval is calculated. The correlation coefficient is calculated based on data pairs within the sub-interval, for example, in the range [0.05, 0.35] W / mm². 3 Within each sub-interval, a correlation coefficient r = 0.87 was calculated using 500 sample points. This correlation coefficient was used as a weight value to adjust the contribution of the data points within that interval. For each sub-interval, a linear interpolation method was used to fit the weighted data points, establishing a linear mapping function of the form T = a × E + b, where T represents temperature, E represents electromagnetic energy loss, and a and b are undetermined coefficients. The coefficients for each sub-interval were determined using the least squares method, for example, in the range [0.05, 0.35] W / mm 3 Within the interval, we obtain a = 42.5℃·mm 3 / W, b=25.3℃.

[0087] To ensure the continuity of the overall mapping function, the coefficients of the mapping functions in adjacent sub-intervals need to be adjusted. The slope difference between the mapping functions of adjacent sub-intervals at the segmentation point is calculated. If the difference exceeds a preset threshold (e.g., 10%), the coefficients are adjusted. The adjustment method involves modifying the slope of the linear function in the higher energy loss interval while keeping the function value at the segmentation point unchanged, so that the slope difference between it and the lower energy loss interval is within an acceptable range. For example, for a segmentation point of 0.35W / mm... 3 At the point where the slope of the mapping function in the first interval is 42.5 and the initial slope of the second interval is 50.8, the difference is 19.5%, exceeding the threshold. By adjusting the slope of the second interval to 46.2 and the intercept term, the functions in the two intervals maintain continuity at the segmentation point while controlling the slope difference to within 8.7%, below the threshold. After performing similar processing on all segmentation points, a complete piecewise linear mapping relationship is obtained.

[0088] After acquiring real-time electromagnetic field energy loss data of semiconductor devices, this data is substituted into an established piecewise linear mapping relationship to predict temperature. For example, the real-time energy loss at a certain test point is 0.78 W / mm². 3 It belongs to the third sub-interval [0.35, 1.2] W / mm 3 Using the mapping function T = 46.2 × E + 22.5 for this interval, the predicted temperature is 58.6℃. Based on the predicted temperature, a pre-established material parameter-temperature relationship database is consulted to obtain the corresponding material parameter correction factor. This database is built based on measured values ​​of parameters such as thermal conductivity and specific heat capacity of materials at different temperatures. For example, the thermal conductivity correction factor for a certain semiconductor material at 58.6℃ is 0.92. The real-time energy loss data is 0.78 W / mm². 3 Multiplying by this correction factor of 0.92 yields the corrected heat source power density of 0.7176 W / mm².3 As input to the thermal analysis model, it can more accurately reflect the influence of material parameters on heat conduction as temperature changes.

[0089] In this embodiment, by introducing piecewise linear interpolation and adaptive filtering mechanisms, a more accurate and dynamically adjustable mapping relationship between electromagnetic field energy loss and temperature is achieved, making the estimated heat source power density more consistent with the energy transfer patterns under actual operating conditions. By identifying the inflection point of energy loss distribution and establishing a continuous and smooth sub-interval mapping function, the resolution and stability of temperature prediction are significantly improved. Combined with the introduction of material parameter correction factors, dynamic calibration of energy and heat coupling characteristics is achieved, enhancing the system's adaptability to the thermal response characteristics of different materials, thereby providing high-precision basic data support for subsequent thermal stability control and material optimization.

[0090] Figure 2 This figure shows a comparison of the efficiency of three electromagnetic field energy loss data processing methods in four different scenarios. The adaptive filtering method performs best in all scenarios, especially in high-energy scenarios where it achieves a processing efficiency of 91.2%, which is 11.7 percentage points higher than the traditional parameter correction method. The piecewise linear mapping method is at an intermediate level, showing relatively stable performance in medium-to-high energy scenarios. The traditional parameter correction method has the lowest efficiency in multi-field interference scenarios, at only 69.2%. The results indicate that the adaptive filtering combined with piecewise linear mapping method proposed in this paper can better adapt to different electromagnetic field environments, especially maintaining high processing efficiency in complex interference environments, providing more reliable technical support for semiconductor device temperature prediction and heat source power density calculation.

[0091] In one optional implementation, the heat source power density is input into a multi-scale decomposition network to obtain heat distribution characteristic data. Based on the heat distribution characteristic data, the location information of hotspot regions is determined, and time-series temperature data of the hotspot regions is established, including:

[0092] The heat source power density is input into a multi-scale decomposition network for signal decomposition to obtain a decomposition coefficient matrix group, which is then divided into multiple scale levels.

[0093] Local statistical features are calculated for the decomposition coefficient matrices at multiple scale levels to obtain the feature variance matrix and feature mean matrix. A dynamic decision function is constructed based on the feature variance matrix and feature mean matrix. The decomposition coefficient matrix set is compared with the dynamic decision function to obtain heat distribution feature data.

[0094] The location relationships of hotspot areas are calculated based on the heat distribution characteristic data. Density clustering is used to analyze and process the location relationships to obtain the location information of the hotspot areas. Temperature data is collected based on the location information of the hotspot areas, and the temperature data is organized in chronological order to establish time-series temperature data of the hotspot areas.

[0095] The process of inputting the heat source power density into a multi-scale decomposition network to obtain heat distribution characteristic data, determining the location information of hotspot areas based on the heat distribution characteristic data, and establishing time-series temperature data of hotspot areas is as follows.

[0096] In this embodiment, heat source power density data of the electronic device is collected. This data can be acquired using a thermal imager or a temperature sensor array. The acquired heat source power density data is a two-dimensional matrix, where each element represents the power density value at the corresponding location, expressed in watts per square centimeter (W / cm²). For example, the acquired power density matrix is ​​128×128 pixels, covering an area of ​​10cm×10cm, with values ​​ranging from 0.01 to 2.5 W / cm².

[0097] The acquired heat source power density matrix is ​​input into a pre-designed multi-scale decomposition network for processing. This multi-scale decomposition network consists of a combination of five convolutional and pooling layers to decompose the input signal into different frequency components. The first layer of the network uses 64 3×3 convolutional kernels with a stride of 1, followed by a 2×2 max-pooling layer; the second layer uses 128 3×3 convolutional kernels with a stride of 1, followed by a 2×2 max-pooling layer; the third layer uses 256 3×3 convolutional kernels with a stride of 1; the fourth layer uses 256 3×3 convolutional kernels with a stride of 1; and the fifth layer uses 128 3×3 convolutional kernels with a stride of 1. After processing by this network, decomposition coefficient matrix sets at five different scale levels are obtained.

[0098] For the first scale level (highest frequency component), the decomposition coefficient matrix size is 64×64×64, representing the finest heat distribution characteristics; for the second scale level, the decomposition coefficient matrix size is 32×32×128; for the third scale level, the decomposition coefficient matrix size is 32×32×256; for the fourth scale level, the decomposition coefficient matrix size is 32×32×256; for the fifth scale level (lowest frequency component), the decomposition coefficient matrix size is 32×32×128, representing the coarsest heat distribution characteristics.

[0099] Local statistical features are calculated for the decomposition coefficient matrix at each scale level. Specifically, for the decomposition coefficient matrix at each scale level, a 9×9 sliding window with a step size of 3 is used to calculate the variance and mean of the coefficients within the window, forming a feature variance matrix and a feature mean matrix, respectively. For example, for the first scale level, the generated feature variance matrix and feature mean matrix are both 19×19×64 in size; for the fifth scale level, the generated feature variance matrix and feature mean matrix are both 9×9×128 in size.

[0100] A dynamic decision function is constructed based on the feature variance matrix and the feature mean matrix. This function is obtained by thresholding each element in the feature variance matrix and then weighting it according to the corresponding value in the feature mean matrix. Specifically, for an element in the feature variance matrix of the i-th scale level, if the element value is greater than 1.5 times the average variance of that level, it is marked as a hotspot candidate region and assigned a value of twice the feature mean at that location; otherwise, it is assigned a value of 0. In this way, a binary decision matrix is ​​constructed for each scale level.

[0101] The decomposition coefficient matrix set is compared with the dynamic decision function to obtain the heat distribution feature data. Specifically, for each scale level, the decomposition coefficient matrix is ​​multiplied element-wise with the corresponding decision matrix to obtain feature data highlighting heat anomaly areas. Then, the feature data from the five scale levels are weighted and fused with weights of 0.1, 0.15, 0.2, 0.25, and 0.3 respectively to obtain the final heat distribution feature data.

[0102] The location relationships of hotspot areas are calculated based on heat distribution characteristic data. The heat distribution characteristic data is normalized to a value between 0 and 1, and points with values ​​greater than 0.6 are extracted as candidate hotspot locations. For the example 128×128 matrix, approximately 200-300 candidate hotspot locations may be identified.

[0103] Density-based spatial clustering with noise (DBSCAN) was used to perform density clustering analysis on hotspot candidate locations. A clustering distance threshold of 5 pixels and a minimum sample size of 8 were set to aggregate adjacent hotspot candidate locations into hotspot regions. In this way, three main hotspot regions were identified in the 128×128 power density matrix, located near coordinates (32, 45), (78, 62), and (105, 30), with a radius of approximately 10-15 pixels.

[0104] Temperature sensors are deployed based on the identified hotspot location information to collect temperature data. Specifically, five temperature sensors are placed within each hotspot area, sampling at a frequency of 2Hz and recording continuously for 30 minutes, obtaining 3600 temperature data points from each sensor. For example, in the hotspot area at location (32, 45), the initial temperatures of the five sensors are 42.5℃, 43.1℃, 41.8℃, 42.9℃, and 41.5℃, respectively, and after 30 minutes, the temperatures rise to 56.2℃, 57.4℃, 54.9℃, 57.1℃, and 53.8℃, respectively.

[0105] The collected temperature data were organized chronologically to create time-series temperature data for hotspot areas. For each hotspot area, the average temperature values ​​from five sensors at each time point were calculated, forming a time-series temperature sequence containing 3600 data points. Simultaneously, the rate of temperature change, i.e., the temperature difference between adjacent time points, was recorded to form a rate of temperature change sequence. For example, in the hotspot area at location (78, 62), the recorded average temperature increased at a rate of 0.12℃ / min in the first 5 minutes, at a rate of 0.25℃ / min in the 10-15 minute interval, and stabilized in the last 5 minutes with fluctuations less than 0.05℃ / min. This time-series temperature data can be used for subsequent hotspot evolution prediction and heat dissipation system optimization design.

[0106] In this embodiment, by inputting the heat source power density into a multi-scale decomposition network, heat distribution characteristics at different scales can be accurately extracted, enhancing the analytical capability for complex heat conduction behavior. A dynamic judgment function constructed using feature variance and mean effectively improves the identification accuracy of heat anomaly regions. Density clustering is employed to analyze heat distribution characteristics, enabling accurate spatial location of hotspot regions. Furthermore, time-series temperature data of these hotspot regions are collected and constructed, providing a reliable basis for subsequent temperature evolution trend prediction and thermal stability control, thereby significantly improving the thermal management efficiency and stability of semiconductor devices under high-frequency operating conditions.

[0107] In one optional implementation, time-series temperature data is input into a recurrent neural network to obtain the temperature change trend of hotspot areas, and a temperature gradient matrix is ​​generated based on the temperature change trend, including:

[0108] Feature enhancement processing is performed on time-series temperature data to extract temperature values, temperature changes, temperature change rates, and hotspot volume, and to construct temperature feature vectors.

[0109] Based on the temperature feature vector, calculate the temperature difference between the hot spot area and the adjacent area, obtain the shape features and location information of the hot spot area, and generate spatial correlation features.

[0110] The temperature feature vector and spatial correlation features are input into a two-layer recurrent neural network. The first layer of the network extracts the short-term temperature change features and temperature evolution patterns. The short-term temperature change features and temperature evolution patterns are then input into the second layer of the network to establish a long-term temperature dependency relationship and obtain the temperature change trend of the hotspot area.

[0111] The magnitude and direction of the spatial temperature gradient are calculated based on the temperature change trend. The distribution information of gradient singular points is extracted, and the magnitude and direction of the spatial temperature gradient and the distribution information of singular points are combined to generate a temperature gradient matrix.

[0112] This embodiment provides a method for inputting time-series temperature data into a recurrent neural network to obtain the temperature change trend of hotspot areas, and generating a temperature gradient matrix based on the temperature change trend.

[0113] In practical applications, temperature data of conductor devices under different operating conditions can be collected, including temperature values ​​at multiple measuring points on the device surface. For example, 20 temperature sensors can be arranged on a 100W conductor device, collecting data every 5 seconds for 30 minutes to obtain temperature data at 360 time points. This raw temperature data needs feature enhancement processing to extract more valuable feature information. Specifically, four types of features are extracted from the time-series temperature data: temperature value, temperature change, temperature change rate, and hotspot volume. The temperature value is obtained directly from the sensor readings; the temperature change is obtained by calculating the temperature difference between adjacent time points, such as the current temperature minus the previous temperature; the temperature change rate is the temperature change divided by the time interval, reflecting how fast the temperature changes; the hotspot volume refers to the size of the area where the temperature exceeds a certain threshold (e.g., 85℃), which can be estimated by calculating the number of measuring points exceeding the threshold and combining this with the measuring point distribution density. These four types of features combine to form a temperature feature vector, with one feature vector corresponding to each time point, and the vector dimension being four times the number of measuring points.

[0114] The next crucial step is to calculate the temperature difference between the hot spot region and its adjacent regions based on temperature feature vectors. A hot spot region is typically defined as a continuous area where the temperature exceeds a certain threshold, such as a region with a temperature above 85°C. Adjacent regions are those in direct contact with the hot spot region but with temperatures below the threshold. By calculating the temperature difference between each measuring point within the hot spot region and measuring points in adjacent regions, the boundary information of the hot spot region can be obtained. For example, if a hot spot region contains 5 measuring points with an average temperature of 90°C, and its adjacent region contains 8 measuring points with an average temperature of 80°C, then the temperature difference between the hot spot and its adjacent regions is 10°C. This temperature difference distribution can be used to analyze the shape characteristics of the hot spot region, such as whether it exhibits a circular, elliptical, or irregular shape. Simultaneously, the location coordinates of the hot spot region on the device are recorded to obtain location information. Combining shape features and location information, spatial correlation features are generated. These features describe the spatial distribution characteristics of the hot spot region, helping to understand the conduction and diffusion patterns of heat in conductor devices.

[0115] The core of temperature trend prediction lies in inputting temperature feature vectors and spatial correlation features into a two-layer recurrent neural network. This two-layer recurrent neural network consists of two layers of Long Short-Term Memory (LSTM) networks or gated recurrent units (GRUs). In practice, a first-layer LSTM network with 64 hidden units and a second-layer LSTM network with 32 hidden units can be constructed. The first layer is primarily responsible for extracting short-term temperature variation features and temperature evolution patterns. Short-term variation features reflect temperature fluctuations over several minutes, such as the temperature response characteristics when workload changes; temperature evolution patterns describe temperature change patterns under stable operating conditions, such as heating rate and settling time. Specifically, the temperature feature vector and spatial correlation features at each time step are combined into an input vector, which is then input into the first-layer LSTM network. The network processes the sequence data and outputs the hidden states at each time step, which contain the short-term temperature variation features and temperature evolution patterns. These hidden states are used as input to a second-layer LSTM network, which further establishes long-term temperature dependencies, capturing patterns in temperature changes over longer time scales, such as hourly temperature trends. This two-layer structure allows the network to simultaneously consider short-term temperature fluctuations and long-term temperature trends, improving prediction accuracy. During training, mean squared error is used as the loss function, and the Adam optimizer is employed for parameter updates with a learning rate of 0.001 and 200 training epochs. After training, the network can output the temperature change trend of hotspot areas over a future period, such as the predicted temperature value for each measuring point within the next 30 minutes.

[0116] Calculating the spatial temperature gradient based on temperature change trends is a crucial step in generating the temperature gradient matrix. The spatial temperature gradient describes the rate and direction of temperature change in space, which is essential for understanding heat flow and hotspot formation mechanisms. In practice, the temperature gradient at each spatial point can be calculated based on the predicted temperature distribution. The magnitude of the temperature gradient indicates the severity of the temperature change, and its direction points in the direction of the fastest temperature increase. For example, on a two-dimensional plane, if the temperature at point (x, y) is T(x, y), the temperature gradient in the x-direction at that point can be obtained by calculating T(x + Δx, y) - T(x, y) divided by Δx; the y-direction is calculated similarly. By combining the gradients in the x and y directions, the temperature gradient vector at that point can be obtained. For three-dimensional space, the gradient in the z-direction also needs to be considered. During the calculation, Δx, Δy, and Δz can be set to the actual distance between adjacent measurement points, such as 5 millimeters. Furthermore, it is necessary to extract information on the distribution of gradient singularities. Singularities are locations where the temperature gradient changes significantly, such as points where the gradient magnitude is close to zero or where the direction changes abruptly. These singularities often correspond to heat sources or heat flow obstacles, and are crucial for understanding the thermal characteristics of conductor devices. By analyzing the magnitude and direction of the temperature gradient and the distribution of singularities throughout the entire spatial domain, a temperature gradient matrix can be constructed. This matrix is ​​a multi-dimensional data structure, with each spatial point corresponding to an element. The element value contains the gradient magnitude, direction, and whether it is a singularity at that point. For example, for a 20×20×5 three-dimensional grid, the temperature gradient matrix has dimensions of 20×20×5×4, where the first three dimensions represent spatial coordinates, and the fourth dimension contains the x, y, and z components of the gradient and the singularity identifier.

[0117] The application of temperature gradient matrices is of significant value for the synergistic optimization design of high-frequency electromagnetic characteristics and thermal stability of conductor devices. By analyzing the temperature gradient matrix, hot spots and their evolution trends in conductor devices can be identified, and potential thermal failure risks can be predicted. For example, if the temperature gradient amplitude in a certain region continues to increase, it indicates that the region may become a new hot spot; if the temperature gradient direction of the hot spot region points towards a critical component, that component may face overheating risks. Based on these analytical results, the structural design of conductor devices can be optimized, such as adjusting conductor layout, adding heat dissipation channels, or changing material combinations to improve heat distribution. Simultaneously, the temperature gradient matrix can be combined with electromagnetic field analysis to study the impact of temperature changes on the high-frequency electromagnetic characteristics of conductor devices, such as impedance changes caused by thermal expansion and increased losses due to temperature rise. By synergistically considering thermal and electromagnetic characteristics, the overall performance of conductor devices can be optimized, improving their reliability and efficiency in high-power, high-frequency applications.

[0118] In this embodiment, a deep modeling of the temperature evolution process in hotspot regions using a two-layer recurrent neural network not only enhances the response to short-term temperature fluctuations but also effectively captures the dependence on long-term temperature changes, achieving accurate prediction of hotspot temperature trends. By constructing a multi-dimensional feature vector that integrates temperature values, rates of change, and spatial characteristics, the model's analytical capabilities for complex heat conduction dynamics are enhanced. Furthermore, the generated temperature gradient matrix accurately reflects the intensity direction and anomalous distribution of the spatial thermal field, helping to identify thermal anomaly diffusion paths and potential thermal runaway risk points. This provides crucial information for subsequent thermal conduction structure reconstruction and material parameter adjustment, significantly improving the proactive control capabilities of device thermal design.

[0119] In one optional implementation, the temperature gradient matrix is ​​eigenvalued to obtain thermal conductivity characteristics. Based on these characteristics, the device region is divided and inter-regional thermal conductivity channels are established. The thermal conductivity channels are correlated with the temperature gradient to construct a material parameter mapping relationship. The final dielectric constant and conductivity distributions are generated through parameter sensitivity analysis and boundary continuity optimization, including:

[0120] Eigenvalue decomposition and eigenvector decomposition are performed on the temperature gradient matrix to obtain heat conduction direction feature data. A gradient intensity distribution map is generated based on the heat conduction direction feature data. The device is divided into regions according to the gradient intensity distribution map to obtain material parameter priority region data. A heat conduction channel between adjacent regions is constructed based on the material parameter priority region data.

[0121] The heat conduction channel is correlated with the temperature gradient matrix to establish a mapping relationship between temperature gradient and dielectric constant and between temperature and conductivity. Adaptive weights of material parameters are calculated based on the mapping relationship.

[0122] The parameter sensitivity of each region is calculated using the adaptive weights, and the parameter sensitivity is compared with a preset threshold to obtain the parameter update interval. Local parameter adjustment data is then generated based on the parameter update interval.

[0123] Boundary continuity analysis is performed on the local parameter adjustment data, the parameter gradient and boundary features of adjacent regions are calculated, a boundary feature vector is constructed, the boundary type is classified based on the boundary feature vector, the comprehensive weight of the region is calculated according to the obtained boundary type, the parameters are processed in a hierarchical smoothing manner, and a parameter smoothing transition scheme is generated.

[0124] Material physical constraints are applied to the parameter smooth transition scheme to establish a mapping relationship between the temperature field and material parameters, and the dielectric constant distribution and conductivity distribution of each region of the device are calculated.

[0125] In the implementation process, the temperature gradient matrix of the device is first obtained, which reflects the rate and direction of temperature change at various points on the device. Taking a semiconductor device with a 64×64 grid as an example, temperature data of each grid point is collected using a thermal imager, and the temperature gradient matrix G is calculated. Eigenvalue decomposition is performed on this temperature gradient matrix to obtain the principal eigenvalues ​​λ1=8.76, λ2=3.42, and λ3=0.88, with corresponding eigenvectors v1=[0.72, 0.65, 0.24], v2=[-0.54, 0.62, 0.57], and v3=[0.43, -0.44, 0.79]. These eigenvectors represent the main directions of temperature conduction, while the eigenvalues ​​represent the conduction intensity in those directions. A gradient intensity distribution map is generated based on the eigenvalues ​​and eigenvectors, where the gradient intensity value is calculated by the inner product of the eigenvector and the actual temperature gradient. In the gradient intensity distribution map, high-intensity areas (such as intensity values ​​greater than 7.5) are marked as high-priority areas, medium-intensity areas (intensity values ​​between 3.0 and 7.5) are marked as medium-priority areas, and low-intensity areas (intensity values ​​less than 3.0) are marked as low-priority areas, thus obtaining the material parameter priority area data.

[0126] Based on priority region data, heat conduction channels between adjacent regions are constructed. For each pair of adjacent regions A and B, the heat flux between them is calculated; the flux value is the product of the boundary temperature difference and the boundary area. For example, the boundary temperature difference between region 1 and region 2 is 4.2℃, and the boundary area is 2.1mm². 2 The calculated heat flux is 8.82 W / mm². 2 Similarly, the heat flux between all adjacent regions is calculated to form a heat conduction channel network. In this network, the heat flux is greater than 5 W / mm². 2 The channel is defined as a strong heat conduction channel with a heat flux of 2 W / mm. 2 Up to 5W / mm 2 The space between them is defined as a medium heat conduction channel, less than 2 W / mm. 2 It is defined as a weak heat conduction channel.

[0127] By correlating the heat conduction path with the temperature gradient matrix, mapping relationships between temperature gradient and dielectric constant, and between temperature and conductivity, are established. Specifically, the mapping relationship between dielectric constant ε and temperature gradient amplitude |∇T| is set as ε = ε0 + α|∇T|, where ε0 is the basic dielectric constant with a value of 4.5, and α is an adjustment coefficient with a value of 0.8; the mapping relationship between conductivity σ and temperature T is set as σ = σ0exp(βT), where σ0 is the basic conductivity with a value of 0.02 S / m, and β is a temperature coefficient with a value of 0.03 K. -1The adaptive weights of material parameters are calculated through these mapping relationships. For example, in a region with a temperature gradient amplitude of 5 K / mm, the weight of dielectric constant is 4.5 + 0.8 × 5 = 8.5; in a region with a temperature of 85℃, the weight of conductivity is 0.02 × exp(0.03 × 85) = 0.14 S / m.

[0128] The parameter sensitivity S for each region is calculated using adaptive weights. Sensitivity is defined as the degree to which parameter changes affect the temperature field. In practice, a small perturbation δp (e.g., changing the dielectric constant by 1%) is applied to each region, and the temperature field change δT is observed. The sensitivity S = |δT / δp|. For example, when the dielectric constant of region 3 changes from 5.0 to 5.05 (an increase of 1%), the temperature decreases by an average of 0.35℃. Therefore, the sensitivity of this region to the dielectric constant is 0.35 / 0.05 = 7.0. The parameter sensitivity of each region is compared with a preset threshold (e.g., a sensitivity threshold of 5.0) to obtain the parameter update interval. In the above example, the sensitivity of region 3 is greater than the threshold, so its dielectric constant update interval is determined to be [4.8, 5.3], generating local parameter adjustment data.

[0129] Boundary continuity analysis is performed on the local parameter adjustment data to calculate the parameter gradient and boundary characteristics of adjacent regions. For example, the dielectric constants of regions 4 and 5 are 6.2 and 4.8, respectively, and their boundary length is 1.5 mm. The parameter gradient is (6.2-4.8) / 1.5=0.93 mm. -1 Construct boundary feature vectors containing information such as parameter gradients, boundary lengths, and heat flux. Classify boundary types based on these feature vectors, where the parameter gradient is greater than 0.8 mm. -1 The boundary is a "sharp transition boundary" with a parameter gradient of 0.3 mm. -1 up to 0.8mm -1 The boundary between these two points is a "gradually varying boundary," with a parameter gradient less than 0.3 mm. -1 The boundary is designated as "gentle". The comprehensive weight W for the region is calculated based on the boundary type, taking into account the region area, parameter sensitivity, and boundary characteristics. For region 4, its area is 2.4 mm. 2 With a parameter sensitivity of 6.5, it forms a "sudden change boundary" with region 5. The calculated comprehensive weight W4 = 2.4 × 6.5 × 1.2 = 18.72 (where 1.2 is the weight coefficient of the sudden change boundary).

[0130] The parameters are subjected to hierarchical smoothing to generate a parameter smoothing transition scheme. A weighted average method is used, where the smoothed parameter p' of region i is... i =p i ×(1-γ)+Σ(p j ×w j )×γ / Σw j , where p ip is the original parameter value. j For the parameter values ​​of adjacent region j, w j Let γ be the weight of region j, and γ be the smoothing factor with a value of 0.3. For example, if the original dielectric constant of region 4 is 6.2, and the dielectric constants of its neighboring regions 5 and 6 are 4.8 and 5.5 respectively, with weights of 18.72 and 12.45 respectively, then the smoothed dielectric constant of region 4 is 6.2×(1-0.3)+(4.8×18.72+5.5×12.45) / (18.72+12.45)×0.3=5.83.

[0131] Finally, material physical constraints are imposed on the parameter smoothing transition scheme, such as limiting the dielectric constant range to [3.5, 8.0] and the conductivity range to [0.01 S / m, 0.2 S / m]. A mapping relationship between the temperature field and material parameters is established, and the final dielectric constant and conductivity distributions for each region of the device are obtained through iterative calculations. In this example, the final dielectric constant distribution range for each region of the semiconductor device is 3.8 to 7.6, and the conductivity distribution range is 0.015 S / m to 0.18 S / m, achieving optimized configuration of material parameters and meeting the design requirements for the device's thermal conductivity performance.

[0132] Figure 3 This is a comparison chart of the effects of temperature gradient analysis and material parameter optimization in embodiments of the present invention, as shown below. Figure 3 As shown in the figure, this visually compares the performance of the three methods in temperature gradient analysis and material parameter optimization. The adaptive weighting method performs best in all five evaluation metrics, especially in computational convergence speed and thermal conductivity efficiency, achieving high efficiencies of 92.1% and 94.2%, respectively. The gradient characteristic method is in the middle, significantly better than the traditional method but slightly inferior to the adaptive weighting method. The traditional method performs weakly in all metrics, especially in material parameter sensitivity, which is only 58.9%. This indicates that the new method based on eigenvalue decomposition and adaptive weighting can effectively improve the accuracy and efficiency of temperature field analysis and material parameter optimization, providing more reliable support for device design.

[0133] In existing technologies, the distribution of material parameters for semiconductor devices mostly relies on static thermal analysis or preset models, lacking the ability to dynamically respond to real-time temperature changes and heat conduction behavior, making it difficult to achieve synergistic optimization of the device's thermal field and electromagnetic performance. This application achieves precise partitioning of the device region by performing eigenvalue decomposition on the temperature gradient matrix, extracting heat conduction direction information and gradient intensity distribution, and constructing physically guided heat conduction channels, thereby enhancing the responsiveness to local thermal behavior. Based on this, by correlating the heat conduction channels with temperature gradient data, a dynamic mapping relationship between dielectric constant and temperature gradient, and between conductivity and temperature, is established, effectively overcoming the limitations of traditional static parameter settings. Furthermore, based on parameter sensitivity calculated using adaptive weights and combined with the continuity analysis of local boundaries, a smooth transition strategy for parameter distribution is proposed, ensuring the physical consistency of material parameters at the boundaries of each region. Compared to the thermal discontinuity and electrical performance drift caused by using a uniform material model or coarse-grained parameter tuning in existing technologies, this application achieves spatially differentiated distribution of dielectric constant and conductivity while maintaining the clarity of the heat conduction path. This significantly improves the thermal stability and electromagnetic consistency of the device under high-frequency operating conditions, meeting the requirements of high-performance semiconductor devices operating in complex environments.

[0134] In one optional implementation, boundary continuity analysis is performed on the local parameter adjustment data to calculate the parameter gradient and boundary features of adjacent regions, construct boundary feature vectors, classify boundary types based on the boundary feature vectors, calculate the comprehensive weight of the regions according to the obtained boundary types, and perform hierarchical smoothing processing on the parameters to generate a parameter smoothing transition scheme, including:

[0135] A multi-directional difference template is used to analyze local parameter adjustment data to obtain parameter gradient information. Based on the parameter gradient information, the normal gradient, tangential gradient and boundary curvature are calculated, and boundary features are extracted.

[0136] The boundary features are normalized, and the normalized parameter gradient magnitude, gradient direction, boundary curvature, and parameter jump variables are combined to construct a boundary feature vector.

[0137] A boundary grading matrix is ​​established based on the boundary feature vector. The first boundary grading is determined according to the parameter gradient magnitude, the second boundary grading is determined by combining the gradient direction continuity, and the third boundary grading is determined by the boundary curvature characteristics.

[0138] A weighted correlation matrix is ​​constructed based on the boundary classification. The boundary classification weight, regional area weight, parameter importance weight, and location relevance weight are input into the weighted correlation matrix to calculate the comprehensive regional weight.

[0139] Based on the comprehensive weight of the region, a smoothing coefficient is selected. The first smoothing coefficient is used for the first boundary graded region, the second smoothing coefficient is used for the second boundary graded region, and the third smoothing coefficient is used for the third boundary graded region. The boundary continuity of the smoothing result is evaluated, and when the boundary continuity meets the preset conditions, the parameter smoothing transition scheme is output.

[0140] The detailed implementation method of the parameter smoothing transition scheme can be achieved by performing boundary continuity analysis on the local parameter adjustment data and classifying it based on the boundary feature vector, thereby realizing hierarchical smoothing of parameters.

[0141] For example, analyzing local parameter adjustment data using a multi-directional differential template is the first step in achieving smooth parameter transitions. In the high-frequency electromagnetic and thermal stability optimization design of semiconductor devices, local parameter adjustment data typically includes the adjusted values ​​of key parameters such as resistivity, thermal conductivity, and dielectric constant of the semiconductor material in different regions. The spatial discontinuous distribution of these parameters can lead to distortions in the electromagnetic and thermal fields, affecting device performance. To obtain parameter gradient information, a 3×3 or 5×5 multi-directional differential template can be used to scan and analyze the parameter distribution. Taking a 5×5 template as an example, on the two-dimensional plane of the semiconductor device, the center of the template is aligned with each node of the parameter grid, and the parameter difference values ​​in the four directions—horizontal, vertical, and diagonal—are calculated respectively. For example, regarding the thermal conductivity parameter of a semiconductor material, a 5×5 grid is set up in a certain region, with each grid cell measuring 0.5 mm × 0.5 mm. The thermal conductivity value of adjacent grid points may abruptly change from 120 W / (m·K) to 80 W / (m·K). The parameter gradient at this point can be calculated using a difference template as 80 W / (m·K) / 0.5 mm. Based on the parameter gradient information, the normal gradient, tangential gradient, and boundary curvature can be further calculated. The normal gradient refers to the gradient component perpendicular to the parameter boundary, indicating the degree of drastic change of the parameter at the boundary; the tangential gradient refers to the gradient component along the parameter boundary, indicating the change of the parameter in the boundary direction; the boundary curvature describes the degree of bending of the parameter boundary and can be obtained by calculating the rate of change of the tangential direction between adjacent points on the boundary. These features together constitute the basic description of the boundary characteristics.

[0142] Normalizing boundary features is a crucial step in constructing boundary feature vectors. Normalization ensures that different types of features are within the same scale range, facilitating subsequent analysis and classification. Normalization of parameter gradient magnitudes can employ a max-min normalization method, mapping the original gradient magnitudes to a range of 0 to 1. For example, if the detected thermal conductivity gradient magnitude on a semiconductor device ranges from 0 to 100 W / (m·K) / mm, the gradient magnitude at each location can be normalized by dividing by 100. Gradient direction is typically represented by angles and can be normalized to a range of -1 to 1, representing the degree of deviation from the reference direction. Boundary curvature also needs normalization, which can be achieved by dividing by the maximum curvature value. Parameter jump variables represent the difference in parameter values ​​across the boundary; for example, resistivity from 1.68 × 10⁻⁶ to 10⁻⁶. -8 Ω·m jumps to 2.44 × 10 -8 Ω·m, jump variable is 0.76×10 -8 Ω·m, during normalization, can be divided by the maximum jump variable detected within the device. After normalization, the parameter gradient magnitude, gradient direction, boundary curvature, and parameter jump variable are combined into a four-dimensional vector, namely the boundary feature vector. Each boundary point corresponds to a feature vector, and these vectors constitute the mathematical representation of the boundary features, providing the foundation for subsequent boundary classification.

[0143] Establishing a boundary grading matrix based on boundary feature vectors is the core of boundary classification. The boundary grading matrix is ​​a data structure used to classify boundary types, containing multi-level classification results. The first boundary grading is determined based on the parameter gradient magnitude, which can be divided into multiple intervals corresponding to different boundary levels. For example, in the electromagnetic-thermal coupling analysis of semiconductor devices, the normalized conductivity gradient magnitude can be divided into three intervals: 0-0.3 for low gradient boundaries, 0.3-0.7 for medium gradient boundaries, and 0.7-1.0 for high gradient boundaries. The second boundary grading is determined by combining gradient direction continuity. Directional continuity can be judged by calculating the rate of change of gradient direction between adjacent boundary points. If the gradient direction change between adjacent points is less than 15 degrees, the direction is considered highly continuous; between 15-45 degrees, it is considered moderately continuous; and greater than 45 degrees, it is considered lowly continuous. The third boundary grading is determined by boundary curvature features, which can classify normalized curvature values ​​into three categories: low curvature (0-0.3), medium curvature (0.3-0.7), and high curvature (0.7-1.0). By combining the results of these three levels of classification, a complete boundary classification matrix can be established. For example, a boundary line on a semiconductor device may be classified as a high gradient boundary at the first level, a medium directional continuous boundary at the second level, a low curvature boundary at the third level, and finally marked as a boundary type that requires special processing.

[0144] Constructing a weighted correlation matrix based on boundary grading is fundamental to calculating the overall weight of regions. The weighted correlation matrix is ​​a data structure that integrates multiple weighting factors, mapping different types of weights to overall region weights. Boundary grading weights are weighting factors determined based on the aforementioned boundary grading results; for example, high-gradient boundaries can be assigned a weight of 0.8, medium-gradient boundaries 0.5, and low-gradient boundaries 0.2. Region area weights reflect the impact of region size on parameter smoothing and can be calculated by dividing the region area by the total device area. For example, if a semiconductor region has an area of ​​100 square millimeters and the total device area is 1000 square millimeters, then the area weight of that region is 0.1. Parameter importance weights represent the degree of influence of parameters on device performance. For example, in high-frequency semiconductor design, conductivity may have a more significant impact on electromagnetic properties than thermal conductivity; conductivity can be assigned an importance weight of 0.7, and thermal conductivity 0.4. Location-related weights consider the location importance of a region within the device; for example, regions near heat sources or critical signal paths can be assigned higher location weights. These four types of weights are input into the weighted correlation matrix, and the overall region weight is calculated through weighted averaging or other combinations. For example, if a region has a boundary classification weight of 0.8, an area weight of 0.3, a parameter importance weight of 0.6, and a location relevance weight of 0.7, then its overall weight may be 0.65, indicating that the region has a high priority in parameter smoothing.

[0145] Selecting the smoothing coefficient based on regional comprehensive weighting is key to achieving hierarchical smoothing. Using different smoothing coefficients for different boundary-level regions allows for a smooth spatial transition of parameters. The first boundary-level region (high gradient boundary) uses the first smoothing coefficient; for example, setting the smoothing coefficient to 0.2 means retaining 80% of the original parameter values ​​and replacing 20% ​​with the weighted average of parameters from adjacent regions. In practical applications, for resistivity at the boundary from 1.68 × 10⁻⁶... -8 Ω·m jumps to 2.44 × 10 -8 For the Ω·m case, a gradual transition can be achieved within 5 grid points near the boundary to avoid abrupt changes. A second smoothing coefficient, such as 0.5, is applied to the second boundary hierarchical region (medium gradient boundary), indicating that the original parameter value and the parameters of adjacent regions each account for 50%. A third smoothing coefficient, such as 0.8, is applied to the third boundary hierarchical region (low gradient boundary), indicating that the weighted average of the parameters of adjacent regions is used for most of the transition. After smoothing, the boundary continuity of the results needs to be evaluated. The smoothing effect can be assessed by calculating the rate of change of the parameter gradient before and after processing. For example, if the maximum parameter gradient is reduced by 80%, the average gradient is reduced by 65%, and the first derivative of the parameter distribution is spatially continuous, then the boundary continuity is considered to meet the preset conditions, and a parameter smoothing transition scheme can be output.

[0146] Parameter smoothing transition schemes have significant application value in the synergistic optimization design of high-frequency electromagnetic characteristics and thermal stability of semiconductor devices. By smoothing, discontinuities in the spatial distribution of parameters can be reduced, electromagnetic and thermal field distortions can be decreased, and the accuracy of simulation analysis can be improved. For example, in the design of high-frequency power electronic devices, abrupt changes in conductivity and thermal conductivity at the interfaces of different materials may lead to local current density concentration and hot spot formation. Parameter smoothing transition schemes can achieve a gradual transition of material properties, reducing current distortion and thermal stress concentration. Furthermore, parameter smoothing transition schemes can also be used for gradient material design in semiconductor devices. By continuously adjusting the material composition and performance parameters in space, synergistic optimization of electromagnetic characteristics and thermal stability can be achieved. For example, in the semiconductor heat dissipation structure design of RF power amplifiers, parameter smoothing transition schemes can generate spatial distribution schemes of conductivity and thermal conductivity, ensuring both high-frequency electromagnetic performance and rapid heat dissipation, extending device lifespan. Parameter smoothing transition schemes are not only applicable to static parameter distribution optimization but also to parameter adjustment under dynamic operating conditions, such as considering changes in material properties caused by temperature increases, to achieve optimized design across the entire operating range.

[0147] In existing technologies, parameter smoothing often employs uniform coefficients or simple weighted averaging methods, which struggle to balance boundary continuity with regional physical characteristics. This can lead to issues such as parameter jumps at boundaries and inconsistent thermo-electrical performance, limiting the refined design and performance optimization of devices. This application constructs boundary feature vectors based on multi-directional differential templates, integrating key features such as parameter gradients, gradient directions, and boundary curvature. For the first time, it achieves refined classification and multi-level hierarchical expression of boundary types. By constructing a boundary hierarchical matrix and combining it with multi-dimensional weighting factors, the comprehensive weight of each region is accurately calculated. Then, differentiated smoothing coefficients are used to implement hierarchical processing for different boundary levels, effectively mitigating local physical distortions caused by parameter abrupt changes. Simultaneously, by introducing a boundary continuity evaluation mechanism, the smoothing results are dynamically verified and optimized, ensuring smooth parameter transitions and gradient continuity between regions, improving the physical consistency of the overall temperature field and material parameter field. Compared to traditional methods, this scheme uses boundary physical characteristics as a guide for hierarchical control, exhibiting stronger adaptability and local sensitivity, significantly enhancing the modeling accuracy and engineering applicability of highly complex devices in thermo-electric coupling scenarios.

[0148] A second aspect of the present invention provides an electronic device, comprising:

[0149] processor;

[0150] Memory used to store processor-executable instructions;

[0151] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0152] A third aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0153] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0154] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for synergistic optimization design of high-frequency electromagnetic characteristics and thermal stability of semiconductor devices, characterized in that, include: Electromagnetic field detection points are arranged at the grid nodes of semiconductor devices to collect the electric field strength and magnetic field strength of the electromagnetic field detection points. The electromagnetic field energy loss of the grid nodes is calculated based on the electric field strength and magnetic field strength. A piecewise linear interpolation method is used to establish the mapping relationship between electromagnetic field energy loss and temperature. The electromagnetic field energy loss is converted into heat source power density based on the mapping relationship. The heat source power density is input into a multi-scale decomposition network to obtain heat distribution characteristic data. Based on the heat distribution characteristic data, the location information of hotspot areas is determined, and time-series temperature data of hotspot areas is established. The time-series temperature data is input into a recurrent neural network to obtain the temperature change trend of hotspot areas. Based on the temperature change trend, a temperature gradient matrix is ​​generated. The thermal conduction characteristics are obtained by performing eigenvalue decomposition on the temperature gradient matrix. Based on the thermal conduction characteristics, the device region is divided and the thermal conduction channels between regions are established. The thermal conduction channels are associated with the temperature gradient to construct the material parameter mapping relationship. The dielectric constant distribution and conductivity distribution are generated through parameter sensitivity analysis and boundary continuity optimization. Using the dielectric constant distribution and conductivity distribution as initial parameters, material parameters are configured for each region of the semiconductor device. Based on the configuration results, a new temperature distribution is calculated. The new temperature distribution is compared with a preset temperature threshold. When the temperature distribution meets the preset threshold requirements, the current dielectric constant distribution and conductivity distribution are determined as the optimized structural parameters of the semiconductor device.

2. The method according to claim 1, characterized in that, Electromagnetic field detection points are arranged at the grid nodes of the semiconductor device to collect the electric and magnetic field strengths of the detection points. The electromagnetic field energy loss of the grid nodes is calculated based on the electric and magnetic field strengths, including: Electromagnetic field detection points are arranged at the grid nodes. Each electromagnetic field detection point includes an electric field detector and a magnetic field detector. The output signals of the electric field detector and the magnetic field detector are sampled to obtain electromagnetic field sampling data. The electromagnetic field sampling data is filtered to obtain the electric field strength and magnetic field strength at the electromagnetic field detection point. The temperature data of the semiconductor device is obtained, the material parameters are determined based on the temperature data, the dielectric loss value is calculated based on the electric field strength, the magnetic loss value is calculated based on the magnetic field strength, and the conduction loss value is calculated based on the electric field strength and the material parameters. The electromagnetic energy loss of the grid node is obtained by adding the dielectric loss value, magnetic loss value and conduction loss value.

3. The method according to claim 1, characterized in that, A piecewise linear interpolation method is used to establish the mapping relationship between electromagnetic field energy loss and temperature. Based on this mapping relationship, the electromagnetic field energy loss is converted into heat source power density, including: Adaptive filtering is performed on the electromagnetic field energy loss data. The filtering parameters are dynamically adjusted according to the signal change amplitude to obtain the filtered electromagnetic field energy loss data. The distribution probability of the filtered electromagnetic field energy loss data is calculated, and the inflection point of the probability change is determined as the segmentation point. The range of electromagnetic field energy loss is divided into multiple sub-intervals. Within each sub-interval, the correlation coefficient between electromagnetic field energy loss data and temperature data is calculated. The correlation coefficient is used as a weight value, and a piecewise linear interpolation method is used to establish a sub-interval mapping function for the weighted data. Calculate the slope difference of the mapping function of adjacent sub-intervals at the segmentation point, adjust the mapping function coefficients according to the slope difference, so that the mapping function of adjacent sub-intervals is continuous at the segmentation point, and obtain the piecewise linear mapping relationship between electromagnetic field energy loss and temperature. Real-time electromagnetic field energy loss data of semiconductor devices is acquired. The real-time electromagnetic field energy loss data is substituted into the piecewise linear mapping relationship to obtain the temperature prediction value. The corresponding material parameter correction factor is obtained based on the temperature prediction value. The real-time electromagnetic field energy loss data is multiplied by the material parameter correction factor to obtain the heat source power density.

4. The method according to claim 1, characterized in that, The heat source power density is input into a multi-scale decomposition network to obtain heat distribution characteristic data. Based on the heat distribution characteristic data, the location information of hotspot regions is determined, and time-series temperature data of the hotspot regions is established, including: The heat source power density is input into a multi-scale decomposition network for signal decomposition to obtain a decomposition coefficient matrix group, which is then divided into multiple scale levels. Local statistical features are calculated for the decomposition coefficient matrices at multiple scale levels to obtain the feature variance matrix and feature mean matrix. A dynamic decision function is constructed based on the feature variance matrix and feature mean matrix. The decomposition coefficient matrix set is compared with the dynamic decision function to obtain heat distribution feature data. The location relationships of hotspot areas are calculated based on the heat distribution characteristic data. Density clustering is used to analyze and process the location relationships to obtain the location information of the hotspot areas. Temperature data is collected based on the location information of the hotspot areas, and the temperature data is organized in chronological order to establish time-series temperature data of the hotspot areas.

5. The method according to claim 1, characterized in that, Time-series temperature data is input into a recurrent neural network to obtain the temperature change trend of hotspot areas. A temperature gradient matrix is ​​then generated based on this trend, including: Feature enhancement processing is performed on time-series temperature data to extract temperature values, temperature changes, temperature change rates, and hotspot volume, and to construct temperature feature vectors. Based on the temperature feature vector, calculate the temperature difference between the hot spot area and the adjacent area, obtain the shape features and location information of the hot spot area, and generate spatial correlation features. The temperature feature vector and spatial correlation features are input into a two-layer recurrent neural network. The first layer of the network extracts the short-term temperature change features and temperature evolution patterns. The short-term temperature change features and temperature evolution patterns are then input into the second layer of the network to establish a long-term temperature dependency relationship and obtain the temperature change trend of the hotspot area. The magnitude and direction of the spatial temperature gradient are calculated based on the temperature change trend. The distribution information of gradient singular points is extracted, and the magnitude and direction of the spatial temperature gradient and the distribution information of singular points are combined to generate a temperature gradient matrix.

6. The method according to claim 1, characterized in that, Eigenvalue decomposition of the temperature gradient matrix yields thermal conductivity characteristics. Based on these characteristics, device regions are divided and inter-regional thermal conduction channels are established. The thermal conduction channels are correlated with the temperature gradient to construct a material parameter mapping relationship. Through parameter sensitivity analysis and boundary continuity optimization, the final dielectric constant and conductivity distributions are generated, including: Eigenvalue decomposition and eigenvector decomposition are performed on the temperature gradient matrix to obtain heat conduction direction feature data. A gradient intensity distribution map is generated based on the heat conduction direction feature data. The device is divided into regions according to the gradient intensity distribution map to obtain material parameter priority region data. A heat conduction channel between adjacent regions is constructed based on the material parameter priority region data. The heat conduction channel is correlated with the temperature gradient matrix to establish a mapping relationship between temperature gradient and dielectric constant and between temperature and conductivity. Adaptive weights of material parameters are calculated based on the mapping relationship. The parameter sensitivity of each region is calculated using the adaptive weights, and the parameter sensitivity is compared with a preset threshold to obtain the parameter update interval. Local parameter adjustment data is then generated based on the parameter update interval. Boundary continuity analysis is performed on the local parameter adjustment data, the parameter gradient and boundary features of adjacent regions are calculated, a boundary feature vector is constructed, the boundary type is classified based on the boundary feature vector, the comprehensive weight of the region is calculated according to the obtained boundary type, the parameters are processed in a hierarchical smoothing manner, and a parameter smoothing transition scheme is generated. Material physical constraints are applied to the parameter smooth transition scheme to establish a mapping relationship between the temperature field and material parameters, and the dielectric constant distribution and conductivity distribution of each region of the device are calculated.

7. The method according to claim 6, characterized in that, Boundary continuity analysis is performed on the local parameter adjustment data to calculate the parameter gradient and boundary features of adjacent regions, construct boundary feature vectors, classify boundary types based on the boundary feature vectors, calculate the comprehensive weight of the regions according to the obtained boundary types, perform hierarchical smoothing processing on the parameters, and generate a parameter smoothing transition scheme, including: A multi-directional difference template is used to analyze local parameter adjustment data to obtain parameter gradient information. Based on the parameter gradient information, the normal gradient, tangential gradient and boundary curvature are calculated, and boundary features are extracted. The boundary features are normalized, and the normalized parameter gradient magnitude, gradient direction, boundary curvature, and parameter jump variables are combined to construct a boundary feature vector. A boundary grading matrix is ​​established based on the boundary feature vector. The first boundary grading is determined according to the parameter gradient magnitude, the second boundary grading is determined by combining the gradient direction continuity, and the third boundary grading is determined by the boundary curvature characteristics. A weighted correlation matrix is ​​constructed based on the boundary classification. The boundary classification weight, regional area weight, parameter importance weight, and location relevance weight are input into the weighted correlation matrix to calculate the comprehensive regional weight. Based on the comprehensive weight of the region, a smoothing coefficient is selected. The first smoothing coefficient is used for the first boundary graded region, the second smoothing coefficient is used for the second boundary graded region, and the third smoothing coefficient is used for the third boundary graded region. The boundary continuity of the smoothing result is evaluated, and when the boundary continuity meets the preset conditions, the parameter smoothing transition scheme is output.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images