A cross-scale electro-thermal performance prediction method based on double-machine learning model data generation

Abstract

This invention discloses a method for predicting cross-scale electrothermal performance based on data generated from dual machine learning models. The method includes: determining actual input parameters and iterative input parameter ranges, such as semiconductor device type, package type, semiconductor device-package level structural parameters, electrical excitation, and environmental factors; collecting electrical and thermal datasets; selecting and training electrical and thermal machine learning models respectively; analyzing the electrical and thermal datasets and setting iteration rules; conducting electrothermal coupling iteration based on the dual machine learning models to generate an electrothermal coupling dataset; and selecting and training the electrothermal coupling machine learning model to achieve rapid prediction of the cross-scale electrothermal performance of semiconductor devices from actual input parameters. This invention effectively solves the problem of complex cross-platform data conversion and transmission between traditional device-level and package-level simulations, significantly improving data generation efficiency and final model training accuracy.

Description

Technical Field

[0001] This invention relates to the field of semiconductor device electrothermal performance prediction technology, and specifically to a cross-scale electrothermal performance prediction method based on dual machine learning model data generation. Background Technology

[0002] Power semiconductor devices are widely used in automotive, avionics, and renewable energy fields. To meet practical application requirements, these devices typically require packaging to provide mechanical support, electrical interconnection, and thermal management. However, cross-scale self-heating effects at the device-package level can significantly impact the device's electrothermal behavior, leading to performance degradation and consequently affecting system stability and reliability. As the power density of semiconductor devices continues to increase, cross-scale self-heating effects become increasingly pronounced. Evaluating the changes in electrothermal performance influenced by device-package structural parameters and operating environment factors has become a crucial foundation for achieving synergistic optimization of device-package and improving system performance.

[0003] In existing technologies, device-level electrothermal simulations typically neglect the influence of package-level structural parameters and environmental factors on the device's electrothermal behavior. Furthermore, package-level thermal simulations do not address the feedback effect of internal device structural parameters and their temperature on electrical performance, making it difficult to comprehensively reflect the device-package level electrothermal coupling characteristics under actual application conditions. Some studies have proposed joint simulations using device-level electrical simulation software and package-level thermal simulation software, evaluating cross-scale electrothermal effects through iterative coupling of power loss and chip temperature. However, due to the lack of a unified data interface between multiple simulation platforms, power loss and temperature data need frequent transfer and conversion between different simulation software, resulting in cumbersome data interaction, complex simulation processes, and low computational efficiency, hindering engineering applications. To improve the efficiency of cross-scale electrothermal performance prediction, invention CN118364728B discloses a device electrothermal stress extraction method based on machine learning and cross-scale iterative coupling. This method constructs a high-quality dataset to train a machine learning model, achieving device-package cross-scale electrothermal performance prediction. This method effectively improves prediction accuracy, but it still has the following shortcomings: due to the dependence of machine learning models on a large amount of high-quality training data, the data generation stage still needs to repeatedly perform device-package level joint calculation iterations, resulting in long data collection cycles and low efficiency. Furthermore, it cannot complete calculations under extreme working conditions or when some inputs are missing, which limits the wide applicability of this method. Summary of the Invention

[0004] The purpose of this invention is to provide a cross-scale electrothermal performance prediction method based on dual machine learning model data generation. By coupling and iterating the electrical and thermal dual machine learning models, the traditional device-level and package-level simulation process is replaced, avoiding the frequent transfer and conversion of complex data between simulation platforms. This method can quickly generate multiple sets of high-quality electrothermal coupling data in a short time, significantly improving the data collection efficiency and prediction accuracy of the electrothermal coupling machine learning model.

[0005] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for predicting cross-scale electrothermal performance based on data generated from dual machine learning models, the method comprising the following steps:

[0007] S1, determine the input parameters related to the semiconductor device-level structure, package-level structure, electrical excitation and environmental factors, and the data range of each input parameter, which are related to the cross-scale electrothermal performance of semiconductor devices;

[0008] S2, under different combinations of device-level structural parameters, electrical excitation and chip temperature, simulates the electrical performance parameters of semiconductor devices and generates an electrical dataset; under different combinations of package-level structural parameters, environmental factors and power loss, simulates the chip temperature of semiconductor devices and generates a thermal dataset.

[0009] S3. Construct and train an electrical machine learning model and a thermal machine learning model respectively. The input parameters of the electrical machine learning model include the semiconductor device type, device-level structural parameters, electrical excitation, and chip temperature. The output parameter is the electrical performance parameters of the semiconductor device. Based on the electrical performance parameter values ​​of the semiconductor device and the set electrical excitation, the power loss value is obtained through the power loss calculation formula. The input parameters of the thermal machine learning model include the semiconductor device package type, package-level structural parameters, environmental factors, and power loss value. The output parameter is the chip temperature of the semiconductor device. Among them, the electrical performance parameters of the semiconductor device include the initial current or initial voltage and the steady-state current or steady-state voltage.

[0010] S4. Analyze the electrical and thermal datasets obtained in step S2 and set iteration rules. Based on the dual machine learning model of electrical and thermal systems, carry out electrothermal coupling iteration to generate an electrothermal coupling dataset.

[0011] S5. Based on the electrothermal coupling dataset, select and train an electrothermal coupling machine learning model. The input parameters of the electrothermal coupling machine learning model are the device-level structure, packaging-level structure, electrical excitation and environmental factors of the semiconductor device. The output parameters are the cross-scale electrothermal performance of the semiconductor device, including temperature state, steady-state chip temperature, initial current or initial voltage and steady-state current or steady-state voltage.

[0012] S6 imports the input parameters from the actual operation process into the trained electrothermal coupling machine learning model to quickly predict the cross-scale electrothermal performance of semiconductor devices.

[0013] Further, in step S1, the device-level structural parameters include all structural parameters covered by nano / micron-level devices, including device type, drift region doping concentration, P-Well doping concentration, drift region thickness, channel length, and chip area; the package-level structural parameters include millimeter / centimeter-level package parameters, including package type, adhesive layer thermal conductivity, molding compound thermal conductivity, and chip area, as well as external device structural parameters affecting heat dissipation, including but not limited to heat sinks and printed circuit boards; electrical excitation refers to the electrical conditions applied to the semiconductor device, including the voltage bias or current bias experienced by the semiconductor device; environmental factors include but are not limited to ambient temperature, ambient radiation intensity, and ambient convection state.

[0014] Furthermore, in step S2, the electrical performance parameters include, but are not limited to, the semiconductor device current value, the semiconductor device voltage value, and the semiconductor device on-resistance value.

[0015] Furthermore, in step S3, the electrical machine learning model and the thermal machine learning model adopt regression models, including but not limited to one or a combination of deep neural networks, Gaussian process regression, support vector machines and random forests.

[0016] Step S4 further includes:

[0017] S41. Based on the type of semiconductor device, determine the temperature limit threshold of the corresponding type of semiconductor device and use it as the temperature threshold for determining the over-temperature state during the iteration process.

[0018] S42, set an iteration threshold to determine the convergence condition of the iteration process. When the ratio of the difference between two consecutive predicted temperature values ​​to the predicted temperature value of the previous round is less than the iteration threshold, it is determined that the iteration process has reached convergence and the iteration calculation is stopped.

[0019] S43, Based on the data range of each input parameter determined in step S1, generate multiple sets of actual input parameter combinations;

[0020] S44, based on the generated actual input parameter combination, calls the trained electrical machine learning model, predicts the corresponding semiconductor device electrical performance parameter values ​​based on device type, device-level structural parameters, electrical excitation and chip temperature, with the initial value of chip temperature being the ambient temperature;

[0021] S45. Based on the obtained electrical performance parameter values ​​of the semiconductor device and the set electrical excitation, the power loss value is obtained through the power loss calculation formula.

[0022] S46, based on the package type, package-level structural parameters, environmental factor parameters and calculated power loss value, calls the trained thermal machine learning model to predict the chip temperature;

[0023] S47, based on the obtained chip temperature, implements adaptive cross-scale electrothermal coupling adjustment according to a set temperature criterion, specifically including:

[0024] S471, determine whether the obtained chip temperature is greater than the preset temperature threshold. If it is greater, output an over-temperature status label and proceed to step S48; if it is less than, execute step S472.

[0025] S472, determine whether the difference between the obtained chip temperature and the chip temperature in the previous round is less than the set iteration threshold. If it is less, output the normal temperature state label, and record the steady-state chip temperature, initial current or initial voltage and steady-state current or steady-state voltage, and proceed to step S48. If it is greater, use the chip temperature value as input again in step S44 and continue iterating.

[0026] S48. Repeat steps S44 to S47 until all actual input parameter combinations generated in step S43 have completed iteration, thus completing the generation of the electrothermal coupling dataset.

[0027] Step S5 further includes:

[0028] S51, a regression model is selected as the electrothermal coupling machine learning model. The input parameters of the electrothermal coupling machine learning model are the device type, package type, device-level structural parameters, package-level structural parameters, electrical excitation and environmental factors of the semiconductor device. The output parameters are the temperature state, steady-state chip temperature, initial current or initial voltage and steady-state current or steady-state voltage.

[0029] The electrothermal coupling machine learning model adopts a multi-branch structure. Based on the input characteristics, the device type, device-level structural parameters, and electrical excitation are used as the first input branch, which is connected to the first fully connected layer group. The package type, package-level structural parameters, and environmental factors are used as the second input branch, which is connected to the second fully connected layer group. Then, the outputs of the two branches are connected to the unified fusion layer, and finally output the temperature state, steady-state chip temperature, initial current or initial voltage, and steady-state current or steady-state voltage. Residual connection modules and attention mechanism modules are embedded in the first fully connected layer group, the second fully connected layer group, and the unified fusion layer to enhance the model's ability to extract dynamic weights for complex coupling relationships and key features. The number, embedding position, and specific form of the residual connection modules and attention mechanism modules are configured according to the data complexity, model depth, and task type, and are applied individually or in combination to each branch or the unified fusion layer.

[0030] S52, the electrothermal coupling dataset obtained in step S4 is divided into training set, validation set and test set according to the set ratio;

[0031] S53, based on the type of electrothermal coupling dataset obtained in step S4, standardize, normalize or logize all data in the dataset; among them, the normal temperature and over-temperature states are converted into two different values ​​respectively.

[0032] S54, the electrothermal coupling machine learning model is trained using the processed training set data to obtain the machine learning model weight values; the trained electrothermal coupling machine learning model is verified or tested using the processed validation set data or test set data.

[0033] Step S6 further includes:

[0034] S61 standardizes, normalizes, or logs the actual operating data related to the device type, package type, device-level structural parameters, package-level structural parameters, electrical excitation, and environmental factors of the input semiconductor device.

[0035] S62 inputs the processed actual operating data into the trained electrothermal coupling machine learning model to predict the chip temperature state, initial current or initial voltage, steady-state current or steady-state voltage, and steady-state chip temperature.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0037] First, the cross-scale electrothermal performance prediction method based on dual machine learning model data generation of the present invention considers generating electrical and thermal datasets based on the principles of electrical and thermal performance, and constructing electrical and thermal machine learning models respectively. Through model iteration, the data generation efficiency can be improved and the data collection time can be significantly shortened. At the same time, based on the proposed dual machine learning model coupled iterative data generation method, the structure of electrical and thermal machine learning models is further optimized to improve the model's ability to identify and filter coupled feature data, reduce abnormal data interference, and make the generated data more suitable for the coupled iterative process. This effectively solves the technical problem that the two types of data are constructed independently, which easily contain some data that is not suitable for coupled iteration, and directly using them for coupled iteration can easily lead to non-convergence of iteration and large prediction errors.

[0038] Second, the cross-scale electrothermal performance prediction method based on dual machine learning model data generation of the present invention, based on the characteristics of semiconductor devices, sets reasonable over-temperature thresholds and iteration stopping criteria to ensure the convergence and prediction stability of the iteration process, effectively solving the problems of iteration instability and low efficiency.

[0039] Third, the cross-scale electrothermal performance prediction method based on dual machine learning model data generation of this invention can effectively simulate the cross-scale electrothermal coupling effect between nanoscale / micrometer-scale device structures and millimeter / centimeter-scale packaged heat dissipation structures through iterative coupling calculation of electrical and thermal machine learning models. This avoids the problems caused by complex data transfer and format conversion between simulation platforms in existing technologies, and significantly improves the generation efficiency and quality of electrothermal coupling data. In particular, residual connections and attention mechanisms are innovatively introduced into the electrical and thermal machine learning models, which enhances the model's ability to extract features of complex coupling relationships, effectively alleviates the overfitting or loss of key information problems that are prone to occur in single-layer networks, and further improves the stability and accuracy of the prediction results.

[0040] Fourth, the cross-scale electrothermal performance prediction method based on dual machine learning model data generation of this invention constructs an electrothermal coupled machine learning model that can rapidly and accurately predict the cross-scale electrothermal performance of semiconductor devices based on device type, package type, device-level structural parameters, package-level structural parameters, electrical excitation, and environmental factors. Furthermore, it innovatively sets up a multi-branch structure, inputting micron-level device structure and electrical excitation parameters, package structure, and environmental factor parameters into independent branch networks, and then completing feature fusion in a unified fusion layer, improving the model's adaptability to the differentiated expression of multi-scale features. Based on the actual data distribution characteristics, residual connections and attention mechanisms are flexibly configured in each branch and fusion layer, further enhancing the model's ability to fit complex electrothermal coupling laws and effectively improving the model's prediction accuracy and generalization ability.

[0041] Fifth, the cross-scale electrothermal performance prediction method based on dual machine learning model data generation of the present invention, while fully considering the structural parameters of nano / micron-level devices and the heat dissipation structural parameters of millimeter / centimeter-level packages, also achieves a full-process speed improvement from data collection to evaluation, providing an efficient tool for cross-scale electrothermal performance evaluation at the device-package level, which is conducive to device-package co-design optimization. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the cross-scale electrothermal performance prediction method based on dual machine learning model data generation according to the present invention.

[0043] Figure 2 The diagram shows a silicon-based VDMOSFET device-package structure according to an embodiment of the present invention; wherein, (a) is a device-level structure diagram of the embodiment, (b) is a TO-247 package-level structure diagram, (c) is a TO-263 package-level structure diagram, (d) is a TO-220 package-level structure diagram, and (e) is a cross-sectional view of the package structure of the embodiment.

[0044] Figure 3The diagram shows the structure of the dual machine learning model in the data generation process of the dual machine learning model in the embodiment of the present invention; wherein, (a) is a schematic diagram of the structure of the electrical machine learning model and (b) is a schematic diagram of the structure of the thermal machine learning model.

[0045] Figure 4 This is a schematic diagram illustrating the comparison of average time obtained from different datasets through dual machine learning models and simulations in an embodiment of the present invention.

[0046] Figure 5 This is a schematic diagram of the electrothermal coupling machine learning model structure in an embodiment of the present invention.

[0047] Figure 6 This is a schematic diagram showing the average time comparison results of electrothermal performance evaluation under different datasets using an electrothermal coupling machine learning model and simulation, according to an embodiment of the present invention. Detailed Implementation

[0048] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0049] like Figure 1 As shown, a method for predicting cross-scale electrothermal performance based on data generated from dual machine learning models includes the following steps.

[0050] Step 1: Determine the actual input parameters and iterative input parameter ranges, including semiconductor device type, package type, semiconductor device-package level structural parameters, electrical excitation, and environmental factors. Specifically, this includes:

[0051] Based on semiconductor device type and package type, this study analyzes the impact of actual and iterative input parameters, such as semiconductor device-package level structural parameters, electrical excitation, and environmental factors, on the cross-scale electrothermal performance of semiconductor devices through theoretical knowledge or numerical simulation. The study identifies parameters with significant impact on electrothermal performance and their data ranges. Actual input parameters include semiconductor device type, package type, semiconductor device-package level structural parameters, electrical excitation, and environmental factors. Device-level structural parameters include all structural parameters covered by nanometer / micrometer-level devices. Package-level structural parameters include all structural parameters of millimeter / centimeter-level packages and their external heat dissipation effects (including but not limited to heat sinks and printed circuit boards). Electrical excitation refers to the electrical conditions applied to the semiconductor device, including the voltage bias or current it experiences. Environmental factors include, but are not limited to, ambient temperature (specifically, the initial chip temperature), ambient radiation levels, and ambient convection conditions. Iterative input parameters are the iterative input parameters of the electrical machine learning model and the thermal machine learning model during the electrothermal coupling iteration process, specifically including the chip temperature input parameter of the electrical machine learning model and the power loss input parameter of the thermal machine learning model. For example, for silicon-based VDMOSFETs, which are micrometer-scale devices, structural parameters include, but are not limited to, gate oxide thickness, channel length, channel doping concentration, drift region thickness, and drift region doping concentration. Package-level structural parameters include all structural parameters of the millimeter / centimeter-scale package and its external factors affecting heat dissipation (including but not limited to heat sinks and printed circuit boards). For the TO-247 package type, which is a millimeter-scale package, structural parameters include, but are not limited to, the thermal conductivity of the molding compound, the thermal conductivity of the adhesive layer, the lead frame material, and the lead material; for pure aluminum heat sinks, structural parameters include, but are not limited to, the length, width, and height of the heat sink, the number of fins, and the fin spacing; for printed circuit boards, structural parameters include, but are not limited to, the length, width, and height of the printed circuit board and the percentage of copper material.

[0052] This embodiment uses a silicon-based VDMOSFET as an example, which is a micrometer-scale device. In this embodiment, based on the semiconductor device type, the package-level structure is selected from three forms: millimeter-scale TO-220, TO263, and TO-247. The heat sink and printed circuit board are not yet included. A schematic diagram of the semiconductor device-package-level structure in this embodiment is shown below. Figure 2 As shown. The structure of a silicon-based VDMOSFET device is as follows. Figure 2 As shown in (a): it includes a drain electrode, an N+ substrate, an N-drift region, a P-type doped drift region, a heavily doped P-type region, a heavily doped N-type drift region, a gate electrode, and a source electrode. The semiconductor device is packaged in TO-220, TO263, and TO-247 packages, as shown in... Figure 2 As shown in (b), (c) and (d) in the figure, its cross-sectional view is as follows. Figure 2As shown in (e) in the diagram. The structures of the three package types include leads (three-pin electrodes, from left to right: gate (G), drain (D), and source (S)), a molding compound, and a chip (for...). Figure 2 The device structure shown in (a) consists of multiple parallel connections, an adhesive layer, and a lead frame. It is worth noting that the package leads have a relatively small impact on temperature and are therefore ignored in this embodiment. Through theoretical knowledge and numerical simulation analysis, the actual input parameters such as the silicon-based VDMOSFET device-package level structural parameters, electrical excitation, and environmental factors, as well as the parameters with a significant impact on electrothermal activity and their data ranges among the iterative input parameters, were determined, as shown in Table 1.

[0053] Table 1. Relevant parameters and data ranges for this embodiment.

[0054]

[0055] Step 2: Collect electrical and thermal datasets.

[0056] In this embodiment, the electrical dataset is a dataset of current values ​​for corresponding semiconductor devices obtained based on device type, device-level structural parameters, electrical excitation, and chip temperature; the thermal dataset is a dataset of chip temperature values ​​for corresponding semiconductor devices obtained based on package type, package-level structural parameters, ambient temperature, and power loss. The process of acquiring the electrical dataset is as follows: based on the device type, under different combinations of device-level structural parameters, electrical excitation, and chip temperature, device-level electrical performance parameter values ​​are obtained through device-level electrical calculations, including but not limited to circuit simulation software, TCAD tools, and mathematical or physical calculation formulas, to form the electrical dataset. The process of acquiring the thermal dataset is as follows: based on the package type, under different combinations of package-level structural parameters, environmental factors, and power loss, package-level thermal calculations, including but not limited to thermal simulation software and thermal network models, are used to obtain semiconductor device chip temperature values ​​to form the thermal dataset.

[0057] In this embodiment, firstly, within the parameter range given in step 1, based on the combination of different device-level structural parameters, electrical excitation, and chip temperature, the semiconductor device current values ​​are obtained using the Silvaco software in the TCAD tool, totaling 1,401,148 data points, forming an electrical dataset. Then, within the parameter range given in step 1, based on the combination of different package-level structural parameters, ambient temperature, and power loss, the semiconductor device chip temperature values ​​are obtained using the thermal simulation tool ANSYS Icepak, totaling 11,760 data points, forming a thermal dataset.

[0058] Step 3: Based on the electrical and thermal datasets obtained in Step 2, select and train the electrical machine learning model and the thermal machine learning model. In this embodiment, the specific steps include:

[0059] Step 3.1: Selecting a machine learning model: Based on the electrical and thermal datasets obtained in Step 2, the selected electrical and thermal machine learning models are both regression models, including but not limited to one or a combination of deep neural networks, Gaussian process regression, support vector machines, and random forests.

[0060] Step 3.2, Dataset Classification: The electrical and thermal datasets obtained in Step 2 are divided into training, validation, and test sets in a ratio of 7:1.5:1.5.

[0061] Step 3.3, Dataset Processing: Based on the type of electrical and thermal datasets obtained in Step 2, standardize, normalize, and logize all data in the electrical and thermal datasets.

[0062] Step 3.4, Machine Learning Model Construction: Based on the machine learning model selected in Step 3.1, the structures and input / output parameters of the electrical and thermal machine learning models are set respectively. The input parameters of the electrical machine learning model include device type, device-level structural parameters, electrical excitation, and chip temperature; the output parameter is the semiconductor device current value. The input parameters of the thermal machine learning model include package type, package-level structural parameters, ambient temperature, and power loss; the output parameter is the semiconductor device chip temperature value. Depending on the complexity of the problem and data, two fully connected layers, one residual block layer, and one attention mechanism layer are constructed in the electrical and thermal machine learning models respectively to dynamically adjust the feature importance weights, improving the model convergence speed and performance. The constructed model structures are as follows: Figure 3As shown. The electrical machine learning model structure includes an input layer, an output layer, a fully connected layer, a residual connection module, and an attention mechanism module. The input layer receives the device type, device-level structural parameters, electrical excitation, and chip temperature. The output layer predicts the electrical performance parameters of the semiconductor device. Based on actual needs, a multi-layer structure consisting of a fully connected layer, a residual connection module, and an attention mechanism module is set up. The fully connected layer is used to extract the basic nonlinear mapping features from the input parameters; the residual connection module is used to capture the complex coupling relationship between device type, device-level structural parameters, electrical excitation, and chip temperature, avoiding gradient vanishing and information loss in deep networks, and improving the model's ability to fit high-dimensional complex features; the attention mechanism module dynamically allocates weights according to the different importance of different input features to the electrical performance prediction results, focusing on input features that significantly affect the electrical performance of the semiconductor device, improving the model's feature expression ability and prediction accuracy. The thermal machine learning model structure includes an input layer, an output layer, a fully connected layer, a residual connection module, and an attention mechanism module. The input layer receives the package type, package-level structural parameters, environmental factors, and power loss. The output layer predicts the chip temperature value of the semiconductor device. Based on actual needs, a multi-layer structure consisting of a fully connected layer, a residual connection module, and an attention mechanism module is set up. The fully connected layer extracts the basic nonlinear relationship features from the input parameters; the residual connection module is used to capture the complex coupling characteristics between package type, package-level structural parameters, environmental factors, and power loss, improving the model's ability to fit the multi-physics thermal coupling effect; the attention mechanism module dynamically adjusts the input feature weights based on the sensitivity of different input parameters to chip temperature changes, highlighting parameters that have a greater impact on temperature and improving the model's prediction accuracy. In the electrical and thermal machine learning models, the number, embedding position, and specific form of the fully connected layer, residual connection module, and attention mechanism module can be flexibly configured according to data characteristics, model depth, and prediction requirements, and can be applied individually or in combination to the electrical or thermal machine learning models.

[0063] Step 3.5, Machine learning model training: Train the machine learning model constructed in step 3.4 separately using the training set data processed in step 3.3 to obtain the machine learning model weight values.

[0064] Step 3.6, Machine Learning Model Validation and Testing: Validate or test the trained electrical and thermal machine learning models using the validation or test set data processed in Step 3.3.

[0065] Step 4: Analyze the electrical and thermal datasets obtained in Step 2 and set the selection rules. Based on the dual machine learning models of electrical and thermal systems, carry out electrothermal coupling iteration to generate an electrothermal coupling dataset.

[0066] In this embodiment, based on the application characteristics of semiconductor devices, a temperature threshold (denoted as T) is set to determine the over-temperature state during the iteration process. th This ensures that the temperature threshold can reasonably reflect the thermal reliability requirements of this type of device under actual application conditions, and guarantees that the prediction results have good engineering applicability and reliability. An iteration threshold (denoted as ε) is set to determine the convergence condition of the iteration process. When the ratio of the difference between two adjacent predicted temperature values ​​to the previous predicted temperature value is less than the iteration threshold, the iteration process is determined to have converged and the iteration calculation is stopped, thereby ensuring that the prediction process has good convergence and rationality.

[0067] This embodiment focuses on silicon-based VDMOSFETs, which have a maximum withstand temperature of 150°C. Therefore, the temperature threshold T during the iteration process is... th Set the temperature to 150℃; set the iteration threshold ε during the iteration process to 1×10. -3 This ensures that the prediction process has good convergence.

[0068] In this embodiment, the aforementioned electrothermal coupling dataset is a dataset obtained from actual input parameters, namely semiconductor device type, package type, semiconductor device-package level structural parameters, electrical excitation, and ambient temperature, to obtain corresponding cross-scale electrothermal performance, namely temperature state, initial current, steady-state current, and steady-state chip temperature. Specifically, it includes the following steps:

[0069] Step 4.1: Based on the actual input parameter range determined in Step 1, generate multiple sets of actual input parameter combinations. In this embodiment, taking one set of actual input parameters from the parameter range given in Table 1 as an example, the following steps are performed as shown in Table 2.

[0070] Table 2. First set of actual input parameters

[0071]

[0072] Step 4.2: Taking the set of input parameters given in Step 4.1 as an example, based on the device-level structural parameters, electrical excitation, and chip temperature (at this time, the ambient temperature), Using the electrical machine learning model trained in step 3, predict the current value of the semiconductor device (denoted as I). DS At this time, I DS =0.95A.

[0073] Step 4.3: Based on the semiconductor device current value obtained in Step 4.2, and combined with the set electrical excitation conditions (applying a constant gate and drain voltage bias V), GS =10V, V DS =0.8V), the power loss value (expressed as P) is obtained through the power loss calculation formula. loss ), Ploss =I DS ×V DS =0.95A×0.8V=0.76W.

[0074] Step 4.4: Based on the package-level structural parameters determined in Step 4.1, the ambient temperature parameters, and the power loss value obtained in Step 4.3, call the thermal machine learning model trained in Step 3 to predict the chip temperature (denoted as...). ),at this time .

[0075] Step 4.5: Based on the chip temperature obtained in Step 4.4, adaptive cross-scale electrothermal coupling adjustment is implemented according to the set temperature criterion, specifically including:

[0076] Step 451: Since this embodiment only considers one device type, the corresponding T th The value will only be 150℃;

[0077] The chip temperature value obtained in steps 452 and 4.4 is 89.11℃, which is less than the temperature threshold, so we continue to step 453.

[0078] Step 453: In this embodiment, the deviation threshold ε is 1×10 -3 Calculate the difference between the chip temperature obtained in step 4.4 and the chip temperature in the previous round. This value is greater than 1×10 -3 Therefore, the chip temperature value of 89.11℃ is continued to be input into step 4.2 as the input parameter for the electrical machine learning model. This continues until the difference between the chip temperature and the previous chip temperature is less than 1×10. -3 In the output parameter temperature state, "normal temperature" is output; in the steady-state chip temperature, "77.07℃" is output; in the initial current, "0.95A" is output; and in the steady-state current, "0.76A" is output, thus completing the collection of a set of data.

[0079] Step 4.6 continues until all actual input parameter combinations from Step 4.1 have been iterated through, thus completing the generation of the electrothermal coupling dataset.

[0080] In this embodiment, the average time for obtaining different datasets through dual machine learning models and simulation is compared, such as... Figure 4 As shown.

[0081] Step 5: Select and train the electrothermal coupling machine learning model based on the dataset obtained in Step 4. In this embodiment, the specific steps include:

[0082] Step 5.1: Selecting a machine learning model: Based on the electrothermal coupling dataset obtained in Step 4, the selected electrothermal coupling machine learning model is a deep neural network model.

[0083] Step 5.2, Dataset Classification: The electrothermal coupling dataset obtained in Step 4 is divided into training set, validation set and test set in a ratio of 7:1.5:1.5.

[0084] Step 5.3, Dataset Processing: Based on the type of the electrothermal coupling dataset obtained in Step 4, standardize and logarithmize all data in the dataset. Specifically, since the temperature state parameters were recorded as "normal temperature" and "over-temperature" during the collection of the electrothermal coupling dataset in Step 4, these two state records need to be converted from text to numerical values ​​for training the model. In this embodiment, the "normal temperature" label is recorded as the value "1", and the "over-temperature" label is recorded as the value "0".

[0085] Step 5.4, Machine Learning Model Construction: Based on the machine learning model selected in Step 5.1, the electrothermal coupling machine learning model structure and input / output parameters are set. The input parameters of the electrothermal coupling machine learning model are device type, package type, semiconductor device-package level structural parameters, electrical excitation, and ambient temperature. The output parameters are temperature state, steady-state chip temperature, initial current, and steady-state current. Depending on the complexity of the problem and data, a three-layer fully connected structure is constructed in the electrothermal coupling machine learning model. Device type / micron-level device structural parameters / electrical excitation serve as the first input branch, connected to the first fully connected layer group. Package type / package structural parameters / ambient temperature serve as the second input branch, connected to the second fully connected layer group. The outputs of both branches are then connected to a unified fusion layer, ultimately outputting the temperature state, steady-state chip temperature, initial current, and steady-state current. This model structure can better distinguish input parameters at different scales, avoiding information coupling interference caused by large data differences, which can lead to model inaccuracies. The constructed model structure is as follows: Figure 5 As shown.

[0086] Step 5.5, Machine learning model training: Train the machine learning model constructed in step 5.4 using the training set data processed in step 5.3.

[0087] Step 5.6, Machine Learning Model Validation and Testing: Validate or test the electrothermal coupling machine learning model trained in Step 5.5 using the validation set data or test set data processed in Step 5.3.

[0088] Step 6: Using the electrothermal coupling machine learning model described in Step 5, achieve rapid prediction of the cross-scale electrothermal performance of semiconductor devices from actual input parameters. In this embodiment, the specific steps include:

[0089] Step 6.1 Input Parameter Data Processing: In this embodiment, a set of actual input parameters within the range of Table 1 are randomly selected, as shown in Table 3. The input semiconductor device-package level structural parameters, electrical excitation, and ambient temperature are standardized and logarithmized.

[0090] Table 3. Second set of actual input parameters

[0091]

[0092] Step 6.2, Device electrothermal performance prediction: Input the parameters processed in Step 6.1 into the electrothermal coupling machine learning model trained in Step 5 to quickly predict the corresponding chip temperature state, initial current or initial voltage, steady-state current or steady-state voltage, and steady-state chip temperature. That is, the temperature state is "normal temperature", the steady-state chip temperature is "56.94℃", the initial current is "0.25A", and the steady-state current is "0.23A".

[0093] In this embodiment, the average time for obtaining different datasets using an electrothermal coupling machine learning model and simulation is compared, such as... Figure 6 As shown in the figure, the proposed method for predicting cross-scale electrothermal performance based on dual machine learning model data generation fully considers the cross-scale electrothermal coupling effect between the device and package levels, and effectively solves the problem of complex cross-platform data conversion and transmission in traditional device-level and package-level simulations, significantly improving data generation efficiency and final model training accuracy. The finally constructed electrothermal coupling machine learning model can predict cross-scale electrothermal performance from semiconductor device-package level structural parameters, electrical excitation, and environmental factors. It has advantages such as high prediction accuracy, fast calculation speed, and reduced experimental manpower and costs, which is beneficial for joint optimization design of the device and package levels.

[0094] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0095] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for predicting cross-scale electrothermal performance based on data generated from dual machine learning models, characterized in that, The method includes the following steps: S1, determine the input parameters related to the semiconductor device-level structure, package-level structure, electrical excitation and environmental factors, and the data range of each input parameter, which are related to the cross-scale electrothermal performance of semiconductor devices; S2, under different combinations of device-level structural parameters, electrical excitation and chip temperature, simulates the electrical performance parameters of semiconductor devices and generates an electrical dataset; under different combinations of package-level structural parameters, environmental factors and power loss, simulates the chip temperature of semiconductor devices and generates a thermal dataset. S3. Construct and train an electrical machine learning model and a thermal machine learning model respectively. The input parameters of the electrical machine learning model include the semiconductor device type, device-level structural parameters, electrical excitation, and chip temperature. The output parameter is the electrical performance parameters of the semiconductor device. Based on the electrical performance parameter values ​​of the semiconductor device and the set electrical excitation, the power loss value is obtained through the power loss calculation formula. The input parameters of the thermal machine learning model include the semiconductor device package type, package-level structural parameters, environmental factors, and power loss value. The output parameter is the chip temperature of the semiconductor device. Among them, the electrical performance parameters of the semiconductor device include the initial current or initial voltage and the steady-state current or steady-state voltage. S4. Analyze the electrical and thermal datasets obtained in step S2 and set iteration rules. Based on the dual machine learning model of electrical and thermal systems, carry out electrothermal coupling iteration to generate an electrothermal coupling dataset. S5. Based on the electrothermal coupling dataset, select and train an electrothermal coupling machine learning model. The input parameters of the electrothermal coupling machine learning model are the device-level structure, packaging-level structure, electrical excitation and environmental factors of the semiconductor device. The output parameters are the cross-scale electrothermal performance of the semiconductor device, including temperature state, steady-state chip temperature, initial current or initial voltage and steady-state current or steady-state voltage. S6 imports the input parameters from the actual operation process into the trained electrothermal coupling machine learning model to quickly predict the cross-scale electrothermal performance of semiconductor devices.

2. The method for predicting cross-scale electrothermal performance based on dual machine learning model data as described in claim 1, characterized in that, In step S1, the device-level structural parameters include all structural parameters of nano / micron-level devices, including device type, drift region doping concentration, P-Well doping concentration, drift region thickness, channel length, and chip area; the package-level structural parameters include millimeter / centimeter-level package parameters, including package type, adhesive layer thermal conductivity, molding compound thermal conductivity, and chip area, as well as external device structural parameters that affect heat dissipation, including but not limited to heat sinks and printed circuit boards; electrical excitation refers to the electrical conditions applied to the semiconductor device, including the voltage or current bias that the semiconductor device withstands; environmental factors include but are not limited to ambient temperature, ambient radiation intensity, and ambient convection conditions.

3. The method for predicting cross-scale electrothermal performance based on dual machine learning model data as described in claim 1, characterized in that, In step S2, the electrical performance parameters include, but are not limited to, the semiconductor device current value, the semiconductor device voltage value, and the semiconductor device on-resistance value.

4. The method for predicting cross-scale electrothermal performance based on dual machine learning model data as described in claim 1, characterized in that, In step S3, the electrical machine learning model and the thermal machine learning model adopt regression models, including but not limited to one or a combination of deep neural networks, Gaussian process regression, support vector machines and random forests.

5. The method for predicting cross-scale electrothermal performance based on dual machine learning model data generation according to claim 1, characterized in that, Step S4 further includes: S41. Based on the type of semiconductor device, determine the temperature limit threshold of the corresponding type of semiconductor device and use it as the temperature threshold for determining the over-temperature state during the iteration process. S42, set an iteration threshold to determine the convergence condition of the iteration process. When the ratio of the difference between two consecutive predicted temperature values ​​to the predicted temperature value of the previous round is less than the iteration threshold, it is determined that the iteration process has reached convergence and the iteration calculation is stopped. S43, Based on the data range of each input parameter determined in step S1, generate multiple sets of actual input parameter combinations; S44, based on the generated actual input parameter combination, calls the trained electrical machine learning model, predicts the corresponding semiconductor device electrical performance parameter values ​​based on device type, device-level structural parameters, electrical excitation and chip temperature, with the initial value of chip temperature being the ambient temperature; S45. Based on the obtained electrical performance parameter values ​​of the semiconductor device and the set electrical excitation, the power loss value is obtained through the power loss calculation formula. S46, based on the package type, package-level structural parameters, environmental factor parameters and calculated power loss value, calls the trained thermal machine learning model to predict the chip temperature; S47, based on the obtained chip temperature, implements adaptive cross-scale electrothermal coupling adjustment according to a set temperature criterion, specifically including: S471, determine whether the obtained chip temperature is greater than the preset temperature threshold. If it is greater, output an over-temperature status label and proceed to step S48; if it is less than, execute step S472. S472, determine whether the difference between the obtained chip temperature and the chip temperature in the previous round is less than the set iteration threshold. If it is less, output the normal temperature state label, and record the steady-state chip temperature, initial current or initial voltage and steady-state current or steady-state voltage, and proceed to step S48. If it is greater, use the chip temperature value as input again in step S44 and continue iterating. S48. Repeat steps S44 to S47 until all actual input parameter combinations generated in step S43 have completed iteration, thus completing the generation of the electrothermal coupling dataset.

6. The method for predicting cross-scale electrothermal performance based on dual machine learning model data as described in claim 1, characterized in that, Step S5 further includes: S51, a regression model is selected as the electrothermal coupling machine learning model. The input parameters of the electrothermal coupling machine learning model are the device type, package type, device-level structural parameters, package-level structural parameters, electrical excitation and environmental factors of the semiconductor device. The output parameters are the temperature state, steady-state chip temperature, initial current or initial voltage and steady-state current or steady-state voltage. The electrothermal coupling machine learning model adopts a multi-branch structure. Based on the input characteristics, the device type, device-level structural parameters, and electrical excitation are used as the first input branch, which is connected to the first fully connected layer group. The package type, package-level structural parameters, and environmental factors are used as the second input branch, which is connected to the second fully connected layer group. Then, the outputs of the two branches are connected to the unified fusion layer, and finally output the temperature state, steady-state chip temperature, initial current or initial voltage, and steady-state current or steady-state voltage. Residual connection modules and attention mechanism modules are embedded in the first fully connected layer group, the second fully connected layer group, and the unified fusion layer to enhance the model's ability to extract dynamic weights for complex coupling relationships and key features. The number, embedding position, and specific form of the residual connection modules and attention mechanism modules are configured according to the data complexity, model depth, and task type, and are applied individually or in combination to each branch or the unified fusion layer. S52, the electrothermal coupling dataset obtained in step S4 is divided into training set, validation set and test set according to the set ratio; S53, based on the type of electrothermal coupling dataset obtained in step S4, standardize, normalize or logize all data in the dataset; among them, the normal temperature and over-temperature states are converted into two different values ​​respectively. S54, the electrothermal coupling machine learning model is trained using the processed training set data to obtain the machine learning model weight values; the trained electrothermal coupling machine learning model is verified or tested using the processed validation set data or test set data.

7. The method for predicting cross-scale electrothermal performance based on dual machine learning model data as described in claim 1, characterized in that, Step S6 further includes: S61 standardizes, normalizes, or logs the actual operating data related to the device type, package type, device-level structural parameters, package-level structural parameters, electrical excitation, and environmental factors of the input semiconductor device. S62 inputs the processed actual operating data into the trained electrothermal coupling machine learning model to predict the chip temperature state, initial current or initial voltage, steady-state current or steady-state voltage, and steady-state chip temperature.

Images