AI-based automatic extraction and optimization method of diode model parameters

Abstract

The application discloses an AI-based diode model parameter automatic extraction and optimization method, and belongs to the technical field of electronic design.The application solves the problem that the existing diode model parameter extraction method is seriously dependent on manual experience, the process is complicated and time-consuming, and it is difficult to stably obtain high-precision and strong generalization capability model parameters from nonlinear data under the influence of temperature and process deviation.The application constructs an intelligent mapping model from macro electrical characteristics to microscopic model parameters by adaptively denoising and multidimensional feature engineering on measured data, and finally generates a high-fidelity diode electronic design model by combining global optimization of physical constraints, so as to realize high-precision, high-efficiency and full-automatic extraction and optimization of key model parameters such as diode saturation current, ideal factor and series resistance, thereby improving the efficiency, automation level and cross-condition robustness of model extraction.

Description

Technical Field

[0001] This invention relates to the field of electronic design technology, specifically to an AI-based method for automatic extraction and optimization of diode model parameters. Background Technology

[0002] With the continuous advancement of semiconductor technology and the increasing complexity of integrated circuits, accurate device models are crucial for circuit design, simulation, and reliability analysis. As a fundamental and widely used semiconductor device, the accuracy of diodes' compact models (such as SPICE models) directly impacts the design performance of analog, RF, and power circuits. These models typically include multiple physical parameters such as saturation current, ideality factor, series resistance, and parallel resistance, which need to be accurately extracted from the measurement data of actual devices. Traditional model parameter extraction methods mainly rely on manual fitting or iterative optimization based on analytical formulas. These processes are cumbersome, time-consuming, and heavily dependent on engineers' professional experience. Furthermore, because diode characteristics are significantly affected by factors such as temperature and process variations, traditional methods often struggle to quickly and stably obtain globally optimal model parameters when dealing with multi-condition and nonlinear data, resulting in insufficient model accuracy and generalization ability. Therefore, these methods do not meet current needs. To address this, we propose an AI-based automatic extraction and optimization method for diode model parameters. Summary of the Invention

[0003] The purpose of this invention is to provide an AI-based method for automatic extraction and optimization of diode model parameters. By constructing an intelligent parameter extraction and optimization framework that integrates physical knowledge and data-driven approaches, the invention combines the fast nonlinear mapping capability of machine learning models to macroscopic features with the precise optimization capability of global optimization algorithms in the physical model space. This achieves high-precision, high-efficiency, and high-robust automatic extraction of key physical parameters of compact diode models from multi-condition measured data, thus solving the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an AI-based method for automatic extraction and optimization of diode model parameters, comprising the following steps:

[0005] Step 1: Obtain the measured current-voltage characteristic data of the target diode under different temperatures and bias voltages to form the original training dataset. Then, normalize, filter noise, and remove outliers to form the preprocessed standard dataset.

[0006] Step 2: Extract multi-class features from the standard dataset, including analytical features based on semiconductor physical equations, as well as geometric and statistical features extracted from the current-voltage curve shape. Then, fuse the extracted features to construct a hybrid feature vector to characterize the electrical behavior of the diode under specific operating conditions.

[0007] Step 3: Select a compact diode model containing multiple parameters to be extracted as the initial electronic design model. The parameters to be extracted include at least saturation current, ideality factor, series resistance, and parallel resistance.

[0008] Step 4: Using the hybrid feature vector as input and the parameter vector of the initial electronic design model as output, construct a machine learning model based on support vector regression. Train the machine learning model through supervised learning to establish a nonlinear mapping relationship between the macroscopic electrical characteristics of the diode and the parameters of the compact model.

[0009] Step 5: Perform the preprocessing and feature extraction of the measured current-voltage data of the diode to be modeled under the given test conditions in Step 1 and Step 2 to obtain the corresponding mixed feature vector and input it into the trained machine learning model. The machine learning model directly outputs a set of predicted model parameter values ​​as the initial parameter values ​​of the diode electronic design model.

[0010] Step Six: Using the initial parameter values ​​obtained in Step Five as the starting point, construct an optimization problem based on the initial electronic design model. The objective function is to minimize the error between the simulated data and the measured data. Use a global optimization algorithm to iteratively optimize the model parameters until the error function converges to the preset threshold, and obtain the final optimized model parameters.

[0011] Step 7: Substitute the final model parameters obtained in Step 6 into the initial electronic design model to generate an electronic design model file for the target diode, which can then be used in the circuit simulation environment.

[0012] Furthermore, the noise filtering and outlier removal in step one is specifically a joint denoising method using adaptive weighted dynamic sliding window filtering and physical constraints, including the following steps:

[0013] A dynamic evaluation window is established for each data point in the original current-voltage sequence. The size of this window is adaptively adjusted based on the local first derivative variance of the neighboring region of the data point.

[0014] Within the dynamic evaluation window, the Shockley ideal diode equation is used as a benchmark to calculate the deviation between the measured value and the theoretical value for each data point, and weights are assigned based on this deviation.

[0015] Based on the local statistical characteristics and weights of each data point, a weighted median filter is constructed to filter the current sequence, removing random noise and physically unreasonable outliers.

[0016] The filtered current-voltage sequence is subjected to a global continuity check. The continuity of the first derivative is verified by spline interpolation fitting. Discontinuities are smoothed and corrected, and a standard dataset is output.

[0017] Furthermore, in step two, the process of extracting and fusing multi-class features to construct a hybrid feature vector specifically includes:

[0018] Based on the nonlinear characteristics of the diode current-voltage curve, the voltage axis is adaptively divided into three characteristic physical regions: low bias region, medium bias region and high bias region.

[0019] Within each characteristic physical region, analytical features are extracted based on semiconductor physical equations, specifically as follows:

[0020] In the low bias region, the first analytical feature and the second analytical feature are extracted based on the linear segment of the logarithmic current and voltage, where the first analytical feature is the ideal factor estimate and the second analytical feature is the saturation current estimate.

[0021] In the mid-bias region, the third and fourth analytical features are extracted based on the first and second derivative curves of current versus voltage. The third analytical feature is the voltage at the maximum curvature point, and the fourth analytical feature is the dynamic resistance at the inflection point.

[0022] In the high bias region, the fifth and sixth analytical features are extracted based on the linear segment of the current-voltage curve. The fifth analytical feature is the estimated value of the series resistance, and the sixth analytical feature is the curve linearity deviation factor.

[0023] For the entire current-voltage curve, at least three global morphological features are extracted, including the curvature integral in global logarithmic coordinates, the equivalent area of ​​charge in a specified voltage range, and the parallelism offset statistics between the families of current-voltage curves at different temperatures.

[0024] The aforementioned analytical features and global morphological features are input into a lightweight gated attention network. The gated attention network assigns fusion weights to each input feature and generates weighted basic features.

[0025] The weighted basic features are connected across layers with the first to sixth analytical features and the global morphological features. The connected features are then subjected to a nonlinear transformation to form higher-order abstract features. The top-level feature vector containing these higher-order abstract features is then used as the hybrid feature vector.

[0026] Furthermore, in step four, when constructing and training the machine learning model based on support vector regression, a multi-core hierarchical weighted support vector regression framework is adopted, specifically including:

[0027] For the sub-feature sets with different physical meanings and dimensions in the hybrid feature vector, dedicated kernel functions are initialized respectively. Specifically, a radial basis kernel function is used for the analytical feature subset based on semiconductor physical equations, a polynomial kernel function is used for the geometric and statistical feature subset extracted from the current-voltage curve shape, and a linear kernel function is introduced.

[0028] Construct a two-level cascaded regression structure as follows:

[0029] The first layer is a parallel kernel mapping and prediction layer, which simultaneously inputs the mixed feature vectors into the support vector regression sub-models corresponding to the radial basis function kernel function, the multinomial kernel function and the linear kernel function, and each support vector regression sub-model independently outputs a set of preliminary predicted values ​​of model parameters.

[0030] The second layer is an adaptive weight fusion layer connected to a lightweight meta-learner. The lightweight meta-learner dynamically assigns fusion weights to the preliminary predicted values ​​output by each support vector regression sub-model in the first layer based on the input fusion feature vector, and uses the weighted fusion result as the final model parameter vector output by the machine learning model.

[0031] During the training phase of the machine learning model, an objective function is constructed that includes a prediction error term and a physical consistency regularization term. The physical consistency regularization term is calculated as follows:

[0032] Substitute the preliminary prediction value of any support vector regression submodel in the first layer into the initial electronic design model to generate a simulated current-voltage curve, and calculate the morphological difference measure between the simulated curve and the real curve.

[0033] Furthermore, in step five, before inputting the mixed feature vector of the diode to be modeled into the trained machine learning model, a feature domain adaptive calibration step is performed, specifically including:

[0034] Calculate the maximum mean difference between the mixed feature vector of the target diode and the mixed feature vector of the standard dataset in the space formed by the hidden layers of the machine learning model;

[0035] Determine whether the maximum mean difference is greater than a preset offset threshold;

[0036] When the maximum mean difference is greater than the offset threshold, the mixed feature vectors and model parameter labels corresponding to at least three sets of measured current-voltage data of the target diode are used to fine-tune the parameters of the last layer of the trained machine learning model to complete the model calibration.

[0037] The mixed feature vector of the target diode is input into a calibrated machine learning model, which outputs the predicted initial values ​​of the model parameters.

[0038] Furthermore, in step five, after the machine learning model outputs a set of predicted initial values ​​for the model parameters, a verification and correction step is performed, specifically including:

[0039] Establish an expert rule base based on semiconductor physics. The rule base defines the reasonable range of values ​​for each model parameter, the mutual constraints between parameters, and the constraints on the changing trends at different temperatures.

[0040] The initial values ​​predicted by the machine learning model are automatically compared with the expert rule base to check whether the values ​​of each parameter fall within their physically reasonable hard boundaries. Then, the logical relationship between key parameter pairs is verified to conform to physical laws, such as whether the series resistance is non-negative, whether the ideal factor is within the typical range, and whether the temperature coefficient conforms to the material properties.

[0041] If the initial value passes all rule checks, it is output directly. If a rule violation is found, the corresponding correction mechanism is activated according to the type of rule violation. Specifically:

[0042] For parameters that are slightly out of bounds, force the parameters that are slightly out of bounds back to the nearest reasonable boundary value;

[0043] In cases where there are serious conflicts between parameters, the key parameters that are more sensitive to the impact on the global error are adjusted first, and the parameters that have mutual constraints with them are recalculated until the set of initial parameter values ​​output satisfies all physical constraints, forming the final physically consistent initial parameter values ​​for subsequent optimization.

[0044] Furthermore, in step six, the iterative optimization of the global optimization algorithm adopts a two-stage hybrid optimization strategy, specifically including:

[0045] In the first stage, the initial parameter values ​​obtained in step five are used as the initial population center, and an adaptive covariance matrix evolution strategy is executed in the parameter space for global sampling.

[0046] Simultaneously, a Gaussian process surrogate model is constructed and dynamically updated to fit the mapping relationship between the model parameter combination and the simulation error. In this stage, the optimization process executes the acquisition function optimization guided by the surrogate model. The acquisition function comprehensively balances the predicted mean and predicted variance of the surrogate model.

[0047] In the second stage, the region formed by the optimal parameter points obtained in the first stage is switched to the quasi-Newton method for local optimization. In this stage, physical constraints on the parameter update direction are constructed based on the semiconductor physical relationships of the initial electronic design model.

[0048] Furthermore, in step six, during the iterative process of the two-stage hybrid optimization strategy, an adaptive adjustment mechanism for dynamic error weights and convergence criteria is integrated, specifically including:

[0049] Based on the measured current-voltage curves, the low bias region, medium bias region, and high bias region are divided. Dynamically adjustable weight coefficients are assigned to the error terms corresponding to each region. The adjustment of the weight coefficients is based on the magnitude of the fitting residuals of each region in the current iteration.

[0050] A composite convergence criterion is set up, which simultaneously monitors three indicators, including the relative improvement rate of the objective function value, the norm of change of the key model parameter vector, and the morphological similarity index between the simulated curve and the measured curve.

[0051] The optimization process is considered to have converged when the relative improvement rate of the objective function value, the norm of change of the key model parameter vector, and the morphological similarity index are all lower than their respective set thresholds in a series of preset iterations.

[0052] Furthermore, in step six, after determining that the optimization process has converged, and before generating the electronic design model file in step seven, a model confidence assessment step based on physical-data dual consistency is performed. This model confidence assessment step specifically includes:

[0053] Obtain the final optimized model parameters output from step six, and use the initial electronic design model to calculate the simulated current and simulated conductance values ​​at all test voltage points in the standard dataset;

[0054] Based on the measured current-voltage characteristic data in the standard dataset, the measured conductance values ​​at all test voltage points are calculated. The measured conductance values ​​are obtained by performing a first-order central difference operation on the measured current-voltage characteristic data.

[0055] Using the following formula for calculating the physical-data dual consistency confidence score, a comprehensive score index characterizing the physical reliability of the model is calculated. :

[0056]

[0057] in, It is a dimensionless comprehensive scoring index; This represents the total number of test voltage points. The index number of the voltage point; For the first The simulated current values ​​at each voltage point are in amperes. For the first The measured current values ​​at each voltage point are in amperes. For the first Simulated conductance values ​​at each voltage point, in Siemens units; For the first Measured conductance values ​​at each voltage point, in Siemens units; To prevent numerical stability constants with a denominator of zero, the unit must be consistent with the unit of the physical quantity in the denominator; These are the weighting coefficients of the data fitting term. Let be the weighting coefficient of the physical form term, and satisfy . and The sum equals 1; It is an exponential function with the natural constant e as its base; This is the physical boundary sensitivity coefficient, with a unit of 1; This is a parameter out-of-bounds penalty term. When all parameters of the final optimized model are within the preset reasonable range for semiconductor physics, The value is 0 when there are parameters that exceed the physically reasonable range of the semiconductor. The value is the sum of the normalized distances of all out-of-bounds parameters from their respective physical boundaries, and its specific calculation formula is:

[0058]

[0059] Among them, the single-parameter deviation function Defined as:

[0060]

[0061] The parameters in the above formula are defined as follows:

[0062] This represents the total number of model parameters to be extracted.

[0063] This is the index number of the model parameter, and its value range is... to ;

[0064] For the first Predicted or optimized values ​​of each model parameter;

[0065] For the first The lower bound thresholds of each model parameter are defined in the semiconductor physics rule base;

[0066] For the first The upper bound thresholds of each model parameter are defined in the semiconductor physics rule base;

[0067] and These represent the absolute values ​​of the lower and upper bound thresholds, respectively.

[0068] This is a numerical stability constant used to prevent the denominator from being zero; its value is [value missing]. .

[0069] Judge the comprehensive score index obtained from the calculation. If the parameters exceed the preset release threshold, the final optimized model parameters are deemed qualified and step seven is executed; otherwise, the final optimized model parameters are deemed to have physical consistency defects, the physical constraint weights of the objective function in step six are adjusted, and step six is ​​returned to iterate and optimize again.

[0070] Furthermore, while performing the model confidence assessment step, an adversarial stress test step based on virtual boundary extension is also performed in parallel to verify the robustness of the model in the non-sampling region. The adversarial stress test step specifically includes:

[0071] Based on the voltage coverage range of the measured current-voltage characteristic data in the standard dataset, mathematical extensions are performed to the positive high-level region and the reverse cutoff region of the voltage axis to generate several virtual voltage probe points that exceed the measured range, thus forming a virtual test set.

[0072] Input the final optimized model parameters output from step six and the virtual test set into the initial electronic design model to calculate the corresponding virtual response current sequence;

[0073] The virtual response current sequence is subjected to a first derivative sign test and a second derivative convexity test to verify whether the virtual response current sequence remains monotonically increasing in the positive high-level region without inflection point oscillation, and whether the leakage current remains flat or conforms to the monotonically rapid increasing trend of avalanche breakdown in the reverse cutoff region.

[0074] If the verification finds that the virtual response current sequence violates the physical constraints of monotonically increasing or convexity at any virtual voltage probe point, it is determined that the current model has an overfitting risk. The voltage interval corresponding to the virtual voltage probe point that violates the physical constraints is marked as a high-risk interval, and the high-risk interval is fed back to step six. In the global optimization algorithm of step six, an additional smoothing constraint regularization term is applied to the high-risk interval, and then the optimization iteration is restarted until the model passes the adversarial stress test.

[0075] Compared with the prior art, the beneficial effects of the present invention are:

[0076] This invention constructs a complete AI-driven parameter extraction and optimization process by introducing feature engineering that integrates multiple categories of physical and morphological features, intelligent parameter initial value prediction based on multi-core hierarchical weighted support vector regression machine, and a two-stage global optimization strategy that combines surrogate model guidance and physical constraints. This method achieves high-precision and fully automatic extraction of diode model parameters, reduces the dependence of traditional methods on professional experience, and effectively overcomes the problems of easily getting trapped in local optima and poor robustness under multi-condition nonlinear data. This significantly improves the extraction efficiency, accuracy, and generalization ability of model parameters under complex working conditions, thus providing a more reliable device model foundation for high-performance integrated circuit design. Attached Figure Description

[0077] Figure 1 This is a flowchart of the AI-based automatic extraction and optimization method for diode model parameters according to the present invention;

[0078] Figure 2 This is an execution diagram of the AI-based automatic extraction and optimization method for diode model parameters according to the present invention. Detailed Implementation

[0079] 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.

[0080] To address the problems of existing diode model parameter extraction methods, which heavily rely on manual experience, are cumbersome and time-consuming, and struggle to reliably extract high-precision, high-generalization-ability model parameters from nonlinear data under the influence of temperature and process variations, please refer to [link to relevant documentation]. Figures 1-2 This embodiment provides the following technical solution:

[0081] An AI-based method for automatic extraction and optimization of diode model parameters includes the following steps:

[0082] Step 1: Obtain the measured current-voltage characteristic data of the target diode under different temperatures and bias voltages to form the original training dataset. Then, normalize, filter noise, and remove outliers to form the preprocessed standard dataset.

[0083] Step 2: Extract multi-class features from the standard dataset, including analytical features based on semiconductor physical equations, as well as geometric and statistical features extracted from the current-voltage curve shape. Then, fuse the extracted features to construct a hybrid feature vector to characterize the electrical behavior of the diode under specific operating conditions.

[0084] Step 3: Select a compact diode model containing multiple parameters to be extracted as the initial electronic design model. The parameters to be extracted include at least saturation current, ideality factor, series resistance, and parallel resistance.

[0085] Step 4: Using the hybrid feature vector as input and the parameter vector of the initial electronic design model as output, construct a machine learning model based on support vector regression. Train the machine learning model through supervised learning to establish a nonlinear mapping relationship between the macroscopic electrical characteristics of the diode and the parameters of the compact model.

[0086] Step 5: Perform the preprocessing and feature extraction of the measured current-voltage data of the diode to be modeled under the given test conditions in Step 1 and Step 2 to obtain the corresponding mixed feature vector and input it into the trained machine learning model. The machine learning model directly outputs a set of predicted model parameter values ​​as the initial parameter values ​​of the diode electronic design model.

[0087] Step Six: Using the initial parameter values ​​obtained in Step Five as the starting point, construct an optimization problem based on the initial electronic design model. The objective function is to minimize the error between the simulated data and the measured data. Use a global optimization algorithm to iteratively optimize the model parameters until the error function converges to the preset threshold, and obtain the final optimized model parameters.

[0088] Step 7: Substitute the final model parameters obtained in Step 6 into the initial electronic design model to generate an electronic design model file for the target diode, which can then be used in the circuit simulation environment.

[0089] The technical effects of the above solution are as follows: By utilizing preprocessing and multi-dimensional feature fusion technology, a hybrid feature vector that comprehensively reflects the electrical behavior of diodes is constructed from measured data. Then, a nonlinear mapping from macroscopic features to model parameters is established using support vector regression, which can quickly provide high-quality initial values ​​of model parameters, thereby reducing the dependence of traditional methods on experience and iteration counts. Furthermore, by combining a global optimization algorithm to perform fine-grained iterative optimization of the initial parameter values, the final model can accurately fit the measured characteristics under various working conditions. This improves modeling efficiency and automation level while ensuring the reliability and practicality of the model. In this way, a reliable and scalable automatic diode model generation solution is provided for circuit simulation. In summary, by deeply integrating artificial intelligence algorithms with traditional semiconductor modeling processes, efficient, accurate, and automated extraction of diode model parameters is achieved.

[0090] The noise filtering and outlier removal in step one is specifically a joint denoising method combining adaptive weighted dynamic sliding window filtering and physical constraints, including the following steps:

[0091] A dynamic evaluation window is established for each data point in the original current-voltage sequence. The size of this window is adaptively adjusted based on the local first derivative variance of the neighboring region of the data point.

[0092] Within the dynamic evaluation window, the Shockley ideal diode equation is used as a benchmark to calculate the deviation between the measured value and the theoretical value for each data point, and weights are assigned based on this deviation.

[0093] Based on the local statistical characteristics and weights of each data point, a weighted median filter is constructed to filter the current sequence, removing random noise and physically unreasonable outliers.

[0094] The filtered current-voltage sequence is subjected to a global continuity check. The continuity of the first derivative is verified by spline interpolation fitting. Discontinuities are smoothed and corrected, and a standard dataset is output.

[0095] The technical effects of the above solution are as follows: by combining dynamic adaptive windowing, weight evaluation based on semiconductor physical equations and weighted median filtering, it is possible to achieve accurate and adaptive cleaning of diode test data without the need for a preset fixed noise model. This not only effectively suppresses random noise, but also identifies and removes physically unreasonable outliers caused by measurement errors or non-ideal effects. At the same time, global continuity correction ensures the smoothness and physical consistency of the data curves, thereby providing a high-quality and highly reliable standard dataset for subsequent feature extraction and model training, and thus improving the accuracy of the overall parameter extraction process.

[0096] Step two, the process of extracting and fusing multi-class features to construct a hybrid feature vector, specifically includes:

[0097] Based on the nonlinear characteristics of the diode current-voltage curve, the voltage axis is adaptively divided into three characteristic physical regions: low bias region, medium bias region and high bias region.

[0098] Within each characteristic physical region, analytical features are extracted based on semiconductor physical equations, specifically as follows:

[0099] In the low bias region, the first analytical feature and the second analytical feature are extracted based on the linear segment of the logarithmic current and voltage, where the first analytical feature is the ideal factor estimate and the second analytical feature is the saturation current estimate.

[0100] In the mid-bias region, the third and fourth analytical features are extracted based on the first and second derivative curves of current versus voltage. The third analytical feature is the voltage at the maximum curvature point, and the fourth analytical feature is the dynamic resistance at the inflection point.

[0101] In the high bias region, the fifth and sixth analytical features are extracted based on the linear segment of the current-voltage curve. The fifth analytical feature is the estimated value of the series resistance, and the sixth analytical feature is the curve linearity deviation factor.

[0102] For the entire current-voltage curve, at least three global morphological features are extracted, including the curvature integral in global logarithmic coordinates, the equivalent area of ​​charge within a specified voltage range, and the parallelism offset statistics between the families of current-voltage curves at different temperatures; wherein, the curvature integral in global logarithmic coordinates ( The formula for calculating ) is:

[0103]

[0104] In the formula, and These are the minimum and maximum voltage values ​​measured under the current bias direction, respectively, in volts (V). The measured current value is given at the corresponding voltage, in amperes (A). It is the natural logarithm operator. The equivalent area of ​​the charge ( The formula for calculating ) is:

[0105]

[0106] In the formula, and The starting and ending voltages for a specific physical observation range specified by the user, in volts (V); This represents the measured functional relationship between current and voltage.

[0107] The aforementioned analytical features and global morphological features are input into a lightweight gated attention network. The gated attention network assigns fusion weights to each input feature and generates weighted basic features.

[0108] The weighted basic features are connected across layers with the first to sixth analytical features and the global morphological features. The connected features are then subjected to a nonlinear transformation to form higher-order abstract features. The top-level feature vector containing these higher-order abstract features is then used as the hybrid feature vector.

[0109] The technical effects of the above solution are as follows: By adaptive voltage partitioning and combining semiconductor physics principles, analytical features with clear physical meaning are extracted from different working regions, thereby comprehensively capturing the core electrical behavior of the diode. At the same time, global features describing the overall shape of the curve are introduced to make up for the limitations of single-region features. Furthermore, by utilizing a lightweight gated attention network and cross-layer connections, adaptive evaluation and weighted fusion of the importance of different features can be achieved. This not only strengthens the representation ability of key features, but also generates high-order abstract features that can more profoundly reflect the complex characteristics of the diode through nonlinear transformation. This constructs an information-rich and robust hybrid feature vector, thus laying a solid and reliable data foundation for the subsequent machine learning model to establish an accurate mapping from macroscopic characteristics to model parameters.

[0110] In step four, when constructing and training the machine learning model based on support vector regression, a multi-core hierarchical weighted support vector regression framework is adopted, specifically including:

[0111] For the sub-feature sets with different physical meanings and dimensions in the hybrid feature vector, dedicated kernel functions are initialized respectively. Specifically, a radial basis kernel function is used for the analytical feature subset based on semiconductor physical equations, a polynomial kernel function is used for the geometric and statistical feature subset extracted from the current-voltage curve shape, and a linear kernel function is introduced.

[0112] Construct a two-level cascaded regression structure, specifically as follows:

[0113] The first layer is a parallel kernel mapping and prediction layer, which simultaneously inputs the mixed feature vectors into the support vector regression sub-models corresponding to the radial basis function kernel function, the multinomial kernel function and the linear kernel function, and each support vector regression sub-model independently outputs a set of preliminary predicted values ​​of model parameters.

[0114] The second layer is an adaptive weight fusion layer connected to a lightweight meta-learner. The lightweight meta-learner dynamically assigns fusion weights to the preliminary predicted values ​​output by each support vector regression sub-model in the first layer based on the input fusion feature vector, and uses the weighted fusion result as the final model parameter vector output by the machine learning model.

[0115] During the training phase of the machine learning model, an objective function is constructed that includes a prediction error term and a physical consistency regularization term. The physical consistency regularization term is calculated as follows:

[0116] Substitute the preliminary prediction value of any support vector regression submodel in the first layer into the initial electronic design model to generate a simulated current-voltage curve, and calculate the morphological difference measure between the simulated curve and the real curve.

[0117] The technical effects of the above solution are as follows: By designing a multi-kernel hierarchical weighted support vector regression framework, the most suitable kernel function can be adaptively selected for parallel mapping and prediction of different types of features, thereby fully mining the unique information of each subset in the mixed feature vector. Furthermore, by leveraging a two-level cascaded structure and a lightweight meta-learner, intelligent dynamic fusion of prediction results from multiple sub-models can be achieved, effectively integrating the prediction advantages from different kernel function perspectives, thereby improving the overall robustness and prediction accuracy of the model. At the same time, a physical consistency regularization term is introduced during training, using the physical characteristics of semiconductor devices as prior knowledge to constrain the machine learning process, so that the parameters predicted by the model can not only fit the data, but also generate current-voltage curves that conform to real physical behavior, thus ensuring the accuracy of the extracted parameters.

[0118] In step five, before inputting the mixed feature vector of the diode to be modeled into the trained machine learning model, a feature domain adaptive calibration step is performed, which specifically includes:

[0119] Calculate the maximum mean difference between the mixed feature vector of the target diode and the mixed feature vector of the standard dataset in the space formed by the hidden layers of the machine learning model;

[0120] Determine whether the maximum mean difference is greater than a preset offset threshold;

[0121] When the maximum mean difference is greater than the offset threshold, the mixed feature vectors and model parameter labels corresponding to at least three sets of measured current-voltage data of the target diode are used to fine-tune the parameters of the last layer of the trained machine learning model to complete the model calibration.

[0122] The mixed feature vector of the target diode is input into a calibrated machine learning model, which outputs the predicted initial values ​​of the model parameters.

[0123] The technical effect of the above solution is as follows: by calculating and evaluating the maximum mean difference in feature distribution between the target diode and the original training set, it can intelligently identify the data distribution shift caused by process fluctuations or measurement condition differences, and only activate a lightweight model fine-tuning mechanism when the shift is significant. It can quickly calibrate the last layer of the model using a small amount of measured data from the target device itself, thereby effectively improving the generalization ability and prediction accuracy of the trained machine learning model when facing new batches or new types of diodes with extremely low additional data and computational costs.

[0124] In step five, after the machine learning model outputs a set of predicted initial model parameter values, a verification and correction step is performed, which specifically includes:

[0125] Establish an expert rule base based on semiconductor physics. The rule base defines the reasonable range of values ​​for each model parameter (such as saturation current, ideality factor, series resistance), the mutual constraints between parameters, and the constraints on the changing trends at different temperatures.

[0126] The initial values ​​predicted by the machine learning model are automatically compared with the expert rule base to check whether the values ​​of each parameter fall within their physically reasonable hard boundaries. Then, the logical relationship between key parameter pairs is verified to conform to physical laws, such as whether the series resistance is non-negative, whether the ideal factor is within the typical range, and whether the temperature coefficient conforms to the material properties.

[0127] If the initial value passes all rule checks, it is output directly. If a rule violation is found, the corresponding correction mechanism is activated according to the type of rule violation. Specifically:

[0128] For parameters that are slightly out of bounds, force the parameters that are slightly out of bounds back to the nearest reasonable boundary value;

[0129] In cases where there are serious conflicts between parameters, the key parameters that are more sensitive to the impact on the global error are adjusted first, and the parameters that have mutual constraints with them are recalculated until the set of initial parameter values ​​output satisfies all physical constraints, forming the final physically consistent initial parameter values ​​for subsequent optimization.

[0130] The technical effects of the above solution are as follows: By introducing an expert rule base based on semiconductor physics to automatically verify and correct the initial parameter values ​​predicted by the machine learning model, it can effectively identify and correct physically unreasonable predictions caused by limitations of training data or model generalization bias. This ensures that the initial parameter set is more in line with the basic physical laws and process knowledge of diodes, thereby providing a high-quality and physically meaningful search starting point for subsequent global optimization. This significantly improves the convergence speed of the optimization process and the physical reliability of the final model parameters, thus avoiding the risk of optimization failure or the generation of an incorrect model due to incorrect initial values.

[0131] In step six, the iterative optimization of the global optimization algorithm adopts a two-stage hybrid optimization strategy, specifically including:

[0132] In the first stage, the initial parameter values ​​obtained in step five are used as the initial population center, and an adaptive covariance matrix evolution strategy is executed in the parameter space for global sampling.

[0133] Simultaneously, a Gaussian process surrogate model is constructed and dynamically updated to fit the mapping relationship between the model parameter combination and the simulation error. In this stage, the optimization process executes the acquisition function optimization guided by the surrogate model. The acquisition function comprehensively balances the predicted mean and predicted variance of the surrogate model.

[0134] In the second stage, the region formed by the optimal parameter points obtained in the first stage is switched to the quasi-Newton method for local optimization. In this stage, physical constraints on the parameter update direction are constructed based on the semiconductor physical relationships of the initial electronic design model.

[0135] In the iterative process of the two-stage hybrid optimization strategy, an adaptive adjustment mechanism integrating dynamic error weights and convergence criteria is implemented, specifically including:

[0136] Based on the measured current-voltage curves, the low bias region, medium bias region, and high bias region are divided. Dynamically adjustable weight coefficients are assigned to the error terms corresponding to each region. The adjustment of the weight coefficients is based on the magnitude of the fitting residuals of each region in the current iteration.

[0137] A composite convergence criterion is set up, which simultaneously monitors three indicators, including the relative improvement rate of the objective function value, the norm of change of the key model parameter vector, and the morphological similarity index between the simulated curve and the measured curve.

[0138] The optimization process is considered to have converged when the relative improvement rate of the objective function value, the norm of change of the key model parameter vector, and the morphological similarity index are all lower than their respective set thresholds in a series of preset iterations.

[0139] The technical effects of the above solution are as follows: By adopting a two-stage hybrid optimization strategy and introducing a Gaussian process surrogate model to efficiently guide parameter space sampling, the optimization efficiency and the possibility of finding the global optimum can be improved. At the same time, by integrating a dynamic error weight mechanism, the fitting focus of different working regions can be intelligently adjusted to ensure that the final model is balanced and accurate across the entire voltage range. The constraints based on semiconductor physics provide correct directional guidance for local search, thereby enhancing the physical rationality of the optimization. In addition, the designed composite convergence criterion comprehensively judges convergence from three dimensions: objective function, parameter stability, and curve shape. This can avoid premature convergence or invalid iterations that may be caused by a single criterion, thus achieving robustness and automation of the convergence process while ensuring high accuracy of the optimization results.

[0140] Working principle: By adaptively denoising and extracting features from measured current-voltage data, a hybrid feature vector integrating physical analysis and morphological statistical information is constructed. Then, a machine learning model using multi-kernel hierarchical support vector regression is used to learn the complex nonlinear mapping relationship from these macroscopic features to compact model parameters, thereby quickly providing high-quality, physically consistent initial parameter values ​​for the diode under test. Starting from these initial values, a two-stage optimization strategy combining global exploration and local fine-grained search is adopted, and iterative optimization is performed with dynamically weighted multi-region errors as the target until convergence, ultimately obtaining a high-precision diode simulation model. This invention significantly improves the automation and accuracy of parameter extraction, reduces the dependence on professional experience, and ensures that the final model parameters meet both data fitting requirements and semiconductor physical laws.

[0141] This embodiment details a model confidence evaluation step based on physical-data dual consistency, performed after the global optimization iteration (step six) has converged and before the final generation of the electronic design model file (step seven). This step aims to solve the problems of pseudo-convergence and physical distortion commonly found in existing semiconductor device modeling techniques. In traditional parameter extraction processes, model quality is often judged solely based on current fitting errors (such as RMSE), which can easily lead to the algorithm getting trapped in local minima. Although the mathematical error is minimal, the extracted parameters (such as series resistance, ideality factor, etc.) may violate semiconductor physics, or the model's first derivative (conductance) characteristics may deviate from the actual device. If such a model is applied to radio frequency (RF) circuit or precision analog circuit simulation, it can lead to simulation non-convergence or harmonic analysis errors. Therefore, this embodiment constructs a rigorous dual evaluation system.

[0142] In order to perform the numerical calculations and logical judgments described in this embodiment, the method of the present invention is typically run on a high-performance computing workstation or server cluster.

[0143] At the hardware level, the system configuration includes, but is not limited to: a central processing unit (CPU), preferably using a multi-core high-frequency architecture (such as Intel Xeon or AMD EPYC series) to support multi-threaded parallel matrix operations and differential calculations; random access memory (RAM), with a capacity of 32GB or more recommended to accommodate large-scale experimental datasets (containing massive IV points under different temperatures and biases) and population history records during the optimization process; a graphics processing unit (GPU) to accelerate kernel function mapping of support vector regression (SVR) and vectorization operations of large-scale floating-point numbers; and a high-speed solid-state drive (SSD) for fast reading and writing of IV data and intermediate model files.

[0144] At the software level, the system runs on a Linux or Windows Server operating system, and the underlying computing engine integrates the BLAS / LAPACK linear algebra library to ensure numerical accuracy. The evaluation algorithm described in this embodiment is embedded in the entire EDA (Electronic Design Automation) modeling toolchain, existing as an independent quality control module.

[0145] When the global optimization algorithm in step six (using a two-stage hybrid optimization strategy) meets the preliminary convergence criterion (such as the rate of change of the objective function being less than...), And output a set of final optimized model parameters (denoted as ) After that, the system does not immediately allow passage, but instead triggers the evaluation process described in this embodiment, which specifically includes:

[0146] The system first performs bidirectional mapping and derivation calculations on the data:

[0147] The system reads the measured current-voltage data of the target diode at the current temperature from a standard dataset. Let the set of test voltage points be... The corresponding set of measured currents is .

[0148] To evaluate the dynamic characteristics of the model, the system extracts measured conductance values. Conductance (or conductivity) is a fundamental electrical property of a system. In physics, it is defined as the first derivative of current with respect to voltage. Because measured current data inevitably contains quantization noise, thermal noise, and white noise from the testing instrument, directly using a simple two-point difference method ( Noise is amplified dramatically by the differentiation operation, resulting in a violent, non-physical sawtooth-like fluctuation in the obtained conductance curve. Therefore, this embodiment employs a five-point center difference combined with the Savitzky-Golay smoothing filter.

[0149] Specifically, for each voltage point The system selects two neighboring points before and after it. The least squares method is used to fit a local cubic polynomial, and then the polynomial is calculated in the... The analytical derivative at the point is used as the measured conductivity value. For boundary points, the process automatically degrades to a three-point forward or backward differential method. This approach effectively filters out high-frequency noise while perfectly preserving the diode's abrupt nonlinear changes near the turn-on voltage, ensuring the physical accuracy of the measured conductance.

[0150] Next, the system calls the pre-loaded initial electronic design model in memory. This model is typically a compact model based on physical equations, such as a modified version of the classic Shockley equations, a PSP model diode module, or a JUNCAP model. The model equations are usually in the form of... ,in The parameter set to be evaluated.

[0151] The system will Substituting into the model equations, for For each voltage point, calculate the theoretical set of simulated current values. .

[0152] Simultaneously, the system calculates the simulated conductance value. To ensure the rigor of the evaluation, this embodiment uses the model equations to accurately calculate the analytical partial derivatives of the voltage, i.e. The analytical derivative can reflect whether there are minute physical oscillations or discontinuities in the parameter combination.

[0153] After building After four sets of core data sequences, the system uses the following calculation formula from this invention to quantify the quality of the model:

[0154]

[0155] The following provides a detailed explanation of each variable, constant, and its physical meaning in the formula:

[0156] Relative error architecture and numerical stability constant :

[0157] The core error term in the formula adopts This is because diodes are typical strongly nonlinear devices with extremely wide current ranges (from picoamperes to amperes, spanning 12 orders of magnitude). If absolute error is used, the error at large current points will mask the error at small current points; if ordinary relative error is used, when the measured current approaches 0 (such as zero bias or reverse leakage), the division factor tends to zero, leading to calculation overflow.

[0158] This embodiment introduces It is a numerical stability constant, and its unit is consistent with that of current (or conductance). In engineering practice, The value of is usually set to the background noise limit of the testing instrument, for example... (1pA). This ensures that the error calculation is numerically stable across the entire voltage range, and that the weighting is balanced.

[0159] Weighting coefficient The weights representing the accuracy of static current fitting. The weights representing the accuracy of dynamic conductance fitting, and always satisfying These two coefficients are adaptively adjusted according to the application scenario of the device, as detailed below:

[0160] Scenario A (Power Rectification): If the target device is a power diode, the user is concerned with the forward voltage drop and heat dissipation, and the system automatically configures it. It focuses on DC IV characteristics.

[0161] Scenario B (RF / High-Speed ​​Switching): If the target device is used in a mixer, detector, or high-speed switching circuit, its small-signal impedance ( This directly determines the insertion loss and isolation. At this point, the system automatically configures... The forced optimization algorithm prioritizes ensuring an accurate fit of conductivity.

[0162] Physical punishment mechanism ( and ):

[0163] The second half of the formula It is used to prevent physical distortion of the model. (Parameter out-of-bounds penalty): The system has a built-in expert rule base covering various semiconductor materials (Si, Ge, GaAs, InP, SiC, GaN), which defines the physically reasonable ranges for each model parameter. For example, for silicon processes, the ideality factor... Typically between 1.0 and 2.0; series resistance Must be greater than 0; saturation current The device area and doping concentration must conform to the order of magnitude relationship. The system iterates through all final parameters, and when a parameter is found to be out of bounds (e.g., ...), ... ,or (Ohms), the system calculates its normalized distance from the physical boundary and accumulates it to If all parameters are reasonable, .

[0164] (Physical boundary sensitivity coefficient): It is usually set to a large positive number (such as 10 to 50).

[0165] The exp function: Due to the properties of the exponential function, as long as Slightly greater than 0 (i.e., there is a physical violation). This would become a large negative number, leading to The value drops sharply from 1 and approaches 0. This is equivalent to... A multiplication penalty is applied. This means that no matter how high the data fit score is in the first half, if the physical parameters are unreasonable, the final score will be affected. It will also be dragged down to an unacceptable level.

[0166] The system calculates Then, it is compared with a preset release threshold (such as 0.9).

[0167] like The threshold is set, the model is deemed qualified and approved for release, and the process proceeds to step seven to generate the netlist file.

[0168] like A threshold is set to determine if the model has consistency defects. The system then initiates a feedback correction mechanism, as detailed below:

[0169] Scenario 1: If the low score is mainly due to the exponential term, it indicates that a physical out-of-bounds error has occurred. The system feeds this information back to step six, and in the next round of optimization, applies an infinite potential barrier penalty to the out-of-bounds parameter, forcibly pulling it back to a reasonable range.

[0170] Scenario 2: If the low score is mainly due to the conductivity item ( (Partial) This indicates that the model has not accurately captured the inflection point characteristics. In the objective function of step six, the system dynamically increases the weight coefficient of the first derivative error to guide the optimization algorithm to focus on optimizing the slope of the curve.

[0171] This embodiment details another key step in the invention, performed in parallel with confidence assessment: adversarial stress testing. This step primarily addresses the overfitting and extrapolation failure problems common in data-driven modeling methods, ensuring the robustness of the model under unknown conditions.

[0172] In practical engineering, the provided measured data (training set) is often limited. For example, the forward voltage is only scanned up to 1.5V, and the reverse voltage is only scanned up to -10V. However, when circuit simulators solve nonlinear equations, especially when dealing with transient processes such as inductive load backflush, power supply surges, or electrostatic discharge, the voltage across the diode may momentarily reach 5V or even -100V.

[0173] If the model is only fitted within the measured range (especially when using high-order polynomials or neural network regression), the model equations are highly susceptible to Runge phenomena (high-order oscillations) outside the measured range. These oscillations manifest as a decrease in current as voltage increases (negative resistance), discontinuous derivatives, or numerical explosions. This anomaly can cause the simulator's time step to shrink to its limit. Therefore, this embodiment introduces an adversarial testing mechanism.

[0174] The system first intelligently scans the standard dataset to determine the voltage coverage range. Then, a virtual probe is generated using a mathematical extension algorithm. The system starts from... Initially, it extends in the direction of positive voltage to... (For example, extending to 5V). Within this range, the system generates a series of virtual voltage points. To simulate a smooth transition of the large injection effect, the distribution density of the points adopts a logarithmic decay mode, i.e., near... The density is higher in some areas and lower in others. The system originates from... Initially, it extends towards the direction of negative voltage. (Estimated breakdown voltage). These two sets of points constitute a virtual test set. It should be noted that these points do not have corresponding measured current values; they are virtual probes used to detect the model's behavioral logic. The parameters are defined as follows: This represents the maximum positive value of the measured voltage in the standard dataset, expressed in volts (V). The positive overload extension coefficient is a dimensionless constant, and its value range is set to [value range missing]. to It is used to simulate power surge or overshoot conditions; This is a theoretical estimate of the reverse breakdown voltage of a diode, expressed in volts (V). Here, is the reverse breakdown elongation coefficient, which is a dimensionless constant with a value range set to . to It is used to cover the region from reverse leakage current to critical breakdown.

[0175] Next, the system substitutes the parameters output from step six into the model to calculate the virtual response current sequence under the virtual test set. Since there are no measured values ​​as a standard answer, this embodiment uses physical morphology rules for unsupervised verification, specifically including:

[0176] Semiconductor physics axioms state that for a non-tunneling diode, its IV characteristic must be monotonic; that is, as voltage increases, current can only increase (forward) or increase in absolute value (reverse). The system calculates the first derivative of the virtual response sequence. If a virtual point in the forward extension region is found... The presence of a negative differential resistance indicates model failure. This is typically due to an incorrect numerical coupling between the high injection inflection point parameter and the series resistance, causing the model equations to oscillate in the extension region.

[0177] In a semi-logarithmic coordinate system, the forward current curve of a diode should exhibit a strictly concave or linear characteristic (series resistance-dominated region), and should not show wavy inflection points. The system calculates the second derivative of the virtual response sequence. If frequent sign flips of the second derivative are detected in the forward extension region (i.e., the curve is distorted like an "S"), the model is considered to be at risk of overfitting. In the reverse extension region, the leakage current should remain flat or monotonically increasing slightly. If the leakage current is found to decrease with increasing reverse voltage (tending towards 0), the model is considered to be physically violated.

[0178] If the model is found to violate the laws of physics in the above tests, the system will not simply report an error, but will execute targeted corrective measures, including:

[0179] The system will determine the voltage range of the virtual voltage point where the violation occurred (e.g., The high-risk interval is marked as such. The system generates a smoothing constraint regularization term for this interval. For example, a penalty function is constructed. The goal is to minimize the curvature energy of this interval. This information is fed back to the global optimizer in step six. Upon receiving the feedback, the optimizer introduces this new regularization term in an additive manner into the original objective function (fitting error) and increases the sampling weight of this high-risk interval.

[0180] In addition, the optimizer uses the current parameters as a starting point to restart the local search. Under the constraints of the regularization term, the algorithm is forced to adjust the parameters (usually fine-tuning higher-order nonlinear parameters) to eliminate oscillations in the virtual region.

[0181] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0182] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

Claims

1. An AI-based method for automatic extraction and optimization of diode model parameters, characterized in that, Includes the following steps: Step 1: Obtain the measured current-voltage characteristic data of the target diode under different temperatures and bias voltages to form the original training dataset. Then, normalize, filter noise, and remove outliers to form the preprocessed standard dataset. Step 2: Extract multi-class features from the standard dataset, including analytical features based on semiconductor physical equations, as well as geometric and statistical features extracted from the current-voltage curve shape. Then, fuse the extracted features to construct a hybrid feature vector to characterize the electrical behavior of the diode under specific operating conditions. Step 3: Select a compact diode model containing multiple parameters to be extracted as the initial electronic design model. The parameters to be extracted include at least saturation current, ideality factor, series resistance, and parallel resistance. Step 4: Using the hybrid feature vector as input and the parameter vector of the initial electronic design model as output, construct a machine learning model based on support vector regression. Train the machine learning model through supervised learning to establish a nonlinear mapping relationship between the macroscopic electrical characteristics of the diode and the parameters of the compact model. Step 5: Perform the preprocessing and feature extraction of the measured current-voltage data of the diode to be modeled under the given test conditions in Step 1 and Step 2 to obtain the corresponding mixed feature vector and input it into the trained machine learning model. The machine learning model directly outputs a set of predicted model parameter values ​​as the initial parameter values ​​of the diode electronic design model. Step Six: Using the initial parameter values ​​obtained in Step Five as the starting point, construct an optimization problem based on the initial electronic design model. The objective function is to minimize the error between the simulated data and the measured data. Use a global optimization algorithm to iteratively optimize the model parameters until the error function converges to the preset threshold, and obtain the final optimized model parameters. Step 7: Substitute the final model parameters obtained in Step 6 into the initial electronic design model to generate an electronic design model file for the target diode, which can then be used in the circuit simulation environment.

2. The AI-based automatic extraction and optimization method for diode model parameters according to claim 1, characterized in that, The noise filtering and outlier removal in step one is specifically a joint denoising method combining adaptive weighted dynamic sliding window filtering and physical constraints, including the following steps: A dynamic evaluation window is established for each data point in the original current-voltage sequence. The size of this window is adaptively adjusted based on the local first derivative variance of the neighboring region of the data point. Within the dynamic evaluation window, the Shockley ideal diode equation is used as a benchmark to calculate the deviation between the measured value and the theoretical value for each data point, and weights are assigned based on this deviation. Based on the local statistical characteristics and weights of each data point, a weighted median filter is constructed to filter the current sequence, removing random noise and physically unreasonable outliers. The filtered current-voltage sequence is subjected to a global continuity check. The continuity of the first derivative is verified by spline interpolation fitting. Discontinuities are smoothed and corrected, and a standard dataset is output.

3. The AI-based automatic extraction and optimization method for diode model parameters according to claim 1, characterized in that, Step two, the process of extracting and fusing multi-class features to construct a hybrid feature vector, specifically includes: Based on the nonlinear characteristics of the diode current-voltage curve, the voltage axis is adaptively divided into three characteristic physical regions: low bias region, medium bias region and high bias region. Within each characteristic physical region, analytical features are extracted based on semiconductor physical equations, specifically as follows: In the low bias region, the first analytical feature and the second analytical feature are extracted based on the linear segment of the logarithmic current and voltage, where the first analytical feature is the ideal factor estimate and the second analytical feature is the saturation current estimate. In the mid-bias region, the third and fourth analytical features are extracted based on the first and second derivative curves of current versus voltage. The third analytical feature is the voltage at the maximum curvature point, and the fourth analytical feature is the dynamic resistance at the inflection point. In the high bias region, the fifth and sixth analytical features are extracted based on the linear segment of the current-voltage curve. The fifth analytical feature is the estimated value of the series resistance, and the sixth analytical feature is the curve linearity deviation factor. For the entire current-voltage curve, at least three global morphological features are extracted, including the curvature integral in global logarithmic coordinates, the equivalent area of ​​charge in a specified voltage range, and the parallelism offset statistics between the families of current-voltage curves at different temperatures. The aforementioned analytical features and global morphological features are input into a lightweight gated attention network. The gated attention network assigns fusion weights to each input feature and generates weighted basic features. The weighted basic features are connected across layers with the first to sixth analytical features and the global morphological features. The connected features are then subjected to a nonlinear transformation to form higher-order abstract features. The top-level feature vector containing these higher-order abstract features is then used as the hybrid feature vector.

4. The method for automatic extraction and optimization of diode model parameters based on AI according to claim 1, characterized in that, In step four, when constructing and training the machine learning model based on support vector regression, a multi-core hierarchical weighted support vector regression framework is adopted, specifically including: For the sub-feature sets with different physical meanings and dimensions in the hybrid feature vector, dedicated kernel functions are initialized respectively. Specifically, a radial basis kernel function is used for the analytical feature subset based on semiconductor physical equations, a polynomial kernel function is used for the geometric and statistical feature subset extracted from the current-voltage curve shape, and a linear kernel function is introduced. Construct a two-level cascaded regression structure, specifically as follows: The first layer is a parallel kernel mapping and prediction layer, which simultaneously inputs the mixed feature vectors into the support vector regression sub-models corresponding to the radial basis function kernel function, the multinomial kernel function and the linear kernel function, and each support vector regression sub-model independently outputs a set of preliminary predicted values ​​of model parameters. The second layer is an adaptive weight fusion layer connected to a lightweight meta-learner. The lightweight meta-learner dynamically assigns fusion weights to the preliminary predicted values ​​output by each support vector regression sub-model in the first layer based on the input fusion feature vector, and uses the weighted fusion result as the final model parameter vector output by the machine learning model. During the training phase of the machine learning model, an objective function is constructed that includes a prediction error term and a physical consistency regularization term. The physical consistency regularization term is calculated as follows: Substitute the preliminary prediction value of any support vector regression submodel in the first layer into the initial electronic design model to generate a simulated current-voltage curve, and calculate the morphological difference measure between the simulated curve and the real curve.

5. The AI-based automatic extraction and optimization method for diode model parameters according to claim 1, characterized in that, In step five, before inputting the mixed feature vector of the diode to be modeled into the trained machine learning model, a feature domain adaptive calibration step is performed, which specifically includes: Calculate the maximum mean difference between the mixed feature vector of the target diode and the mixed feature vector of the standard dataset in the space formed by the hidden layers of the machine learning model; Determine whether the maximum mean difference is greater than a preset offset threshold; When the maximum mean difference is greater than the offset threshold, the mixed feature vectors and model parameter labels corresponding to at least three sets of measured current-voltage data of the target diode are used to fine-tune the parameters of the last layer of the trained machine learning model to complete the model calibration. The mixed feature vector of the target diode is input into a calibrated machine learning model, which outputs the predicted initial values ​​of the model parameters.

6. The AI-based automatic extraction and optimization method for diode model parameters according to claim 1, characterized in that, In step five, after the machine learning model outputs a set of predicted initial model parameter values, a verification and correction step is performed, which specifically includes: Establish an expert rule base based on semiconductor physics. The rule base defines the reasonable range of values ​​for each model parameter, the mutual constraints between parameters, and the constraints on the changing trends at different temperatures. The initial values ​​predicted by the machine learning model are automatically compared with the expert rule base to check whether the values ​​of each parameter fall within their physically reasonable hard boundaries. Then, the logical relationship between key parameter pairs is verified to conform to physical laws, such as whether the series resistance is non-negative, whether the ideal factor is within the typical range, and whether the temperature coefficient conforms to the material properties. If the initial value passes all rule checks, it is output directly. If a rule violation is found, the corresponding correction mechanism is activated according to the type of rule violation. Specifically: For parameters that are slightly out of bounds, force the parameters that are slightly out of bounds back to the nearest reasonable boundary value; In cases where there are serious conflicts between parameters, the key parameters that are more sensitive to the impact on the global error are adjusted first, and the parameters that have mutual constraints with them are recalculated until the set of initial parameter values ​​output satisfies all physical constraints, forming the final physically consistent initial parameter values ​​for subsequent optimization.

7. The AI-based automatic extraction and optimization method for diode model parameters according to claim 1, characterized in that, In step six, the iterative optimization of the global optimization algorithm adopts a two-stage hybrid optimization strategy, specifically including: In the first stage, the initial parameter values ​​obtained in step five are used as the initial population center, and an adaptive covariance matrix evolution strategy is executed in the parameter space for global sampling. Simultaneously, a Gaussian process surrogate model is constructed and dynamically updated to fit the mapping relationship between the model parameter combination and the simulation error. In this stage, the optimization process executes the acquisition function optimization guided by the surrogate model. The acquisition function comprehensively balances the predicted mean and predicted variance of the surrogate model. In the second stage, the region formed by the optimal parameter points obtained in the first stage is switched to the quasi-Newton method for local optimization. In this stage, physical constraints on the parameter update direction are constructed based on the semiconductor physical relationships of the initial electronic design model.

8. The AI-based automatic extraction and optimization method for diode model parameters according to claim 7, characterized in that, In step six, during the iterative process of the two-stage hybrid optimization strategy, an adaptive adjustment mechanism for dynamic error weights and convergence criteria is integrated, specifically including: Based on the measured current-voltage curves, the low bias region, medium bias region, and high bias region are divided. Dynamically adjustable weight coefficients are assigned to the error terms corresponding to each region. The adjustment of the weight coefficients is based on the magnitude of the fitting residuals of each region in the current iteration. A composite convergence criterion is set up, which simultaneously monitors three indicators, including the relative improvement rate of the objective function value, the norm of change of the key model parameter vector, and the morphological similarity index between the simulated curve and the measured curve. The optimization process is considered to have converged when the relative improvement rate of the objective function value, the norm of change of the key model parameter vector, and the morphological similarity index are all lower than their respective set thresholds in a series of preset iterations.

9. The method for automatic extraction and optimization of diode model parameters based on AI according to claim 8, characterized in that, In step six, after determining that the optimization process has converged and before generating the electronic design model file in step seven, a model confidence assessment step based on physical-data dual consistency is performed. This model confidence assessment step specifically includes: Obtain the final optimized model parameters output from step six, and use the initial electronic design model to calculate the simulated current and simulated conductance values ​​at all test voltage points in the standard dataset; Based on the measured current-voltage characteristic data in the standard dataset, the measured conductance values ​​at all test voltage points are calculated. The measured conductance values ​​are obtained by performing a first-order central difference operation on the measured current-voltage characteristic data. Using the following formula for calculating the physical-data dual consistency confidence score, a comprehensive score index characterizing the physical reliability of the model is calculated. : in, It is a dimensionless comprehensive scoring index; This represents the total number of test voltage points. The index number of the voltage point; For the first The simulated current values ​​at each voltage point are in amperes. For the first The measured current values ​​at each voltage point are in amperes. For the first Simulated conductance values ​​at each voltage point, in Siemens units; For the first Measured conductance values ​​at each voltage point, in Siemens units; To prevent numerical stability constants with a denominator of zero, the unit must be consistent with the unit of the physical quantity in the denominator; These are the weighting coefficients of the data fitting term. Let be the weighting coefficient of the physical form term, and satisfy . and The sum equals 1; It is an exponential function with the natural constant e as its base; This is the physical boundary sensitivity coefficient, with a unit of 1; This is a parameter out-of-bounds penalty term. When all parameters of the final optimized model are within the preset reasonable range for semiconductor physics, The value is 0 when there are parameters that exceed the physically reasonable range of the semiconductor. The value is the sum of the normalized distances of all out-of-bounds parameters from their respective physical boundaries; Judge the comprehensive score index obtained from the calculation. If the parameters exceed the preset release threshold, the final optimized model parameters are deemed qualified and step seven is executed; otherwise, the final optimized model parameters are deemed to have physical consistency defects, the physical constraint weights of the objective function in step six are adjusted, and step six is ​​returned to iterate and optimize again.

10. The AI-based automatic extraction and optimization method for diode model parameters according to claim 9, characterized in that, While performing the model confidence assessment step, an adversarial stress test step based on virtual boundary extension is also performed in parallel to verify the robustness of the model in the non-sampling region. The adversarial stress test step specifically includes: Based on the voltage coverage range of the measured current-voltage characteristic data in the standard dataset, mathematical extensions are performed to the positive high-level region and the reverse cutoff region of the voltage axis to generate several virtual voltage probe points that exceed the measured range, thus forming a virtual test set. Input the final optimized model parameters output from step six and the virtual test set into the initial electronic design model to calculate the corresponding virtual response current sequence; The virtual response current sequence is subjected to a first derivative sign test and a second derivative convexity test to verify whether the virtual response current sequence remains monotonically increasing in the positive high-level region without inflection point oscillation, and whether the leakage current remains flat or conforms to the monotonically rapid increasing trend of avalanche breakdown in the reverse cutoff region. If the verification finds that the virtual response current sequence violates the physical constraints of monotonically increasing or convexity at any virtual voltage probe point, it is determined that the current model has an overfitting risk. The voltage interval corresponding to the virtual voltage probe point that violates the physical constraints is marked as a high-risk interval, and the high-risk interval is fed back to step six. In the global optimization algorithm of step six, an additional smoothing constraint regularization term is applied to the high-risk interval, and then the optimization iteration is restarted until the model passes the adversarial stress test.

Images