An automated extraction method of semiconductor device model parameters

By combining two-stage sensitivity analysis and the Kriging surrogate model with the multi-objective gray wolf optimization algorithm, the high-dimensional, multi-objective optimization problem in semiconductor device model parameter extraction was solved, achieving efficient and robust automated parameter extraction, reducing computational costs and improving optimization efficiency and accuracy.

CN121859749BActive Publication Date: 2026-06-09HANGZHOU DIANZI UNIV +1
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202610320001.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-03-17
Publication Date
2026-06-09
Estimated Expiration
2046-03-17

Smart Images

  • Figure CN121859749B_ABST
    Figure CN121859749B_ABST
Patent Text Reader

Abstract

The application discloses a kind of automation extraction methods of semiconductor device model parameters, by the key parameter acquisition of semiconductor device model multiple target response, proxy model is constructed and trained to predict target response mean and uncertainty parameter;Build proxy model assisted multi-objective optimization collaborative cycle framework;According to the position updating mechanism of multi-objective optimization algorithm, candidate parameter population is generated in key parameter, and the trained proxy model is used to predict candidate parameter, and the real simulation evaluation is carried out according to the prediction result, and the obtained real sample is added to the non-dominated file, and the file is updated based on the non-dominated relationship, and the real sample is also added to the training set to dynamically update the proxy model, generate potential search trajectory, and select representative representative pareto approximate solution from it, and update the position of population and the direction of next generation search trajectory according to the selected representative solution in the file;Termination cycle, output solution set in file and check.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of semiconductor device technology, and specifically relates to an automated method for extracting semiconductor device model parameters. Background Technology

[0002] Automated extraction of model parameters for semiconductor devices refers to the automatic optimization of device model parameters through algorithms without relying on extensive human experience and manual parameter tuning, enabling accurate reproduction of multiple sets of electrical characteristic curves of the device. However, this problem is essentially a high-dimensional, multi-objective, and costly black-box optimization problem, and existing automated methods still have significant shortcomings. On the one hand, device models often contain dozens of coupled physical parameters, and direct optimization in the complete parameter space leads to enormous computational overhead and excessively long runtime. On the other hand, there are often conflicting optimization objectives between different characteristic curves, and single-objective optimization or simple weighted methods cannot simultaneously guarantee accurate fitting of multiple curves. Furthermore, each real simulation or measurement requires significant computational or experimental resources, and without effective proxy and sampling strategies, the optimization process is prone to getting trapped in local optima or leading to an unbalanced Pareto solution set distribution.

[0003] Specifically, researchers have attempted to introduce various intelligent optimization methods into this field. Genetic algorithms (GA), with their excellent global search capabilities, were initially used for model parameter fitting tasks. Differential Evolution (DE) is robust in handling highly coupled parameters and complex search spaces, and was applied early on to parameter extraction for models such as SOIMOSFET. Particle Swarm Optimization (PSO), due to its simple structure and ease of implementation, is widely used for modeling high-frequency devices such as GaNHEMT. Least Squares (LM) performs well in linear models, but is susceptible to local optima due to initial conditions when dealing with nonlinear device behavior. In recent years, Bayesian optimization (BO), as an efficient black-box optimization method, has also gradually gained attention in device modeling; furthermore, some studies combine LM with unsupervised neural networks (NN) to accelerate convergence and improve fitting accuracy. Machine learning methods such as Artificial Neural Networks (ANN) have also been used to build predictive models of device behavior, supporting the modeling process by fitting complex mappings between input parameters and output responses; however, they rely on large amounts of high-quality training data, and their generalization performance is not stable enough in practical applications.

[0004] While these methods have made progress in improving modeling efficiency, most of them focus on a single objective. When multiple sets of characteristic curves need to be matched simultaneously, conflicts often arise between different objective functions, causing the optimization results to fail to take into account overall accuracy. At the same time, the models generally contain a large number of redundant or low-sensitivity parameters. Directly optimizing in the complete high-dimensional space not only increases computational overhead but also wastes search resources.

[0005] Specifically, existing methods for extracting parameters from semiconductor device models still suffer from the following technical shortcomings when faced with high-dimensional, multi-objective, and computationally expensive fitting problems:

[0006] 1) The search space has high dimensionality and lacks parameter filtering and dimensionality reduction mechanisms.

[0007] Existing device models typically contain dozens of coupling parameters, but most optimization methods directly search within the entire parameter space without incorporating sensitivity analysis to screen for key parameters. This results in a high-dimensional search space, heavy computational burden, slow algorithm convergence, and optimization results that are easily affected by insensitive parameters, thus reducing overall optimization efficiency and stability.

[0008] 2) Insufficient handling of multi-objective trade-offs.

[0009] There are often conflicts between different electrical characteristic curves. Many existing methods use weighted sums or one-by-one optimization, which makes it difficult to simultaneously achieve the global optimum for multiple objectives.

[0010] 3) The proxy model is not fully applied or the update mechanism is too simple.

[0011] While some "surrogate model + intelligent optimization" methods can reduce the number of real simulations, their surrogate model accuracy and update strategy are insufficient, which can easily lead to local overfitting or accumulation of surrogate errors, thus affecting the overall robustness.

[0012] In existing technologies, several studies have combined surrogate models with evolutionary algorithms to reduce simulation calls, such as surrogate-assisted optimization based on RBFN or Kriging, and general surrogate-assisted frameworks like GPSAF. These methods typically involve initial sampling to construct the surrogate, using surrogate predictions to filter candidates, using partial real evaluation to update the surrogate, and repeating the optimization process. Further, for example:

[0013] 1. Combine Kriging (or RBFN) with a multi-objective evolutionary algorithm (such as NSGA-II, MOPSO, MOGWO variants) to reduce the number of actual evaluations through surrogate prediction;

[0014] 2. Apply heuristic algorithms (GA / PSO / DE, etc.) to perform single-objective or multi-objective fitting for the problem of extracting device model parameters;

[0015] 3. Use single-stage sensitivity analysis or empirical dimensionality reduction in engineering cases to reduce the number of parameters.

[0016] This invention aims to propose an efficient and robust automated method for extracting device parameters for engineering applications. While ensuring the accuracy of multi-objective fitting, dimensionality reduction is achieved through systematic sensitivity analysis. Furthermore, the Kriging surrogate model is combined with the Multi-Objective Gray Wolf Optimization Algorithm (MOGWO) to form a surrogate-assisted SAMOGWO framework, thereby reducing the number of actual simulations and improving optimization convergence performance. Summary of the Invention

[0017] To address the shortcomings of existing technologies and achieve the goals of ensuring parameter extraction accuracy while reducing computation time and manual intervention, improving parameter optimization efficiency, reducing the number of real simulations and enhancing optimization convergence performance, as well as improving robustness and reproducibility, this invention adopts the following technical solution:

[0018] An automated method for extracting parameters from a semiconductor device model includes the following steps:

[0019] Step 1: Select key parameters for the semiconductor device model;

[0020] Step 2: Determine the optimization objective, identify the key DC parameters to be extracted from the original device model, and set a physically reasonable range of values ​​for each key parameter;

[0021] Step 3: Obtain multi-target responses through key parameters, construct and train a Kriging surrogate model to predict the mean and uncertainty parameters of the target responses with a limited number of key parameter samples;

[0022] Step 4: Based on the surrogate model-assisted multi-objective optimization algorithm, construct a surrogate-assisted multi-objective optimization collaborative loop framework; according to the position update mechanism of the multi-objective optimization algorithm, generate a candidate parameter population in the key parameters, use the trained surrogate model to predict the candidate parameters, select candidate parameters for real simulation evaluation based on the prediction results, add the obtained real samples to the non-dominated archive, update the archive based on the non-dominated relationship, and add the real samples to the training set to dynamically update the surrogate model, generate potential search trajectories, and select representative Pareto approximate solutions from them, update the position of the population and the direction of the next generation search trajectory based on the representative solutions selected in the archive;

[0023] Step 5: When the termination condition is met, terminate the loop, output the non-dominated solution set in the final file as the optimization result, select the most suitable Pareto value, and verify the final solution using real simulation.

[0024] Furthermore, in step 1, sensitivity analysis is performed on the parameters of the semiconductor device model to select key parameters that contribute significantly to the multi-objective response as decision variables for optimization, including the following steps:

[0025] Step 1.1: Sensitivity analysis (initial screening);

[0026] A reasonable physical range is set for all extracted model parameters, and the absolute mean of the influence of each parameter μ* is calculated using the Morris method. Low-sensitivity parameters are eliminated based on the magnitude of the absolute mean of the influence μ*, and only the set of candidate parameters that have a significant impact on the model output is retained to evaluate their influence on the overall performance index.

[0027] Step 1.2: Global Sensitivity Analysis (Fine Screening);

[0028] Perform a global sensitivity analysis on the candidate parameter set selected by the sensitivity analysis, and calculate the overall sensitivity index S. T According to the overall sensitivity index S T The magnitude of the value is used to select several key parameters that contribute the most to the multi-objective response as decision variables for subsequent optimization.

[0029] The original parameter set of the device model is reduced in dimensionality by using a two-stage sensitivity analysis (e.g., Morris initial screening combined with FAST global sensitivity analysis). This mechanism first identifies low-impact parameters through local perturbation and then ranks the remaining parameters by importance using a global sensitivity metric, thereby determining the optimal subset of key parameters. This process effectively reduces the dimensionality of the high-dimensional search space, improves optimization convergence efficiency, and reduces the interference of insensitive parameters on the stability of model fitting.

[0030] Two-stage sensitivity analysis can theoretically and practically solve the problems of high computational cost and difficult parameter identification caused by high-dimensional search. The reason is that sensitivity analysis quantitatively characterizes the contribution of different parameters to the target response, eliminating parameters with minimal or redundant impact on multi-target output, thus significantly compressing the search dimensionality while retaining the main influencing factors. After dimensionality compression, the optimizer searches in the low-dimensional space, reducing the overall sample size and computation time required for a single simulation, while improving the stability and interpretability of parameter identification and reducing the probability of solution set oscillations or pseudo-optimal solutions caused by low-sensitivity parameters. Therefore, sensitivity analysis, as a preprocessing step, fundamentally improves the efficiency and reliability of the parameter extraction process.

[0031] Furthermore, in step 1.1, variables that do not significantly affect the model output are eliminated based on the magnitude of their influence on the absolute mean μ*, and only parameters that have a strong influence and / or contribute to the uncertainty of modeling accuracy and target conflict are retained, providing a candidate set for subsequent global analysis.

[0032] Furthermore, in step 3, several initial training points are generated in the selected key parameter subspace using Latin hypercube sampling (LHS), and the sampling points are evaluated using real simulation to obtain the multi-objective function response value (multi-objective loss vector).

[0033] Furthermore, in step 3, based on the training sample dataset, a Kriging surrogate model is established for each target, a covariance matrix is ​​constructed, the correlation strength and smoothness of the two key parameters in the corresponding dimensions are controlled by hyperparameters, and the hyperparameters are determined by maximum likelihood estimation to obtain the predicted mean and predicted variance of the Kriging surrogate model for each target, with the predicted variance used as the uncertainty parameter.

[0034] Furthermore, in step 4, the main loop of the framework is constructed, and the following sub-steps are executed in each iteration of the loop until the termination condition is met:

[0035] Step 4.1: Generate a candidate solution population {Xi} in the key parameter subspace based on the MOGWO position update mechanism (α / β / δ guided update);

[0036] Step 4.2: The surrogate model predicts and assesses the uncertainty of candidate solutions, and the candidate solutions are screened based on the results of the surrogate model (surrogate-oriented).

[0037] The trained Kriging surrogate model is used to predict all candidate solutions, and the multi-objective output of each candidate solution is predicted, including the mean as the objective vector and the corresponding variance as the uncertainty parameter.

[0038] Step 4.3: Perform a real evaluation and update the files of the selected candidate solutions;

[0039] The selected candidate solutions are evaluated using real simulation to obtain the true value F(X); the newly added real evaluation samples are added to the external non-dominated archive, and the archive is updated according to the non-dominated relations;

[0040] Step 4.4: Dynamically update the surrogate model based on real evaluations to generate potential search trajectories, and select representative surrogate Pareto approximation solutions as supplementary candidates.

[0041] Add new real samples to the training set, either by sliding window or incrementally, and retrain the Kriging surrogate model; if the cross-validation error exceeds the threshold, trigger the surrogate model to be retrained.

[0042] Perform several evolutionary steps on the surrogate model (i.e. run MOGWO or similar evolutionary processes on the surrogate model) to generate potential search trajectories, and select representative surrogate Pareto approximate solutions as supplementary candidates to enhance boundary exploration capabilities.

[0043] Step 4.5: Update the population position and the next generation search direction based on the representative solutions (α, β, δ) selected from the archive, and continue the loop.

[0044] This invention constructs a collaborative loop framework within the dimensionality-reduced key parameter space, deeply coupling the Kriging surrogate model with the multi-objective grey wolf optimization (MOGWO) algorithm to form a closed-loop optimization mechanism (SAMOGWO) of prediction-screening-evaluation-update. This addresses the engineering challenges of multi-objective conflict and high real evaluation costs, significantly reducing the number of real simulations and improving convergence performance in multi-objective optimization.

[0045] The Kriging surrogate model provides not only the predicted mean but also the predicted uncertainty for a limited number of samples. Based on this, the optimizer can selectively allocate the real evaluation budget to candidate solutions—prioritizing the verification of points with high potential or high uncertainty, thus balancing exploration and exploitation within a limited number of simulations. By combining surrogate prediction, probabilistic / uncertainty-oriented screening strategies, and external non-dominated archive updates, the algorithm maintains MOGWO's global search and Pareto maintenance capabilities while significantly reducing the frequency of real simulation calls.

[0046] Furthermore, in step 4.2, based on the predicted mean and variance, a probabilistic screening and / or uncertainty threshold strategy is adopted to select some high-potential candidate solutions for real simulation evaluation.

[0047] Furthermore, in step 4.3, the remaining candidate solutions that are not selected retain the surrogate model predictions as approximate fitness to participate in the contemporary non-dominated ranking and crowding assessment; in step 4.4, if the file limit is exceeded, the crowding (distance) strategy is used to maintain diversity.

[0048] Furthermore, in step 4.3, if the number of files exceeds the limit, a grid culling strategy is used to maintain diversity.

[0049] Furthermore, in step 2, the optimization objective is the target curve to be fitted (e.g., Id-Vd, Gm-Vg) and the corresponding loss function or error metric (RMS); in step 5, the optimization result is the final list of fitting parameters, and the final fitted curve is generated.

[0050] The advantages and beneficial effects of this invention are as follows:

[0051] This invention achieves parameter dimensionality reduction through sensitivity analysis, effectively narrowing the search space and retaining only parameters that significantly affect output characteristics, thereby improving optimization efficiency. It combines the Kriging surrogate model with the Multi-Objective Grey Wolf Optimization (MOGWO) algorithm to construct a surrogate-assisted multi-objective grey wolf optimization framework (SAMOGWO), reducing the number of real simulations and improving optimization convergence performance. This invention achieves automatic parameter extraction for multiple objectives under a limited real simulation budget, and experimental verification demonstrates its practical effects in engineering scenarios such as DC parameter extraction for semiconductor devices, significantly reducing computation time and manual intervention while maintaining the accuracy of multi-objective electrical characteristic curve fitting.

[0052] This invention integrates sensitivity analysis dimensionality reduction, surrogate model construction, and multi-objective optimization into a unified framework, achieving efficient multi-objective search under dimensionality reduction-first, surrogate-guided, and finite budget conditions. It improves the robustness and reproducibility of the algorithm, makes the parameter extraction process more automated and engineered, and is applicable to modeling and optimization tasks of various semiconductor device models. Attached Figure Description

[0053] Figure 1 This is a flowchart of an automated extraction method for semiconductor device model parameters according to an embodiment of the present invention.

[0054] Figure 2 This is a comparison chart (Id–Vd curve) of measurement data and simulation data results of the 6×125 μm GaN HEMT model in the embodiment of the present invention.

[0055] Figure 3 This is a comparison chart (Id–Vg curve) of measurement data and simulation data results of the 6×125 μm GaN HEMT model in this embodiment of the invention.

[0056] Figure 4 This is a comparison chart (Gm–Vg curve) of measurement data and simulation data results of the 6×125 μm GaN HEMT model in the embodiment of the present invention.

[0057] Figure 5 This is a comparison chart (Id–Vd curve) of measurement data and simulation data results of the 8×75μm GaN HEMT model in the embodiment of the present invention.

[0058] Figure 6 This is a comparison chart (Id–Vg curve) of measurement data and simulation data results of the 8×75μm GaN HEMT model in the embodiment of the present invention.

[0059] Figure 7 This is a comparison chart (Gm–Vg curve) of measurement data and simulation data results of the 8×75μm GaN HEMT model in the embodiment of the present invention.

[0060] Figure 8 This is a comparison chart (Id–Vd curve) of measurement data and simulation data results of the 8×200μm GaN HEMT model in the embodiment of the present invention.

[0061] Figure 9 This is a comparison chart (Id–Vg curve) of measurement data and simulation data results of the 8×200μm GaN HEMT model in the embodiment of the present invention.

[0062] Figure 10 This is a comparison chart (Gm–Vg curve) of measurement data and simulation data results of the 8×200μm GaN HEMT model in the embodiment of the present invention. Detailed Implementation

[0063] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0064] like Figure 1 As shown, an automated method for extracting semiconductor device model parameters employs a proxy-assisted multi-objective gray wolf algorithm, specifically including the following steps:

[0065] Step 1: Parameter sensitivity analysis;

[0066] By performing sensitivity analysis on the parameters of the device model, key parameters are screened, specifically including the following steps:

[0067] Step 1.1: Morris sensitivity analysis (initial screening);

[0068] First, a reasonable physical range is set for all extracted model parameters, and the absolute mean μ* of the influence of each parameter is calculated using the Morris method. Based on the magnitude of μ*, low-sensitivity parameters are eliminated, and only the set of candidate parameters that have a significant impact on the model output is retained to evaluate their influence on the overall performance index.

[0069] Furthermore, based on the magnitude of μ*, variables that have almost no significant effect on the model output are quickly eliminated, and only parameters that have a strong impact on modeling accuracy and objective conflict or contribute to uncertainty are retained, providing a candidate set for subsequent global analysis.

[0070] Step 1.2: FAST global sensitivity analysis (fine screening);

[0071] Perform FAST analysis on the subset of candidate parameters selected by Morris screening to calculate the overall sensitivity index S. T According to S TThe size of the parameters is determined by selecting several key parameters that contribute most to the multi-objective response as decision variables for subsequent optimization.

[0072] Step 2: Initialization phase;

[0073] Define the target curve to be fitted (e.g., Id-Vd, Gm-Vg) and its corresponding loss function or error metric (RMS) as the optimization objective. Determine the DC parameters to be extracted from the original model and set a physically reasonable range for each parameter. Specifically, set hyperparameters such as the experimental / simulation budget (maximum number of true evaluations), the population size of SAMOGWO, the maximum number of generations, and the external archiving limit.

[0074] Step 3: Initial surrogate design and Kriging model training, which includes the following steps:

[0075] Step 3.1: Initial Sample Design;

[0076] In the selected key parameter subspace, Latin hypercube sampling (LHS) is used to generate several initial training points, and these sampling points are evaluated by real simulation to obtain the multi-objective function response values ​​(multi-objective loss vector).

[0077] Step 3.2: Kriging model construction and training;

[0078] Based on the training sample dataset DB={X,f(X)}, a Kriging surrogate model is built for each target: the covariance matrix R is constructed as follows:

[0079]

[0080] Where X represents the set of input variables for the training samples, that is, the combination of device model parameters that ultimately participate in the optimization after being screened by sensitivity analysis, and each sample point x i For a specific set of parameter configurations, for example, when the selected optimization parameters include threshold voltage, mobility, source resistance, and drain resistance, the sample point can be represented as different combinations of values ​​(Vth, μ, Rs, Rd), where f(X) represents the set of objective function response values ​​corresponding to the input sample X, and x i,k This represents the value of the i-th sample on the k-th design variable, where k represents the dimension of the k-th device model parameter involved in the optimization. j,k Let represent the value of the j-th sample on the k-th design variable, d represent the dimension of the input variables, corresponding to the number of parameters in the optimization problem or physical model, exp(·) represents the exponential function, and θ k > 0 and p k∈[1,2] both represent hyperparameters, which control the correlation strength parameter and smoothness parameter on the k-th dimension, respectively.

[0081] Subsequently, maximum likelihood estimation (MLE) was used to determine the hyperparameters to obtain the predicted mean of the Kriging surrogate model for each objective. With prediction variance .

[0082] Step 4: Based on the surrogate-assisted multi-objective optimization algorithm, construct a surrogate-assisted multi-objective optimization collaborative loop framework to obtain the main loop of the surrogate-assisted multi-objective greywolf optimizer algorithm SAMOGWO, for surrogate-assisted multi-objective greywolf optimization. In each iteration of the loop, execute the following sub-steps until the termination condition is met (reaching the preset number of generations), specifically including the following steps:

[0083] Step 4.1: Generate candidate solutions based on the MOGWO multi-objective algorithm;

[0084] The MOGWO multi-objective algorithm is used to generate a new candidate population {Xi} in the key parameter subspace using the position update mechanism (α / β / δ guided update).

[0085] Step 4.2: Proxy model prediction and uncertainty assessment, and candidate solutions are selected based on the Proxy model results (Proxy-oriented).

[0086] The surrogate model prediction and uncertainty assessment utilizes a trained Kriging surrogate model to predict all candidate solutions, forecasting the multi-objective output of each candidate solution, including the objective vector (mean). and the corresponding uncertainty (variance) .

[0087] The selection of candidate solutions is based on the predicted mean and variance, and a probabilistic screening or uncertainty threshold strategy is adopted to select a portion of "high-potential" candidate solutions for real simulation evaluation; the remaining candidate solutions that are not selected can retain the surrogate model prediction values ​​as approximate fitness to participate in contemporary non-dominated ranking and crowding evaluation.

[0088] Step 4.3: Realistic assessment and archive update;

[0089] The selected candidate solutions are evaluated using real simulation to obtain the true value F(X); the newly added real evaluation samples are added to the external non-dominated archive and the archive is updated according to the non-dominated relationship. If the archive exceeds the limit, the crowding (distance) strategy or grid culling strategy is used to maintain diversity.

[0090] Step 4.4: Dynamic update of the agent model and agent trajectory sampling;

[0091] The dynamic update of the surrogate model involves adding new real samples to the training set using a sliding window or incremental method, and retraining the Kriging surrogate model. If the cross-validation error exceeds the threshold, the surrogate model is retrained.

[0092] Proxy trajectory sampling involves performing several evolutionary steps on the proxy model (i.e., running MOGWO or a similar evolutionary process on the proxy model) to generate potential search trajectories, and selecting representative proxy Pareto approximate solutions as supplementary candidates to enhance boundary exploration capabilities.

[0093] Step 4.5: Population and Leader Update;

[0094] The population position and the next search direction are updated based on the representative solutions (α, β, δ) selected from the archive, and the cycle continues.

[0095] Step 5: Termination and Result Output;

[0096] Once the true evaluation budget or convergence condition is met, the algorithm terminates and outputs the non-dominated solution set in the final external archive as the optimization result, i.e., the final list of fitting parameters. The most suitable Pareto value is selected, and the final solution is verified by real simulation and a fitting curve is generated (compared with the measurement data).

[0097] An embodiment of the present invention is as follows:

[0098] 1. Experimental subjects and data sources;

[0099] Using the ZDT (Zitzler-Deb-Thiele) multi-objective test function set, the performance of the SAMOGWO algorithm on multi-objective optimization problems was evaluated by comparing it with three other algorithms—MOGWO, MOPSO, and NSGA-II—in terms of inverse generation distance (IGD) and hypervolume (HV), as shown in Tables 1 and 2.

[0100] Table 1. Statistical results of IGD on the ZDT test set (excluding ZDT5)

[0101]

[0102] Table 2. Statistical results of HV on the ZDT test set (excluding ZDT5)

[0103]

[0104] Overall, SAMOGWO outperforms the comparative algorithms in average IGD across all five test functions, demonstrating a significant advantage on ZDT1–ZDT4. Particularly in ZDT1 and ZDT2, its average IGD is 0.0115 and 0.0119, respectively, representing reductions of approximately 89% and 83% compared to the traditional MOGWO. This indicates that the proposed surrogate-assisted mechanism excels in early-stage convergence guidance, rapidly approaching the true Pareto front. Furthermore, SAMOGWO also achieves a significant improvement in HV metrics, with average HV values ​​of 1.23 and 7.48 on ZDT3 and ZDT4, respectively, far exceeding the comparative algorithms, validating its advantage in maintaining the uniformity of the Pareto solution set distribution. In complex multimodal or degenerate boundary problems (such as ZDT3 and ZDT4), SAMOGWO maintains solution set diversity more stably than MOPSO and NSGA-II, exhibiting a significantly lower HV standard deviation, demonstrating good stability and robustness. In challenging problems like ZDT6 with non-uniform distribution, SAMOGWO also achieves lower IGD and higher HV, demonstrating strong generalization ability and search efficiency.

[0105] Overall, considering both IGD and HV metrics, SAMOGWO outperforms other comparative algorithms on the test function, significantly improving solution set quality and distribution breadth while ensuring convergence, thus verifying the effectiveness of the surrogate modeling mechanism and the co-design of gray wolf optimization.

[0106] This invention uses a GaN HEMT device model as an example for verification. The structure is the same, with geometric dimensions of 6×125 μm, 8×75 μm, and 8×200 μm, respectively. The measurement data includes multiple sets of electrical characteristic curves (e.g., Id-Vd, Gm-Vg). These measurement data are used to compare with simulation results to evaluate the accuracy of model parameter extraction.

[0107] 2. Sensitivity analysis and parameter dimensionality reduction;

[0108] This invention takes an 8×200 μm GaN HEMT device model as an example. The GaN HEMT device model contains 25 DC parameters. After screening using sensitivity analysis, only 5 key parameters (knp, knd, vdscale, voff, and Lambda) that significantly affect the output characteristics are retained, significantly reducing the search space.

[0109] To verify the efficiency improvement brought about by parameter selection in the optimization task, the results of the sensitivity analysis before and after are as follows: Under exactly the same settings, the SAMOGWO algorithm was executed on both the full parameter space and the dimensionality-reduced parameter space, with the actual number of evaluations fixed at 500 for both, to compare the difference in optimization efficiency under the two settings. The experimental results are shown in Table 3.

[0110] Table 3 Comparison before and after sensitivity analysis

[0111]

[0112] The results show that, while maintaining a consistent number of actual evaluations, sensitivity analysis significantly reduced the total runtime, decreasing time overhead by approximately 83.3%. This result fully demonstrates the important role of parameter dimensionality reduction in improving optimization efficiency.

[0113] 3. Proxy model construction and optimization configuration;

[0114] Initial sample points were generated in the parameter space using Latin hypercube sampling (LHS), and the corresponding objective function values ​​were obtained through real simulation to train the initial Kriging surrogate model. This model was then combined with the Multi-Objective Grey Wolf Optimization Algorithm (MOGWO) to construct the SAMOGWO optimization framework.

[0115] The optimization objective is to minimize the RMS error between measurement and simulation, with the target including the fitting accuracy of the Id-Vd curve and the Gm-Vg curve.

[0116] 4. Optimization results and performance verification;

[0117] To more clearly compare the screening results for the three sizes, Table 4 summarizes the final selected optimization parameters for each size of device.

[0118] Table 4 Final Optimized Parameters for Devices of Various Sizes

[0119]

[0120] After completing parameter dimensionality reduction based on sensitivity analysis, this invention uses the selected key parameters as variables and employs surrogate-assisted multi-objective gray wolf optimization (SAMOGWO) to extract DC parameters. Figures 2 to 10 Fitting results for three different geometric sizes (6×125 μm, 8×75 μm, and 8×200 μm) are presented. The optimization process used multi-objective fitting of the Id–Vd and Gm–Vg curves as the optimization objective, while the Ig–Vd curve was not included in the optimization; instead, it was uniformly placed... Figure 3 The modeling effectiveness of the proposed method is further validated in the following sections.

[0121] Taking a 6×125 μm device as an example, such as Figures 2 to 4 As shown, the simulation results and the measured data are correlated on the Id–Vd curve ( Figure 2 ) and Gm–Vg curve ( Figure 4 The data closely match the current-voltage characteristics and transconductance characteristics of the device under different biases, and can accurately reproduce these characteristics. Figure 3 The results of the Id–Vg curve are shown. Although the curve was not directly involved in the optimization, it still maintained a good fit, which enhances the completeness of the results presentation.

[0122] Further comparison of the results of the three sizes ( Figures 2 to 10 It can be observed that the algorithm can stably achieve good curve fitting results for devices of different sizes. This indicates that the method proposed in this invention has strong robustness and adaptability for devices of different sizes.

[0123] In summary, the experimental results fully verify the feasibility and superiority of this invention. This method significantly reduces computational costs while maintaining fitting accuracy, and possesses good versatility and engineering application value.

[0124] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An automated method for extracting parameters from a semiconductor device model, characterized in that... Includes the following steps: Step 1: Select key parameters for the semiconductor device model; Step 2: Determine the optimization objective and identify the key parameters to be extracted from the original device model; Step 3: Obtain multi-target responses through key parameters, build and train a surrogate model to predict the mean and uncertainty parameters of the target responses; Step 4: Based on the surrogate model-assisted multi-objective optimization algorithm, construct a surrogate-assisted multi-objective optimization collaborative loop framework; according to the position update mechanism of the multi-objective optimization algorithm, generate a candidate parameter population in the key parameters, use the trained surrogate model to predict the candidate parameters, select candidate parameters for real simulation evaluation based on the prediction results, add the obtained real samples to the non-dominated archive, update the archive based on the non-dominated relationship, and add the real samples to the training set to dynamically update the surrogate model, generate potential search trajectories, and select representative Pareto approximate solutions from them, update the position of the population and the direction of the next generation search trajectory based on the representative solutions selected in the archive; Specifically, the main loop of the framework is constructed, and the following sub-steps are executed in each iteration of the loop until the termination condition is met: Step 4.1: Generate a candidate solution population from the key parameters based on the position update mechanism of the multi-objective gray wolf optimization algorithm; Step 4.2: The surrogate model predicts and assesses the uncertainty of candidate solutions, and the candidate solutions are screened based on the results of the surrogate model; Step 4.3: Perform a real evaluation and update the files of the selected candidate solutions; Step 4.4: Dynamically update the surrogate model based on real evaluations to generate potential search trajectories, and select representative surrogate Pareto approximation solutions as supplementary candidates. Step 4.5: Update the population position and next-generation search direction based on the representative solutions selected from the archives; Step 5: When the termination condition is met, terminate the loop, output the non-dominated solution set in the final file as the optimization result, select the most suitable Pareto value, and verify the final solution using real simulation.

2. The automated extraction method for semiconductor device model parameters according to claim 1, characterized in that: In step 1, sensitivity analysis is performed on the parameters of the semiconductor device model to select key parameters that contribute significantly to the multi-objective response as decision variables for optimization. This includes the following steps: Step 1.1: Sensitivity analysis; Set reasonable physical ranges for all extracted model parameters, calculate the absolute mean of the influence of each parameter, and remove low-sensitivity parameters based on the magnitude of the absolute mean of the influence, retaining only the set of candidate parameters that have a significant impact on the model output; Step 1.2: Global Sensitivity Analysis; A global sensitivity analysis is performed on the candidate parameter set after sensitivity analysis to calculate the total sensitivity index. Based on the magnitude of the total sensitivity index, several key parameters that contribute the most to the multi-objective response are selected as decision variables for subsequent optimization.

3. The automated extraction method for semiconductor device model parameters according to claim 2, characterized in that: In step 1.1, parameters that have a strong influence and / or uncertainty contribution on modeling accuracy and target conflict are retained based on the magnitude of their influence on the absolute mean.

4. The automated extraction method for semiconductor device model parameters according to claim 1, characterized in that: In step 3, several initial training points are generated using Latin hypercube sampling in the selected key parameter subspace, and the sampling points are evaluated using real simulation to obtain multi-objective response values.

5. The automated extraction method for semiconductor device model parameters according to claim 1, characterized in that: In step 3, a Kriging surrogate model is established for each target based on the training sample dataset. The correlation strength and smoothness of the two key parameters in the corresponding dimensions are controlled by hyperparameters. The hyperparameters are determined by maximum likelihood estimation to obtain the predicted mean and predicted variance of the Kriging surrogate model for each target. The predicted variance is used as the uncertainty parameter.

6. The automated extraction method for semiconductor device model parameters according to claim 1, characterized in that: In step 4.2, based on the predicted mean and variance, a probabilistic screening and / or uncertainty threshold strategy is used to select some high-potential candidate solutions for real simulation evaluation.

7. The automated extraction method for semiconductor device model parameters according to claim 1, characterized in that: In step 4.3, the remaining candidate solutions that are not selected retain the surrogate model predictions as approximate fitness to participate in the current non-dominated ranking and crowding assessment; in step 4.4, if the file exceeds the limit, the crowding strategy is used to maintain diversity.

8. The automated extraction method for semiconductor device model parameters according to claim 1, characterized in that: In step 4.3, if the number of files exceeds the limit, a grid culling strategy is used to maintain diversity.

9. The automated extraction method for semiconductor device model parameters according to claim 1, characterized in that: In step 2, the optimization objective is the target curve to be fitted and the corresponding loss function or error metric; in step 5, the optimization result is the final list of fitting parameters, and the final fitted curve is generated.

Citation Information

Patent Citations

  • Double-layer optimization heterogeneous agent model assisted multi-objective evolutionary optimization calculation method

    CN116579371A

  • S-shaped air inlet channel multi-objective optimization design method

    CN117216879A