A device model parameter automatic extraction method based on multivariate tree structure Ponce estimation and sensitivity collaborative optimization

Abstract

The application discloses a device model parameter automatic extraction method based on multivariate tree structure Ponce estimation and sensitivity collaborative optimization, which is applied to the field of semiconductor device modeling and parameter extraction, and aims at the problem of insufficient parameter extraction efficiency and stability of existing device model parameter extraction methods. The application firstly acquires device test or simulation reference data to be fitted, and determines an electrical characteristic curve to be fitted according to device modeling requirements; then, device model parameters to be extracted are determined, and the value range of each model parameter is set; then, a tree structure Ponce estimation method is adopted to perform adaptive search on the device model parameter space; and a sensitivity analysis method is introduced to dynamically adjust the value range of the parameters; when the iteration stopping condition is met, the current search result is output as the optimal parameter combination; and the method can effectively improve the automation degree and optimization efficiency of device model parameter extraction.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor device modeling and parameter extraction technology, and specifically relates to an automatic parameter extraction technology for semiconductor device models. Background Technology

[0002] Device model parameter extraction is a crucial bridge connecting device development and circuit application, occurring after device optimization and before circuit simulation in the entire Device Design and Process Co-optimization (DTCO) process. The goal of device model parameter extraction is to adjust the parameters in the device model based on the device's electrical characteristic curves and relevant simulation or test data, thereby obtaining a set of parameters that accurately characterizes the device's electrical behavior. The accuracy of device model parameter extraction directly affects the accuracy and evaluation results of subsequent circuit simulations; therefore, its extraction effectiveness is of great significance to the PDK development process.

[0003] Existing device model parameter extraction methods largely rely on engineers' experience, requiring manual setting of parameter ranges, adjustment of optimization strategies, and repeated comparison of fitting results. This is not only time-consuming but also susceptible to initial value selection and local optima issues in high-dimensional parameter spaces, leading to insufficient efficiency and stability in parameter extraction. Meanwhile, many existing studies often focus on parameter extraction for a single device size, fitting and optimizing only the electrical characteristics of devices at a specific size, without fully considering the unified representation capability of the device model across the entire size range. However, in actual PDK development, the most time-consuming and critical part of device model parameter extraction is not the local fitting of a single-size device, but rather the global model extraction across the entire device size range. This requires simultaneously considering the consistency and accuracy of device electrical behavior under different channel lengths, widths, or other structural dimensions. As device structures and model complexity continue to increase, the coupling relationships between parameters become more complex, and traditional manual trial-and-error or experience-driven parameter extraction methods are no longer sufficient to meet the high-efficiency, high-precision development requirements across the entire size range. Therefore, how to reduce the dependence of the parameter extraction process on the engineer's experience, reduce time costs, improve the efficiency of global model extraction, and accelerate the PDK development process has become a key issue in device model parameter extraction.

[0004] To improve the automation and optimization efficiency of parameter extraction, the parameter extraction process not only requires the use of reasonable evaluation methods to quantitatively measure the fitting accuracy between the model curve and the target curve, but also necessitates efficient search and optimization within a large high-dimensional parameter space. Compared to traditional Gaussian process Bayesian optimization methods, the Tree-structured Parzen Estimator (TPE) method exhibits better adaptability in handling high-dimensional search spaces, complex parameter distributions, and conditional parameter relationships, enabling effective exploration and selection of parameter combinations with lower computational cost. Therefore, the TPE-based automatic parameter optimization method can improve the efficiency and stability of device model parameter extraction while ensuring fitting accuracy, thus providing more reliable model support for subsequent circuit simulation and device performance evaluation. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing an automatic device model parameter extraction method based on tree-structured Pazen estimation. This method automatically extracts device model parameters after the device optimization step and before circuit simulation. By constructing a fitting accuracy evaluation mechanism and combining it with the tree-structured Pazen estimation method for efficient searching and optimization of the high-dimensional parameter space, it effectively reduces reliance on engineer experience, improves parameter extraction efficiency and fitting accuracy, enhances the stability and automation level of the parameter extraction process, thereby improving the accuracy of subsequent circuit simulation and evaluation results, accelerating the PDK development process, and demonstrating promising application prospects.

[0006] The objective of this invention is achieved as follows: an automatic extraction method for device model parameters based on tree-structured Pazen estimation, comprising: S1: Obtain the test or simulation reference data of the device to be fitted, and determine the electrical characteristic curve to be fitted according to the device modeling requirements; S2: Determine the parameters of the device model to be extracted, and set the parameter space for each model parameter; S3: Construct a loss function based on the fitting error of the typical characteristic curves corresponding to the model parameters; S4: Based on the loss function constructed in step S3, the tree-structured Pazen estimation method is used to adaptively search the parameter space of the current model parameters; S5: Set a first counter to count the cumulative parameter space sampling times of the tree structure Pazen estimation method. When the value of the first counter reaches the threshold, a sensitivity analysis method is introduced to dynamically adjust the parameter space of each model parameter. Then, after clearing the first counter, proceed to step S6; otherwise, proceed directly to step S6. S6: Set a second counter to count the cumulative parameter space sampling times of the tree structure Pazen estimation method. If the value of the second counter reaches the set upper limit, output the search result of the current step S4 as the optimal model parameter and complete the extraction of device model parameters; otherwise, return to step S4.

[0007] The parameter space of each model parameter introduced in step S5 by using sensitivity analysis to dynamically adjust the parameter space specifically involves: evaluating the influence of each model parameter on the loss function, and dynamically adjusting the parameter search range based on the sensitivity results.

[0008] The beneficial effects of this invention are as follows: The overall process of the method includes: a) selecting curve data such as capacitance, current, and threshold voltage that can characterize the key electrical characteristics of the device as fitting objects; b) determining the device model parameters to be extracted, and setting the value range of each model parameter according to the model manual; c) introducing a multivariable tree structure Pazen estimation algorithm, grouping the device model parameters according to the coupling relationship between them, and performing joint probability modeling and collaborative optimization on each group; d) constructing a loss function, which characterizes the overall fitting accuracy of the device model during the optimization process, and setting the threshold voltage fitting error not exceeding 10 mV and the current fitting error not exceeding 3% under different size conditions as penalty terms introduced into the optimization process; e) introducing a sensitivity analysis method to evaluate the influence of each model parameter on the loss function, and dynamically adjusting the parameter search range according to the sensitivity results; f) when the iteration stopping condition is met, outputting the corresponding optimal model file to complete the extraction. Compared with the prior art, this invention has the following advantages: 1) Compared with existing technologies, this invention employs a tree-structured Pazen estimation method for adaptive search of the device model parameter space. It can simultaneously handle discrete, continuous, and conditional parameters without requiring the objective function to be differentiable or continuous, making it particularly suitable for device model parameter extraction scenarios with high parameter dimensionality, complex objective functions, and high evaluation costs. Furthermore, the multivariate tree-structured Pazen estimation fully considers the coupling relationships between model parameters, achieving collaborative optimization of parameter sets through joint probabilistic modeling, thereby improving parameter search efficiency and the stability of optimization results.

[0009] 2) This invention introduces a parameter sensitivity analysis method to evaluate the influence of different model parameters on the loss function, thereby prioritizing the optimization of highly sensitive parameters, reducing the interference of low-sensitivity parameters on the overall search space, effectively achieving dimensionality reduction in the parameter optimization process, and further improving parameter extraction efficiency and convergence performance.

[0010] 3) After adjusting the parameter range based on sensitivity analysis and restarting the optimizer, the existing sampling history of the optimizer is filtered, and historical sampling data that is still within the updated parameter range is retained. Optimization is then carried out based on the retained sampling results to avoid the complete discarding of existing effective sampling information and reduce the computational overhead caused by repeated searches. Attached Figure Description

[0011] Figure 1 This is a flowchart of the present invention.

[0012] Figure 2 Example graphs of IV and CV characteristic curves generated by the model under different parameter combinations; Among them, (a) is a comparison chart of the CV characteristic curve data corresponding to the parameter combination of iteration A and the target data, (b) is a comparison chart of the IV characteristic curve data corresponding to the parameter combination of iteration B and the target data, and (c) is a comparison chart of the IV characteristic curve data corresponding to the parameter combination of iteration C and the target data. Detailed Implementation

[0013] See Figure 1 The automatic parameter extraction method of the present invention includes the following steps: 1. Obtain test or simulation reference data of the device to be fitted, and determine the electrical characteristic curve to be fitted according to the device modeling requirements; generally, select curve data such as capacitance, current and threshold voltage that can characterize the key electrical characteristics of the device as the fitting object. 2. Based on the device type, structural features, and electrical characteristics of the device to be fitted, select a device model version that matches the device. The device model version may include, but is not limited to, compact model versions applicable to BSIM4, BSIM-CMG, or other semiconductor device structures. The model parameter names, number of parameters, and parameter expression forms may differ in different device model versions. According to the reference parameter extraction process in the appendix of the model manual corresponding to the target device model version, determine the device model parameters to be extracted, and set the value range of each model parameter as the parameter space based on the reference range given in the parameter list at the end of the model manual, to ensure the rationality and feasibility of the parameter extraction process. 3. Before the optimization iteration process, determine the loss function to characterize the fitting error between the device model output and the target curve. This loss function includes at least a threshold voltage error term and a current error term. A penalty constraint is introduced into the loss function, ensuring that the device threshold voltage fitting error does not exceed 10 mV and the current fitting error does not exceed 3% under different size conditions. Specifically, this includes: 3.1: Based on the target curve of the device Comparison with simulation output curve The difference between them is used to construct a loss function, and the loss function is calculated by the following equation (1). The loss value corresponding to secondary parameter sampling:

[0014] in, This represents the error term in the device current curve fitting. This represents the error term in the device capacitance curve fitting. This represents the penalty term incurred when the electrical characteristic index exceeds the fitting error standard. , and These are the weighting coefficients corresponding to each error term. In this embodiment, the weighting coefficients are set according to the importance of different electrical characteristics to the accuracy of the device model, with the fitting accuracy of key electrical characteristics being a key factor. The highest priority is given to IV curve fitting accuracy, followed by the accuracy of the curve fitting. Finally, the fitting accuracy of the CV curve and constraint terms is calculated. Therefore, the weighting coefficients are arranged according to... : : Set the ratio to 2:1:3.

[0015] 3.2: Typical industrial accuracy requirements include: threshold voltage fitting error not exceeding 10 mV, and on-state current fitting error not exceeding 3%. A penalty term is constructed based on the degree to which the fitting error exceeds the industrial accuracy requirements according to the electrical characteristics of devices of different sizes. And expressed by the following formula (2):

[0016] in, This indicates the number of device sizes involved in the fitting process. Indicates the first Each size device in parameter combination The threshold voltage fitting error is below. Indicates the first Each size device in parameter combination Current fitting error under the given conditions and This is the penalty coefficient. In this embodiment, considering that the accuracy of threshold voltage fitting and current fitting are equally important for device model parameter extraction, the penalty coefficient is... and The settings are configured in a 1:1 ratio to simultaneously constrain the fitting error of the threshold voltage and current characteristics.

[0017] 4. Introducing a multivariate tree-structured Pazen estimation algorithm, grouping devices based on the coupling relationships between their model parameters, and performing joint probabilistic modeling and collaborative optimization on each group, specifically including: 4.1: Group the parameters to be extracted according to the physical coupling relationship between them. These parameter groups include, but are not limited to: carrier mobility-related parameter groups. Saturation velocity related parameter group Threshold voltage related parameter group Size effect related parameter set Capacitor-related parameter group In practical applications, additional parameter groups can be added as needed. Parameters not assigned to any parameter group or with weak coupling to other parameters are grouped into other parameter sets. .

[0018] 4.2: Sampling is performed in the parameter space to obtain the parameter vector. and corresponding loss function value ,in:

[0019] Based on the loss value Historical parameter vector samples from smallest to largest Sort the samples and select the top 20% as the good sample set. The remaining samples are considered as the bad sample set. .

[0020] 4.3: For each parameter group, establish a Gaussian kernel density distribution model for each group of parameters in a good sample set. For other parameter sets... Independent parameters in, let:

[0021] in, Indicates the first The first set of other parameters in the complete historical sample Each independent parameter This indicates the number of independent parameters in the other parameter set.

[0022] For the Each independent parameter is present in the good sample set for each historical sample point. Construct a one-dimensional Gaussian kernel function:

[0023] in, This indicates the value of the candidate parameter to be evaluated. The first term is adaptively determined by the historical sample distribution. The variance corresponding to each independent parameter.

[0024] Subsequently, a weighted average is performed on multiple one-dimensional Gaussian kernel functions in the good sample set to form the first... Independent probability density distribution of good samples corresponding to each independent parameter:

[0025] in, Representing a good sample set The number of complete historical samples. Similarly, establish the first [sample] in the bad sample set. Independent probability density distribution of bad samples corresponding to each independent parameter .

[0026] For model parameter set The parameters that are coupled with each other in each historical sample point in the good sample set. Construct a multidimensional Gaussian distribution:

[0027] in, , This indicates the number of items to be evaluated. The candidate parameter group values ​​for the group model parameters. This represents the parameter set adaptively determined by the historical sample distribution. The corresponding covariance matrix.

[0028] Then, a weighted average of multiple Gaussian distributions is performed to form the joint probability density distribution of the corresponding parameter groups:

[0029] Similarly, for the parameter set Establish the corresponding joint probability density distribution for the set of bad samples. .

[0030] 4.4: For the parameter set participating in joint modeling, from Mid-sampling; for independent parameters in other parameter sets, from The parameters are sampled and then combined into a complete candidate parameter vector. The combined probability density ratio is then calculated.

[0031] The first term represents the probability density ratio of the parameter group participating in the joint probability modeling, and the second term represents the probability density ratio of each independent parameter in the other parameter sets. The overall probability density ratio is selected. The largest candidate complete parameter vector is used as the sampling result for the next round of device model simulation and loss function evaluation.

[0032] 5. Introduce sensitivity analysis to evaluate the impact of each model parameter on the loss function, and iterate according to a preset interval. Based on the historical sampling results up to the current iteration and their corresponding loss function values, the sensitivity index of each model parameter is calculated, and the parameter search range is scaled and adjusted according to the sensitivity results. The parameter range adjustment interval... This setting can be customized as needed, with a reference value of 200. This means that after every 200 parameter samplings and corresponding loss function evaluations, a sensitivity analysis and parameter search range adjustment will be triggered. The specific steps for scaling and adjusting the parameter search range include: 5.1: Parameter combinations obtained from historical sampling and their corresponding loss function values Sensitivity analysis was performed on each model parameter. The first... The complete parameter vector obtained from the sampling is represented as follows:

[0033] in, This represents the total number of model parameters after the complete parameter vector to be extracted is expanded. Indicates the first In the second sampling The values ​​of each model parameter, .

[0034] To calculate the first The impact of the i-th model parameter on the loss function, specifically for the i-th Secondary sampling parameter combination Construct the corresponding control parameter combination The control parameter combination Combined with original parameters Except for the first Keep all other parameters except the first model parameter consistent, only keep the first parameter consistent. Each model parameter is derived from the sampled values. Replace with reference parameter value Generally, the first one from the historical good samples is taken. The mean of each model parameter, i.e.:

[0035] Combine the control parameters Input the target device model for simulation, and use the same... Using the same loss function calculation method, we obtained the control loss function value. Therefore, the first The sensitivity index corresponding to each model parameter is:

[0036] 5.2: Dynamically scale the parameter search range based on the sensitivity index corresponding to each model parameter.

[0037] Let the first The current search range for each model parameter is ,in, and They represent the first The lower and upper bounds of the current search range for each model parameter. Let the parameter combination that minimizes the loss function value during the current optimization process be the optimal parameter combination. , of which The values ​​of each model parameter are: Based on the sensitivity index of each model parameter. The sensitivity index is linearly mapped to the search range scaling factor. ,in:

[0038] in, This represents the minimum value among the sensitivity indices of each model parameter. This represents the maximum value among the sensitivity indices for each model parameter. This indicates the preset minimum search range proportion. If the denominator... If the value is less than a preset threshold (i.e., close to 0), the division by zero problem is avoided, and the normalization term is not calculated. This ensures that the scaling factor for the search range of each model parameter is equal to the preset threshold. In this invention, the preset threshold value is 1e-30.

[0039] Based on the search range scaling factor Calculate the first Search radius after adjusting model parameters:

[0040] And with the current optimal parameter value Centered on, update the first Search range for each model parameter:

[0041]

[0042] in, and They represent the first The lower and upper limits of the search range after adjusting the model parameters. The result of modifying the parameter range in this way is: for parameters with high sensitivity, a larger search range is retained; for parameters with low sensitivity, the search range is narrowed to the vicinity of the current optimal parameter.

[0043] 5.3: After completing the parameter search range adjustment, the existing historical sampling results of the tree structure Pazen estimator are filtered, and the historical sampling data that are still within the updated parameter range are retained. Based on the retained historical sampling results, the probability distribution model is constructed and the parameter optimization is carried out as described in step 4.

[0044] 6: When the maximum number of sampling steps involved in step 4 is reached, output the corresponding optimal model parameters to complete the extraction of device model parameters. Generally, the number of sampling steps is set to 1000.

[0045] The above methods can effectively improve the automation and optimization efficiency of device model parameter extraction, reduce the reliance on experience in traditional manual parameter tuning, and ensure that the extracted model parameters have high fitting accuracy under multiple device sizes and various electrical characteristics, thereby meeting the accuracy and robustness requirements of industrial-grade modeling applications.

[0046] The present invention will be further described below with reference to specific embodiments and accompanying drawings.

[0047] See Figure 2 This embodiment takes the parameter extraction process of a semiconductor device model at a 40nm process node as an example. First, CV / IV characteristic curve data of a semiconductor device with a width of 0.5μm and a channel length of 0.5μm are collected. Then, based on the commercial BSIM4 device model, more than 20 commonly used device model parameters are selected as parameters to be optimized, and a parameter search space is constructed. Among them, the carrier mobility-related parameter group... Including u0, ua, ub, and saturation velocity related parameter groups Including ags, a0, vsat, threshold voltage related parameter group Including k1, vth0, cit, and size effect related parameter groups Including LLC, DLC, XL, LINT, and capacitor-related parameter groups. Including toxe, phin, cgso, voffcv, acde, moin, noff, and other parameters. This includes rdsw and pdiblc2. Then, the device model parameters are automatically optimized using the tree-structured Pazen estimation method. The parameter symbols provided in this embodiment are those known in the appendix of the model manual corresponding to the commercial BSIM4 device model; specific parameter symbols will not be described in detail here.

[0048] During the optimization process, the IV and CV characteristic curves generated by the model under different parameter combinations are compared with the target data to calculate the corresponding fitting errors. A loss function is then constructed using a predefined penalty term to evaluate the merits of the current parameter combination. The target data is the actual measured data to be fitted. Figure 2 In this context, Measured represents the model parameter combination, and the IV and CV characteristic curves generated through simulation are represented by HSPICE. The calculated fitting error is the root mean square error. Figure 2 The RMSRE representation is used. After 500 iterations of optimization, a set of device model parameters with good fitting effect was obtained, enabling the model to accurately characterize the IV and CV characteristics of the device, thus providing a reliable model basis for subsequent circuit simulation and performance evaluation.

[0049] like Figure 2 As shown, a comparison chart of the IV characteristic curves and CV characteristic curves with the target data for several different parameter combinations is presented, along with the corresponding fitting error results. Figure 2 The comparison results of parameter combinations for all iterations are not given. Only the comparison results of parameter combinations for some iterations are given as examples to facilitate understanding of the content of this invention.

[0050] The above description is only a preferred embodiment of the present invention. Modifications may be made within the scope defined by the claims of the present invention, but all such modifications shall fall within the protection scope of the present invention.

Claims

1. A method for automatically extracting device model parameters based on tree-structured Pazen estimation, characterized in that, include: S1: Obtain the test or simulation reference data of the device to be fitted, and determine the electrical characteristic curve to be fitted according to the device modeling requirements; S2: Determine the parameters of the device model to be extracted, and set the parameter space for each model parameter; S3: Construct a loss function based on the fitting error of the typical characteristic curves corresponding to the model parameters; S4: Based on the loss function constructed in step S3, the tree-structured Pazen estimation method is used to adaptively search the parameter space of the current model parameters; S5: Set a first counter to count the cumulative parameter space sampling times of the tree structure Pazen estimation method. When the value of the first counter reaches the threshold, a sensitivity analysis method is introduced to dynamically adjust the parameter space of each model parameter. Then, after clearing the first counter, proceed to step S6; otherwise, proceed directly to step S6. S6: Set a second counter to count the cumulative parameter space sampling times of the tree structure Pazen estimation method. If the value of the second counter reaches the set upper limit, output the search result of the current step S4 as the optimal model parameter and complete the extraction of device model parameters; otherwise, return to step S4.

2. The method for automatic extraction of device model parameters based on tree-structured Pazen estimation according to claim 1, characterized in that, The loss function expression for step S3 is: ， in, This represents the threshold voltage fitting error term. This represents the current fitting error term. Indicates a penalty item; , and These are the weighting coefficients corresponding to each error term.

3. The method for automatic extraction of device model parameters based on tree-structured Pazen estimation according to claim 2, characterized in that, Penalty items Built to meet the industrial precision requirements of devices of different sizes.

4. The method for automatic extraction of device model parameters based on tree-structured Pazen estimation according to claim 3, characterized in that, Step S4 includes the following sub-steps: S41. Based on the coupling relationship between the device model parameters, the device model parameters determined in step S2 are grouped to obtain multiple parameter groups; The device model parameters to be extracted that are not classified into any parameter group will be assigned to other parameter sets. ; S42. Sample in the parameter space to obtain the parameter vector and the corresponding loss function value; and add the currently sampled parameter vector to the historical sample set. S43. Divide the historical sample set into a good sample set and a bad sample set according to the loss function value; S44: For other parameter sets The independent parameters in the good sample set are used to construct a one-dimensional Gaussian kernel function for each historical sample point; A weighted average of multiple one-dimensional Gaussian kernel functions in the good sample set is used to form the independent probability density distribution of the good samples corresponding to the independent parameter. At the same time, establish the independent probability density distribution of bad samples corresponding to this independent parameter in the bad sample set; For each parameter in the parameter set, a multidimensional Gaussian distribution is constructed at each historical sample point in the good sample set; then, a weighted average of multiple Gaussian distributions is performed to form the joint probability density distribution of the good samples of the parameter set. At the same time, a joint probability density distribution of bad samples of this parameter group is established in the bad sample set; For the parameters in the parameter set, sample from the joint probability density distribution; for the other parameter sets... The independent parameters in the vector are sampled from the independent probability density distribution of the corresponding good samples and then combined into a complete candidate parameter vector. Based on other parameter sets Calculate the ratio of the combined probability density of good samples and bad samples for the independent parameters in the parameter set, as well as the ratio of good samples and bad samples for the parameters in the parameter set. S45: Select the candidate parameter vector with the largest comprehensive probability density ratio as the search result for the current iteration.

5. The method for automatic extraction of device model parameters based on tree-structured Pazen estimation according to claim 4, characterized in that, The parameter space of each model parameter introduced in step S5 by using sensitivity analysis to dynamically adjust the parameter space specifically involves: evaluating the influence of each model parameter on the loss function, and dynamically adjusting the parameter search range based on the sensitivity results.

6. The method for automatic extraction of device model parameters based on tree-structured Pazen estimation according to claim 1 or 5, characterized in that, Step S5 It includes the following steps: S51: For samples in the historical sample set, based on the parameter vector and their corresponding loss function values The degree of influence of changes in the statistical parameter set on the change in the loss function: ， in, Indicates the first Sensitivity index of each model parameter Indicates the number of historical samples. Indicates the first In the second sampling The possible values ​​of each parameter This indicates the corresponding reference parameter value. Indicates parameter combination The corresponding loss function value, Indicates only for the first The loss function value after changing each parameter; S52: Dynamically adjust the parameter search range based on the sensitivity index corresponding to each model parameter; retain the search range for parameters with high sensitivity; and narrow the search range for parameters with low sensitivity.

7. The method for automatic extraction of device model parameters based on tree-structured Pazen estimation according to claim 6, characterized in that, Step S52 is as follows: S521, Regarding the first Secondary sampling parameter combination Construct the corresponding control parameter combination The combination of control parameters Combined with original parameters Except for the first Model parameters Keep all other parameters consistent, and keep the first Each model parameter is derived from the sampled values. Replace with reference parameter value ; S522. Based on the comparison parameter combination, perform simulation and calculate the corresponding loss function value, denoted as... ; S523, The sensitivity index corresponding to the i-th model parameter is: ； in, Indicates the first Secondary sampling parameter combination The corresponding loss function value, This indicates the number of sampled parameter combinations used to calculate parameter sensitivity during this round of parameter space adjustment; S524. Take the parameter combination with the minimum loss function value as the optimal parameter combination, and adjust the parameter space of the optimal parameter combination according to the following steps: S5241, based on the sensitivity index of each model parameter The sensitivity index is linearly mapped to the parameter space scaling factor. : ； in, This represents the minimum value among the sensitivity indices of each model parameter. This represents the maximum value among the sensitivity indices for each model parameter. This indicates the preset minimum parameter space ratio; S5242, Based on the space scaling factor of the parameters Calculate the first Search radius after adjusting model parameters: ； in, and They represent the first The lower and upper limits of the parameter space before parameter adjustment for each model; S5243, using the current optimal parameter values Centered on, update the first The parameter space of each model parameter: ； ； in, and They represent the first The lower and upper limits of the parameter space after adjusting the model parameters.

8. The method for automatic extraction of device model parameters based on tree-structured Pazen estimation according to claim 7, characterized in that, Step S5 also includes: after completing the parameter space adjustment, filtering the current historical sample set and retaining the sample data that is still in the updated parameter space.

Images