A method for automatically extracting parameters of a local nerve modification device model

Abstract

The application provides a local neural correction device model parameter automatic extraction method, relates to the technical field of semiconductor device modeling and parameter extraction, and comprises the following steps: obtaining test data and dividing the test data into different work areas; determining a device model parameter set and establishing a device model parameter space; generating an initial parameter sample set by using a Latin hypercube sampling method; obtaining a high-confidence parameter area of each device model parameter and an initial optimal parameter solution by iterative search through Bayesian optimization; constructing a local supervision training set; constructing a training error-aware neural network; obtaining a predicted correction amount according to the initial optimal parameter solution, and generating a device model parameter file. By using the above method, global iteration is not needed, and the overall calculation resource consumption is reduced. The fitting effect balance of multiple work areas is achieved, the overall fitting precision and robustness of the device model under multiple size and multiple bias conditions are improved, and the local rapid correction of parameters is achieved.

Description

Technical Field

[0001] This invention relates to the field of semiconductor device modeling and parameter extraction technology, specifically to an automatic extraction method for parameters of a local neural correction device model. Background Technology

[0002] Device model parameter extraction is a crucial link between device modeling and circuit simulation. Its core objective is to optimize and calibrate the parameters in the device model based on the device's electrical test data or simulation data, thereby obtaining a set of model parameters that accurately reflects the actual electrical behavior of the device. The results of device model parameter extraction not only directly affect the fitting accuracy of the device model under different bias conditions and size ranges, but also determine the model reliability in subsequent circuit simulation, performance evaluation, and process design co-optimization processes. Therefore, it is of great significance in the PDK development process.

[0003] As the complexity of device structures continues to increase at advanced process nodes, modern device models often contain a large number of strongly coupled model parameters. Different parameters have varying degrees of influence on the electrical behavior of devices, including the subthreshold region, the strong inversion region encompassing linear and saturation regions, capacitance characteristics, and size effects. In the traditional device model parameter extraction process, engineers typically need to continuously adjust parameter ranges based on experience, analyze the sources of curve errors, and repeatedly perform simulation verification. This not only requires a long development cycle but is also susceptible to the influence of initial parameter value selection and local optima in high-dimensional parameter spaces, resulting in low parameter extraction efficiency and insufficient stability. Furthermore, when performing unified model extraction across multiple device size ranges, the error contribution and parameter sensitivity between different working regions often differ significantly, making it difficult for traditional parameter extraction methods relying on manual experience to achieve a balanced fit between different working regions.

[0004] To improve the automation of device model parameter extraction, existing research has gradually introduced machine learning methods such as Bayesian optimization and neural networks to automatically optimize device model parameters. Bayesian optimization can achieve efficient searching in high-dimensional parameter spaces with fewer simulations, reducing reliance on human experience in the parameter search process. Neural network methods, on the other hand, can accelerate the parameter extraction process by learning the nonlinear relationship between device electrical characteristics and model parameters. However, most existing neural network-based methods directly establish a global mapping relationship between device electrical curves and model parameters, requiring the construction of a large-scale training dataset across the entire high-dimensional parameter space. This not only results in high training costs but also, due to the strong nonlinearity and multiple solutions inherent in the device model parameter space, easily leads to difficulties in neural network training and insufficient generalization ability.

[0005] Therefore, how to combine the global search capability of Bayesian optimization in high-dimensional parameter space with the rapid prediction capability of neural networks in local nonlinear mapping modeling, while reducing the training sample requirements and the number of simulations, to achieve rapid parameter correction for high-confidence parameter regions, thereby improving the efficiency and stability of device model parameter extraction, has become a key issue in the field of automatic device model extraction. Summary of the Invention

[0006] To overcome the existing technical problems, this invention provides an automatic extraction method for local neural correction device model parameters based on Bayesian optimization.

[0007] The present invention adopts the following technical solution.

[0008] An automatic extraction method for parameters of a local neural correction device model includes the following steps: S1. Acquire test data that can characterize the key electrical properties of the device, and divide the test data into different working areas; S2. Determine the set of device model parameters to be extracted and establish the device model parameter space; S3. The Latin hypercube sampling method is used to perform hierarchical random sampling in the parameter space of the device model to generate an initial parameter sample set for Bayesian optimization initialization; S4. Construct a Bayesian model based on the initial parameter sample set, calculate the comprehensive loss function value according to the error between the test curve and the first simulation curve generated by the device model, and obtain the high-confidence parameter region and the initial better parameter solution of each device model parameter through Bayesian optimization iterative search. S5. Using the initial optimal parameter solution as the center, locally perturb the parameters of each device model within the high-confidence parameter region to generate local training parameter samples, and construct a local supervised training set. S6. Construct a training error-aware neural network based on the local supervised training set; S7. Input the initial optimized parameter solution into the device model to obtain the second simulation curve. Extract error features based on the second simulation curve and the test curve, substitute them into the error-aware neural network to obtain the prediction correction amount, and generate the device model parameter file.

[0009] As a further improvement of the present invention, the test data includes electrical characteristic curves under different bias conditions and device sizes, and the working region includes one or more of the following: subthreshold region, linear region, saturation region and capacitance characteristic region.

[0010] As a further improvement of the present invention, the specific steps of S2 include: selecting a device model version corresponding to the device to be fitted according to the device type of the device to be fitted; selecting model parameters related to the electrical characteristics to be fitted as the set of device model parameters to be extracted according to the model manual corresponding to the device model version; and determining the initial search range according to the default values ​​and reference value ranges given in the model manual to form the device model parameter space. The parameter set of the device model is represented as follows: , in, It is a set of device model parameters. Indicates the first One model parameter to be extracted. This represents the total number of model parameters to be extracted. , They are the first The lower limit and upper limit of the initial search range for each model parameter to be extracted.

[0011] As a further improvement of the present invention, the specific steps of S3 include: S31. Set the initial sampling quantity as... In the normalized device model parameter space In this process, each model parameter is divided into... 1. Equal probability subintervals; S32, regarding the first A model parameter is generated from a set of parameters. A random permutation sequence consisting of n subintervals is represented as: , in, It is the first A random permutation sequence of model parameters, It is the first The initial parameter sample at the th ... Sub-intervals corresponding to each model parameter dimension; S33, No. The initial parameter sample is randomly selected from the corresponding sub-interval, and a normalized sample value is obtained to obtain the first... The normalized sampled values ​​of each initial parameter sample on each model parameter, wherein the normalized sampled values ​​are expressed as: , in, It is the first The initial parameter sample at the th ... Normalized sampled values ​​of each model parameter, No. The initial parameter sample at the th ... Sub-intervals corresponding to each model parameter dimension It is located in Random numbers within; S34, according to the... The normalized sampled values ​​of each initial parameter sample on each model parameter are mapped according to the initial search range of each model parameter to obtain the corresponding actual parameter values. The actual parameter values ​​under the same initial parameter sample are integrated to obtain the initial parameter sample vector of the corresponding initial parameter sample. The initial parameter sample vector of all initial parameter samples is integrated to obtain the initial parameter sample set. The expression for the actual parameter value is: , It is the first The initial parameter sample at the th ... The actual parameter values ​​of each model parameter. It is the first The lower bound of the initial search range for each model parameter. These are normalized sample values. It is the first The upper limit of the initial search range for each model parameter.

[0012] As a further improvement of the present invention, the specific steps of S4 include: inputting the initial parameter sample vector from the initial parameter sample set into the device model to generate a first simulation curve, calculating the normalized local error between the test curve and the first simulation curve in each working area, calculating the comprehensive loss function value of the corresponding initial parameter sample vector, and combining the comprehensive loss function value, the initial parameter sample vector and the normalized local error of each working area to form a historical observation sample. Based on the changes in model parameters in historical observation samples and the changes in normalized local errors in different working areas, the first... The local error of the normalization of the first working area and the first The correlation between model parameters Calculate the first The search priority of each model parameter, wherein the search priority is calculated using the following expression: , in, It is the first The search priority of each model parameter. Indicates the first The weights corresponding to each work area It is the first The local error of the normalization of the first working area and the first The correlation between model parameters ; Based on the search priority, a region relevance guide is constructed. The expression for the region relevance guide is as follows: , in, It is the value of the regional correlation guide item. This indicates the current optimal parameter value. It is the first The search priority of each model parameter. It is the first among the candidate parameter points to be evaluated. The values ​​of each model parameter; Exploration of constructing a Bayesian optimizer based on the region correlation guide term - utilizing the balanced sampling function and performing Bayesian optimization iterative search, exploration - using the balanced sampling function expression: , in, This represents the mean of the predicted loss output by the Bayesian model. To predict uncertainty, Indicates the exploration weight coefficient. This represents the regional correlation weighting coefficient. It is the value of the regional correlation guide item; When the comprehensive loss function value reaches the preset coarse extraction accuracy threshold, or the number of Bayesian optimization iterations reaches the preset upper limit, the search stops. All parameter samples are sorted from low to high according to the comprehensive loss function value. Parameter samples with lower comprehensive loss function values ​​are selected according to the preset confidence threshold or confidence ratio. The distribution range of all parameter samples is statistically analyzed to obtain the high confidence parameter region of the device model parameters. The parameter sample with the smallest comprehensive loss function value is selected as the initial better parameter solution.

[0013] As a further improvement of the present invention, the specific steps for calculating the normalized local error between the test curve and the first simulation curve in each working region include: for the subthreshold region, the logarithmic error function is used to characterize the fitting error in the subthreshold region, and the expression of the local loss function corresponding to the subthreshold region is as follows: , in, Indicates subthreshold error. Indicates the number of sampling points in the subthreshold region. and They represent the first Simulated current value and test current value corresponding to each sampling point; For the linear and saturated regions, a relative error function is used to characterize the fitting error in the linear and saturated regions. The expressions for the local loss functions in the corresponding linear and saturated regions are as follows: , in, Indicates the error in the linear region and the saturation region. Indicates the number of sampling points in the linear region and the saturation region. and They represent the first Simulated current value and test current value corresponding to each sampling point This represents a stable term to prevent the denominator from being zero. For the capacitance characteristic region, the mean square error function is used to characterize the capacitance characteristic fitting error, and the expression for the local loss function corresponding to the capacitance characteristic region is as follows: , in, Indicates the error in the characteristic region of the capacitor. Indicates the number of sampling points in the capacitance characteristic region. and These represent the simulated capacitance value and the tested capacitance value at the corresponding sampling points, respectively. The median value of the local error of each working area in the initial parameter sample set is selected as the reference error of the corresponding working area. The local errors of each working area are normalized to a reference error, which is expressed as: , in, This represents the local error of the normalization. It is the local error of the r-th working region. It is the reference error.

[0014] As a further improvement of the present invention, the specific steps for calculating the comprehensive loss function value of the corresponding initial parameter sample vector include: substituting the normalized local error into a preset comprehensive loss function to calculate the comprehensive loss function value, the expression of which is: , in, Represents the comprehensive loss function. Indicates the number of work areas. Indicates the first The weights corresponding to each work area This represents the local error of the normalization.

[0015] As a further improvement of the present invention, the specific steps of constructing the local supervised training set include: inputting each of the local training parameter samples into the device model to generate a third simulation curve, and subtracting the third simulation curve from the test curve point by point at the same sampling point to obtain a first error feature vector; The difference between the initial optimal parameter solution and the local training parameter samples is calculated to obtain the true parameter correction amount. The first error feature vector and the corresponding true parameter correction amount are integrated to form a local supervised training set.

[0016] As a further improvement of the present invention, the specific steps of S6 include: using the first error feature vector in the local supervised training set as the input feature of the error-aware neural network, the error-aware neural network outputting the predicted parameter correction amount, and training based on the true parameter correction amount in the local supervised training set. The neural network training loss function expression of the error-aware neural network is as follows: , in, Indicates the first The amount of correction to the prediction parameters corresponding to each model parameter. This represents the correction amount for the corresponding actual parameters. This indicates the total number of model parameters to be extracted.

[0017] As a further improvement of the present invention, the specific steps in S7 of extracting error features based on the second simulation curve and the test curve, substituting them into the error-aware neural network to obtain the predicted correction amount, and generating the device model parameter file include: subtracting the second simulation curve and the test curve point by point at the same sampling point to obtain the second error feature vector; inputting the second error feature vector into the trained error-aware neural network to obtain the corresponding predicted correction amount; updating the initial optimal parameter solution based on the predicted correction amount to obtain the corrected device model parameters; and generating the device model parameter file.

[0018] The beneficial effects of this invention are as follows: 1. This invention employs a fully automated parameter extraction process combining Latin hypercube sampling, Bayesian optimization iterative search, and error-aware neural network local correction. Compared to traditional trial-and-error parameter tuning methods that rely on manual experience, this invention can automatically complete the entire parameter search and correction process within the parameter space of a high-dimensional device model, eliminating the need for repeated manual adjustments, significantly shortening the parameter extraction cycle, and improving the automation level and consistency of the extraction process. Compared to traditional global optimization-based parameter extraction methods, this invention first constructs a Bayesian model based on an initial parameter sample set, uses Bayesian optimization methods to perform a coarse search on the device model parameter space, and combines a region correlation guiding term to construct an exploration-utilizing balanced sampling function to dynamically guide the parameter search direction. This eliminates the need for long-term, high-precision global iterations across the entire parameter space, quickly obtaining high-confidence parameter regions and initial optimal parameter solutions, effectively reducing the number of simulation calls for high-cost devices and lowering overall computational resource consumption.

[0019] 2. This invention constructs a region-correlation-guided sampling mechanism based on working region error correlation. On the one hand, it prioritizes Bayesian optimization iterative search towards parameters strongly correlated with the current high-error working region, effectively improving the search efficiency and parameter convergence speed in the high-dimensional parameter space. On the other hand, it constructs corresponding local loss functions for the differences in electrical characteristics of the subthreshold region, linear region, saturation region, and capacitance characteristic region, and uses benchmark error normalization to unify the dimensions of local errors in each working region. While preserving the error evaluation characteristics of each working region, it improves the comparability of errors in different working regions in the comprehensive loss function, achieves a balance in the fitting effect of multiple working regions, and improves the overall fitting accuracy and robustness of the device model under multiple size and bias conditions.

[0020] 3. This invention employs an error-aware neural network to achieve rapid local parameter correction, solving the problems of high fitting difficulty, high sample requirements, and insufficient local accuracy in traditional global neural network parameter extraction methods. This invention first determines a high-confidence parameter region through Bayesian optimization iterative search. Then, centered on the initial optimal parameter solution, it generates local training parameter samples within the high-confidence parameter region through local perturbation, constructing a local supervised training set. This allows the network to learn the mapping relationship between error features and parameter correction amounts within the local region. Since the network only needs to fit a low-complexity mapping within the local parameter space, it does not need to perform global fitting of the entire parameter space, effectively reducing the network training difficulty and the number of training samples required, and improving network training stability and local parameter correction accuracy. The finally trained error-aware neural network can quickly output the predicted correction amount based on the second error feature vector, completing the correction of the initial optimal parameter solution. This avoids the computational overhead of a large amount of repetitive simulation and iterative search in the traditional optimization process, further improving parameter extraction efficiency and providing a high-precision device model parameter foundation for subsequent circuit simulation. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating the present invention; Figure 2 The accompanying drawing is a verification diagram of the effect of Embodiment 1 of the present invention; Figure 3 The accompanying drawings are for verifying the effect of Embodiment 2 of the present invention; Figure 4 The accompanying drawing is a verification diagram of the effect of Embodiment 3 of the present invention. Detailed Implementation

[0023] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent. To better illustrate this embodiment, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions of the product.

[0024] It will be understood by those skilled in the art that certain well-known structures and their descriptions may be omitted in the accompanying drawings. The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0025] Reference Figures 1 to 4 As can be seen, an automatic extraction method for parameters of a local neural correction device model includes the following steps: S1. Acquire test data that can characterize the key electrical properties of the device, and divide the test data into different working areas; As a further improvement of the present invention, the test data includes electrical characteristic curves under different bias conditions and device sizes, and the working region includes one or more of the following: subthreshold region, linear region, saturation region and capacitance characteristic region.

[0026] Electrical characteristic curves, for example, are shown below. Curves, and linearity and saturation Curves, for example Figures 2 to 4 It can be seen that linearity and saturation are... The curve can be divided into the subthreshold region, the linear region, and the saturation region. Since acquiring test data and dividing it into different working regions is a common technique used by those skilled in the art, it will not be elaborated upon here.

[0027] S2. Determine the set of device model parameters to be extracted and establish the device model parameter space; As a further improvement of the present invention, the specific steps of S2 include: selecting a device model version corresponding to the device to be fitted according to the device type of the device to be fitted; selecting model parameters related to the electrical characteristics to be fitted as the set of device model parameters to be extracted according to the model manual corresponding to the device model version; and determining the initial search range according to the default values ​​and reference value ranges given in the model manual to form the device model parameter space. The parameter set of the device model is represented as follows: , in, It is a set of device model parameters. Indicates the first One model parameter to be extracted. This represents the total number of model parameters to be extracted. , They are the first The lower limit and upper limit of the initial search range for each model parameter to be extracted.

[0028] S3. The Latin hypercube sampling method is used to perform hierarchical random sampling in the parameter space of the device model to generate an initial parameter sample set for Bayesian optimization initialization; As a further improvement of the present invention, the specific steps of S3 include: S31. Set the initial sampling quantity as... In the normalized device model parameter space In this process, each model parameter is divided into... 1. Equal probability subintervals; The expression for each equally probable subinterval: , S32, regarding the first A model parameter is generated from a set of parameters. A random permutation sequence consisting of n subintervals is represented as: , in, It is the first A random permutation sequence of model parameters, It is the first The initial parameter sample at the th ... Sub-intervals corresponding to each model parameter dimension; S33, No. The initial parameter sample is randomly selected from the corresponding sub-interval, and a normalized sample value is obtained to obtain the first... The normalized sampled values ​​of each initial parameter sample on each model parameter, wherein the normalized sampled values ​​are expressed as: , in, It is the first The initial parameter sample at the th ... Normalized sampled values ​​of each model parameter, No. The initial parameter sample at the th ... Sub-intervals corresponding to each model parameter dimension It is located in Random numbers within; Therefore, it can be ensured that each model parameter is uniformly sampled in layers within its normalized value range, while random combinations are formed between different model parameter dimensions, thus obtaining high-dimensional initial parameter samples with good coverage.

[0029] S34, according to the... The normalized sampled values ​​of each initial parameter sample on each model parameter are mapped according to the initial search range of each model parameter to obtain the corresponding actual parameter values. The actual parameter values ​​under the same initial parameter sample are integrated to obtain the initial parameter sample vector of the corresponding initial parameter sample. The initial parameter sample vector of all initial parameter samples is integrated to obtain the initial parameter sample set. The expression for the actual parameter value is: , It is the first The initial parameter sample at the th ... The actual parameter values ​​of each model parameter. It is the first The lower bound of the initial search range for each model parameter. These are normalized sample values. It is the first The upper limit of the initial search range for each model parameter.

[0030] Therefore, the first The initial parameter sample vector is represented as follows: , in, Indicates by the first The first to the second A complete set of device model parameters is composed of the values ​​of each model parameter. Repeat the above process to obtain the initial parameter sample set: .

[0031] S4. Construct a Bayesian model based on the initial parameter sample set, calculate the comprehensive loss function value according to the error between the test curve and the first simulation curve generated by the device model, and obtain the high-confidence parameter region and the initial better parameter solution of each device model parameter through Bayesian optimization iterative search. As a further improvement of the present invention, the specific steps of S4 include: inputting the initial parameter sample vector from the initial parameter sample set into the device model to generate a first simulation curve, calculating the normalized local error between the test curve and the first simulation curve in each working area, calculating the comprehensive loss function value of the corresponding initial parameter sample vector, and combining the comprehensive loss function value, the initial parameter sample vector and the normalized local error of each working area to form a historical observation sample. Based on the changes in model parameters in historical observation samples and the changes in normalized local errors in different working areas, the first... The local error of the normalization of the first working area and the first The correlation between model parameters Calculate the first The search priority of each model parameter, wherein the search priority is calculated using the following expression: , in, It is the first The search priority of each model parameter. Indicates the first The weights corresponding to each work area It is the first The local error of the normalization of the first working area and the first The correlation between model parameters ; Based on the search priority, a region relevance guide is constructed. The expression for the region relevance guide is as follows: , in, It is the value of the regional correlation guide item. This indicates the current optimal parameter value. It is the first The search priority of each model parameter. It is the first among the candidate parameter points to be evaluated. The values ​​of each model parameter, This indicates that the candidate parameter point is relative to the current better parameter point at the th... The offset in the direction of each model parameter.

[0032] By setting a regional correlation guide, Bayesian optimization can prioritize searching for candidate sampling points with large parameter offsets in the direction of parameters strongly correlated with the current high-error working area.

[0033] Exploration of constructing a Bayesian optimizer based on the region correlation guide term - utilizing the balanced sampling function and performing Bayesian optimization iterative search, exploration - using the balanced sampling function expression: , in, This represents the mean of the predicted loss output by the Bayesian model. To predict uncertainty, Indicates the exploration weight coefficient. This represents the regional correlation weighting coefficient. It is the value of the regional correlation guide item; Specifically, explore the weighting coefficients. The regional correlation weighting coefficient is used to adjust the contribution ratio of the prediction uncertainty term to the sampling function. This enhances the search priority for parameters strongly correlated with the high-error working region, thereby improving the search efficiency and parameter convergence capability of Bayesian optimization in the high-dimensional device model parameter space. The ratio of λ to β can be dynamically adjusted based on the device model parameter space dimension, historical sampling distribution, and optimization stage. Typically, λ ranges from 0.1 to 2, with a default value of 1; β ranges from 0.01 to 1, with a default value of 0.2.

[0034] It should be noted that since Bayesian optimization iterative search is a common technique used by those skilled in the art, this invention will not describe the specific scheme of Bayesian optimization iterative search. However, it should be noted that during the iteration process, the candidate parameter points with the highest score will also be added to the historical observation samples to participate in the next round of iteration after simulation and calculation of the comprehensive loss function value.

[0035] When the comprehensive loss function value reaches the preset coarse extraction accuracy threshold, or the number of Bayesian optimization iterations reaches the preset upper limit, the search stops. All parameter samples are sorted from low to high according to the comprehensive loss function value. Parameter samples with lower comprehensive loss function values ​​are selected according to the preset confidence threshold or confidence ratio. The distribution range of all parameter samples is statistically analyzed to obtain the high confidence parameter region of the device model parameters. The parameter sample with the smallest comprehensive loss function value is selected as the initial better parameter solution.

[0036] As a further improvement of the present invention, the specific steps for calculating the normalized local error between the test curve and the first simulation curve in each working region include: for the subthreshold region, the logarithmic error function is used to characterize the fitting error in the subthreshold region, and the expression of the local loss function corresponding to the subthreshold region is as follows: , in, Indicates subthreshold error. Indicates the number of sampling points in the subthreshold region. and They represent the first Simulated current value and test current value corresponding to each sampling point; It should be noted that the logarithmic error function is used because the drain current exhibits an exponential change characteristic. Number of sampling points in the subthreshold region. The recommended value range is 5 to 20.

[0037] For the linear and saturated regions, a relative error function is used to characterize the fitting error in the linear and saturated regions. The expressions for the local loss functions in the corresponding linear and saturated regions are as follows: , in, Indicates the error in the linear region and the saturation region. Indicates the number of sampling points in the linear region and the saturation region. and They represent the first Simulated current value and test current value corresponding to each sampling point This represents a stable term to prevent the denominator from being zero. It should be noted that the relative error function is used because the current magnitude varies over a large range; therefore, the relative error function is used to characterize the fitting error in the strong inversion region. The number of sampling points in the linear and saturation regions... The value ranges from 5 to 20. The stable term takes the minimum value 1e-20.

[0038] For the capacitance characteristic region, the mean square error function is used to characterize the capacitance characteristic fitting error, and the expression for the local loss function corresponding to the capacitance characteristic region is as follows: , in, Indicates the error in the characteristic region of the capacitor. Indicates the number of sampling points in the capacitance characteristic region. and These represent the simulated capacitance value and the tested capacitance value at the corresponding sampling points, respectively. It should be noted that the mean square error function is used because the capacitance parameter changes relatively smoothly.

[0039] To eliminate the differences in error functions across different working areas on numerical scales, baseline error normalization is necessary.

[0040] The median value of the local error of each working area in the initial parameter sample set is selected as the reference error of the corresponding working area. The local errors of each working area are normalized to a reference error, which is expressed as: , in, This represents the local error of the normalization. It is the local error of the r-th working region. It is the reference error.

[0041] This benchmark error normalization process does not change the evaluation form of the error function for each working area; it is only used to make the errors of different working areas comparable in the comprehensive loss function.

[0042] As a further improvement of the present invention, the specific steps for calculating the comprehensive loss function value of the corresponding initial parameter sample vector include: substituting the normalized local error into a preset comprehensive loss function to calculate the comprehensive loss function value, the expression of which is: , in, Represents the comprehensive loss function. Indicates the number of work areas. Indicates the first The weights corresponding to each work area This represents the local error of the normalization.

[0043] Since the subthreshold characteristics, on-state characteristics, and capacitance characteristics of the device are of similar importance, the weights are empirically assigned as 1:1:1.

[0044] S5. Solve using the initial optimal parameters. Centered on the high-confidence parameter region, local perturbations are performed on the parameters of each device model to generate local training parameter samples, and a local supervised training set is constructed. , in, Indicates the first A local training parameter sample, Indicates the first In the nth local training parameter sample The values ​​of each model parameter, For the Each model parameter has a local perturbation value range based on the high-confidence parameter region. , can be represented as: , , in, Indicates the first The local perturbation radius of each model parameter This represents the local disturbance scaling factor, with a value range of [value missing]. Generally, 0.5 is chosen.

[0045] As a further improvement of the present invention, the specific steps for constructing the local supervised training set include: inputting each of the local training parameter samples into the device model to generate a third simulation curve, and subtracting the third simulation curve from the test curve point by point at the same sampling point to obtain a first error feature vector. ; First error feature vector expression: , , in, Indicates the first Error feature vectors corresponding to local training parameter samples Indicates the first Curve error at each sampling point This indicates the number of sampling points in the device curve. Indicates the first The local training parameter samples correspond to the third simulation curve in the... Simulated values ​​at each sampling point Indicates the test curve at the 1st Test values ​​at each sampling point.

[0046] The difference between the initial optimal parameter solution and the local training parameter samples is calculated to obtain the true parameter correction amount. The first error feature vector and the corresponding true parameter correction amount are integrated to form a local supervised training set.

[0047] For the The expression for the correction amount of each model parameter and the actual parameter:

[0048] in, Indicates the first In the nth local training sample The amount of correction to the actual parameters corresponding to each model parameter. It is the first The initial optimal parameter solutions for each model parameter. Based on this, a locally supervised training set is constructed:

[0049] in, This represents the locally supervised training set used to train the error-aware neural network. Indicates the first Error feature vectors corresponding to local training parameter samples This indicates the number of local training parameter samples.

[0050] S6. Construct a training error-aware neural network based on the local supervised training set; As a further improvement of the present invention, the specific steps of S6 include: using the first error feature vector in the local supervised training set as the input feature of the error-aware neural network, the error-aware neural network outputting the predicted parameter correction amount, and training based on the true parameter correction amount in the local supervised training set. The neural network training loss function expression of the error-aware neural network is as follows: , in, Indicates the first The amount of correction to the prediction parameters corresponding to each model parameter. This represents the correction amount for the corresponding actual parameters. This indicates the total number of model parameters to be extracted.

[0051] This technical solution enables the learning of the mapping relationship between device error characteristics and model parameter correction directions.

[0052] S7. Input the initial optimized parameter solution into the device model to obtain the second simulation curve. Extract error features based on the second simulation curve and the test curve, substitute them into the error-aware neural network to obtain the prediction correction amount, and generate the device model parameter file.

[0053] As a further improvement of the present invention, the specific steps in S7 of extracting error features based on the second simulation curve and the test curve, substituting them into the error-aware neural network to obtain the predicted correction amount, and generating the device model parameter file include: subtracting the second simulation curve and the test curve point by point at the same sampling point to obtain the second error feature vector; inputting the second error feature vector into the trained error-aware neural network to obtain the corresponding predicted correction amount; updating the initial optimal parameter solution based on the predicted correction amount to obtain the corrected device model parameters; and generating the device model parameter file.

[0054] The expression for the corresponding prediction correction is obtained as follows: , in, This represents the trained error-aware neural network. Represents network parameters, This represents the amount of prediction correction in the network output.

[0055] The prediction correction amount can be expressed as:

[0056] in, Indicates the first The prediction correction amount corresponding to each model parameter. This indicates the number of model parameters to be extracted.

[0057] The expression for updating the initial better parameter solution based on the predicted correction amount:

[0058] in, This represents the final device model parameters after local correction by the neural network. It is the initial, relatively optimal parameter solution.

[0059] See attached document Figure 2-4 The accompanying drawings are for verifying the effects of various embodiments of the present invention.

[0060] For appendix Figure 2This is a comparison chart of the gate capacitance-gate voltage (Cgg-Vgs) characteristic fitting under zero-drain bias, used to verify the fitting accuracy of the capacitance characteristic region. The chart contains three parallel subplots, corresponding to GAA devices with gate lengths Lg=39nm, Lg=50nm, and Lg=60nm, respectively; the horizontal axis represents the gate-source voltage Vgs, and the vertical axis represents the gate capacitance Cgg. Hollow circles represent the capacitance data obtained from actual device testing, while solid lines represent the capacitance curve output from the device model simulation after parameters are extracted using the method of this invention.

[0061] For appendix Figure 3 This figure presents a comparison of the drain-gate voltage (Ids-Vgs) characteristic fitting in the low-drain bias linear region, used to verify the current fitting accuracy in the linear operating region. The figure contains three parallel subplots, corresponding to GAA devices with gate lengths Lg=39nm, Lg=50nm, and Lg=60nm, respectively; the horizontal axis represents the gate-source voltage Vgs, and the vertical axis represents the drain current Id. Hollow circles represent current data obtained from actual device testing, while solid lines represent the current curves output by the device model simulation after parameters are extracted using the method of this invention.

[0062] For appendix Figure 4 This figure presents a comparison of the drain current-gate voltage (Ids-Vgs) characteristic fitting in the high-drain bias saturation region, used to verify the current fitting accuracy in the saturation operating region. The figure contains three parallel subplots, corresponding to GAA devices with gate lengths Lg=39nm, Lg=50nm, and Lg=60nm, respectively; the horizontal axis represents the gate-source voltage Vgs, and the vertical axis represents the drain current Id. Hollow circles represent current data obtained from actual device testing, while solid lines represent the current curves output from device model simulations after parameters are extracted using the method of this invention.

[0063] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. An automatic extraction method for parameters of a local neural correction device model, characterized in that, Includes the following steps: S1. Acquire test data that can characterize the key electrical properties of the device, and divide the test data into different working areas; S2. Determine the set of device model parameters to be extracted and establish the device model parameter space; S3. The Latin hypercube sampling method is used to perform hierarchical random sampling in the parameter space of the device model to generate an initial parameter sample set for Bayesian optimization initialization; S4. Construct a Bayesian model based on the initial parameter sample set, calculate the comprehensive loss function value according to the error between the test curve and the first simulation curve generated by the device model, and obtain the high-confidence parameter region and the initial better parameter solution of each device model parameter through Bayesian optimization iterative search. S5. Using the initial optimal parameter solution as the center, locally perturb the parameters of each device model within the high-confidence parameter region to generate local training parameter samples, and construct a local supervised training set. S6. Construct a training error-aware neural network based on the local supervised training set; S7. Input the initial optimized parameter solution into the device model to obtain the second simulation curve. Extract error features based on the second simulation curve and the test curve, substitute them into the error-aware neural network to obtain the prediction correction amount, and generate the device model parameter file.

2. The method for automatically extracting model parameters of a local neural correction device according to claim 1, characterized in that, The test data includes electrical characteristic curves under different bias conditions and device sizes, and the working region includes one or more of the following: subthreshold region, linear region, saturation region, and capacitance characteristic region.

3. The method for automatically extracting model parameters of a local neural correction device according to claim 1, characterized in that, The specific steps of S2 include: selecting a device model version corresponding to the device to be fitted according to the device type; selecting model parameters related to the electrical characteristics to be fitted as the set of device model parameters to be extracted according to the model manual corresponding to the device model version; and determining the initial search range according to the default values ​​and reference value ranges given in the model manual to form the device model parameter space. The parameter set of the device model is represented as follows: ， in, It is a set of device model parameters. Indicates the first One model parameter to be extracted. This represents the total number of model parameters to be extracted. , They are the first The lower limit and upper limit of the initial search range for each model parameter to be extracted.

4. The method for automatically extracting model parameters of a local neural correction device according to claim 3, characterized in that, The specific steps of S3 include: S31. Set the initial sampling quantity as... In the normalized device model parameter space In this process, each model parameter is divided into...

1. Equal probability subintervals; S32, regarding the first A model parameter is generated from a set of parameters. A random permutation sequence consisting of n subintervals is represented as: ， in, It is the first A random permutation sequence of model parameters, It is the first The initial parameter sample at the th ... Sub-intervals corresponding to each model parameter dimension; S33, No. The initial parameter sample is randomly selected from the corresponding sub-interval, and a normalized sample value is obtained to obtain the first parameter. The normalized sampled values ​​of each initial parameter sample on each model parameter, wherein the normalized sampled values ​​are expressed as: ， in, It is the first The initial parameter sample at the th ... Normalized sampled values ​​of each model parameter, No. The initial parameter sample at the th ... Sub-intervals corresponding to each model parameter dimension It is located in Random numbers within; S34, according to the... The normalized sampled values ​​of each initial parameter sample on each model parameter are mapped according to the initial search range of each model parameter to obtain the corresponding actual parameter values. The actual parameter values ​​under the same initial parameter sample are integrated to obtain the initial parameter sample vector of the corresponding initial parameter sample. The initial parameter sample vector of all initial parameter samples is integrated to obtain the initial parameter sample set. The expression for the actual parameter value is: ， It is the first The initial parameter sample at the th ... The actual parameter values ​​of each model parameter. It is the first The lower bound of the initial search range for each model parameter. These are normalized sample values. It is the first The upper limit of the initial search range for each model parameter.

5. The method for automatically extracting model parameters of a local neural correction device according to claim 1, characterized in that, The specific steps of S4 include: inputting the initial parameter sample vector from the initial parameter sample set into the device model to generate a first simulation curve; calculating the normalized local error between the test curve and the first simulation curve in each working area; calculating the comprehensive loss function value of the corresponding initial parameter sample vector; and combining the comprehensive loss function value, the initial parameter sample vector, and the normalized local error of each working area to form a historical observation sample. Based on the changes in model parameters in historical observation samples and the changes in normalized local errors in different working areas, the first... The local error of the normalization of the first working area and the first The correlation between model parameters Calculate the first The search priority of each model parameter, wherein the search priority is calculated using the following expression: ， in, It is the first The search priority of each model parameter. Indicates the first The weights corresponding to each work area It is the first The local error of the normalization of the first working area and the first The correlation between model parameters ; Based on the search priority, a region relevance guide is constructed. The expression for the region relevance guide is as follows: ， in, It is the value of the regional correlation guide item. This indicates the current optimal parameter value. It is the first The search priority of each model parameter. It is the first among the candidate parameter points to be evaluated. The values ​​of each model parameter; Exploration of constructing a Bayesian optimizer based on the region correlation guide term - utilizing the balanced sampling function and performing Bayesian optimization iterative search, exploration - using the balanced sampling function expression: ， in, This represents the mean of the predicted loss output by the Bayesian model. To predict uncertainty, Indicates the exploration weighting coefficient. This represents the regional correlation weighting coefficient. It is the value of the regional correlation guide item; When the comprehensive loss function value reaches the preset coarse extraction accuracy threshold, or the number of Bayesian optimization iterations reaches the preset upper limit, the search stops. All parameter samples are sorted from low to high according to the comprehensive loss function value. Parameter samples with lower comprehensive loss function values ​​are selected according to the preset confidence threshold or confidence ratio. The distribution range of all parameter samples is statistically analyzed to obtain the high confidence parameter region of the device model parameters. The parameter sample with the smallest comprehensive loss function value is selected as the initial better parameter solution.

6. The method for automatically extracting model parameters of a local neural correction device according to claim 5, characterized in that, The specific steps for calculating the normalized local error between the test curve and the first simulation curve in each working region include: For the subthreshold region, the logarithmic error function is used to characterize the fitting error in the subthreshold region, and the expression for the local loss function corresponding to the subthreshold region is as follows: ， in, Indicates subthreshold error. Indicates the number of sampling points in the subthreshold region. and They represent the first Simulated current value and test current value corresponding to each sampling point; For the linear and saturated regions, a relative error function is used to characterize the fitting error in the linear and saturated regions. The expressions for the local loss functions in the corresponding linear and saturated regions are as follows: ， in, Indicates the error in the linear region and the saturation region. Indicates the number of sampling points in the linear region and the saturation region. and They represent the first Simulated current value and test current value corresponding to each sampling point This represents a stable term to prevent the denominator from being zero. For the capacitance characteristic region, the mean square error function is used to characterize the capacitance characteristic fitting error, and the expression for the local loss function corresponding to the capacitance characteristic region is as follows: ， in, Indicates the error in the characteristic region of the capacitor. Indicates the number of sampling points in the capacitance characteristic region. and These represent the simulated capacitance value and the tested capacitance value at the corresponding sampling points, respectively. The median value of the local error of each working area in the initial parameter sample set is selected as the reference error of the corresponding working area. The local errors of each working area are normalized to a reference error, which is expressed as: ， in, This represents the local error of the normalization. It is the local error of the r-th working region. It is the reference error.

7. The method for automatically extracting model parameters of a local neural correction device according to claim 6, characterized in that, The specific steps for calculating the comprehensive loss function value of the corresponding initial parameter sample vector include: substituting the normalized local error into the preset comprehensive loss function to calculate the comprehensive loss function value. The expression for the comprehensive loss function is: ， in, Represents the comprehensive loss function. Indicates the number of work areas. Indicates the first The weights corresponding to each work area This represents the local error of the normalization.

8. The method for automatically extracting model parameters of a local neural correction device according to claim 1, characterized in that, The specific steps for constructing the local supervised training set include: inputting each local training parameter sample into the device model to generate a third simulation curve, and subtracting the third simulation curve from the test curve at the same sampling point to obtain a first error feature vector. The difference between the initial optimal parameter solution and the local training parameter samples is calculated to obtain the true parameter correction amount. The first error feature vector and the corresponding true parameter correction amount are integrated to form a local supervised training set.

9. The method for automatically extracting model parameters of a local neural correction device according to claim 1, characterized in that, The specific steps of S6 include: using the first error feature vector in the local supervised training set as the input feature of the error-aware neural network; the error-aware neural network outputs a predicted parameter correction amount; training is performed based on the true parameter correction amount in the local supervised training set; and the neural network training loss function expression of the error-aware neural network is as follows: ， in, Indicates the first The amount of correction to the prediction parameters corresponding to each model parameter. This represents the correction amount for the corresponding actual parameters. This indicates the total number of model parameters to be extracted.

10. The method for automatically extracting model parameters of a local neural correction device according to claim 1, characterized in that, The specific steps in S7 of extracting error features from the second simulation curve and the test curve, substituting them into the error-aware neural network to obtain the predicted correction amount, and generating the device model parameter file include: subtracting the second simulation curve and the test curve point by point at the same sampling point to obtain the second error feature vector; inputting the second error feature vector into the trained error-aware neural network to obtain the corresponding predicted correction amount; updating the initial optimal parameter solution according to the predicted correction amount to obtain the corrected device model parameters; and generating the device model parameter file.

Images