A method and device for intelligently extracting parameters of semiconductor devices

Abstract

The present invention provides a method and device for intelligently extracting parameters of semiconductor devices, and relates to the field of integrated circuit technology. This method comprises: firstly, inputting the acquired target data of the semiconductor device into the extractor of the initial neural network to obtain the current model parameter combination. Then, the current target data of the current model parameter combination is outputted using a simulator, and a plurality of first electrical targets corresponding to the current target data are calculated. The first error is calculated to determine whether the first error is less than a preset error. If not, the current model parameter combination is adjusted and the first error is re-simulated and calculated until the first error is less than the preset error. According to the current model parameter combination and the preset range, a model parameter range is generated. Finally, a data set is generated according to the model parameter range, and the target model parameter combination is outputted using the extractor of the neural network trained by the data set. The efficiency and accuracy of parameter extraction of semiconductor devices can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of integrated circuit technology, and in particular to a method and device for intelligently extracting parameters of a semiconductor device. Background Art

[0002] In the semiconductor field, the extraction of transistor compact model parameters is the core of achieving seamless integration between circuit design and semiconductor device technology. The extraction of transistor compact model parameters requires extracting a set of parameters from the physical properties of semiconductor devices. These parameters can accurately describe the behavior of semiconductor devices and are used in circuit simulation software to predict the performance of semiconductor devices under various operating conditions. The accuracy of semiconductor device parameter extraction is crucial to the performance prediction and optimization of circuit design, and has a decisive impact on the success of integrated circuit design.

[0003] At present, traditional semiconductor device parameter extraction methods usually involve grouping parameters and associating them with different curve regions. However, traditional semiconductor device parameter extraction methods not only require multiple iterations, but also rely on the expertise of domain experts. Especially for complex models, such as Berkeley Short-channel IGFET Model-Compact Model for 3D Gate Stack Transistors (BSIM-CMG), the extraction of semiconductor device parameters may take days or even weeks to complete, which consumes a lot of time and makes the efficiency of semiconductor device parameter extraction low. The extraction of semiconductor device parameters requires the deep expertise of experts, and various physical effects and analysis of the structure of semiconductor devices must be considered comprehensively. If the factors analyzed by experts are not very detailed, it will affect the accuracy of semiconductor device parameter extraction, making the accuracy of semiconductor device parameter extraction low. Summary of the invention

[0004] The purpose of the embodiments of the present invention is to provide a method and apparatus for intelligently extracting semiconductor device parameters, so as to solve the problems of low efficiency and low accuracy of semiconductor device parameter extraction.

[0005] To solve the above technical problems, the embodiments of the present invention provide the following technical solutions:

[0006] A first aspect of the present invention provides a method for intelligently extracting parameters of a semiconductor device, the method comprising:

[0007] Acquire target data of the semiconductor device, the target data being a relationship curve between terminal characteristics of the semiconductor device;

[0008] Inputting the target data into an extractor of an initial neural network so that the extractor of the initial neural network outputs a current model parameter combination of the semiconductor device, wherein the extractor of the initial neural network is an extractor trained using preset target data and a preset model parameter combination corresponding to the preset target data;

[0009] Inputting the current model parameter combination into the simulator, simulating the current target data of the current model parameter combination, and calculating a plurality of first electrical targets corresponding to the current target data;

[0010] Calculating a first error according to a plurality of first electrical targets and an actual electrical target, wherein the actual electrical target is an electrical target calculated according to the target data;

[0011] Determining whether the first error is less than a preset error;

[0012] If not, a probability factor is determined based on the gradient of the neural network, the scaling factor and the first error;

[0013] According to the probability factor, the current model parameter combination is adjusted to obtain an updated model parameter combination, the updated model parameter combination is used as the current model parameter combination, and the current model parameter combination is input into the simulator to simulate the current target data of the current model parameter combination, and the steps of calculating a plurality of first electrical targets corresponding to the current target data are performed;

[0014] If yes, a model parameter range is generated according to the current model parameter combination and the preset range, and a Monte Carlo simulation is performed within the model parameter range to generate a data set, the data set including a first target data set and a corresponding first model parameter combination;

[0015] The target data is input into the extractor of the trained neural network so that the extractor of the trained neural network outputs a target model parameter combination. The extractor of the trained neural network is an extractor obtained by inputting a first target data set and a first model parameter combination into the extractor of the initial neural network to train the extractor of the initial neural network.

[0016] A second aspect of the present application provides a semiconductor device parameter intelligent extraction device, the semiconductor device parameter intelligent extraction device comprising:

[0017] An acquisition module, used for acquiring target data of the semiconductor device, wherein the target data is a relationship curve between terminal characteristics of the semiconductor device;

[0018] A current model parameter combination extraction module, used for inputting target data into an extractor of an initial neural network so that the extractor of the initial neural network outputs a current model parameter combination of the semiconductor device, wherein the extractor of the initial neural network is an extractor trained using preset target data and a preset model parameter combination corresponding to the preset target data;

[0019] A simulation module, used for inputting the current model parameter combination into the simulator, simulating the current target data of the current model parameter combination, and calculating a plurality of first electrical targets corresponding to the current target data;

[0020] An error calculation module, used to calculate a first error according to a plurality of first electrical targets and an actual electrical target, wherein the actual electrical target is an electrical target calculated according to the target data;

[0021] A judging module, used for judging whether the first error is less than a preset error;

[0022] a determination module, for determining a probability factor according to a gradient of the neural network, a scaling factor and a first error if no;

[0023] An adjustment module, used for adjusting the current model parameter combination according to the probability factor to obtain an updated model parameter combination, taking the updated model parameter combination as the current model parameter combination, and returning the step of inputting the current model parameter combination into the simulator, simulating the current target data of the current model parameter combination, and calculating a plurality of first electrical targets corresponding to the current target data;

[0024] A generating module, for generating a model parameter range according to the current model parameter combination and the preset range, and performing Monte Carlo simulation within the model parameter range to generate a data set, the data set including a first target data set and a corresponding first model parameter combination;

[0025] The target model parameter combination extraction module is used to input the target data into the extractor of the trained neural network so that the extractor of the trained neural network outputs the target model parameter combination. The extractor of the trained neural network is an extractor obtained by inputting the first target data set and the first model parameter combination into the extractor of the initial neural network to train the extractor of the initial neural network.

[0026] Compared with the prior art, the semiconductor device parameter intelligent extraction method and device provided by the present invention adjust the current model parameter combination according to the probability factor to obtain an updated model parameter combination, and can automatically determine the updated model parameter combination, and then when the first error is less than the preset error, generate the model parameter range according to the current model parameter combination and the preset range, and can automatically determine the model parameter range, thereby improving the efficiency of determining the model parameter range, and thereby improving the efficiency of semiconductor device parameter extraction; when the first error is not less than the preset error, determine the probability factor according to the gradient of the neural network, the proportional factor and the first error, and can adjust the current model parameter combination according to the probability factor, so that the updated model parameter combination is more accurate, thereby making the accuracy of semiconductor device parameter extraction higher. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] By reading the detailed description below with reference to the accompanying drawings, the above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

[0028] Figure 1 The flowchart of the method for intelligently extracting semiconductor device parameters is schematically shown;

[0029] Figure 2 A schematic diagram schematically shows target data simulated under a wide parameter range;

[0030] Figure 3 A schematic diagram schematically shows target data simulated under a narrow parameter range;

[0031] Figure 4 The structure diagram of the semiconductor device parameter intelligent extraction device is schematically shown. DETAILED DESCRIPTION

[0032] The exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0033] It should be noted that, unless otherwise specified, the technical terms or scientific terms used in the present invention should have the common meanings understood by those skilled in the art to which the present invention belongs.

[0034] The method in the embodiment of the present invention is described in detail below.

[0035] Figure 1The flowchart of the method for intelligently extracting semiconductor device parameters in an embodiment of the present invention is schematically shown. Figure 1 As shown, the method may include:

[0036] S101 , acquiring target data of a semiconductor device.

[0037] The target data is a relationship curve between terminal characteristics of a semiconductor device.

[0038] The relationship curves between the terminal characteristics of the semiconductor device include, but are not limited to, the current-voltage relationship curve of the semiconductor device, the capacitance-voltage relationship curve of the semiconductor device, and the scattering parameter-frequency relationship curve of the semiconductor device.

[0039] The target data provides real semiconductor device behavior information by obtaining the actual measured current-voltage relationship curve of the semiconductor device, the capacitance-voltage relationship curve of the semiconductor device, and the scattering parameter-frequency relationship curve of the semiconductor device. These target data can cover different working conditions and semiconductor device samples.

[0040] S102, inputting the target data into the extractor of the initial neural network so that the extractor of the initial neural network outputs the current model parameter combination of the semiconductor device.

[0041] The extractor of the initial neural network is an extractor trained using a combination of preset target data and preset model parameters corresponding to the preset target data.

[0042] The current model parameter combinations include: gate work function, interface trap parameters, SiO 2 Equivalent gate dielectric thickness, zero bias source to drain extension resistance per unit width, low field mobility, saturation velocity in the saturation region, long channel saturation drain voltage parameter, smoothing function factor of saturation drain voltage, and parameters related to drain induced barrier lowering effect.

[0043] The extractor of the initial neural network is a model parameter combination extractor based on the gated neural network. The extractor of the initial neural network includes multiple expert networks, multiple gated networks and multiple tower networks. Each expert network is connected to multiple gated networks respectively. The number of multiple gated networks and multiple tower networks is the same, and one gated network corresponds to one tower network, and each gated network is connected to the corresponding tower network. Each expert network includes 3 fully connected layers, and each tower network includes 3 fully connected layers. The expert network is used to connect the input and extract the features of each input. Each expert network focuses on processing a specific type of data. The tower network is used to extract the output features of the gated network and build a hierarchical feature representation based on the output features to improve the model's ability to process complex data.

[0044] A wide range of parameter range data is used for preliminary training to obtain an initial neural network extractor. The training process of the initial neural network extractor involves two main data sources, which are preset target data in the simulation IV data pool and preset model parameter combinations corresponding to the preset target data.

[0045] The simulation IV data pool is used to store a large amount of simulation data generated by semiconductor device models and simulation tools. These simulation data mainly come from: simulation data generated by Monte Carlo method using the built-in semiconductor device model of the simulator, which simulates various behaviors of semiconductor devices through a wide range of parameter changes. The generated simulation data is a combination of preset target data and preset model parameters corresponding to the preset target data.

[0046] The preset target data in the simulation IV data pool and the preset model parameter combination corresponding to the preset target data are generated through a wide range of parameter changes, providing rich simulated semiconductor device behavior information.

[0047] The data that the extractor of the initial neural network needs to train is a combination of preset target data and a combination of preset model parameters corresponding to the preset target data.

[0048] The extractor of the initial neural network is trained using the preset target data and the preset model parameter combination corresponding to the preset target data. The extractor of the initial neural network can learn a wide range of semiconductor device behavior characteristics, thereby extracting a preliminary semiconductor device model parameter combination. The extractor of the initial neural network can handle a variety of different semiconductor device states and conditions.

[0049] S103, inputting the current model parameter combination into the simulator, simulating the current target data of the current model parameter combination, and calculating a plurality of first electrical targets corresponding to the current target data.

[0050] Among them, multiple first electrical targets include a first drain current, a first maximum transconductance, a first linear region transconductance, a threshold voltage of the first linear region, a first saturation drain current, a first cutoff drain current and a first subthreshold slope, and actual electrical targets include actual drain current, actual maximum transconductance, actual linear region transconductance, actual linear region threshold voltage, actual saturation drain current, actual cutoff drain current and actual subthreshold slope.

[0051] Simulators include HSPICE, Advanced Design System (ADS) and other simulators.

[0052] S104: Calculate a first error according to the multiple first electrical targets and the actual electrical targets.

[0053] The actual electrical target is the electrical target calculated based on the target data.

[0054] The first error is an error between the plurality of first electrical targets and the actual electrical target.

[0055] Specifically, calculating the first error according to the plurality of first electrical targets and the actual electrical targets includes:

[0056] Step A1: Calculate a first error based on the first drain current, the first maximum transconductance, the first linear region transconductance, the threshold voltage of the first linear region, the first saturation drain current, the first cutoff drain current and the first subthreshold slope, as well as the actual drain current, the actual maximum transconductance, the actual linear region transconductance, the threshold voltage of the actual linear region, the actual saturation drain current, the actual cutoff drain current and the actual subthreshold slope.

[0057] Specifically, the first error is calculated according to the first drain current, the first maximum transconductance, the first linear region transconductance, the threshold voltage of the first linear region, the first saturated drain current, the first cutoff drain current and the first subthreshold slope, and the actual drain current, the actual maximum transconductance, the actual linear region transconductance, the threshold voltage of the actual linear region, the actual saturated drain current, the actual cutoff drain current and the actual subthreshold slope, including:

[0058] The first error is calculated using the following first formula according to the first drain current, the first maximum transconductance, the first linear region transconductance, the threshold voltage of the first linear region, the first saturated drain current, the first cutoff drain current and the first subthreshold slope, as well as the actual drain current, the actual maximum transconductance, the actual linear region transconductance, the threshold voltage of the actual linear region, the actual saturated drain current, the actual cutoff drain current and the actual subthreshold slope:

[0059] ;

[0060] in, is the first error, is the first drain current, is the RMS difference of the first drain current, is the first maximum transconductance, is the actual maximum transconductance, is the transconductance in the first linear region, is the actual linear region transconductance, is the threshold voltage of the first linear region, is the threshold voltage in the actual linear region, is the first saturation leakage current, is the actual saturation leakage current, is the first cut-off leakage current, is the actual cut-off leakage current, is the first subthreshold slope, is the actual subthreshold slope, is the weight factor, , For the The first electrical target.

[0061] What is calculated is the RMS difference between the first drain current and the actual drain current.

[0062] S105: Determine whether the first error is smaller than a preset error.

[0063] The preset error is 5 percent of the first error.

[0064] It is determined whether the first error is less than a preset error. If not, step S106 is executed. If so, step S108 is executed.

[0065] S106. Determine a probability factor according to the gradient of the neural network, the proportional factor and the first error.

[0066] The scaling factor includes a first scaling factor and a second scaling factor.

[0067] The neural network can be a multi-layer perceptron neural network or a gated neural network.

[0068] The neural network-based correlation analysis specifically determines the probability factor based on the gradient, scale factor and first error of the neural network. The neural network-based correlation analysis uses the inherent ability of neural networks to generate high-precision approximate functions for multivariate parameters, even if there are correlations and nonlinearities between the multivariate parameters. Once the training process is completed, the neural network-based correlation analysis has no additional tasks and is very fast in calculation.

[0069] Before determining the probability factor according to the gradient, proportional factor and first error of the neural network, the method further includes: inputting a plurality of first electrical targets into the neural network so that the neural network outputs a gradient.

[0070] The neural network is a neural network trained by training data. The training data of the neural network is the same as the training data of the extractor of the initial neural network, which is a combination of preset target data and preset model parameters corresponding to the preset target data.

[0071] The gradient of the neural network output is calculated as follows:

[0072] ;

[0073] in, is the gradient, is the partial derivative of the first electrical objective, is the partial derivative of the current model parameter combination, For the The first electrical target, is the first combination of current model parameters The current model parameters.

[0074] The gradient represents the rate at which the output changes with the input, which reveals the control strength of the current model parameter combination on the first electrical target and the control source of the change of the first electrical target.

[0075] Specifically, the probability factor is determined according to the gradient, the scaling factor and the first error of the neural network, including:

[0076] Step B1: Determine a first transformation equation according to the gradient and the first proportional factor.

[0077] Specifically, according to the gradient and the first proportional factor, a first transformation equation is determined, including:

[0078] According to the gradient and the first proportional factor, the first transformation equation is determined using the following second formula:

[0079] ;

[0080] in, is the first transformation equation, is the gradient, is the first scaling factor, For the The first electrical target, is the total number of the first electrical targets, is the first combination of current model parameters The current model parameters.

[0081] The first scale factor is set to 2.

[0082] Step B2: Determine a probability factor according to the first transformation equation, the second proportional factor and the first error.

[0083] Specifically, determining the probability factor according to the first transformation equation, the second proportional factor and the first error includes:

[0084] According to the first transformation equation, the second proportional factor and the first error, the probability factor is determined using the following third formula:

[0085] ;

[0086] ;

[0087] in, is the probability factor, is the second scale factor, is the second transformation equation, is the first error, is the first transformation equation, is the total number of current model parameters of the current model parameter combination, For the The first electrical target, is the total number of the first electrical targets, is the first combination of current model parameters The current model parameters.

[0088] The second scale factor is set to 10.

[0089] The probability factor is then used in the crossover and mutation process of the genetic algorithm to guide of the search process.

[0090] S107. According to the probability factor, adjust the current model parameter combination to obtain an updated model parameter combination, use the updated model parameter combination as the current model parameter combination, and return to the step of inputting the current model parameter combination into the simulator, simulating the current target data of the current model parameter combination, and calculating multiple first electrical targets corresponding to the current target data.

[0091] According to the probability factor, the current model parameter combination is adjusted to obtain an updated model parameter combination, the updated model parameter combination is used as the current model parameter combination, and the process returns to step S103 to continue execution.

[0092] Specifically, according to the probability factor, the current model parameter combination is adjusted to obtain an updated model parameter combination, including:

[0093] Step C1: Compare each current model parameter in the current model parameter combination with the probability factor to determine a first current model parameter combination and a second current model parameter combination.

[0094] Among them, the first current model parameter combination is a combination in which all first current model parameters in the first current model parameter combination are greater than the probability factor, and the second current model parameter combination is a combination in which all second current model parameters in the second current model parameter combination are not greater than the probability factor.

[0095] The current model parameter combination includes a first current model parameter combination and a second current model parameter combination.

[0096] Compare each current model parameter in the current model parameter combination with the probability factor, determine all first current model parameters that are greater than the probability factor, and the combination containing all first current model parameters is the first current model parameter combination, determine all second current model parameters that are not greater than the probability factor, and the combination containing all second current model parameters is the second current model parameter combination.

[0097] Step C2: Using the first current model parameter combination as an updated model parameter combination.

[0098] Each current model parameter in the current model parameter combination is compared with the probability factor, and the model parameter combination corresponding to all model parameters greater than the probability factor, i.e., the first current model parameter combination, is not modified and is directly used as the updated model parameter combination.

[0099] Step C3: performing cross-mutation processing on the second current model parameter combination to obtain an updated second model parameter combination, and using the updated second model parameter combination as the updated model parameter combination.

[0100] Each current model parameter in the current model parameter combination is compared with the probability factor, and the model parameter combination corresponding to all model parameters not greater than the probability factor, i.e., the second current model parameter combination, is subjected to cross-mutation processing to obtain an updated second model parameter combination, and the updated second model parameter combination is used as the updated model parameter combination.

[0101] The updated model parameter combination includes a first current model parameter combination and an updated second model parameter combination.

[0102] After performing cross-mutation processing on the second current model parameter combination to obtain an updated second model parameter combination, and using the updated second model parameter combination as the updated model parameter combination, the updated model parameter combination is used as the current model parameter combination, and the process returns to step S103 to continue executing and perform an iterative operation.

[0103] Here, the process of natural selection is simulated by a genetic algorithm, specifically by changing the population through crossover mutation processing, that is, changing the training data of the extractor for training the initial neural network, in order to find the optimal solution.

[0104] Returning to step S103 to continue the operation is to use the updated model parameter combination as the current model parameter combination, the current model parameter combination is historical data, the historical data is input into the simulator, the current target data of the current model parameter combination is simulated, and multiple first electrical targets corresponding to the current target data are calculated, and then steps S104 to S107 are continued until it is determined that the first error is less than the preset error, and steps S108 to S109 are executed.

[0105] Since historical data is close to the real situation, it has high quality and provides rich information about semiconductor device behavior for simulation. Historical data takes advantage of the excellent ability of neural networks in storing such data, which helps to reduce repeated calculations when encountering similar target data.

[0106] The second current model parameter combination is subjected to cross-mutation processing to continuously optimize the second current model parameter combination and reduce the error.

[0107] S108. Generate a model parameter range according to the current model parameter combination and the preset range, and perform Monte Carlo simulation within the model parameter range to generate a data set.

[0108] The data set includes: a first target data set and a corresponding first model parameter combination.

[0109] It is determined whether the first error is less than a preset error. If so, a model parameter range is generated according to the current model parameter combination and the preset range, and a Monte Carlo simulation is performed within the model parameter range to generate a data set.

[0110] The preset range is a 20% change in the parameter value corresponding to each model parameter combination in the current model parameter combination. That is, the parameter value corresponding to each model parameter combination in the current model parameter combination plus 20% of the parameter value is used as the maximum value of the preset range, and the parameter value corresponding to each model parameter combination in the current model parameter combination minus 20% of the parameter value is used as the minimum value of the preset range. The range formed by the minimum value and the maximum value is the preset range. For example, a parameter value corresponding to a current model parameter combination is 1, and the preset range is 0.8-1.2.

[0111] The size range of the number of data sets can be set between 10K and 60K.

[0112] The resulting dataset is a narrow range dataset.

[0113] The data set generated by a large number of Monte Carlo simulations within the model parameter range is a refined data set, which is used to capture the impact of small changes in parameters on the behavior of semiconductor devices.

[0114] Figure 2 Schematic diagram showing target data simulated under a wide range of parameters, Figure 3 A schematic diagram schematically shows the target data simulated under a narrow parameter range, Figure 2 and Figure 3 The horizontal axis in the table is the gate-source voltage, in units of V , the vertical axis is the drain current, the unit is uA , Figure 2 The wide parameter range refers to the model parameter combination that is not optimized over the parameter range. Figure 3 The narrow parameter range is a 20% change in the parameter value corresponding to each model parameter in the updated model parameter combination. The line connected by the circles in the figure is used as a curve, and the target data of the fin field effect transistor with a gate length of 18nm and a fin number of 4 is simulated. Figure 2 and Figure 3 The upper curves are all for drain voltage of 0.05 V The linear region curve of the lower curve is when the drain voltage is 0.9 V The saturation curve. Figure 2 The multiple vertical lines in the graph represent the range of target data covered by the wide parameter range simulation without parameter range optimization. It can be clearly seen that the curve covers a large range and it is not easy to highlight the characteristics of the target data. Figure 3 The multiple vertical lines in the figure represent the target data coverage range corresponding to the narrow parameter range simulation after parameter range optimization, and the IV characteristics can be clearly seen. Figure 2 and Figure 3 The drain voltage is 0.05 V The drain current is low over the entire gate voltage range when Figure 2 and Figure 3 The drain voltage is 0.9 V , the drain current is higher over the entire gate voltage range and the saturation current is more pronounced.

[0115] S109, inputting the target data into the extractor of the trained neural network, so that the extractor of the trained neural network outputs the target model parameter combination.

[0116] The trained neural network extractor is an extractor obtained by inputting the first target data set and the first model parameter combination into the initial neural network extractor to train the initial neural network extractor.

[0117] The extractor of the trained neural network is an extractor obtained by training the extractor of the initial neural network. The extractor of the trained neural network has the same architecture as the extractor of the initial neural network.

[0118] The gated neural network in step S106 has the same architecture as the extractor of the trained neural network, and the gated neural network also includes multiple expert networks, multiple gated networks and multiple tower networks. Each expert network is connected to multiple gated networks respectively, and the number of multiple gated networks and multiple tower networks is the same, and one gated network corresponds to one tower network, and each gated network is connected to the corresponding tower network. Each expert network includes 3 fully connected layers, and each tower network includes 3 fully connected layers. The expert network is used to connect the input and extract the features of each input. Each expert network focuses on processing a specific type of data. The tower network is used to extract the output features of the gated network, and a hierarchical feature representation is constructed based on the output features to improve the model's ability to process complex data. The structural parameters of the model of the gated neural network and the extractor of the trained neural network are different, that is, the learning rate of the hidden layer neurons is different. The training data of the gated neural network and the extractor of the trained neural network are different. The training data of the gated neural network is data without reducing the parameter range, and the training data of the extractor of the trained neural network is data with reduced parameter range. The data with reduced parameter range is the data set generated in step S108.

[0119] Since the current model parameter combination is historical data, a model parameter range is generated according to the current model parameter combination and a preset range, and a Monte Carlo simulation is performed within the model parameter range to generate a data set, which is a collection of historical data.

[0120] The first target data set and the first model parameter combination are used as training data, and the training data includes historical data. The first target data set including historical data and the first model parameter combination are input into the extractor of the initial neural network to obtain the extractor of the trained neural network. The target data is input into the extractor of the trained neural network so that the extractor of the trained neural network outputs the target model parameter combination, and the target model parameter combination is the optimal model parameter combination.

[0121] The target model parameter combination is used to perform circuit simulation.

[0122] The following Table 1 shows the errors of electrical performance indicators. The first column of Table 1 is without reduced parameter range (WithoutShrink, W / O Shrink) and reduced parameter range (With Shrink, W / Shrink). The reduced parameter range specifically refers to generating a model parameter combination range according to the current model parameter combination and the preset range when it is judged that the first error is less than the preset error. The generated model parameter combination range is a narrow range. is the drain current, is the threshold voltage, is the maximum transconductance, is the saturation leakage current. It can be seen that the electrical target error of the multilayer perceptron neural network with reduced parameter range is lower than that of the multilayer perceptron neural network without reduced parameter range; the electrical target error of the gated neural network with reduced parameter range is lower than that of the gated neural network without reduced parameter range.

[0123] Table 1 Errors of electrical performance indicators

[0124]

[0125] The following Table 2 shows the maximum mean absolute error of the model parameters during the training process. Table 2 contains the mean absolute error (MAE) and model parameters. The model parameters include the gate work function PHIG, the interface trap parameter CIT, and the SiO 2 Equivalent gate dielectric thickness EOT, saturation velocity VSAT in saturation region, long channel saturation drain voltage parameter KSATIV, smoothing function factor MEXP of saturation drain voltage, zero bias source to drain extension resistance RDSW per unit width, low field mobility U0, and related parameters of drain induced barrier lowering effect PDIBL2. MAE is the mean of the current model parameters, and the error index of each parameter extraction is much lower than 1%, which proves the reliability and effectiveness of the method of the present invention, indicating that the automatically generated model parameter range and the extractor of the trained neural network provide an efficient and accurate way for semiconductor device parameter extraction.

[0126] Table 2 Maximum mean absolute error of model parameters during training

[0127]

[0128] Table 3 below shows the normalized gradient of the electrical target under each model parameter, where the model parameters include gate work function PHIG, interface trap parameter CIT, SiO 2 Equivalent gate dielectric thickness EOT, saturation velocity VSAT in the saturation region, long channel saturation drain voltage parameter KSATIV, smoothing function factor MEXP of saturation drain voltage, zero bias source to drain extension resistance RDSW per unit width, low field mobility U0, and related parameters of drain induced barrier lowering effect PDIBL2; is the maximum transconductance, is the linear region transconductance, is the threshold voltage in the linear region, is the saturation leakage current, is the cut-off leakage current, is the subthreshold slope, is the sum of the drain currents, is the sum of the logarithms of the drain current. It can be seen that the subthreshold slope Mainly affected by the interface trap parameter CIT, the interface trap parameter CIT can be adjusted to control the subthreshold slope changes to achieve the required performance indicators.

[0129] Table 3 Normalized gradient of electrical target under various model parameters

[0130]

[0131] Based on the above Figure 1 It can be seen from the implementation method that the embodiment of the present invention obtains target data of the semiconductor device; inputs the target data into the extractor of the initial neural network so that the extractor of the initial neural network outputs the current model parameter combination of the semiconductor device; inputs the current model parameter combination into the simulator, simulates the current target data of the current model parameter combination, and calculates multiple first electrical targets corresponding to the current target data; calculates the first error according to the multiple first electrical targets and the actual electrical targets; determines whether the first error is less than the preset error; if not, determines the probability factor according to the gradient, scale factor and first error of the neural network; according to the probability factor, adjusts the current model parameter combination to obtain an updated model parameter combination, uses the updated model parameter combination as the current model parameter combination, and returns to the step of inputting the current model parameter combination into the simulator, simulating the current target data of the current model parameter combination, and calculating multiple first electrical targets corresponding to the current target data; if yes, generates a model parameter range according to the current model parameter combination and the preset range, and performs Monte Carlo simulation within the model parameter range to generate a data set; inputs the target data into the extractor of the trained neural network so that the extractor of the trained neural network outputs the target model parameter combination. In this way, according to the probability factor, the current model parameter combination is adjusted to obtain an updated model parameter combination, and the updated model parameter combination can be automatically determined. Then, when the first error is less than the preset error, the model parameter range is generated according to the current model parameter combination and the preset range, and the model parameter range can be automatically determined, thereby improving the efficiency of determining the model parameter range, thereby improving the efficiency of semiconductor device parameter extraction; when the first error is not less than the preset error, the probability factor is determined according to the gradient of the neural network, the proportional factor and the first error, and the current model parameter combination can be adjusted according to the probability factor, so that the updated model parameter combination is more accurate, thereby making the accuracy of semiconductor device parameter extraction higher.

[0132] Based on the same inventive concept, as an implementation of the above-mentioned method for intelligently extracting parameters of a semiconductor device, an embodiment of the present invention further provides an apparatus for intelligently extracting parameters of a semiconductor device. Figure 4 FIG. 1 is a structural diagram of a semiconductor device parameter intelligent extraction device in an embodiment of the present invention, see Figure 4 As shown, the semiconductor device parameter intelligent extraction device may include:

[0133] An acquisition module 401 is used to acquire target data of a semiconductor device, where the target data is a relationship curve between terminal characteristics of the semiconductor device;

[0134] A current model parameter combination extraction module 402 is used to input the target data into an extractor of the initial neural network so that the extractor of the initial neural network outputs a current model parameter combination of the semiconductor device, and the extractor of the initial neural network is an extractor trained using the preset target data and the preset model parameter combination corresponding to the preset target data;

[0135] The simulation module 403 is used to input the current model parameter combination into the simulator, simulate the current target data of the current model parameter combination, and calculate a plurality of first electrical targets corresponding to the current target data;

[0136] An error calculation module 404 is used to calculate a first error according to a plurality of first electrical targets and an actual electrical target, where the actual electrical target is an electrical target calculated according to the target data;

[0137] A judging module 405 is used to judge whether the first error is less than a preset error;

[0138] A determination module 406 is used to determine a probability factor according to the gradient of the neural network, the scale factor and the first error if no;

[0139] An adjustment module 407 is used to adjust the current model parameter combination according to the probability factor to obtain an updated model parameter combination, use the updated model parameter combination as the current model parameter combination, and return the step of inputting the current model parameter combination into the simulator, simulating the current target data of the current model parameter combination, and calculating a plurality of first electrical targets corresponding to the current target data;

[0140] A generating module 408 is used to generate a model parameter range according to the current model parameter combination and the preset range, and perform Monte Carlo simulation within the model parameter range to generate a data set, wherein the data set includes a first target data set and a corresponding first model parameter combination;

[0141] The target model parameter combination extraction module 409 is used to input the target data into the extractor of the trained neural network so that the extractor of the trained neural network outputs the target model parameter combination. The extractor of the trained neural network is an extractor obtained by inputting the first target data and the first model parameter combination into the extractor of the initial neural network to train the extractor of the initial neural network.

[0142] The multiple first electrical targets in the simulation module 403 include the current drain current, the current maximum transconductance, the current linear region transconductance, the threshold voltage of the current linear region, the current saturation drain current, the current cutoff drain current and the current subthreshold slope, and the actual electrical targets include the actual drain current, the actual maximum transconductance, the actual linear region transconductance, the threshold voltage of the actual linear region, the actual saturation drain current, the actual cutoff drain current and the actual subthreshold slope.

[0143] The error calculation module 404 is specifically used to calculate the first error according to the first drain current, the first maximum transconductance, the first linear region transconductance, the threshold voltage of the first linear region, the first saturation drain current, the first cutoff drain current and the first subthreshold slope, and the actual drain current, the actual maximum transconductance, the actual linear region transconductance, the threshold voltage of the actual linear region, the actual saturation drain current, the actual cutoff drain current and the actual subthreshold slope.

[0144] The error calculation module 404 calculates a first error according to the first drain current, the first maximum transconductance, the first linear region transconductance, the threshold voltage of the first linear region, the first saturated drain current, the first cut-off drain current and the first subthreshold slope, and the actual drain current, the actual maximum transconductance, the actual linear region transconductance, the threshold voltage of the actual linear region, the actual saturated drain current, the actual cut-off drain current and the actual subthreshold slope, including: calculating the first error according to the first drain current, the first maximum transconductance, the first linear region transconductance, the threshold voltage of the first linear region, the first saturated drain current, the first cut-off drain current and the first subthreshold slope, and the actual drain current, the actual maximum transconductance, the actual linear region transconductance, the threshold voltage of the actual linear region, the actual saturated drain current, the actual cut-off drain current and the actual subthreshold slope using the following first formula:

[0145] ;

[0146] in, is the first error, is the first drain current, is the RMS difference of the first drain current, is the first maximum transconductance, is the actual maximum transconductance, is the transconductance in the first linear region, is the actual linear region transconductance, is the threshold voltage of the first linear region, is the threshold voltage in the actual linear region, is the first saturation leakage current, is the actual saturation leakage current, is the first cut-off leakage current, is the actual cut-off leakage current, is the first subthreshold slope, is the actual subthreshold slope, is the weight factor, , For the The first electrical target.

[0147] The determination module 406 is specifically configured to determine a first transformation equation according to the gradient and the first proportional factor; and determine a probability factor according to the first transformation equation, the second proportional factor and the first error, wherein the proportional factor includes a first proportional factor and a second proportional factor.

[0148] The determination module 406 determines the first transformation equation according to the gradient and the first proportional factor, including: determining the first transformation equation according to the gradient and the first proportional factor using the following second formula:

[0149] ;

[0150] in, is the first transformation equation, is the gradient, is the first scaling factor, For the The first electrical target, is the total number of the first electrical targets, is the first combination of current model parameters The current model parameters.

[0151] The determination module 406 determines the probability factor according to the first transformation equation, the second proportional factor and the first error, including: determining the probability factor according to the first transformation equation, the second proportional factor and the first error using the following third formula:

[0152] ;

[0153] ;

[0154] in, is the probability factor, is the second scale factor, is the second transformation equation, is the first error, is the first transformation equation, is the total number of current model parameters of the current model parameter combination, For the The first electrical target, is the total number of the first electrical targets, is the first combination of current model parameters The current model parameters.

[0155] The adjustment module 407 adjusts the current model parameter combination according to the probability factor to obtain an updated model parameter combination, including: comparing each current model parameter in the current model parameter combination with the probability factor to determine a first current model parameter combination and a second current model parameter combination, the first current model parameter combination being a combination in which all first current model parameters in the first current model parameter combination are greater than the probability factor, and the second current model parameter combination being a combination in which all second current model parameters in the second current model parameter combination are not greater than the probability factor; using the first current model parameter combination as the updated model parameter combination; performing cross-mutation processing on the second current model parameter combination to obtain an updated second model parameter combination, and using the updated second model parameter combination as the updated model parameter combination.

[0156] The current model parameter combination in the current model parameter combination extraction module 402 includes: gate work function, interface trap parameter, SiO 2 Equivalent gate dielectric thickness, zero bias source to drain extension resistance per unit width, low field mobility, saturation velocity in the saturation region, long channel saturation drain voltage parameter, smoothing function factor of saturation drain voltage, and parameters related to drain induced barrier lowering effect.

[0157] It should be pointed out here that the above description of the embodiment of the semiconductor device parameter intelligent extraction device is similar to the description of the embodiment of the semiconductor device parameter intelligent extraction method, and has similar beneficial effects as the embodiment of the semiconductor device parameter intelligent extraction method. For technical details not disclosed in the embodiment of the semiconductor device parameter intelligent extraction device of the embodiment of the present invention, please refer to the description of the embodiment of the semiconductor device parameter intelligent extraction method of the present invention for understanding.

[0158] The above are only specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for intelligently extracting semiconductor device parameters, characterized in that: The semiconductor device parameter intelligent extraction method comprises: Acquiring target data of a semiconductor device, wherein the target data is a relationship curve between terminal characteristics of the semiconductor device; Inputting the target data into an extractor of an initial neural network so that the extractor of the initial neural network outputs a current model parameter combination of the semiconductor device, wherein the extractor of the initial neural network is an extractor trained using preset target data and a preset model parameter combination corresponding to the preset target data; Inputting the current model parameter combination into a simulator, simulating current target data of the current model parameter combination, and calculating a plurality of first electrical targets corresponding to the current target data; Calculating a first error according to the plurality of first electrical targets and an actual electrical target, wherein the actual electrical target is an electrical target calculated according to the target data; Determining whether the first error is less than a preset error; If not, determining a probability factor according to the gradient of the neural network, a scaling factor and the first error, the scaling factor comprising a first scaling factor and a second scaling factor, the first scaling factor and the second scaling factor being different; According to the probability factor, the current model parameter combination is adjusted to obtain an updated model parameter combination, the updated model parameter combination is used as the current model parameter combination, and the current model parameter combination is input into the simulator to simulate the current target data of the current model parameter combination, and a plurality of first electrical targets corresponding to the current target data are calculated; If yes, generating a model parameter range according to the current model parameter combination and the preset range, and performing Monte Carlo simulation within the model parameter range to generate a data set, the data set including a first target data set and a corresponding first model parameter combination; Inputting the target data into an extractor of a trained neural network so that the extractor of the trained neural network outputs a target model parameter combination, wherein the extractor of the trained neural network is an extractor obtained by inputting the first target data set and the first model parameter combination into the extractor of the initial neural network to train the extractor of the initial neural network; Determining the probability factor according to the gradient of the neural network, the scaling factor and the first error includes: Determining a first transformation equation according to the gradient and the first proportional factor; determining the probability factor according to the first transformation equation, the second proportional factor and the first error; Determining a first transformation equation according to the gradient and the first proportional factor includes: The first transformation equation is determined according to the gradient and the first proportional factor using the following second formula: in, is the first transformation equation, is the gradient, α is the first proportional factor, j is the jth first electrical target, q is the total number of the first electrical targets, and k is the kth current model parameter of the current model parameter combination; The determining of the probability factor according to the first transformation equation, the second proportional factor and the first error comprises: The probability factor is determined according to the first transformation equation, the second proportional factor and the first error using the following third formula: Wherein, ρk is the probability factor, β is the second proportional factor, is the second transformation equation, Ej is the first error, is the first transformation equation, m is the total number of current model parameters of the current model parameter combination, j is the jth first electrical target, q is the total number of the first electrical targets, and k is the kth current model parameter of the current model parameter combination.

2. The method for intelligently extracting semiconductor device parameters according to claim 1, characterized in that: The multiple first electrical targets include a first drain current, a first maximum transconductance, a first linear region transconductance, a threshold voltage of the first linear region, a first saturation drain current, a first cutoff drain current and a first subthreshold slope, and the actual electrical targets include an actual drain current, an actual maximum transconductance, an actual linear region transconductance, an actual linear region threshold voltage, an actual saturation drain current, an actual cutoff drain current and an actual subthreshold slope.

3. The method for intelligently extracting semiconductor device parameters according to claim 2, characterized in that: The calculating the first error according to the plurality of first electrical targets and the actual electrical target comprises: The first error is calculated based on the first drain current, the first maximum transconductance, the first linear region transconductance, the threshold voltage of the first linear region, the first saturated drain current, the first cut-off drain current and the first subthreshold slope, and the actual drain current, the actual maximum transconductance, the actual linear region transconductance, the threshold voltage of the actual linear region, the actual saturated drain current, the actual cut-off drain current and the actual subthreshold slope.

4. The method for intelligently extracting semiconductor device parameters according to claim 3, characterized in that: The calculating the first error according to the first drain current, the first maximum transconductance, the first linear region transconductance, the threshold voltage of the first linear region, the first saturated drain current, the first cut-off drain current and the first subthreshold slope, and the actual drain current, the actual maximum transconductance, the actual linear region transconductance, the threshold voltage of the actual linear region, the actual saturated drain current, the actual cut-off drain current and the actual subthreshold slope, comprises: The first error is calculated using the following first formula according to the first drain current, the first maximum transconductance, the first linear region transconductance, the threshold voltage of the first linear region, the first saturated drain current, the first cut-off drain current and the first subthreshold slope, and the actual drain current, the actual maximum transconductance, the actual linear region transconductance, the threshold voltage of the actual linear region, the actual saturated drain current, the actual cut-off drain current and the actual subthreshold slope: Among them, Fitness is the first error, Ids is the first drain current, RMSE (Ids) is the root mean square error of the first drain current, Gmpeak1 is the first maximum transconductance, Gmpeak2 is the actual maximum transconductance, Gmlin1 is the first linear region transconductance, Gmlin2 is the actual linear region transconductance, Vthlin1 is the threshold voltage of the first linear region, Vthlin2 is the threshold voltage of the actual linear region, Idsat1 is the first saturation leakage current, Idsat2 is the actual saturation leakage current, Ioff1 is the first off-state leakage current, Ioff2 is the actual off-state leakage current, SS1 is the first subthreshold slope, SS2 is the actual subthreshold slope, Ij is the weight factor, j=1,…,8, j is the jth first electrical target.

5. The method for intelligently extracting semiconductor device parameters according to claim 1, characterized in that: The step of adjusting the current model parameter combination according to the probability factor to obtain an updated model parameter combination includes: Compare each current model parameter in the current model parameter combination with the probability factor to determine a first current model parameter combination and a second current model parameter combination, wherein the first current model parameter combination is a combination in which all first current model parameters in the first current model parameter combination are greater than the probability factor, and the second current model parameter combination is a combination in which all second current model parameters in the second current model parameter combination are not greater than the probability factor; using the first current model parameter combination as the updated model parameter combination; The second current model parameter combination is subjected to cross-mutation processing to obtain an updated second model parameter combination, and the updated second model parameter combination is used as the updated model parameter combination.

6. The method for intelligently extracting semiconductor device parameters according to claim 1, characterized in that: The current model parameter combination includes: gate work function, interface trap parameter, SiO2 equivalent gate dielectric thickness, zero bias source to drain extension resistance per unit width, low field mobility, saturation velocity in the saturation region, long channel saturation drain voltage parameter, smoothing function factor of saturation drain voltage and related parameters of drain induced barrier lowering effect.

7. A semiconductor device parameter intelligent extraction device, characterized in that: The semiconductor device parameter intelligent extraction device comprises: An acquisition module, used for acquiring target data of a semiconductor device, wherein the target data is a relationship curve between terminal characteristics of the semiconductor device; a current model parameter combination extraction module, used for inputting the target data into an extractor of an initial neural network so that the extractor of the initial neural network outputs a current model parameter combination of the semiconductor device, wherein the extractor of the initial neural network is an extractor trained using preset target data and a preset model parameter combination corresponding to the preset target data; A simulation module, used for inputting the current model parameter combination into a simulator, simulating current target data of the current model parameter combination, and calculating a plurality of first electrical targets corresponding to the current target data; an error calculation module, configured to calculate a first error according to the plurality of first electrical targets and an actual electrical target, wherein the actual electrical target is an electrical target calculated according to the target data; A judging module, used for judging whether the first error is less than a preset error; a determination module, configured to, if not, determine a probability factor according to a gradient of the neural network, a scaling factor and the first error, wherein the scaling factor includes a first scaling factor and a second scaling factor, and the first scaling factor and the second scaling factor are different; An adjustment module, used for adjusting the current model parameter combination according to the probability factor to obtain an updated model parameter combination, using the updated model parameter combination as the current model parameter combination, and returning the step of inputting the current model parameter combination into a simulator, simulating current target data of the current model parameter combination, and calculating a plurality of first electrical targets corresponding to the current target data; a generating module, configured to generate a model parameter range according to the current model parameter combination and the preset range, and perform Monte Carlo simulation within the model parameter range to generate a data set, wherein the data set includes a first target data set and a corresponding first model parameter combination; a target model parameter combination extraction module, used for inputting the target data into an extractor of a trained neural network so that the extractor of the trained neural network outputs a target model parameter combination, wherein the extractor of the trained neural network is an extractor obtained by inputting the first target data set and the first model parameter combination into the extractor of the initial neural network to train the extractor of the initial neural network; The determination module is specifically configured to determine a first transformation equation according to the gradient and the first proportional factor; and determine the probability factor according to the first transformation equation, the second proportional factor and the first error; The determining module, determining the first transformation equation according to the gradient and the first proportional factor, includes: determining the first transformation equation according to the gradient and the first proportional factor using the following second formula: in, is the first transformation equation, is the gradient, α is the first proportional factor, j is the jth first electrical target, q is the total number of the first electrical targets, and k is the kth current model parameter of the current model parameter combination; The determining module, determining the probability factor according to the first transformation equation, the second proportional factor and the first error, includes: determining the probability factor according to the first transformation equation, the second proportional factor and the first error using the following third formula: Wherein, ρk is the probability factor, β is the second proportional factor, is the second transformation equation, Ej is the first error, is the first transformation equation, m is the total number of current model parameters of the current model parameter combination, j is the jth first electrical target, q is the total number of the first electrical targets, and k is the kth current model parameter of the current model parameter combination.

Images