Method and device for extracting model parameters of semiconductor device

Abstract

The invention provides a model parameter extraction method and device for a semiconductor device, and the method comprises the steps: determining a device model parameter corresponding to a target measurement current through a successive approximation method in the process of extracting the model parameter of the semiconductor device, and enabling the value range of the parameter to be successive approximation in the process of determining the device parameter, compared with a parameter determination method of a single model, the method has the advantages that the model parameter extraction precision of the semiconductor device is higher, and in the parameter value interval and model parameter determination process, under the condition that the output parameter precision requirement of the convolutional neural network is not high, the data size of the used training sample is small, the training speed is high, and the training efficiency is high. Parameters used by the Gaussian regression model, the random forest model and the greedy algorithm are all determined in a gradually-reduced parameter value interval, the obtained data size is small, the calculation speed is high, and the parameter determination speed is guaranteed.

Description

Technical Field

[0001] The present application relates to the field of semiconductor device parameter extraction, and in particular to a method and device for extracting model parameters of a semiconductor device. Background Art

[0002] With the continuous breakthroughs in semiconductor industry technology nodes and the rapid development of computer science and technology, establishing accurate equivalent circuit models of semiconductor devices has become the key to ensuring the success of semiconductor circuit design and the reliability of simulation.

[0003] The extraction of model parameters of semiconductor devices is the key to establishing an equivalent circuit model, which is determined based on the physical principles of the device and the actual measurement results. The simulation software can simulate the measured electrical characteristics of the device based on the device model. With the development of semiconductor technology, semiconductor devices have entered the nano era, and various effects in the devices have become more and more complex. Taking the metal oxide field effect transistor as an example, the number of parameters exceeds 300. It usually takes a lot of time to adjust a complete set of model parameter cards, and the fine-tuning of the manufacturing process requires the extraction of parameters again, resulting in a high time cost for the extraction of model parameters of semiconductor devices.

[0004] In the prior art, model parameters are extracted mainly through intelligent parameter extraction algorithms to shorten the time of model parameter extraction. Intelligent parameter extraction algorithms include CNN and genetic algorithms. For devices with a large number of parameters, CNN requires a large number of training samples to ensure the accuracy of the algorithm, and the genetic algorithm requires multiple calls to the simulation software to determine the parameters one by one by setting the fitting current for the model parameters. There is a technical problem that the extraction accuracy and speed cannot be taken into account at the same time. Summary of the invention

[0005] The present application provides a method and device for extracting model parameters of a semiconductor device, so as to solve the problem of insufficient accuracy and slow speed in extracting model parameters of a semiconductor device.

[0006] In a first aspect, the present application provides a method for extracting model parameters of a semiconductor device, the method comprising:

[0007] Obtaining target measurement current;

[0008] Obtaining a first confidence interval of a device model parameter according to the target measured current and the trained convolutional neural network model;

[0009] Based on the multiple groups of parameters in the first confidence interval and the preset device model, the first model and the second model are trained respectively to obtain a Gaussian process regression model and a random forest model, and the first confidence interval is updated according to the target measurement current, the Gaussian process regression model and the random forest model to obtain a parameter sampling value interval of the preset device model;

[0010] In the parameter sampling value range, a greedy algorithm is used to determine the device model parameters for generating the target measurement current.

[0011] In the above technical solution, during the extraction of the model parameters of a semiconductor device, the device model parameters corresponding to the target measurement current are determined by the method of successive approximation. Since during the process of determining the device parameters, the value range of the parameters is successively approximated. Compared with the parameter determination method of a single model, the extraction accuracy of the model parameters of the semiconductor device is higher. And during the process of the parameter value range and the determination of the model parameters, when the accuracy requirement for the output parameters of the convolutional neural network is not high, the amount of data of the training samples used is small and the training speed is fast. The parameters used by the Gaussian regression model, the random forest model, and the greedy algorithm are all determined in the gradually shrinking parameter value range, and the amount of data obtained is small and the calculation speed is fast to ensure the speed of parameter determination.

[0012] Optionally, obtaining a first confidence interval of the model parameters according to the target measurement current and the trained convolutional neural network model specifically includes:

[0013] Input the target measurement current into the convolutional neural network model to obtain a first mean value and a first variance of the device model parameters;

[0014] According to the first mean value and the first variance, obtain a first confidence interval of the device model parameters.

[0015] Optionally, before obtaining a first confidence interval of the device model parameters according to the target measurement current and the trained convolutional neural network model, the method further includes:

[0016] When a semiconductor device is loaded with multiple groups of measurement voltage values, obtain the corresponding drain measurement current; wherein, each group of measurement voltage values includes a gate voltage and a drain voltage;

[0017] For each group of measurement voltage values, based on the drain measurement current corresponding to the measurement voltage values, adjust the parameters of the device model to determine the device model parameters corresponding to the semiconductor device; wherein, the current fitted by the device model when setting the device model parameters is the drain measurement current corresponding to the measurement voltage values;

[0018] Based on each group of the drain measurement currents, the corresponding measurement voltage values, and the corresponding device model parameters, construct a training set;

[0019] Use the training set to train the original convolutional neural network to obtain the trained convolutional neural network; the loss function of the convolutional neural network includes:

[0020]

[0021] Among them, i represents the identifier of each parameter in the parameter set, D represents the number of parameters, and θ i represents the i-th parameter obtained by the convolutional neural network when inputting the preset drain current, represents the i-th preset parameter corresponding to the preset drain current in the training set, and σ i represents the variance of the i-th parameter.

[0022] In the above technical solution, during the training process of the convolutional neural network, a training set with richer corresponding relationships is created by using the corresponding relationships among the drain voltage, gate voltage, and drain current. During the process of training the convolutional neural network using this training set, it can learn the structured information of the device characteristics, improving the accuracy of the convolutional neural network and the diversity of input parameters.

[0023] Optionally, training a first model and a second model respectively based on multiple groups of parameters in the first confidence interval and a preset device model to obtain a Gaussian process regression model and a random forest model, and updating the first confidence interval according to the target measurement current, the Gaussian process regression model, and the random forest model to obtain the parameter sampling value range of the preset device model, including:

[0024] When the number of updates of the first confidence interval is less than the first preset target number, training a first model and a second model respectively based on multiple groups of randomly sampled parameters in the first confidence interval and a preset device model to obtain a Gaussian process regression model and a random forest model;

[0025] Inputting the target measurement current into the Gaussian process regression model to obtain the Gaussian process confidence interval of the device model parameters;

[0026] Inputting the target measurement current into the random forest model to obtain the random forest confidence interval of the device model parameters;

[0027] Updating the first confidence interval based on the Gaussian process confidence interval and the random forest confidence interval until the number of updates of the first confidence interval is equal to the first preset target number;

[0028] Determining the last updated first confidence interval as the parameter sampling value range of the preset device model.

[0029] Optionally, the training a first model and a second model respectively based on multiple groups of randomly sampled parameters in the first confidence interval and a preset device model to obtain a Gaussian process regression model and a random forest model specifically includes:

[0030] Randomly sample within the first confidence interval to obtain multiple sets of sampling parameter sets; each set of sampling parameter sets includes multiple parameters of the device model;

[0031] Set the multiple parameters of each set of sampling parameter sets to the device model, and after inputting multiple sets of preset voltage values to the device model, obtain the simulated current corresponding to each set of preset voltage values;

[0032] Construct a sampling training set by using each set of sampling parameter sets and the corresponding simulated current;

[0033] Train the first model by using the sampling training set to obtain a Gaussian process regression model;

[0034] Train the second model by using the sampling training set to obtain a random forest model.

[0035] Optionally, the updating of the first confidence interval based on the Gaussian process confidence interval and the random forest confidence interval specifically includes:

[0036] When the Gaussian process confidence interval and the random forest confidence interval have an intersection, update the first confidence interval by using the intersection;

[0037] When the Gaussian process confidence interval and the random forest confidence interval do not have an intersection, update the first confidence interval by using the union of the Gaussian process confidence interval and the random forest confidence interval.

[0038] In the above technical solution, the random forest model assisting the Gaussian regression model to update the first confidence interval can enable the Gaussian regression model to converge quickly even when the training data is insufficient, so as to output the confidence interval.

[0039] Optionally, the inputting of the target measured current into the Gaussian process regression model to obtain the Gaussian process confidence interval of the device model parameters specifically includes:

[0040] Input the target measured current into the Gaussian process regression model to obtain the second mean and the second variance of the device model parameters;

[0041] Determine the Gaussian process confidence interval of the device model parameters according to the second mean, the second variance and the confidence interval calculation formula.

[0042] Optionally, the random forest model includes multiple decision trees;

[0043] The inputting of the target measured current into the random forest model to obtain the random forest confidence interval of the device model parameters specifically includes:

[0044] Input the target measured current into the random forest model to obtain the third mean value of the device model parameters;

[0045] Determine the mean square error of each decision tree as the third variance of the device model parameters;

[0046] Determine the Gaussian process confidence interval of the device model parameters according to the third mean value, the third variance and the confidence interval calculation formula.

[0047] Optionally, in the parameter sampling value range, using the greedy algorithm to determine the device model parameters that generate the target measured current specifically includes:

[0048] Randomly sample a set of random parameter sets in the parameter sampling value range; the random parameter sets include the initial random parameters corresponding to multiple variables of the device model, the parameter sampling value range includes the sampling sub-ranges of each variable, and each initial random parameter is located in the corresponding sampling sub-range;

[0049] Repeat, according to the arrangement order of the random parameters in the random parameter sets, sequentially generate multiple sets of random parameter sets corresponding to each variable; among them, the parameters corresponding to this variable in the multiple sets of random parameter sets corresponding to each variable are randomly selected parameters in the sampling sub-range of this variable, and the parameters corresponding to other variables are the same in each set of random parameter sets;

[0050] For each variable, set the multiple random parameters of the corresponding random parameter sets to the device model, input a preset voltage into the device model, and obtain the predicted current corresponding to the preset voltage;

[0051] Determine the simulation error according to the predicted current and the measured current corresponding to the preset voltage, update the sampling sub-range corresponding to the variable according to the simulation error, and adjust the random parameter sets until the adjustment times of the sampling sub-ranges corresponding to each variable in the parameter sampling value range are all the second preset target times, and determine the last adjusted random parameter set as the device model parameters of the target measured current.

[0052] Optionally, the determining the simulation error according to the predicted current and the measured current corresponding to the preset voltage, updating the sampling sub-range corresponding to the variable according to the simulation error, and adjusting the random parameter sets specifically includes:

[0053] Determine the relative error between the predicted current and the measured current as the simulation error of the variable;

[0054] Select the parameter with the smallest simulation error from the multiple simulation errors corresponding to the variable to update the random parameter sets;

[0055] Update the sampling sub - interval according to the maximum and minimum values of the sampling sub - interval of the variable and the updated parameters of the variable.

[0056] Optionally, the step of updating the sampling sub - interval according to the maximum and minimum values of the sampling sub - interval of the variable and the updated parameters of the variable specifically includes:

[0057] Update the sampling sub - interval according to the maximum and minimum values of the sampling sub - interval of the variable, the updated parameters of the variable, and an interval update formula;

[0058] The interval update formula includes:

[0059]

[0060] where θ i represents the i - th parameter after updating of the variable in the set of random parameters, max i represents the maximum value of the sampling sub - interval of the i - th parameter, min i represents the minimum value of the sampling sub - interval of the i - th parameter, and k is a positive number.

[0061] In the above - mentioned technical solution, when using the greedy algorithm to determine the device model parameters, within the sampling sub - interval corresponding to each device variable, by sampling, fitting and comparing with the actual measurement, an analog error is obtained. Based on the analog error, the parameter distribution interval is narrowed by modifying each parameter of the device model one by one to determine the accurate parameters, ensuring the extraction accuracy of the model parameters of the semiconductor device.

[0062] In a second aspect, the present application provides a device model parameter extraction device for a semiconductor device, including:

[0063] An acquisition module, configured to obtain a target measurement current;

[0064] A processing module, configured to obtain a first confidence interval of the device model parameters according to the target measurement current and a trained convolutional neural network model;

[0065] The processing module is further configured to train a first model and a second model respectively based on multiple groups of parameters in the first confidence interval and a preset device model to obtain a Gaussian process regression model and a random forest model, and update the first confidence interval according to the target measurement current, the Gaussian process regression model, and the random forest model to obtain a parameter sampling value interval of the preset device model;

[0066] The processing module is further configured to determine the device model parameters that generate the target measurement current by using a greedy algorithm within the parameter sampling value interval.

[0067] The present application provides a method and apparatus for extracting model parameters of a semiconductor device. An electronic device obtains a target measured current measured when the semiconductor device obtains a preset voltage, and transmits the target measured current to a trained convolutional neural network to obtain a first confidence interval of the device model parameters. Within the first confidence interval, through multiple Gaussian process regressions and random forest regressions, the first confidence interval is gradually narrowed to determine a parameter sampling value range, and a greedy algorithm is used within the parameter sampling value range. In the process of extracting the model parameters of the semiconductor device, through a successive approximation method, the device model parameters corresponding to the target measured current are determined. Since in the process of determining the device parameters, the value range of the parameters is successively approximated, compared with the parameter determination method of a single model, the extraction accuracy of the model parameters of the semiconductor device is higher. And in the process of the parameter sampling value range and the determination of the model parameters, when the accuracy requirement of the output parameters of the convolutional neural network is not high, the amount of training sample data used is small and the training speed is fast. The parameters used in the Gaussian regression model, random forest model, and greedy algorithm are all determined in a gradually shrinking parameter value range, and the amount of data obtained is small and the calculation speed is fast to ensure the speed of parameter determination. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] The accompanying drawings herein are incorporated into the specification and form a part of the specification, showing embodiments consistent with the present application, and are used together with the specification to explain the principles of the present application.

[0069] Figure 1 FIG. is an application scenario diagram of a method for extracting model parameters of a semiconductor device provided according to an exemplary embodiment of the present application;

[0070] Figure 2 FIG. is a flowchart of a method for extracting model parameters of a semiconductor device provided according to an exemplary embodiment of the present application;

[0071] Figure 3 FIG. is a flowchart of a method for training a convolutional neural network provided according to an exemplary embodiment of the present application;

[0072] Figure 4 FIG. is a structural diagram of a convolutional neural network provided according to an exemplary embodiment of the present application;

[0073] Figure 5 FIG. is a flowchart of a method for updating a first confidence interval provided according to an exemplary embodiment of the present application;

[0074] Figure 6 FIG. is a flowchart of a coordinate descent method provided according to an exemplary embodiment of the present application;

[0075] Figure 7A schematic diagram of the structure of a model parameter extraction device for a semiconductor device provided by the present application according to an embodiment;

[0076] Figure 8 A schematic diagram of the structure of an electronic device provided according to an embodiment of the present application.

[0077] The above drawings have shown clear embodiments of the present application, which will be described in more detail later. These drawings and text descriptions are not intended to limit the scope of the present application in any way, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0078] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application. Instead, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0079] With the continuous breakthroughs in semiconductor industry technology nodes and the rapid development of computer science and technology, establishing accurate equivalent circuit models of semiconductor devices has become the key to ensuring the success of semiconductor circuit design and the reliability of simulation.

[0080] The extraction of model parameters of semiconductor devices is the key to establishing an equivalent circuit model, which is determined based on the physical principles of the device and the actual measurement results. The simulation software can simulate the measured electrical characteristics of the device based on the device model. With the development of semiconductor technology, semiconductor devices have entered the nano era, and various effects in the devices have become more and more complex. Taking the metal oxide field effect transistor as an example, the number of parameters exceeds 300. It usually takes a lot of time to adjust a complete set of model parameter cards, and the fine-tuning of the manufacturing process requires the extraction of parameters again, resulting in a high time cost for the extraction of model parameters of semiconductor devices.

[0081] In the prior art, model parameters are extracted mainly through intelligent parameter extraction algorithms to shorten the time of model parameter extraction. Intelligent parameter extraction algorithms include CNN and genetic algorithms. For devices with a large number of parameters, CNN requires a large number of training samples to ensure the accuracy of the algorithm, and the genetic algorithm requires multiple calls to the simulation software to determine the parameters one by one by setting the fitting current for the model parameters. There is a technical problem that the extraction accuracy and speed cannot be taken into account at the same time.

[0082] To solve the above problems, the present application provides a method and device for extracting model parameters of a semiconductor device. The technical concept of the present application is as follows: multiple intelligent models are used to analyze the measured current successively to correspondingly modify the value range of the model parameters corresponding to the measured current until accurate model parameters are obtained, so as to improve the extraction accuracy of the parameters. The data processed by each intelligent model is obtained from the confidence interval, with a small amount of processed data, which can improve the processing speed.

[0083] The following uses specific embodiments to elaborate in detail on the technical solutions of the present application and how the technical solutions of the present application solve the above technical problems. These several specific embodiments below can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below with reference to the accompanying drawings.

[0084] Figure 1 FIG. is an application scenario diagram of a method for extracting model parameters of a semiconductor device provided by the present application according to an exemplary embodiment. In Figure 1 the application scenario shown, it includes a semiconductor device 10, a current measurement device 20, a voltage measurement device 30, and an electronic device 40.

[0085] When the semiconductor device 10 is powered on, the current measurement device 20 and the voltage measurement device 30 measure the electrical signals of each pin of the semiconductor device 10, and the electronic device 40 obtains relevant data of the electrical signals: voltage data and current data, and uses the voltage data and current data to determine the parameters of the semiconductor device model in the simulation software, so that the semiconductor device model can simulate the electrical characteristics of the semiconductor device 10.

[0086] Figure 2 FIG. is a schematic flowchart of a method for extracting model parameters of a semiconductor device provided by the present application according to an exemplary embodiment. As Figure 2 shown, the method includes:

[0087] S101. The electronic device obtains a target measurement current.

[0088] The target measurement current is the current value measured after applying a voltage to the semiconductor device.

[0089] Among them, the voltage applied to the semiconductor device includes a gate voltage or a drain voltage.

[0090] S102. The electronic device obtains a first confidence interval of the device model parameters according to the target measurement current and the trained convolutional neural network model.

[0091] The convolutional neural network model is a model trained with a training set composed of current as the input and device model parameters as the output.

[0092] When the electronic device obtains the first confidence interval, it inputs the target measured current into the trained convolutional neural network model to obtain the first mean and the first variance of the device model parameters. According to the first mean and the first variance, the first confidence interval of the device model parameters is obtained. The first mean is the device model parameters provided by the neural network based on the target measured current, representing the current expectation of the device model parameters; the first variance represents the uncertainty of the device model parameters currently provided by the neural network.

[0093] In one embodiment, the first confidence interval is determined according to the first mean, the first variance and the confidence interval calculation formula. The confidence interval calculation formula includes:

[0094] [θ - 3*σ, θ + 3*σ],

[0095] where θ represents the mean and σ represents the variance.

[0096] S103. The electronic device trains the first model and the second model respectively based on multiple groups of parameters in the first confidence interval and the preset device model to obtain the Gaussian process regression model and the random forest model. According to the target measured current, the Gaussian process regression model and the random forest model, the first confidence interval is updated to obtain the parameter sampling value interval of the preset device model.

[0097] In the process of each update of the first confidence interval, multiple groups of parameters need to be randomly sampled from the first confidence interval, and the parameters obtained by the sampling are used to set the parameters of the preset device model in the simulation software to simulate the current value corresponding to the parameters. The multiple pairs of current values and parameters are used as the training set to train the first model and the second model respectively to obtain the Gaussian process regression model and the random forest model.

[0098] The target measured current obtained in step S101 is input into the trained Gaussian process regression model and random forest model, and the first confidence interval is updated according to the outputs of the two models.

[0099] Through multiple updates of the first confidence interval, the finally determined interval is the parameter sampling value interval of the preset device model.

[0100] S104. The electronic device uses the greedy algorithm to determine the device model parameters that generate the target measured current in the parameter sampling value interval.

[0101] The greedy algorithm is a method of finding the global optimal solution by seeking the local optimal choice in each step of solving the problem. In the process of seeking the local optimal solution, the coordinate descent method can be used. Based on the coordinate descent method to seek the local optimal solution can improve the solution speed and the solution accuracy.

[0102] Using a greedy algorithm to determine the device model parameters for generating the target measurement current means that for each parameter of the device model in the electronic device, the sampling sub-intervals corresponding to each parameter are gradually narrowed within the parameter sampling value range determined in step S103 until the optimal parameter is determined as the device model parameter for the target measurement current.

[0103] In the above technical solution, during the process of extracting the model parameters of a semiconductor device, the device model parameters corresponding to the target measurement current are determined by the method of successive approximation. Since the value range of the parameters is successively approximated during the process of determining the device parameters, compared with the parameter determination method of a single model, the accuracy of extracting the model parameters of the semiconductor device is higher. And during the process of the parameter value range and the determination of the model parameters, when the accuracy requirement for the output parameters of the convolutional neural network is not high, the amount of training sample data used is small and the training speed is fast. The parameters used in the Gaussian regression model, random forest model, and greedy algorithm are all determined within the gradually narrowing parameter value range, and the amount of data obtained is small and the calculation speed is fast to ensure the speed of parameter determination.

[0104] Before step S102, it is also necessary to train the convolutional neural network. Figure 3 This is a schematic flowchart of the training method of the convolutional neural network provided by the present application according to an exemplary embodiment, as Figure 3 shown. The method includes:

[0105] S201. When a semiconductor device is loaded with multiple sets of measured voltage values, obtain the corresponding drain measurement current.

[0106] Among them, according to the terminal setting of the semiconductor device, in one embodiment, each set of measured voltage values includes gate voltage, source voltage, drain voltage, and substrate voltage; in another embodiment, each set of measured voltage values includes gate voltage, source voltage, and drain voltage.

[0107] S202. For each set of measured voltage values, based on the drain measurement current corresponding to the measured voltage values, adjust the parameters of the device model to determine the device model parameters corresponding to the semiconductor device.

[0108] Among them, the current fitted by the device model when setting the device model parameters is the drain measurement current corresponding to the measured voltage values.

[0109] S203. Based on each group of drain measurement currents, the corresponding measured voltage values, and the corresponding device model parameters, construct a training set.

[0110] In the training set, a set of measured voltage values and the corresponding drain measurement current are a set of training data, and the device model parameters corresponding to this set of measured voltage values are the labels of this set of training data.

[0111] S204. Train the original convolutional neural network using the training set to obtain a trained convolutional neural network.

[0112] Among them, the structure of the convolutional neural network is as Figure 4 shown, including the Resnet18 network structure, the feature fusion layer, and the Resnet101 network structure. The Resnet18 network structure includes the first Resnet18 network and the second Resnet18 network.

[0113] The input end of the convolutional neural network can obtain multiple groups of training data and corresponding labels in the training set constructed in step S203. The first Resnet18 network analyzes and extracts features from the change of drain voltage - drain current between multiple groups of training data to obtain the Id - Vd feature; the second Resnet18 network analyzes and extracts features from the change of gate voltage - drain current between multiple groups of training data to obtain the Id - Vg feature. Among them, Id represents the drain current, Vd represents the drain voltage, and Vg represents the gate voltage.

[0114] The feature fusion layer fuses the Id - Vd feature and the Id - Vg feature to obtain multiple features corresponding to each Id, and inputs the fused feature into the Resnet101 network for processing, and outputs the mean and variance that can be determined at the current training level of the network. Among them, the mean is the expectation of the model parameters corresponding to the gate current, and the variance is the uncertainty of the model parameters.

[0115] Calculate the training loss of the network through the preset mean corresponding to the gate current, the mean output by the network, the variance, and the loss function, and use the training loss to adjust the parameters of the network model to promote the convergence of the network training process.

[0116] Among them, the preset mean of the gate current is the parameter corresponding to the drain current in the training set.

[0117] The loss function of the convolutional neural network includes:

[0118]

[0119] Among them, i represents the identifier of each parameter in the parameter set, D represents the number of parameters, θ i represents the i-th parameter obtained by the convolutional neural network when inputting the preset drain current, represents the i-th preset parameter corresponding to the preset drain current in the training set, and σ i represents the variance of the i-th parameter.

[0120] In the above technical solution, during the training of the convolutional neural network, the corresponding relationship between the drain voltage, gate voltage, and drain current is used to create a training set with a richer corresponding relationship. During the training of the convolutional neural network using this training set, the convolutional neural network can learn the structured information of the device characteristics, improving the accuracy of the convolutional neural network and the diversity of input parameters.

[0121] Figure 5 FIG. is a schematic flowchart of a method for updating a first confidence interval according to an exemplary embodiment of the present application. As Figure 5 shown, the method includes:

[0122] S301. When the number of updates of the first confidence interval is less than the first preset target number, train the first model and the second model respectively based on multiple groups of parameters randomly sampled from the first confidence interval and a preset device model to obtain a Gaussian process regression model and a random forest model.

[0123] Randomly sample in the first confidence interval to obtain multiple groups of sampling parameter sets.

[0124] Each group of sampling parameter sets includes multiple parameters of the device model, and each parameter is the data corresponding to each variable in the device model.

[0125] More specifically, the first confidence interval includes multiple first confidence sub-intervals, and each first confidence sub-interval is the parameter value range corresponding to each variable in the device model.

[0126] When obtaining each group of sampling parameter sets, the electronic device randomly samples in the first confidence sub-interval corresponding to each variable to obtain multiple parameters corresponding to each variable, and randomly combines the parameters corresponding to each variable to obtain multiple groups of sampling parameter sets. For example: the device model includes 3 variables, and 2 parameters are sampled in the first confidence sub-interval of each variable, then 8 groups of sampling parameter sets can be formed.

[0127] It should be noted that if this loop is the first loop, the first confidence interval is the first confidence interval determined in step S102; if this loop is not the first loop, the first confidence interval is the first confidence interval updated in step S304 during the previous loop.

[0128] After determining multiple groups of sampling parameter sets, set the multiple parameters of each group of sampling parameter sets in the device model, and after inputting multiple groups of preset voltage values into the device model, obtain the simulated current corresponding to each group of preset voltage values.

[0129] The electronic device calls simulation software, inputs the parameters in the above-obtained groups of sampling parameter sets, and constructs a model corresponding to the semiconductor device.

[0130] Based on the constructed semiconductor model, after setting multiple groups of preset voltage values as input, it will correspondingly output multiple current values fitted by the semiconductor model. This current is the drain current.

[0131] Use each set of sampling parameter sets and the corresponding simulated current to construct a sampling training set.

[0132] Use the sampling training set to train the first model and the second model respectively to obtain a Gaussian process regression model and a random forest model.

[0133] That is to say, use the sampling training set to train the first model to obtain a Gaussian process regression model, and use the sampling training set to train the second model to obtain a random forest model.

[0134] S302: Input the target measured current into the Gaussian process regression model to obtain the Gaussian process confidence interval of the device model parameters.

[0135] S303: Input the target measured current into the random forest model to obtain the random forest confidence interval of the device model parameters.

[0136] When the electronic device determines the Gaussian process confidence interval, it inputs the target measured current into the Gaussian process regression model to obtain the second mean and the second variance of the device model parameters, and determines the Gaussian process confidence interval of the device model parameters according to the second mean, the second variance, and the confidence interval calculation formula.

[0137] When the electronic device determines the random forest confidence interval, it inputs the target measured current into the random forest model to obtain the third mean. Since the random forest model includes multiple decision trees, the mean square error of each decision tree is determined as the third variance of the device model parameters, and the Gaussian process confidence interval of the device model parameters is determined according to the third mean, the third variance, and the confidence interval calculation formula.

[0138] Among them, the confidence interval calculation formula is the same as the calculation formula in step S102, and will not be elaborated here.

[0139] S304: Update the first confidence interval based on the Gaussian process confidence interval and the random forest confidence interval.

[0140] When the Gaussian process confidence interval and the random forest confidence interval have an intersection, use the intersection to update the first confidence interval.

[0141] When the Gaussian process confidence interval and the random forest confidence interval do not have an intersection, use the union of the Gaussian process confidence interval and the random forest confidence interval to update the first confidence interval.

[0142] The random forest model assisting the Gaussian regression model to update the first confidence interval can enable the Gaussian regression model to converge quickly and output the confidence interval even when the training data is insufficient.

[0143] S305. Adjust the update times of the first confidence interval, and determine whether the update times of the first confidence interval are equal to the first preset target times.

[0144] After each update of the first confidence interval is completed, increment the update times by one.

[0145] If the update times of the first confidence interval are equal to the first preset target times, go to step S306; otherwise, go to step S301 to enter a new round of confidence interval update.

[0146] S306. Determine the last updated first confidence interval as the parameter sampling value range of the preset device model.

[0147] Figure 6 This is a flowchart of the greedy algorithm provided by the present application according to an exemplary embodiment. As Figure 6 shown, the method includes:

[0148] S401. Randomly sample a set of random parameter sets in the parameter sampling value range.

[0149] The random parameter set includes initial random parameters corresponding to multiple variables of the device model. The parameter sampling value range includes sampling sub-ranges for each variable, and each initial random parameter is located in the corresponding sampling sub-range.

[0150] S402. According to the arrangement order of the random parameters in the random parameter set, sequentially generate multiple sets of random parameter sets corresponding to each variable.

[0151] Among them, the parameter corresponding to the variable in each set of random parameter sets corresponding to each variable is a parameter randomly selected in the sampling sub-range of the variable, and the parameters corresponding to other variables are the same in each set of random parameter sets.

[0152] That is to say, when generating multiple sets of random parameter sets corresponding to each variable, ensure that the parameters corresponding to other variables remain unchanged, randomly sample multiple parameters in the sampling sub-range corresponding to the variable, and form multiple sets of random parameter sets corresponding to the variable with the parameters of other variables.

[0153] The generation process of the set of random parameters corresponding to each variable is generated in a preset order. After the generation of multiple sets of random parameters for each variable, it enters step S403 to fit the predicted current corresponding to each set of random parameters, and then enters step S404 to update the sampling sub-interval corresponding to the variable and adjust the set of random parameters according to the simulation error. Then, while keeping the parameters corresponding to the variable unchanged, multiple random parameters corresponding to the next variable are generated until all the parameters and sampling sub-intervals corresponding to all variables are updated.

[0154] S403. For each variable, set the multiple random parameters of the corresponding set of random parameters in the device model, input the preset voltage into the device model, and obtain the corresponding predicted current.

[0155] Similar to step S301, after determining the set of random parameters, call the simulation software, set the device model with the parameters in the set of random parameters, and when setting the input voltage to the voltage corresponding to the target measurement current, obtain the predicted current of the simulation.

[0156] S404. Determine the simulation error according to the predicted current and the measurement current corresponding to the preset voltage, and update the sampling sub-interval corresponding to the updated variable and adjust the set of random parameters according to the simulation error.

[0157] The electronic device determines the relative error between the predicted current and the measurement current as the simulation error of the variable, selects the parameter with the smallest simulation error from the multiple simulation errors corresponding to the variable to update the set of random parameters, and updates the sampling sub-interval corresponding to the variable according to the maximum value and the minimum value of the sampling sub-interval corresponding to the variable and the updated parameter of the variable.

[0158] Among them, the update of the sampling sub-interval corresponding to each variable is updated according to the maximum value and the minimum value of the sampling sub-interval of the variable, the updated parameter of the variable, and the interval update formula.

[0159] The interval update formula includes:

[0160]

[0161] Among them, θ i represents the i-th parameter after the update of the variable in the set of random parameters, max i represents the maximum value of the sampling sub-interval of the i-th parameter, min i represents the minimum value of the sampling sub-interval of the i-th parameter, and k is a positive number.

[0162] In some embodiments, k is 2.

[0163] The following explains steps S401 to S404 by way of example:

[0164] There are three variables in the device model: A, B, and C. Each variable corresponds to a sampling sub-interval: [A1, A2], [B1, B2], [C1, C2]. First, randomly sample a parameter in each sampling sub-interval to form a set of random parameter sets [a1, b1, c1].

[0165] Enter the first round of loop. According to the order from left to right in the parameter set, first update the parameter and sampling sub-interval corresponding to variable A, keep the parameters corresponding to variables B and C in the random parameter set unchanged, randomly sample multiple parameters a2, a3, a4,... in the sampling sub-interval corresponding to variable A, and form multiple random parameter sets corresponding to variable A by combining the above parameters with the parameters corresponding to variables B and C respectively: [a2, b1, c1], [a3, b1, c1], [a4, b1, c1],... The electronic device calls the simulation software, sets the parameters of the simulated device model with the parameter pairs in each random parameter set corresponding to variable A, inputs the same voltage, obtains multiple corresponding predicted currents, performs operations on each current and the target measured current respectively, obtains the current with the smallest error, updates the parameter corresponding to variable A in the initial random parameter set with the parameter in the random parameter set corresponding to this current, and updates the sampling sub-interval corresponding to this variable according to the maximum and minimum values of the multiple parameters corresponding to variable A and the parameters of the updated initial random parameter set. If a3 is the parameter corresponding to the current with the smallest error, then the initial parameter set is updated to [a3, b1, c1], the minimum value of the sampling sub-interval corresponding to variable A is a3 - (A2 - A1) / 2, and the maximum value of the sampling sub-interval corresponding to variable A is a3 + (A2 - A1) / 2.

[0166] After variable A is updated, update the parameter and sampling sub-interval corresponding to variable B. Then randomly sample multiple parameters b2, b3, b4,... in the sampling sub-interval corresponding to variable B, and form multiple random parameter sets corresponding to variable B by combining the above parameters with the parameters corresponding to variables A and C respectively: [a3, b2, c1], [a3, b3, c1], [a3, b4, c1],... and update the parameter and sampling sub-interval corresponding to variable B in the initial random parameter set in the same way as the above method for updating the parameter and sampling sub-interval of variable A. If b2 has the smallest error among the parameters corresponding to variable B, then the initial parameter set is updated to [a3, b2, c1].

[0167] Then update the parameter and sampling sub-interval corresponding to variable C according to the above method to complete the first round of loop.

[0168] When entering the second round of loop, still update the corresponding parameters and sampling sub-intervals in the order of parameters ABC. In this round of loop, when randomly sampling the parameters corresponding to each variable, sample in the sampling sub-intervals updated in the previous round of loop.

[0169] And so on until the second loop end condition is met, then the update of each parameter stops.

[0170] S405. Determine whether the adjustment times of the sampling sub-intervals corresponding to each variable in the parameter sampling value range are all the second preset target times.

[0171] For the above example, determine whether the number of completed loops is equal to the second preset target time.

[0172] If satisfied, go to step S406;

[0173] If not satisfied, go to step S402, enter the next loop to update each parameter and its corresponding sampling range for the next round.

[0174] S406. Determine the device model parameters of the target measurement current as the finally adjusted random parameter set.

[0175] More specifically, determine the device model parameters of the target measurement current as the random parameter set updated in the last loop.

[0176] In the above technical solution, when using the coordinate descent method to determine the device model parameters, sampling and fitting are performed for the sampling sub-intervals corresponding to each device variable, the simulation error is obtained by comparing with the actual measurement, and the sampling sub-intervals of each parameter of the device model are modified one by one according to the simulation error to determine the accurate parameters, ensuring the extraction accuracy of the model parameters of the semiconductor device.

[0177] Figure 7 The following is a schematic structural diagram of a device model parameter extraction device for a semiconductor device according to an embodiment of the present application. As Figure 7 shown, the device model parameter extraction device 500 for the semiconductor device includes an acquisition module 501 and a processing module 502.

[0178] The acquisition module 501 is used to obtain the target measurement current;

[0179] The processing module 502 is used to obtain the first confidence interval of the device model parameters according to the target measurement current and the trained convolutional neural network model;

[0180] The processing module 502 is further used to train the first model and the second model respectively based on multiple groups of parameters in the first confidence interval and the preset device model, obtain the Gaussian process regression model and the random forest model, and update the first confidence interval according to the target measurement current, the Gaussian process regression model and the random forest model to obtain the parameter sampling value range of the preset device model;

[0181] The processing module 502 is further configured to determine the device model parameters for generating the target measurement current by using a greedy algorithm within the parameter sampling value range.

[0182] In one embodiment, the processing module 502 is specifically configured to:

[0183] Input the target measurement current into a convolutional neural network model to obtain the first mean and the first variance of the device model parameters;

[0184] Obtain the first confidence interval of the device model parameters according to the first mean and the first variance.

[0185] In one embodiment, the processing module 502 is further configured to:

[0186] When a semiconductor device is loaded with multiple sets of measured voltage values, obtain the corresponding drain measurement current; wherein, each set of measured voltage values includes a gate voltage and a drain voltage;

[0187] For each set of measured voltage values, based on the drain measurement current corresponding to the measured voltage value, adjust the parameters of the device model to determine the device model parameters corresponding to the semiconductor device; wherein, the current fitted by the device model when setting the device model parameters is the drain measurement current corresponding to the measured voltage value; l

[0188] Based on the drain measurement currents of each group and the corresponding measured voltage values and corresponding device model parameters, construct a training set;

[0189] Use the training set to train the original convolutional neural network to obtain a trained convolutional neural network; the loss function of the convolutional neural network includes:

[0190]

[0191] wherein, i represents the identifier of each parameter in the parameter set, D represents the number of parameters, and θ i represents the i-th parameter obtained by the convolutional neural network when inputting a preset drain current, represents the i-th preset parameter corresponding to the preset drain current in the training set, and σ i represents the variance of the i-th parameter.

[0192] In one embodiment, the processing module 502 is specifically configured to:

[0193] When the update times of the first confidence interval are less than the first preset target times, train the first model and the second model respectively based on multiple groups of parameters randomly sampled from the first confidence interval and a preset device model to obtain a Gaussian process regression model and a random forest model;

[0194] Input the target measurement current into the Gaussian process regression model to obtain the Gaussian process confidence interval of the device model parameters;

[0195] Input the target measured current into the random forest model to obtain the random forest confidence interval of the device model parameters;

[0196] Update the first confidence interval based on the Gaussian process confidence interval and the random forest confidence interval until the number of updates of the first confidence interval is equal to the first preset target number;

[0197] Determine the parameter sampling value interval of the preset device model as the finally updated first confidence interval.

[0198] In one embodiment, the processing module 502 is specifically configured to:

[0199] Randomly sample in the first confidence interval to obtain multiple sets of sampling parameter sets; each set of sampling parameter sets includes multiple parameters of the device model;

[0200] Set the multiple parameters of each set of sampling parameter sets to the device model, and after inputting multiple sets of preset voltage values into the device model, obtain the simulated current corresponding to each set of preset voltage values;

[0201] Construct a sampling training set by using each set of sampling parameter sets and the corresponding simulated current;

[0202] Train the first model by using the sampling training set to obtain a Gaussian process regression model;

[0203] Train the second model by using the sampling training set to obtain a random forest model.

[0204] In one embodiment, the processing module 502 is specifically configured to:

[0205] When the Gaussian process confidence interval and the random forest confidence interval have an intersection, update the first confidence interval by using the intersection;

[0206] When the Gaussian process confidence interval and the random forest confidence interval do not have an intersection, update the first confidence interval by using the union of the Gaussian process confidence interval and the random forest confidence interval.

[0207] In one embodiment, the processing module 502 is specifically configured to:

[0208] Input the target measured current into the Gaussian process regression model to obtain the second mean and the second variance of the device model parameters;

[0209] Determine the Gaussian process confidence interval of the device model parameters according to the second mean, the second variance and the confidence interval calculation formula.

[0210] In one embodiment, the processing module 502 is specifically configured to:

[0211] Input the target measured current into the random forest model to obtain the third mean value of the device model parameters;

[0212] Determine the mean square error of each decision tree as the third variance of the device model parameters;

[0213] According to the third mean value, the third variance and the confidence interval calculation formula, determine the Gaussian process confidence interval of the device model parameters;

[0214] Among them, the random forest includes multiple decision trees.

[0215] In one embodiment, the processing module 502 is specifically configured to:

[0216] Randomly sample a set of random parameter sets in the parameter sampling value range; the random parameter set includes the initial random parameters corresponding to multiple variables of the device model, the parameter sampling value range includes the sampling sub-ranges of each variable, and each initial random parameter is located in the corresponding sampling sub-range;

[0217] Repeat the execution, and sequentially generate multiple sets of random parameter sets corresponding to each variable according to the arrangement order of the random parameters in the random parameter set; among them, the parameter corresponding to the variable in the multiple sets of random parameter sets corresponding to each variable is a parameter randomly selected in the sampling sub-range of the variable, and the parameters corresponding to other variables are the same in each set of random parameter sets;

[0218] For each variable, set the multiple random parameters of the corresponding random parameter set to the device model, input a preset voltage into the device model, and obtain the predicted current corresponding to the preset voltage;

[0219] According to the predicted current and the measured current corresponding to the preset voltage, determine the simulation error, and according to the simulation error, update the sampling sub-range corresponding to the variable, and adjust the random parameter set until the adjustment times of the sampling sub-ranges corresponding to each variable in the parameter sampling value range are all the second preset target times, and then determine the random parameter set adjusted subsequently as the device model parameters of the target measured current.

[0220] In one embodiment, the processing module 502 is specifically configured to:

[0221] Determine the relative error between the predicted current and the measured current as the simulation error of the variable;

[0222] Select the parameter with the smallest simulation error from the multiple simulation errors corresponding to the variable to update the random parameter set;

[0223] Update the sampling sub-range according to the maximum and minimum values of the sampling sub-range of the variable and the updated parameter of the variable.

[0224] In one embodiment, the processing module 502 is specifically configured to:

[0225] Update the sampling sub - interval corresponding to the variable according to the maximum and minimum values of the sampling sub - intervals of the variable, the updated parameters of the variable, and the interval update formula;

[0226] The interval update formula includes:

[0227]

[0228] Among them, θ i represents the i - th parameter after the update of the variable in the set of random parameters, max i represents the maximum value of the sampling sub - interval of the i - th parameter, min i represents the minimum value of the sampling sub - interval of the i - th parameter, and k is a positive number.

[0229] Figure 8 FIG. 20 is a schematic structural diagram of an electronic device according to an embodiment of the present application. Among them, the electronic device 600 includes a memory 601 and a processor 602. The memory 601 is used to store computer instructions executable by the processor. The memory 601 may include a high - speed random access memory (Random Access Memory, RAM), and may also include non - volatile storage (Non - Volatile Memory, NVM), such as at least one disk memory, and may also be a USB flash drive, a mobile hard disk, a read - only memory, a magnetic disk, or an optical disc, etc.

[0230] When the processor 602 executes the computer instructions, it implements each step in the method for extracting model parameters of a semiconductor device with the electronic device as the execution subject in the above - mentioned embodiment. Specifically, reference may be made to the relevant descriptions in the foregoing method embodiments. The processor 602 may be a central processing unit (Central Processing Unit, CPU), and may also be other general - purpose processors, digital signal processors (Digital Signal Processor, DSP), application - specific integrated circuits (ApplicationSpecific Integrated Circuit, ASIC), etc. The general - purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The steps of the method disclosed in combination with the invention may be directly embodied as being executed and completed by a hardware processor, or may be executed and completed by a combination of hardware and software modules in the processor.

[0231] Optionally, the above-mentioned memory 601 can be either independent or integrated with the processor 602. When the memory 601 is independently provided, the electronic device 600 further includes a bus for connecting the memory 601 and the processor 602. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, an Extended Industry Standard Architecture (EISA) bus, or the like. The bus can be divided into an address bus, a data bus, a control bus, etc. For the sake of convenience of representation, the buses in the drawings of this application are not limited to only one bus or one type of bus.

[0232] The embodiments of the present application further provide a computer-readable storage medium. Computer instructions are stored in the computer-readable storage medium. When the processor executes the computer instructions, each step in the method for extracting model parameters of a semiconductor device in the above embodiments is implemented.

[0233] The embodiments of the present application further provide a computer program product, including computer instructions. When the computer instructions are executed by the processor, each step in the method for extracting model parameters of a semiconductor device in the above embodiments is implemented.

[0234] Those skilled in the art will readily conceive of other embodiments of the present application after considering the specification and practicing the invention disclosed herein. The present application is intended to cover any variations, uses, or adaptations of the present application, which follow the general principles of the present application and include common general knowledge or conventional technical means in the technical field not disclosed in the present application. The specification and the embodiments are only regarded as exemplary, and the true scope and spirit of the present application are pointed out by the following claims.

[0235] It should be understood that the present application is not limited to the exact structures already described and shown in the drawings, and various modifications and changes can be made without departing from its scope. The scope of the present application is only limited by the appended claims.

Claims

1. A method for extracting model parameters of a semiconductor device, characterized in that, The method includes: Obtaining a target measured current; According to the target measured current and the trained convolutional neural network model, obtaining a first confidence interval of the device model parameters; Based on multiple groups of parameters in the first confidence interval and a preset device model, training a first model and a second model respectively to obtain a Gaussian process regression model and a random forest model, and updating the first confidence interval according to the target measured current, the Gaussian process regression model and the random forest model to obtain a parameter sampling value range of the preset device model; In the parameter sampling value range, using a greedy algorithm to determine the device model parameters that generate the target measured current.

2. The method according to claim 1, characterized in that, The obtaining a first confidence interval of the model parameters according to the target measured current and the trained convolutional neural network model specifically includes: Inputting the target measured current into the convolutional neural network model to obtain a first mean and a first variance of the device model parameters; According to the first mean and the first variance, obtaining a first confidence interval of the device model parameters.

3. The method according to claim 1 or 2, characterized in that, Before obtaining a first confidence interval of the device model parameters according to the target measured current and the trained convolutional neural network model, the method further includes: When a semiconductor device is loaded with multiple groups of measured voltage values, obtaining corresponding drain measured currents; wherein each group of measured voltage values includes a gate voltage and a drain voltage; For each group of measured voltage values, adjusting the parameters of the device model based on the drain measured current corresponding to the measured voltage value to determine the device model parameters corresponding to the semiconductor device; wherein the current fitted by the device model when setting the device model parameters is the drain measured current corresponding to the measured voltage value; Based on each group of the drain measured currents, the corresponding measured voltage values, and the corresponding device model parameters, constructing a training set; Using the training set to train an original convolutional neural network to obtain the trained convolutional neural network; the loss function of the convolutional neural network includes: where \(i\) represents the identifier of each parameter in the parameter set, \(D\) represents the number of parameters, and \(\theta\) i represents the \(i\)-th parameter obtained by the convolutional neural network when the preset drain current is input, represents the \(i\)-th preset parameter corresponding to the preset drain current in the training set, and \(\sigma\) i represents the variance of the \(i\)-th parameter.

4. The method according to claim 1, characterized in that, Based on multiple groups of parameters in the first confidence interval and a preset device model, training a first model and a second model respectively to obtain a Gaussian process regression model and a random forest model, and updating the first confidence interval according to the target measured current, the Gaussian process regression model and the random forest model to obtain a parameter sampling value range of the preset device model, including: When the number of updates of the first confidence interval is less than a first preset target number, training a first model and a second model respectively based on multiple groups of randomly sampled parameters in the first confidence interval and a preset device model to obtain a Gaussian process regression model and a random forest model; Inputting the target measured current into the Gaussian process regression model to obtain a Gaussian process confidence interval of the device model parameters; Inputting the target measured current into the random forest model to obtain a random forest confidence interval of the device model parameters; Updating the first confidence interval based on the Gaussian process confidence interval and the random forest confidence interval until the number of updates of the first confidence interval is equal to the first preset target number; Determine the first confidence interval of the last update as the parameter sampling value range of the preset device model.

5. The method according to claim 4, characterized in that, The training of the first model and the second model based on multiple groups of parameters randomly sampled from the first confidence interval and the preset device model to obtain the Gaussian process regression model and the random forest model specifically includes: Randomly sample in the first confidence interval to obtain multiple groups of sampling parameter sets; each group of the sampling parameter sets includes multiple parameters of the device model. Set the multiple parameters of each group of sampling parameter sets to the device model, and after inputting multiple groups of preset voltage values into the device model, obtain the simulated current corresponding to each group of preset voltage values. Construct a sampling training set by using each group of sampling parameter sets and the corresponding simulated current. Train the first model by using the sampling training set to obtain the Gaussian process regression model. Train the second model by using the sampling training set to obtain the random forest model.

6. The method according to claim 4 or 5, characterized in that, The updating of the first confidence interval based on the Gaussian process confidence interval and the random forest confidence interval specifically includes: when the Gaussian process confidence interval and the random forest confidence interval have an intersection, update the first confidence interval by using the intersection. When the Gaussian process confidence interval and the random forest confidence interval do not have an intersection, update the first confidence interval by using the union of the Gaussian process confidence interval and the random forest confidence interval.

7. The method according to claim 4 or 5, characterized in that, The inputting of the target measurement current into the Gaussian process regression model to obtain the Gaussian process confidence interval of the device model parameters specifically includes: Input the target measurement current into the Gaussian process regression model to obtain the second mean and the second variance of the device model parameters. Determine the Gaussian process confidence interval of the device model parameters according to the second mean, the second variance and the confidence interval calculation formula.

8. The method according to claim 4 or 5, characterized in that, The random forest model includes multiple decision trees. The inputting of the target measurement current into the random forest model to obtain the random forest confidence interval of the device model parameters specifically includes: Input the target measurement current into the random forest model to obtain the third mean of the device model parameters. Determine the mean square error of each decision tree as the third variance of the device model parameters. Determine the Gaussian process confidence interval of the device model parameters according to the third mean, the third variance and the confidence interval calculation formula.

9. The method according to claim 1, characterized in that, The determination of the device model parameters generating the target measurement current by using the greedy algorithm in the parameter sampling value range specifically includes: Randomly sample a group of random parameter sets in the parameter sampling value range; the random parameter set includes the initial random parameters corresponding to multiple variables of the device model, the parameter sampling value range includes the sampling sub-ranges of each variable, and each initial random parameter is located in the corresponding sampling sub-range. Repeat the execution, and sequentially generate multiple groups of random parameter sets corresponding to each variable according to the arrangement order of the random parameters in the random parameter set; among them, the parameter corresponding to the variable in each group of random parameter sets corresponding to each variable is a parameter randomly selected in the sampling sub-range of the variable, and the parameters corresponding to other variables are the same in each group of random parameter sets. For each variable, set multiple random parameters of the corresponding random parameter set to the device model, input a preset voltage into the device model, and obtain a predicted current corresponding to the preset voltage; According to the predicted current and the measured current corresponding to the preset voltage, determine the simulation error. According to the simulation error, update the sampling sub-interval corresponding to the variable, and adjust the random parameter set until the adjustment times of the sampling sub-intervals corresponding to each variable in the parameter sampling value range are all the second preset target times, then determine the finally adjusted random parameter set as the device model parameters of the target measurement current.

10. The method according to claim 9, characterized in that, The step of determining the simulation error according to the predicted current and the measured current corresponding to the preset voltage, updating the sampling sub-interval corresponding to the variable according to the simulation error, and adjusting the random parameter set specifically includes: Determine the relative error between the predicted current and the measured current as the simulation error of the variable; Select the parameter with the smallest simulation error from the multiple simulation errors corresponding to the variable to update the random parameter set; Update the sampling sub-interval according to the maximum value and the minimum value of the sampling sub-interval of the variable and the parameter after the variable is updated.

11. The method according to claim 10, characterized in that, The step of updating the sampling sub-interval according to the maximum value and the minimum value of the sampling sub-interval of the variable and the parameter after the variable is updated specifically includes: Update the sampling sub-interval according to the maximum value and the minimum value of the sampling sub-interval of the variable, the parameter after the variable is updated, and the interval update formula; The interval update formula includes: where, θ i represents the i-th parameter after the update of the variable in the set of random parameters, max i represents the maximum value of the sampling sub-interval of the i-th parameter, min i represents the minimum value of the sampling sub-interval of the i-th parameter, and k is a positive number.

12. A model parameter extraction device for a semiconductor device, characterized in that, Includes: An acquisition module for obtaining a target measurement current; A processing module for obtaining a first confidence interval of device model parameters according to the target measurement current and a trained convolutional neural network model; The processing module is further configured to train a first model and a second model respectively based on multiple groups of parameters in the first confidence interval and a preset device model to obtain a Gaussian process regression model and a random forest model, and update the first confidence interval according to the target measurement current, the Gaussian process regression model and the random forest model to obtain a parameter sampling value range of the preset device model; The processing module is further configured to use a greedy algorithm to determine device model parameters that generate the target measurement current in the parameter sampling value range.

Images