A method for automatically detecting performance parameters of a semiconductor device

Abstract

The application relates to the technical field of semiconductors, in particular to a semiconductor device performance parameter automatic detection method, which comprises the following steps: obtaining an initial detection working interval of a semiconductor device; performing initial random sampling in the initial detection working interval to obtain an initial sampling point set; performing fitting processing on the initial sampling point set by using a Gaussian process regression algorithm to construct a proxy model; and performing analysis processing on the proxy model by using a batch Bayesian optimization algorithm. The asynchronous hardware pipeline capability of a source measurement unit is utilized to perform high-speed parallel sampling, the communication overhead between an upper computer and an instrument is reduced, parallel processing of calculation and measurement tasks is realized, and physical test time consumption is compressed. In this way, the accurate performance parameters can be quickly and safely converged by using far fewer measurement points than in a traditional scanning mode, so that the test efficiency and safety are improved under the premise of ensuring the accuracy.

Description

Technical Field

[0001] This invention relates to the field of semiconductor technology, and in particular to an automatic detection method for semiconductor device performance parameters. Background Technology

[0002] Semiconductor device performance parameter testing is a branch of the field of semiconductor testing and electronic measurement, which uses automated methods to characterize the electrical performance of semiconductor devices (such as transistors, diodes, etc.).

[0003] Current technologies for characterizing semiconductor devices generally rely on a pre-defined, fixed-step, blanket-scanning mode. This mode is essentially an open-loop, brute-force test; to ensure coverage of all key performance characteristics, high-density sampling must be performed across the entire operating range. This inevitably generates massive amounts of redundant measurement data because in regions where device performance is flat, many sampling points do not provide new information. This not only consumes storage resources but also prolongs the testing cycle. Furthermore, since the scan path is pre-defined, it is difficult to dynamically perceive the real-time state of the device during the testing process, making it difficult to avoid risks. For example, when testing the output characteristics of power transistors, blanket scanning may indiscriminately enter the linear region where high voltage and high current coexist, causing thermal damage to the device due to instantaneous power consumption exceeding the safe operating range limit, resulting in waste of R&D samples or production line products. Therefore, improvements are needed. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an automatic detection method for semiconductor device performance parameters.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an automatic detection method for performance parameters of semiconductor devices, comprising the following steps: Obtain the initial detection working range of the semiconductor device; Initial random sampling is performed within the initial detection working interval to obtain an initial set of sampling points; The Gaussian process regression algorithm is used to fit the initial set of sampling points to construct a surrogate model; The surrogate model is analyzed and processed using a batch Bayesian optimization algorithm, and a virtual sampling penalty factor based on the power consumption safety boundary of semiconductor devices is introduced to generate a sequence of predicted sampling points. The predicted sampling point sequence is encapsulated into a list scan instruction; The list scan command is sent to the source measurement unit for asynchronous hardware pipeline sampling processing to obtain a set of device performance sampling data. Determine whether the device performance sampling data set meets the preset convergence condition; If the device performance sampling data set does not meet the preset convergence condition, the initial sampling point set is updated based on the device performance sampling data set, and the step of fitting the initial sampling point set using the Gaussian process regression algorithm to construct a surrogate model is returned. If the preset convergence condition is met, the semiconductor device performance parameters are output based on the device performance sampling data set.

[0006] Preferably, the steps for obtaining the initial detection working range of the semiconductor device are as follows: Obtain the device type identifier of the semiconductor device under test (DUT). Retrieve the corresponding historical test parameter library based on the device type identifier. Query the lower and upper limits of the reference voltage scan in the historical test parameter library. Combine the lower and upper limits of the reference voltage scan to obtain the voltage operating range. Query the lower and upper limits of the reference current scan in the historical test parameter library. Combine the lower and upper limits of the reference current scan to obtain the current operating range. Merge the voltage and current operating ranges to construct a feature scan detection dimension matrix. Determine the initial operating range based on the feature scan detection dimension matrix. Record the coordinates of the boundary nodes in the initial operating range. Verify the legality of the boundary node coordinates within the hardware tolerance range of the semiconductor device test platform. When the boundary node coordinates are legal, lock the initial operating range as the initial detection operating range for the semiconductor device.

[0007] Preferably, the step of performing initial random sampling within the initial detection working interval to obtain the initial sampling point set specifically includes: The voltage and current operating ranges corresponding to the initial detection operating ranges are obtained respectively. The Latin hypercube sampling algorithm is used to perform homogenization and stratification processing on the voltage operating ranges to obtain multiple voltage sampling levels. Similarly, the Latin hypercube sampling algorithm is used to perform homogenization and stratification processing on the current operating ranges to obtain multiple current sampling levels. A first value is randomly selected from the multiple voltage sampling levels, and a second value is randomly selected from the multiple current sampling levels. The first value and the second value are paired to generate multiple initial sampling coordinate pairs. Duplicate coordinate data in the multiple initial sampling coordinate pairs are filtered and removed to obtain a deduplicated sampling coordinate pair set. The driving source measurement unit applies point-by-point electrical signals to the deduplicated sampling coordinate pair set, and the response current and response voltage data fed back by the source measurement unit are recorded. The deduplicated sampling coordinate pair set, the response current data, and the response voltage data are associated and bound to form the initial sampling point set.

[0008] Preferably, the steps for constructing the surrogate model by fitting the initial set of sampling points using the Gaussian process regression algorithm are as follows: Obtain the initial sampling point set, and separate the independent variable matrix and dependent variable vector in the initial sampling point set; initialize the mean function and covariance kernel function of the Gaussian process regression algorithm, wherein the covariance kernel function is configured as a squared exponential kernel function; input the independent variable matrix into the mean function to calculate the prior mean vector, input the independent variable matrix into the covariance kernel function to calculate the prior covariance matrix, and establish a multivariate Gaussian joint distribution function by combining the prior mean vector and the prior covariance matrix; iteratively optimize the hyperparameters in the multivariate Gaussian joint distribution function using the maximum likelihood estimation method to obtain the optimal hyperparameter set; substitute the optimal hyperparameter set into the covariance kernel function to update the covariance kernel function; derive the posterior prediction distribution based on the independent variable matrix, the dependent variable vector, and the updated covariance kernel function, extract the prediction mean expression and prediction variance expression from the posterior prediction distribution, encapsulate the prediction mean expression and prediction variance expression, and construct a surrogate model.

[0009] Preferably, the steps of using a batch Bayesian optimization algorithm to analyze and process the surrogate model, and introducing a virtual sampling penalty factor based on the power consumption safety boundary of semiconductor devices to generate the predicted sampling point sequence are as follows: The process involves: invoking the surrogate model and setting the acquisition function of the batch Bayesian optimization algorithm, wherein the acquisition function is configured as the expected improvement function; calculating the prediction uncertainty distribution of the surrogate model in the global search space, and determining the expected value distribution of the acquisition function based on the prediction uncertainty distribution; using a multi-peak particle swarm optimization algorithm to perform a global extreme value search on the expected value distribution, locating the coordinates of the maximum expected value in the expected value distribution, and using the coordinates of the maximum expected value as the first prediction sampling point; introducing a virtual sampling penalty factor based on the power safety boundary of semiconductor devices to update the expected value distribution of the acquisition function, and performing an extreme value search again on the updated expected value distribution of the acquisition function to locate the updated coordinates of the maximum expected value, and using the updated coordinates of the maximum expected value as the second prediction sampling point; repeatedly executing the steps of introducing a virtual sampling penalty factor based on the power safety boundary of semiconductor devices to update the expected value distribution of the acquisition function and the extreme value search until the total number of prediction sampling points reaches the preset batch sequence length; and sorting all prediction sampling points that have reached the preset batch sequence length according to the order of their generation to form the prediction sampling point sequence.

[0010] Preferably, the step of encapsulating the predicted sampling point sequence into a list scan instruction specifically includes: Extract each predicted sampling point from the predicted sampling point sequence and parse the corresponding operating mode identifier for each predicted sampling point. When the operating mode identifier is voltage forced mode, extract the corresponding target voltage amplitude parameter from each predicted sampling point. When the operating mode identifier is current forced mode, extract the corresponding target current amplitude parameter from each predicted sampling point. Read the standard programmable instrument standard command protocol dictionary and match the basic control string corresponding to the operating mode identifier according to the standard programmable instrument standard command protocol dictionary. Convert the target voltage amplitude parameter and the target current amplitude parameter into hexadecimal machine code data, fill the hexadecimal machine code data into the parameter bits of the basic control string, and generate independent execution action code. Configure the hardware trigger delay time parameter between each predicted sampling point, concatenate the independent execution action code and the hardware trigger delay time parameter according to the arrangement order of the predicted sampling point sequence, add a header check bit and a tail end character, and compile to obtain the list scan instruction.

[0011] Preferably, the step of sending the list scan command to the source measurement unit for asynchronous hardware pipeline sampling processing to obtain the device performance sampling data set is as follows: An asynchronous transmission control protocol connection channel is established between the host computer and the source measurement unit; the list scan instruction in the communication buffer is captured and sent to the underlying hardware first-in-first-out queue of the source measurement unit through the asynchronous transmission control protocol connection channel; the internal state machine of the source measurement unit is triggered to parse the list scan instruction in the underlying hardware first-in-first-out queue, and the hardware trigger of the source measurement unit drives the digital-to-analog converter according to the hardware trigger delay time parameter in the list scan instruction to continuously apply excitation signals to the semiconductor device and synchronously drive the analog-to-digital converter to collect response signals. While the source measurement unit is performing hardware sampling, the central processing unit of the host computer asynchronously executes subsequent data analysis preparation work; after the source measurement unit completes all sampling actions corresponding to the list scan instruction, the collected response signal is converted into a real number result data string; the real number result data string is packaged and sent back to the host computer through the asynchronous transmission control protocol connection channel; the real number result data string is parsed, the real performance data group is extracted, and the real performance data group is stored in the database to form the device performance sampling data set.

[0012] Preferably, the step of determining whether the device performance sampling data set meets the preset convergence condition specifically includes: Extract the device performance sampling data set from the database, calculate the current total number of data in the device performance sampling data set, and compare the current total number of data with a preset maximum sampling threshold; if the current total number of data reaches the preset maximum sampling threshold, it is determined that the preset convergence condition is met; if the current total number of data does not reach the preset maximum sampling threshold, calculate the hyperparameter change rate of the surrogate model between two consecutive prediction batches, and determine whether the hyperparameter change rate is less than a preset stability tolerance; if the hyperparameter change rate is less than the preset stability tolerance, it is determined that the preset convergence condition is met; if the hyperparameter change rate is greater than or equal to the preset stability tolerance, extract the latest performance gradient value and the reference performance gradient value of the previous iteration batch from the device performance sampling data set, substitute the latest performance gradient value and the reference performance gradient value into the convergence determination formula for evaluation, and determine whether the evaluation index value meets the requirements. The formula is: ; in, Indicates the evaluation index value, Represents the gradient weight coefficients. This represents the latest performance gradient value. This represents the reference performance gradient value of the previous iteration batch. This represents a small positive integer used to prevent the denominator from being zero; When the evaluation index value is less than the predetermined tolerance limit, the device performance sampling data set is determined to meet the preset convergence condition; when the evaluation index value is greater than or equal to the predetermined tolerance limit, the device performance sampling data set is determined not to meet the preset convergence condition.

[0013] Preferably, if the device performance sampling data set does not meet the preset convergence condition, the step of updating the initial sampling point set based on the device performance sampling data set and returning to the step of fitting the initial sampling point set using a Gaussian process regression algorithm to construct a surrogate model; if the preset convergence condition is met, the step of outputting semiconductor device performance parameters based on the device performance sampling data set specifically includes: When it is determined that the device performance sampling data set does not meet the preset convergence condition, the device performance sampling data set is appended and merged into the tail data block of the initial sampling point set to generate an expanded sampling point set. The expanded sampling point set is used as a new initial sampling point set, and the process of fitting the initial sampling point set and constructing a surrogate model using the Gaussian process regression algorithm is returned to perform the next round of iterative loop calculation. When the device performance sampling data set is determined to meet the preset convergence condition, the iterative loop operation is terminated. The device performance sampling data set is then smoothed and cleaned by filtering to remove abnormal isolated noise points and retain the core feature data. Curve smoothing interpolation is performed on the core feature data to plot the current-voltage characteristic curve of the semiconductor device. Threshold voltage parameters, leakage current parameters, and breakdown voltage parameters are extracted from the current-voltage characteristic curve of the semiconductor device, and the threshold voltage parameters, leakage current parameters, and breakdown voltage parameters are output as semiconductor device performance parameters.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are as follows: This invention acquires a small amount of representative data through random sampling within the initial detection working range. Based on this, a surrogate model characterizing the global performance and uncertainties of the device is constructed using a Gaussian process regression algorithm. This achieves a mathematical abstraction of the complex electrical behavior of the device. Subsequently, the surrogate model is analyzed using a batch Bayesian optimization algorithm, which can proactively predict and locate multiple sampling points with the most informational value for characterizing performance parameters. This eliminates the blind scanning in traditional testing and improves the directionality and efficiency of data acquisition. Simultaneously, a virtual sampling penalty mechanism based on power consumption safety boundaries is introduced to proactively avoid high-risk areas that may cause over-electrical stress damage to the device during the optimization process, enhancing the safety of the testing process. The generated predicted point sequence is integrated and encapsulated into a single-issue list scan command, and high-speed parallel sampling is performed using the asynchronous hardware pipeline capability of the source measurement unit. This reduces the communication overhead between the host computer and the instrument, enabling parallel processing of computation and measurement tasks and compressing the physical testing time. This allows for rapid and safe convergence to accurate performance parameters with far fewer measurement points than traditional scanning methods, thereby improving testing efficiency and safety while ensuring accuracy. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the steps of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0017] Please see Figure 1 This invention provides a technical solution: an automatic detection method for performance parameters of semiconductor devices, comprising the following steps: Obtain the initial detection working range of the semiconductor device; Initial random sampling is performed within the initial detection working interval to obtain the initial sampling point set; A surrogate model is constructed by fitting the initial set of sampling points using the Gaussian process regression algorithm. The batch Bayesian optimization algorithm is used to analyze and process the surrogate model, and a virtual sampling penalty factor based on the power consumption safety boundary of semiconductor devices is introduced to generate a sequence of predicted sampling points. Encapsulate the predicted sampling point sequence into a list scan instruction; The list scan command is sent to the source measurement unit for asynchronous hardware pipeline sampling processing to obtain a set of device performance sampling data. Determine whether the device performance sampling data set meets the preset convergence conditions; If the device performance sampling data set does not meet the preset convergence condition, the initial sampling point set is updated based on the device performance sampling data set, and the step of fitting the initial sampling point set using the Gaussian process regression algorithm to construct a surrogate model is returned. If the preset convergence condition is met, the semiconductor device performance parameters are output based on the device performance sampling data set.

[0018] In this embodiment, the steps for obtaining the initial detection working range of the semiconductor device are as follows: Obtain the device type identifier of the semiconductor device under test; retrieve the corresponding historical test parameter library based on the device type identifier; query the lower limit and upper limit of the reference voltage scan in the historical test parameter library; combine the lower limit and upper limit of the reference voltage scan to obtain the voltage working range; query the lower limit and upper limit of the reference current scan in the historical test parameter library; combine the lower limit and upper limit of the reference current scan to obtain the current working range; merge the voltage working range and the current working range to construct a feature scan detection dimension matrix; determine the initial working range based on the feature scan detection dimension matrix; record the boundary node coordinates in the initial working range; verify the legality of the boundary node coordinates within the hardware allowable range of the semiconductor device test platform; when the boundary node coordinates are legal, lock the initial working range as the initial detection working range of the semiconductor device.

[0019] Specifically, the device type identifier of the semiconductor device under test (DUT) is obtained, for example, by inputting it from the user interface or reading it from the connected test task management system. This identifier is MOSFET_N_POWER_2N7002. Based on the device type identifier, the corresponding historical test parameter library is retrieved. This library is a structured database storing standard test boundaries for different device types. Records associated with 2N7002 are queried in this library. The lower limit of the reference voltage scan is set to 0V and the upper limit of the reference voltage scan is set to 10V. Combining 0V and 10V yields the voltage operating range. Further queries yield the lower limit of the reference current scan is set to 0A and the upper limit of the reference current scan is set to 5A. Combining 0A and 5A yields the current operating range. The voltage and current operating ranges are merged to construct a two-dimensional feature scan detection dimension matrix. The two dimensions of this matrix correspond to the voltage and current ranges, respectively. Based on the feature scan detection dimension matrix, the initial operating range is determined, which is a rectangular area consisting of a voltage range of 0V-10V and a current range of 0A-5A. The coordinates of the boundary nodes in the initial operating range are recorded, i.e., (0V, 0A), (10V, 0A), (10A, 0A), (10A, 0A), (10A, 0A), (10A, 0A), (10A, 1 ... The coordinates of the four vertices (0A), (0V, 5A), and (10V, 5A) are used to verify the legality of the boundary node coordinates within the hardware tolerance range of the semiconductor device test platform. This process compares the voltage and current values ​​of each boundary point with the datasheet of the source measurement unit (SMU). For example, if the maximum output voltage of the SMU is 40V and the maximum output current is 10A, then 10V is less than 40V and 5A is less than 10A. All boundary points are within the hardware capability range. When the boundary node coordinates are legal, the initial working range is locked as the initial testing working range of the semiconductor device.

[0020] In this embodiment, the specific steps for performing initial random sampling within the initial detection working interval to obtain the initial sampling point set are as follows: First, obtain the voltage working interval and current working interval corresponding to the initial detection working interval. Then, use the Latin hypercube sampling algorithm to perform uniform stratification processing on the voltage working interval to obtain multiple voltage sampling levels. Next, use the Latin hypercube sampling algorithm to perform uniform stratification processing on the current working interval to obtain multiple current sampling levels. Then, randomly select a first value from the multiple voltage sampling levels and a second value from the multiple current sampling levels. Pair the first value with the second value to generate multiple initial sampling coordinate pairs. Filter out duplicate coordinate data in the multiple initial sampling coordinate pairs and remove them to obtain a deduplicated sampling coordinate pair set. Finally, drive the source measurement unit to apply electrical signals point-by-point to the deduplicated sampling coordinate pair set, record the response current data and response voltage data fed back by the source measurement unit, and associate and bind the deduplicated sampling coordinate pair set, the response current data, and the response voltage data to form the initial sampling point set.

[0021] Specifically, the voltage working range (0V to 10V) and current working range (0A to 5A) corresponding to the initial detection working range are obtained respectively. The Latin hypercube sampling algorithm is then used to perform equal division and stratification of the voltage working range. Specifically, the voltage range of 0V to 10V is divided into 20 non-overlapping small intervals, each with a width of 0.5V, i.e., [0, 0.5V), [0.5V, 1.0V) up to [9.5V, ... [10V], resulting in 20 voltage sampling levels. Similarly, the Latin hypercube sampling algorithm is used to evenly divide the current operating range into 20 sub-intervals, with each sub-interval having a width of 0.25A, resulting in 20 current sampling levels. In each of the 20 voltage sampling levels, a voltage value is randomly selected to generate a sequence containing 20 different voltage values. Similarly, in each of the 20 current sampling levels, a current value is randomly selected to generate a sequence containing 20 different current values. The values ​​in these two sequences are randomly paired, for example, the first voltage value is paired with the fifth current value, the second voltage value is paired with the thirteenth current value, and so on, generating 20 initial sampling coordinate pairs. Duplicate coordinate data in multiple initial sampling coordinate pairs are filtered and removed to obtain a deduplicated sampling coordinate pair set. The driving source measurement unit applies a point-by-point electrical signal to the deduplicated sampling coordinate pair set. For example, for the coordinate pair (1.2V, The source measurement unit applies a voltage of 1.2V and measures the response current, or applies a current of 0.8A and measures the response voltage. It records the response current and response voltage data fed back by the source measurement unit. It associates and binds the deduplicated sampling coordinate pair set, the response current data, and the response voltage data. For example, it binds the coordinate (1.2V, 0.8A) with the measured response value (e.g., 0.82A, 1.19V) as a data record. Finally, it summarizes all 20 such data records to form the initial sampling point set.

[0022] In this embodiment, the steps for fitting the initial sampling point set using the Gaussian process regression algorithm and constructing the surrogate model are as follows: First, obtain the initial sampling point set and separate the independent variable matrix and dependent variable vector from it. Second, initialize the mean function and covariance kernel function of the Gaussian process regression algorithm, where the covariance kernel function is configured as a squared exponential kernel function. Third, input the independent variable matrix into the mean function to calculate the prior mean vector, input the independent variable matrix into the covariance kernel function to calculate the prior covariance matrix, and establish a multivariate Gaussian joint distribution function by combining the prior mean vector and the prior covariance matrix. Fourth, iteratively optimize the hyperparameters in the multivariate Gaussian joint distribution function using the maximum likelihood estimation method to obtain the optimal hyperparameter set. Fifth, substitute the optimal hyperparameter set into the covariance kernel function to update it. Sixth, derive the posterior prediction distribution based on the independent variable matrix, dependent variable vector, and the updated covariance kernel function, extract the prediction mean expression and prediction variance expression from the posterior prediction distribution, encapsulate the prediction mean expression and prediction variance expression, and construct the surrogate model.

[0023] Specifically, an initial sampling point set is obtained, containing 20 sets of data consisting of applied coordinates, response voltage, and response current. The independent variable matrix and dependent variable vector in the initial sampling point set are then separated, where the independent variable matrix... The dimension is 20x2, with each row representing the voltage and current coordinates of a sampling point, and the dependent variable vector. The dimension is 20x1, and each row represents a target performance parameter measured at the corresponding sampling point, such as leakage current. The mean function and covariance kernel function of the Gaussian process regression algorithm are initialized. The mean function is usually set to zero, and the covariance kernel function is configured as a squared exponential kernel function, with the specific form as follows: ,in and There are two input vectors. Represents Euclidean distance. It is the signal variance hyperparameter. This involves calculating the length-scale hyperparameters. The independent variable matrix is ​​input into the mean function to calculate the prior mean vector. The independent variable matrix is ​​then input into the covariance kernel function to calculate a 20x20 prior covariance matrix. A multivariate Gaussian joint distribution function is established by combining the prior mean vector and the prior covariance matrix. The maximum likelihood estimation method is then used to iteratively optimize the hyperparameters in the multivariate Gaussian joint distribution function, specifically by maximizing the log-marginal likelihood function to find the optimal value. and and noise variance The expression for the log-marginal likelihood function is: ,in It is the covariance matrix calculated based on the kernel function. It is the identity matrix. The number of sampling points is 20. The maximum value of this function is obtained by using optimization algorithms such as gradient descent to obtain the optimal hyperparameter set. The optimal hyperparameter set is substituted into the covariance kernel function to update the covariance kernel function. Based on the independent variable matrix, dependent variable vector and updated covariance kernel function, the posterior prediction distribution is derived. The prediction mean expression and prediction variance expression are extracted from the posterior prediction distribution. These two expressions are used to predict the performance mean and prediction uncertainty of any new sampling point, respectively. The prediction mean expression and prediction variance expression are encapsulated to construct a surrogate model.

[0024] In this embodiment, the batch Bayesian optimization algorithm is used to analyze and process the surrogate model, and a virtual sampling penalty factor based on the power consumption safety boundary of semiconductor devices is introduced. The specific steps for generating the predicted sampling point sequence are as follows: calling the surrogate model, setting the acquisition function of the batch Bayesian optimization algorithm, wherein the acquisition function is configured as the expected improvement function; calculating the prediction uncertainty distribution of the surrogate model in the global search space, and determining the expected value distribution of the acquisition function based on the prediction uncertainty distribution; using the multi-peak particle swarm optimization algorithm to perform a global extreme value search on the expected value distribution, locating the coordinates of the maximum expected value in the expected value distribution, and taking the coordinates of the maximum expected value as the first predicted sampling point; introducing a virtual sampling penalty factor based on semiconductor device power consumption safety boundary. The expected value distribution of the acquisition function is updated using a virtual sampling penalty factor constructed based on the power safety boundary of the semiconductor device. Then, a multi-peak particle swarm optimization algorithm is executed again to search for extreme values ​​within the updated expected value distribution of the acquisition function, locating the coordinates of the updated maximum expected value. These updated maximum expected value coordinates are then used as the second predicted sampling point. This process of updating the expected value distribution of the acquisition function using a virtual sampling penalty factor constructed based on the power safety boundary of the semiconductor device and searching for extreme values ​​is repeated until the total number of predicted sampling points reaches a preset batch sequence length. Finally, all predicted sampling points that have reached the preset batch sequence length are sorted according to their generation time to form a predicted sampling point sequence.

[0025] Specifically, the surrogate model is invoked, and the acquisition function of the batch Bayesian optimization algorithm is set. This acquisition function is configured as the expected improvement function, which is calculated by multiplying the expected portion of the predicted value that might exceed the current optimal value for each point within the entire unsampled search space. The prediction uncertainty distribution of the surrogate model in the global search space is calculated, i.e., the prediction variance of each unsampled point is calculated using the prediction variance expression in the surrogate model. Based on the prediction uncertainty distribution, the expected value distribution of the acquisition function is determined. A multi-peak particle swarm optimization algorithm is used to perform a global extremum search on the expected value distribution, locating the coordinates of the maximum expected value in the distribution. This coordinate is used as the first prediction sampling point. Subsequently, a virtual sampling penalty factor constructed based on the semiconductor device power consumption safety boundary is introduced to update the expected value distribution of the acquisition function. This penalty factor is a function defined as follows: ,in and These are the voltage and current at the sampling points. It is a preset safe power consumption threshold, which is based on the maximum power consumption value in the device datasheet. Settings, for example , This is a steepness factor, for example, set to 50, used to rapidly decay the function value when the power consumption approaches the safe threshold. The updated acquisition function is the original expected improvement function multiplied by this penalty factor. In the expected value distribution of the updated acquisition function, the multi-peak particle swarm optimization algorithm is executed again to search for extreme values ​​and locate the updated maximum expected value coordinates. Since the expected values ​​of the explored area and the high power consumption area are suppressed, the algorithm will tend to find extreme points in new, safe, and high uncertainty areas. The updated maximum expected value coordinates are used as the second prediction sampling point. The steps of introducing the virtual sampling penalty factor to update the acquisition function distribution and searching for extreme values ​​are executed cyclically until the total number of prediction sampling points reaches the preset batch sequence length, for example, the length is set to 10. These 10 prediction sampling points are sorted according to the order of their generation time to form a prediction sampling point sequence.

[0026] In this embodiment, the specific steps of encapsulating the predicted sampling point sequence into a list scan instruction are as follows: extract each predicted sampling point from the predicted sampling point sequence and parse the working mode identifier corresponding to each predicted sampling point; when the working mode identifier is voltage forced mode, extract the corresponding target voltage amplitude parameter from each predicted sampling point; when the working mode identifier is current forced mode, extract the corresponding target current amplitude parameter from each predicted sampling point; read the standard programmable instrument standard command protocol dictionary and match the basic control string corresponding to the working mode identifier according to the standard programmable instrument standard command protocol dictionary; convert the target voltage amplitude parameter and the target current amplitude parameter into hexadecimal machine code data, fill the hexadecimal machine code data into the parameter bits of the basic control string, and generate an independent execution action code; configure the hardware trigger delay time parameter between each predicted sampling point, concatenate the independent execution action code and the hardware trigger delay time parameter according to the arrangement order of the predicted sampling point sequence, add a header check bit and a tail end character, and compile to obtain the list scan instruction.

[0027] Specifically, each predicted sampling point in the predicted sampling point sequence is extracted. For example, the first point in the sequence is (5.1V, 0.01A, 'voltage_force'). The operating mode identifier of this point is parsed as voltage forced mode, and the corresponding target voltage amplitude parameter of 5.1V is extracted from this point. For another point in the sequence (2.0V, 1.5A, 'current_force'), its operating mode identifier is parsed as current forced mode, and the target current amplitude parameter of 1.5A is extracted. The standard programmable instrument standard command protocol dictionary is read. This is a preset mapping table that stores the correspondence between operating mode identifiers and SCPI basic command strings. For example, 'voltage_force' corresponds to 'SOUR:VOLT %s;MEAS:CURR?', and 'current_force' corresponds to 'SOUR:CURR %s;MEAS:VOLT?'. Based on the dictionary, the basic control string 'SOUR:VOLT' corresponding to the voltage forced mode is matched. `%s;MEAS:CURR?` converts the target voltage amplitude parameter 5.1V and the target current amplitude parameter 1.5A into ASCII string format conforming to the SCPI protocol, and fills them into the parameter bits of the corresponding basic control string. For example, filling 5.1V generates `'SOUR:VOLT 5.1;MEAS:CURR?'` as an independent action code. It configures the hardware trigger delay time parameter between each predicted sampling point. This parameter is set according to the device's response characteristics and settling time experience, for example, set to 10 milliseconds, and generates the corresponding SCPI command, such as `'TRIG:DEL 0.01'`. It then concatenates the independent action codes generated for all 10 predicted sampling points and the hardware trigger delay time parameters between them according to the order of the predicted sampling point sequence, forming a long string, such as `'SOUR:VOLT 5.1;MEAS:CURR?;TRIG:DEL 0.01;SOUR:CURR'`. 1.5;MEAS:VOLT?;...', add an initialization command such as '*RST' as a header checksum at the beginning of the string, and add an execution trigger command such as 'INIT' as a tail terminator at the end, and compile to obtain a list scan instruction.

[0028] In this embodiment, the steps of sending the list scan command to the source measurement unit for asynchronous hardware pipeline sampling processing to obtain the device performance sampling data set are as follows: establishing an asynchronous transmission control protocol connection channel between the host computer and the source measurement unit; capturing the list scan command in the communication buffer and sending the list scan command to the underlying hardware first-in-first-out queue of the source measurement unit through the asynchronous transmission control protocol connection channel; triggering the internal state machine of the source measurement unit to parse the list scan command in the underlying hardware first-in-first-out queue, and using the hardware trigger of the source measurement unit to drive the data according to the hardware trigger delay time parameter in the list scan command. The analog-to-digital converter (ADC) continuously applies excitation signals to the semiconductor device and synchronously drives the ADC to acquire response signals. While the source measurement unit performs hardware sampling, the central processing unit of the host computer asynchronously executes subsequent data analysis preparation work. After the source measurement unit completes all sampling actions corresponding to the list scan command, it converts the acquired response signals into real-valued result data strings. The real-valued result data strings are packaged and transmitted back to the host computer through the asynchronous transmission control protocol connection channel. The host computer parses the real-valued result data strings, extracts the actual performance data sets, and stores the actual performance data sets in the database to form a device performance sampling data set.

[0029] Specifically, an asynchronous transmission control protocol connection channel is established between the host computer and the source measurement unit, for example, through a VISA connection established via GPIB or LXI bus. The list scan instruction in the communication buffer is captured—that is, the long string generated in the previous process containing all sampling point operations. This list scan instruction is then sent in its entirety to the source measurement unit's underlying hardware FIFO queue via the asynchronous transmission control protocol connection channel. This triggers the source measurement unit's internal state machine to start working. The state machine parses the instructions in the FIFO queue one by one. When it parses 'SOUR:VOLT'... At step 5.1, the source measurement unit's hardware trigger, according to the hardware trigger delay time parameter set in the list scan instruction (e.g., 10 milliseconds), drives the digital-to-analog converter (DAC) to output a 5.1V excitation signal to the semiconductor device. Immediately after parsing 'MEAS:CURR?', it synchronously drives the analog-to-digital converter (ADC) to acquire the response current signal. While the source measurement unit autonomously executes this series of hardware pipeline operations, the host computer's central processing unit asynchronously performs subsequent data analysis preparation work, such as pre-allocating memory space to store the returned data or loading algorithm libraries that may be used in the next iteration. After the source measurement unit completes all 10 sampling actions in the list scan instruction, it converts all acquired response signals, such as 10 current and voltage values, into a real-valued result data string, such as a comma-separated string "0.012, The real-valued result data string, 1.45,..., is packaged and transmitted back to the host computer in one go via the asynchronous transmission control protocol connection channel. The real-valued result data string is parsed, the actual performance data set is extracted, and the actual performance data set is associated with the corresponding sampling point coordinates and stored in the local database to form a device performance sampling data set.

[0030] In this embodiment, the step of determining whether the device performance sampling data set meets the preset convergence condition is as follows: Extract the device performance sampling data set from the database, calculate the current total number of data points in the device performance sampling data set, and compare the current total number of data points with the preset maximum sampling threshold; if the current total number of data points reaches the preset maximum sampling threshold, then the preset convergence condition is met; if the current total number of data points does not reach the preset maximum sampling threshold, then calculate the hyperparameter change rate of the surrogate model between two consecutive prediction batches, and determine whether the hyperparameter change rate is less than the preset stability tolerance; if the hyperparameter change rate is less than the preset stability tolerance, then the preset convergence condition is met; if the hyperparameter change rate is greater than or equal to the preset stability tolerance, then extract the latest performance gradient value and the reference performance gradient value of the previous iteration batch from the device performance sampling data set, substitute the latest performance gradient value and the reference performance gradient value into the convergence determination formula for evaluation, and determine whether the evaluation index value meets the requirements. The formula is: ;in, Indicates the evaluation index value, Represents the gradient weight coefficients. This represents the latest performance gradient value. This represents the reference performance gradient value of the previous iteration batch. This represents a small normal number used to prevent the denominator from being zero; when the evaluation index value is less than the predetermined tolerance limit, the device performance sampling data set is determined to meet the preset convergence condition; when the evaluation index value is greater than or equal to the predetermined tolerance limit, the device performance sampling data set is determined not to meet the preset convergence condition.

[0031] Specifically, the device performance sampling data set is extracted from the database, and the current total number of data points in the device performance sampling data set is calculated. For example, if 150 data points have been collected so far, the current total number of data points is compared with the preset maximum sampling threshold. This threshold is a hard stopping condition set to prevent infinite loops and is set based on experience or test budget, for example, 500. If 150 does not reach 500, the process continues. The rate of change of hyperparameters of the surrogate model between two consecutive prediction batches is calculated, specifically by extracting the length-scale hyperparameters obtained from this iteration optimization. The length scale hyperparameter of the previous iteration The normalized difference is calculated, and it is determined whether the rate of change is less than a preset stability tolerance. This tolerance is an empirical value, for example, set to 0.01. If the rate of change is less than 0.01 for three consecutive iterations, the preset convergence condition is met. If the rate of change is greater than or equal to 0.01, the latest performance gradient value and the reference performance gradient value of the previous batch are extracted from the device performance sampling data set. For example, by fitting a curve to the latest batch of data, the slope at a specific voltage point is calculated as the latest performance gradient value. Similarly, the reference performance gradient value is obtained. The latest performance gradient value and the reference performance gradient value are substituted into the convergence criterion formula for evaluation to determine whether the evaluation index value meets the requirements. The formula is as follows: ,in, Indicates the evaluation index value, This represents the gradient weight coefficient, which is used to amplify or reduce the effect of gradient changes. It is set according to the sensitivity requirements for convergence speed, for example, set to 20. This represents the latest performance gradient value. This represents the reference performance gradient value of the previous iteration batch. This represents a tiny positive number used to prevent the denominator from being zero, usually set to a very small value such as 1e-8, when the calculated evaluation index value... When the value is less than a predetermined tolerance limit, such as 0.02, the device performance sampling data set is determined to meet the preset convergence condition. When the evaluation index value is greater than or equal to 0.02, the device performance sampling data set is determined not to meet the preset convergence condition.

[0032] In this embodiment, if the device performance sampling data set does not meet the preset convergence condition, the step of updating the initial sampling point set based on the device performance sampling data set and returning to the step of fitting the initial sampling point set using the Gaussian process regression algorithm to construct a surrogate model is as follows: If the preset convergence condition is met, the step of outputting the semiconductor device performance parameters based on the device performance sampling data set is as follows: When it is determined that the device performance sampling data set does not meet the preset convergence condition, the device performance sampling data set is appended and merged into the tail data block of the initial sampling point set to generate an expanded sampling point set. The expanded sampling point set is used as the new initial sampling point set, and the step of returning to the step of fitting the initial sampling point set using the Gaussian process regression algorithm is executed. The process involves fitting the data and constructing a surrogate model, followed by the next round of iterative calculations. When the device performance sampling data set meets the preset convergence condition, the iterative calculations terminate. The device performance sampling data set is then smoothed and filtered to remove isolated noise points and retain the core feature data. Curve smoothing interpolation is performed on the core feature data to plot the semiconductor device's current-voltage characteristic curves. Threshold voltage parameters, leakage current parameters, and breakdown voltage parameters are extracted from the semiconductor device's current-voltage characteristic curves and output as the semiconductor device's performance parameters.

[0033] Specifically, when the device performance sampling data set does not meet the preset convergence condition, the device performance sampling data set containing 10 data points from the most recent sampling is appended and merged into the tail data block of the existing initial sampling point set containing 150 data points, generating an expanded sampling point set containing 160 data points. This expanded sampling point set is used as the new initial sampling point set. The process then returns to the step of fitting this set of 160 data points using the Gaussian process regression algorithm and constructing a surrogate model, initiating the next round of iterative loop calculation. When the device performance sampling data set meets the preset convergence condition, for example, when the total number of sampling points reaches 320, the iterative loop calculation terminates. The entire 320 data points of the device performance sampling data set stored in the database are then subjected to smoothing filtering and cleaning, specifically using a Savitzky-Golay filter with a window length of 11. The polynomial order is 3. The data is processed to remove measurement noise. Then, the standard deviation of each data point and the local mean of its neighboring data points is calculated. Data points with a standard deviation exceeding 3 times the local standard deviation mean are identified as abnormal isolated noise points and removed. The remaining core feature data are retained. The core feature data are then subjected to curve smoothing interpolation using the cubic spline interpolation method to generate a high-resolution, continuously smooth semiconductor device current-voltage characteristic curve. Key performance parameters are extracted from the semiconductor device current-voltage characteristic curve. For example, the threshold voltage parameter is determined by drawing a tangent line in the linear region of the curve and finding its intersection with the voltage axis. The source-drain current value at zero gate voltage is read as the leakage current parameter. The voltage corresponding to the point on the curve where the current begins to increase sharply and the slope exceeds a predetermined value is found as the breakdown voltage parameter. The extracted threshold voltage parameter, leakage current parameter, and breakdown voltage parameter are output as semiconductor device performance parameters.

[0034] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An automatic detection method for performance parameters of semiconductor devices, characterized in that, Includes the following steps: Obtain the initial detection working range of the semiconductor device; Initial random sampling is performed within the initial detection working interval to obtain an initial set of sampling points; The Gaussian process regression algorithm is used to fit the initial set of sampling points to construct a surrogate model; The surrogate model is analyzed and processed using a batch Bayesian optimization algorithm, and a virtual sampling penalty factor based on the power consumption safety boundary of semiconductor devices is introduced to generate a sequence of predicted sampling points. The predicted sampling point sequence is encapsulated into a list scan instruction; The list scan command is sent to the source measurement unit for asynchronous hardware pipeline sampling processing to obtain a set of device performance sampling data. Determine whether the device performance sampling data set meets the preset convergence condition; If the device performance sampling data set does not meet the preset convergence condition, the initial sampling point set is updated based on the device performance sampling data set, and the step of fitting the initial sampling point set using the Gaussian process regression algorithm to construct a surrogate model is returned. If the preset convergence condition is met, the semiconductor device performance parameters are output based on the device performance sampling data set.

2. The automatic detection method for semiconductor device performance parameters according to claim 1, characterized in that, The specific steps for obtaining the initial detection working range of a semiconductor device are as follows: Obtain the device type identifier of the semiconductor device under test, retrieve the corresponding historical test parameter library based on the device type identifier, query the reference voltage scan lower limit value and reference voltage scan upper limit value in the historical test parameter library, and combine the reference voltage scan lower limit value and the reference voltage scan upper limit value to obtain the voltage operating range; The lower and upper limits of the reference current scan are queried from the historical test parameter database. The lower and upper limits of the reference current scan are combined to obtain the current working range. The voltage working range and the current working range are merged to construct a feature scan detection dimension matrix. The initial working range is determined according to the feature scan detection dimension matrix. The coordinates of the boundary nodes in the initial working range are recorded. The legality of the boundary node coordinates within the hardware tolerance range of the semiconductor device test platform is verified. When the boundary node coordinates are legal, the initial working range is locked as the initial detection working range of the semiconductor device.

3. The automatic detection method for semiconductor device performance parameters according to claim 2, characterized in that, The specific steps for performing initial random sampling within the initial detection working interval to obtain the initial sampling point set are as follows: The voltage working range and the current working range corresponding to the initial detection working range are obtained respectively. The Latin hypercube sampling algorithm is called to perform equalization and stratification processing on the voltage working range to obtain multiple voltage sampling levels. The Latin hypercube sampling algorithm is called to perform equalization and stratification processing on the current working range to obtain multiple current sampling levels. A first value is randomly selected from the multiple voltage sampling levels, and a second value is randomly selected from the multiple current sampling levels. The first value and the second value are paired to generate multiple initial sampling coordinate pairs. Duplicate coordinate data in the multiple initial sampling coordinate pairs are filtered and removed to obtain a set of deduplicated sampling coordinate pairs. The driving source measurement unit applies electrical signals point by point to the deduplicated sampling coordinate pair set, records the response current data and response voltage data fed back by the source measurement unit, and associates and binds the deduplicated sampling coordinate pair set, the response current data, and the response voltage data to form the initial sampling point set.

4. The automatic detection method for semiconductor device performance parameters according to claim 3, characterized in that, The specific steps for constructing the surrogate model by fitting the initial set of sampling points using the Gaussian process regression algorithm are as follows: Obtain the initial sampling point set, and separate the independent variable matrix and dependent variable vector in the initial sampling point set; initialize the mean function and covariance kernel function of the Gaussian process regression algorithm, wherein the covariance kernel function is configured as a squared exponential kernel function; input the independent variable matrix into the mean function to calculate the prior mean vector, input the independent variable matrix into the covariance kernel function to calculate the prior covariance matrix, and establish a multivariate Gaussian joint distribution function by combining the prior mean vector and the prior covariance matrix; use the maximum likelihood estimation method to iteratively optimize the hyperparameters in the multivariate Gaussian joint distribution function to obtain the optimal hyperparameter set; Substitute the optimal hyperparameter set into the covariance kernel function and update the covariance kernel function; Based on the independent variable matrix, the dependent variable vector, and the updated covariance kernel function, the posterior prediction distribution is derived. The prediction mean expression and the prediction variance expression in the posterior prediction distribution are extracted, and the prediction mean expression and the prediction variance expression are encapsulated to construct a surrogate model.

5. The automatic detection method for semiconductor device performance parameters according to claim 4, characterized in that, The surrogate model is analyzed and processed using a batch Bayesian optimization algorithm, and a virtual sampling penalty factor based on the power consumption safety boundary of semiconductor devices is introduced. The specific steps for generating the predicted sampling point sequence are as follows: The process involves: invoking the surrogate model and setting the acquisition function of the batch Bayesian optimization algorithm, wherein the acquisition function is configured as the expected improvement function; calculating the prediction uncertainty distribution of the surrogate model in the global search space, and determining the expected value distribution of the acquisition function based on the prediction uncertainty distribution; using a multi-peak particle swarm optimization algorithm to perform a global extreme value search on the expected value distribution, locating the coordinates of the maximum expected value in the expected value distribution, and using the coordinates of the maximum expected value as the first prediction sampling point; introducing a virtual sampling penalty factor based on the power safety boundary of semiconductor devices to update the expected value distribution of the acquisition function, and performing an extreme value search again on the updated expected value distribution of the acquisition function to locate the updated coordinates of the maximum expected value, and using the updated coordinates of the maximum expected value as the second prediction sampling point; repeatedly executing the steps of introducing a virtual sampling penalty factor based on the power safety boundary of semiconductor devices to update the expected value distribution of the acquisition function and the extreme value search until the total number of prediction sampling points reaches the preset batch sequence length; and sorting all prediction sampling points that have reached the preset batch sequence length according to the order of their generation to form the prediction sampling point sequence.

6. The automatic detection method for semiconductor device performance parameters according to claim 5, characterized in that, The specific steps for encapsulating the predicted sampling point sequence into a list scan instruction are as follows: Extract each predicted sampling point from the predicted sampling point sequence and parse the corresponding operating mode identifier for each predicted sampling point. When the operating mode identifier is voltage forced mode, extract the corresponding target voltage amplitude parameter from each predicted sampling point. When the operating mode identifier is current forced mode, extract the corresponding target current amplitude parameter from each predicted sampling point. Read the standard programmable instrument standard command protocol dictionary and match the basic control string corresponding to the operating mode identifier according to the standard programmable instrument standard command protocol dictionary. Convert the target voltage amplitude parameter and the target current amplitude parameter into hexadecimal machine code data, fill the hexadecimal machine code data into the parameter bits of the basic control string, and generate independent execution action code. Configure the hardware trigger delay time parameter between each predicted sampling point, concatenate the independent execution action code and the hardware trigger delay time parameter according to the arrangement order of the predicted sampling point sequence, add a header check bit and a tail end character, and compile to obtain the list scan instruction.

7. The automatic detection method for semiconductor device performance parameters according to claim 6, characterized in that, The specific steps for issuing the list scan command to the source measurement unit for asynchronous hardware pipeline sampling processing to obtain the device performance sampling data set are as follows: An asynchronous transmission control protocol connection channel is established between the host computer and the source measurement unit; the list scan instruction in the communication buffer is captured and sent to the underlying hardware first-in-first-out queue of the source measurement unit through the asynchronous transmission control protocol connection channel; the internal state machine of the source measurement unit is triggered to parse the list scan instruction in the underlying hardware first-in-first-out queue, and the hardware trigger of the source measurement unit drives the digital-to-analog converter according to the hardware trigger delay time parameter in the list scan instruction to continuously apply excitation signals to the semiconductor device and synchronously drive the analog-to-digital converter to collect response signals. While the source measurement unit is performing hardware sampling, the central processing unit of the host computer asynchronously executes subsequent data analysis preparation work; after the source measurement unit completes all sampling actions corresponding to the list scan instruction, the collected response signal is converted into a real number result data string; the real number result data string is packaged and sent back to the host computer through the asynchronous transmission control protocol connection channel; the real number result data string is parsed, the real performance data group is extracted, and the real performance data group is stored in the database to form the device performance sampling data set.

8. The automatic detection method for semiconductor device performance parameters according to claim 7, characterized in that, The specific steps for determining whether the device performance sampling data set meets the preset convergence conditions are as follows: Extract the device performance sampling data set from the database, calculate the current total amount of data in the device performance sampling data set, and compare the current total amount of data with a preset maximum sampling threshold; If the current total amount of data reaches the preset maximum sampling threshold, then the preset convergence condition is satisfied. If the current total amount of data does not reach the preset maximum sampling threshold, the hyperparameter change rate of the surrogate model between two consecutive prediction batches is calculated, and it is determined whether the hyperparameter change rate is less than the preset stability tolerance; if the hyperparameter change rate is less than the preset stability tolerance, it is determined that the preset convergence condition is met. If the rate of change of the hyperparameter is greater than or equal to the preset stability tolerance, then the latest performance gradient value and the reference performance gradient value of the previous iteration batch are extracted from the device performance sampling data set. The latest performance gradient value and the reference performance gradient value are substituted into the convergence determination formula for evaluation to determine whether the evaluation index value meets the requirements. The formula is as follows: ； in, Indicates the evaluation index value, Represents the gradient weight coefficients. This represents the latest performance gradient value. This represents the reference performance gradient value of the previous iteration batch. This represents a small positive integer used to prevent the denominator from being zero; When the evaluation index value is less than the predetermined tolerance limit, the device performance sampling data set is determined to meet the preset convergence condition; when the evaluation index value is greater than or equal to the predetermined tolerance limit, the device performance sampling data set is determined not to meet the preset convergence condition.

9. The automatic detection method for semiconductor device performance parameters according to claim 8, characterized in that, If the device performance sampling data set does not meet the preset convergence condition, the step of updating the initial sampling point set based on the device performance sampling data set and returning to the step of fitting the initial sampling point set using the Gaussian process regression algorithm to construct a surrogate model is as follows: If the preset convergence condition is met, the step of outputting the semiconductor device performance parameters based on the device performance sampling data set is as follows: When it is determined that the device performance sampling data set does not meet the preset convergence condition, the device performance sampling data set is appended and merged into the tail data block of the initial sampling point set to generate an expanded sampling point set. The expanded sampling point set is used as a new initial sampling point set, and the process of fitting the initial sampling point set and constructing a surrogate model using the Gaussian process regression algorithm is returned to perform the next round of iterative loop calculation. When it is determined that the device performance sampling data set meets the preset convergence condition, the iterative loop operation is terminated, and the device performance sampling data set is smoothed and cleaned by filtering to remove abnormal isolated noise points in the device performance sampling data set and retain the core feature data in the device performance sampling data set. The core feature data is subjected to curve smoothing interpolation to plot the current-voltage characteristic curve of the semiconductor device; the threshold voltage parameter, leakage current parameter and breakdown voltage parameter are extracted from the current-voltage characteristic curve of the semiconductor device, and the threshold voltage parameter, leakage current parameter and breakdown voltage parameter are output as the performance parameters of the semiconductor device.

Images