A field effect tube intelligent testing method and system based on electrical variables

Abstract

The present invention discloses a field effect tube intelligent testing method and system based on electrical variables, and relates to the field of field effect tube testing technology. A field effect tube intelligent testing system based on electrical variables includes: an intelligent testing module and a test analysis module. The present invention can reduce the number of measurements of each field effect tube by screening and dynamically measuring multiple field effect tubes on the wafer to be tested, saving time and resources; automatically select the field effect tube to be tested according to the dynamic changes of the voltage and current characteristic curves, and reduce unnecessary tests and repeated measurements by optimizing the measurement process; use the field effect tube characteristic feature extraction model to strengthen the feature learning process to remove redundant features and ensure that the extracted features have strong correlation; independently collect the strongly correlated features related to each field effect tube to be tested, so as to obtain accurate test results and help make more reliable and scientific decisions.

Description

Technical Field

[0001] The present invention relates to the field of field effect tube testing technology, and in particular to a field effect tube intelligent testing method and system based on electrical variables. Background Art

[0002] A field-effect transistor (FET) is a three-terminal device made of semiconductor material that operates by controlling the flow of current through an electric field. Due to its high input impedance, low power consumption, and high efficiency, FETs are widely used in various electronic devices, including switching circuits, amplifiers, and radio frequency circuits. In practice, FETs are not typically found alone but are arranged in a pattern on a wafer to manufacture integrated circuit chips.

[0003] FET testing is typically performed on a wafer-by-wafer basis, rather than testing each FET individually. Wafer-level testing evaluates the performance of all FETs on the entire wafer. This is done to screen out substandard devices before the wafer is cut into individual chips, thus avoiding waste in subsequent packaging and testing.

[0004] However, when testing a single FET, its performance may be affected by factors such as wafer fabrication non-uniformity, material defects, and production process fluctuations. A single FET tested cannot fully reflect its performance in actual applications, especially when multiple devices work together, where the performance of the FET is affected by the interconnection effects of other devices. Therefore, to fully evaluate the performance of the FET, a detailed analysis must be combined with dynamic electrical variable measurement technology to ensure its reliability and stability in actual applications. Summary of the Invention

[0005] The present invention aims to provide a field effect tube intelligent testing method and system based on electrical variables, which can perform linked intelligent testing on field effect tubes.

[0006] A field effect transistor intelligent testing method based on electrical variables comprises the following steps:

[0007] The wafer to be tested contains a total of n field effect transistors H to be tested. n , n=1, 2, ..., N; each measurement screens M field effect transistors to be tested on the wafer to be tested for dynamic measurement, and obtains the voltage and current characteristic curve X to be analyzed i , i=1, 2, ..., I; traverse all the field effect tubes to be tested and perform dynamic measurements until all the field effect tubes to be tested H n After at least two dynamic measurements have been performed, the dynamic measurement is stopped, and I voltage-current characteristic curve to be analyzed is obtained;

[0008] Based on the voltage and current characteristic curve X to be analyzed iAnalyze the field effect tube characteristic feature extraction model and obtain the voltage and current strong correlation feature Y i The field effect transistor characteristic feature extraction model is used to enhance feature learning of the voltage and current characteristic curves to be analyzed, remove redundant features, and obtain features with strong correlation.

[0009] For the field effect tube H under test n , collect all the FETs H n The related voltage and current strong correlation feature Y i , get the field effect tube test analysis set F n ; Based on the field effect tube test analysis set F n Analyze with the FET intelligent test analysis model to obtain the FET test result G n ; According to the field effect tube test results G n The field effect transistor H under test in the wafer to be tested n Perform subsequent operations.

[0010] As a preferred technical solution of the present invention, the specific steps of screening the field effect transistors to be tested based on the wafers to be tested and performing dynamic measurement include:

[0011] S1. Field effect tube H to be tested n The corresponding test index is Z n Among them, Z n =W1*(2-A n )+W2*B n ; A n Indicates the field effect tube H to be tested n The corresponding number of tests, W1 is A n The corresponding weight; B n Indicates the field effect tube H to be tested n The corresponding prediction test trend, W2 is B n The corresponding weight;

[0012] S2. According to the test index Z n , select M field effect tubes H to be tested n Perform fixed gate voltage measurement and randomly adjust the drain voltage to obtain a fixed gate voltage drain current curve; according to the corresponding M field effect transistors H to be tested n Perform measurements with a fixed drain voltage and randomly adjust the gate voltage to obtain a fixed drain voltage-drain current curve.

[0013] S3. Fusing the fixed gate voltage drain current curve and the fixed drain voltage drain current curve to obtain the voltage-current characteristic curve to be analyzed X i ;

[0014] S4. After completing steps S2-S3, re-output the test index Z based on the field effect tube prediction test model n The size of the test is calculated and the next dynamic measurement is performed until all the tests have been performed for A times. n When both are greater than or equal to 2, stop dynamic prediction.

[0015] As a preferred technical solution of the present invention, the field effect transistor prediction test model includes a frequency change layer, a test trend prediction layer and an index output layer;

[0016] The times change layer is used to change the field effect transistor H to be tested selected in step S2. n A n Perform an add operation to get the new number of tests A n ;

[0017] The test trend prediction layer is used to obtain the H n The test data of the field effect tube to be tested is obtained n ; Based on the test data P of the field effect tube to be tested n and current prediction test trends B n Perform predictive analysis to obtain new predictive test trend B n ;

[0018] Specific steps for training and testing the trend prediction layer:

[0019] Constructing an initial test trend prediction layer based on a BP neural network model; collecting several groups of prediction test training samples, each group of prediction test training samples containing test data, current test trend data, and corresponding test trend labels; using several groups of prediction test training samples to perform model training on the initial test trend prediction layer to obtain a trained test trend prediction layer;

[0020] The exponential output layer is used to calculate the number of times A has been tested. n and new prediction test trend B n Output the changed test index Z n .

[0021] As a preferred technical solution of the present invention, the field effect tube characteristic feature extraction model includes a smooth curve layer, a curve feature extraction layer, a strong correlation feature extraction layer and a feature output layer;

[0022] The smooth curve layer is used to analyze the voltage and current characteristic curve X i Perform smoothing preprocessing to obtain a preprocessing voltage-current characteristic curve;

[0023] The curve feature extraction layer is used to extract corner points of the preprocessed voltage and current characteristic curve using the variable-scale Gaussian kernel function and CSS technology to obtain the voltage and current feature sequence;

[0024] The strong correlation feature extraction layer is used to extract the strong correlation features of the voltage and current feature sequence to obtain the voltage and current strong correlation feature Y i ;

[0025] The feature output layer is used to output the voltage and current strong correlation feature Y i .

[0026] As a preferred technical solution of the present invention, the specific steps of extracting strong correlation features in the strong correlation feature extraction layer include:

[0027] Based on the Bootstrap technology, K voltage and current strongly correlated individuals T are extracted from the voltage and current feature sequence. k , k=1, 2, …, K; set the maximum number of iterations;

[0028] Based on voltage and current to screen strong correlation individuals T k Construct a decision tree and obtain the voltage and current importance decision tree R k ; Based on the voltage and current importance decision tree R k Perform importance assessment and obtain the decision tree performance score Q k ;

[0029] When the maximum number of iterations is reached, the voltage and current strongly correlated individual to be screened corresponding to the maximum performance score of the current decision tree is output, that is, the optimal voltage and current strongly correlated individual to be screened; based on the optimal voltage and current strongly correlated individual to be screened, the voltage and current strongly correlated feature Y is output. i .

[0030] As a preferred technical solution of the present invention, the field effect transistor intelligent test analysis model includes a data preprocessing layer, a feature learning layer and a result output layer;

[0031] The data preprocessing layer is used to analyze the field effect transistor test set F n Perform preprocessing to obtain the preprocessed field effect tube test analysis set F n ';

[0032] The feature learning layer is used to pre-process the field effect transistor test analysis set F n 'Perform feature learning and obtain the field effect tube test result G n ;

[0033] The result output layer is used to output the field effect tube test results G n .

[0034] As a preferred technical solution of the present invention, the specific steps of performing feature learning in the feature learning layer include:

[0035] The feature learning layer consists of d stacked encoder layers and decoder layers. Each encoder layer contains a multi-head attention layer and a fully connected feedforward neural network layer.

[0036] Collecting several groups of defect feature learning training samples; each group of defect feature learning training samples includes annotated test result labels and corresponding feature sets; combining several groups of defect feature learning training samples to obtain a defect feature learning training set;

[0037] The defect feature learning training set is input into the feature learning layer for model training to obtain the initial feature learning layer; the initial feature learning layer is evaluated. If the initial feature learning layer passes the model evaluation, the initial feature learning layer is used as the feature learning layer in the field effect transistor intelligent test and analysis model; otherwise, the defect feature learning training set is used to continue model training.

[0038] An intelligent field effect transistor test system based on electrical variables, comprising:

[0039] The intelligent test module includes an intelligent test unit and a feature extraction unit; the intelligent test unit is used to test a total of n field effect transistors H on the wafer to be tested. n , n=1, 2, ..., N; each measurement screens M field effect transistors to be tested on the wafer to be tested for dynamic measurement, and obtains the voltage and current characteristic curve X to be analyzed i , i=1, 2, ..., I; traverse all the field effect tubes to be tested and perform dynamic measurements until all the field effect tubes to be tested H n After at least two dynamic measurements have been performed, the dynamic measurement is stopped to obtain I voltage and current characteristic curves to be analyzed; the feature extraction unit is used to extract the voltage and current characteristic curves X based on the voltage and current characteristic curves to be analyzed. i Analyze the field effect tube characteristic feature extraction model and obtain the voltage and current strong correlation feature Y i The field effect transistor characteristic feature extraction model is used to enhance feature learning of the voltage and current characteristic curves to be analyzed, remove redundant features, and obtain features with strong correlation.

[0040] The test analysis module includes a test analysis unit; the test analysis unit is used to test the field effect transistor H n , collect all the FETs H n The related voltage and current strong correlation feature Y i , get the field effect tube test analysis set F n ; Based on the field effect tube test analysis set F n Analyze with the FET intelligent test analysis model to obtain the FET test result G n ; According to the field effect tube test results G n The field effect transistor H under test in the wafer to be testedn Perform subsequent operations.

[0041] The present invention has the following advantages:

[0042] 1. The present invention can reduce the number of measurements of each field-effect transistor by screening and dynamically measuring multiple field-effect transistors on the wafer to be tested, saving time and resources; automatically selecting the field-effect transistor to be tested according to the dynamic changes of the voltage-current characteristic curve can reduce unnecessary tests and repeated measurements by optimizing the measurement process; using the field-effect transistor characteristic feature extraction model, the feature learning process is strengthened to remove redundant features and ensure that the extracted features have strong correlation; for each field-effect transistor to be tested, the strongly correlated features related to it are independently collected, which can obtain accurate test results and help make more reliable and scientific decisions.

[0043] 2. The voltage-current characteristic curve to be analyzed obtained by feature fusion in the present invention can comprehensively consider the influence of gate voltage and drain voltage, improve the comprehensiveness of the test, which helps to reduce the errors that may be caused by a single measurement method and ensure more accurate test results; after each test, intelligent adjustments are made based on historical data and predicted trends, which can make the decision of each test more accurate and help reduce uncertainty and errors in the testing process. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a structural diagram of a field effect transistor intelligent testing system based on electrical variables adopted in an embodiment of the present invention. DETAILED DESCRIPTION

[0045] In order to enable those skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0046] Example 1, a field effect transistor intelligent testing method based on electrical variables, comprising the following steps:

[0047] The wafer to be tested contains a total of n field effect transistors H to be tested. n , n=1, 2, ..., N; each measurement screens M field effect transistors to be tested on the wafer to be tested for dynamic measurement, and obtains the voltage and current characteristic curve X to be analyzed i , i=1, 2, ..., I; traverse all the field effect tubes to be tested and perform dynamic measurements until all the field effect tubes to be tested H n After at least two dynamic measurements have been performed, the dynamic measurement is stopped, and I voltage-current characteristic curve to be analyzed is obtained;

[0048] The specific steps for screening the field effect transistors to be tested based on the wafers to be tested and performing dynamic measurements include:

[0049] S1. Field effect tube H to be tested n The corresponding test index is Z n Among them, Z n =W1*(2-A n )+W2*B n ; A n Indicates the field effect tube H to be tested n The corresponding number of tests, W1 is A n The corresponding weight; B n Indicates the field effect tube H to be tested n The corresponding prediction test trend, W2 is B n The corresponding weights; the specific values ​​of W1 and W2 are set by professional technicians based on actual conditions;

[0050] S2. According to the test index Z n , select M field effect tubes H to be tested n Perform fixed gate voltage measurement and randomly adjust the drain voltage to obtain a fixed gate voltage drain current curve; according to the corresponding M field effect transistors H to be tested n Perform measurements with a fixed drain voltage and randomly adjust the gate voltage to obtain a fixed drain voltage-drain current curve.

[0051] S3. Fusing the fixed gate voltage drain current curve and the fixed drain voltage drain current curve to obtain the voltage-current characteristic curve to be analyzed X i ;

[0052] S4. After completing steps S2-S3, re-output the test index Z based on the field effect tube prediction test model n The size of the test is calculated and the next dynamic measurement is performed until all the tests have been performed for A times. n When both are greater than or equal to 2, stop dynamic prediction;

[0053] By calculating the test index Zn, combined with the number of times the field effect tube to be tested has been tested and the predicted test trend, the field effect tube that needs to be measured next is intelligently selected. This weighted selection method can avoid repeated measurements on field effect tubes that have been fully tested, thereby improving measurement efficiency and reducing resource waste; each dynamic measurement is based on the test index Zn. n By understanding the size of the FET and the test progress of the FET, test resources can be allocated more reasonably, giving priority to testing those FETs with larger predicted test trends to ensure uniform test coverage;

[0054] By measuring the current curves at fixed gate voltage and fixed drain voltage respectively and fusing the features of the two current curves, more comprehensive voltage and current characteristic data can be provided. This fusion strategy helps to extract richer features, improve the recognition and accuracy of features, and ensure that the model can better reflect the actual performance of the field effect tube. By combining the number of tests and the predicted test trend, an intelligent prediction mechanism based on historical data and trend analysis can be formed to ensure that the test of the field effect tube to be tested is more accurate and avoid redundant tests. By using the prediction model in step S4 to analyze the test index Z n Recalculation is performed and the measurement sequence is dynamically adjusted to ensure that the number of tests for each FET is properly allocated before reaching two, avoiding overtesting or missing key test points. This can significantly save time and resources and avoid the waste caused by repeated measurements. After each round of measurement, the test index of each FET is re-evaluated and the test strategy is adjusted based on the output of the prediction model. This real-time dynamic optimization makes the measurement process more flexible and efficient.

[0055] The voltage-current characteristic curve to be analyzed, obtained through feature fusion, can comprehensively consider the effects of gate and drain voltages, improving the comprehensiveness of the test. This helps reduce the errors that may be introduced by a single measurement method and ensures more accurate test results. Intelligent adjustments are made after each test based on historical data and predicted trends, making each test decision more accurate and helping to reduce uncertainty and errors during the test process.

[0056] The FET prediction test model includes a frequency change layer, a test trend prediction layer, and an index output layer;

[0057] The times change layer is used to change the field effect transistor H to be tested selected in step S2. n A n Perform an add operation to get the new number of tests A n ;

[0058] The test trend prediction layer is used to obtain the H n The test data of the field effect tube to be tested is obtained n ; Based on the test data P of the field effect tube to be tested n and current prediction test trends B n Perform predictive analysis to obtain new predictive test trend B n ;

[0059] Specific steps for training and testing the trend prediction layer:

[0060] Constructing an initial test trend prediction layer based on a BP neural network model; collecting several groups of prediction test training samples, each group of prediction test training samples containing test data, current test trend data, and corresponding test trend labels; using several groups of prediction test training samples to perform model training on the initial test trend prediction layer to obtain a trained test trend prediction layer;

[0061] The exponential output layer is used to calculate the number of times A has been tested. n and new prediction test trend B n Output the changed test index Z n ;

[0062] Through the frequency change layer and the test trend prediction layer, the model can automatically adjust the test strategy based on historical data after each test, and promptly correct the prediction of the field-effect transistor to be tested, thereby enhancing the intelligence and adaptability of the test process. By training the test trend prediction layer, the test trend of the field-effect transistor can be dynamically predicted based on the test data, so that the test process can adapt to the characteristics of different field-effect transistors in real time. This intelligent prediction mechanism avoids the defects of traditional manual parameter setting and improves prediction accuracy. Through the index output layer, the model can output the changed test index based on the new number of tests and the new predicted test trend, providing an accurate test strategy for each field-effect transistor. This precise decision-making mechanism helps improve the reliability of test results and ensures that the test progress and goals are accurately matched.

[0063] Based on the voltage and current characteristic curve X to be analyzed i Analyze the field effect tube characteristic feature extraction model and obtain the voltage and current strong correlation feature Y i The field effect transistor characteristic feature extraction model is used to enhance feature learning of the voltage and current characteristic curves to be analyzed, remove redundant features, and obtain features with strong correlation.

[0064] The field effect tube characteristic feature extraction model includes a smooth curve layer, a curve feature extraction layer, a strong correlation feature extraction layer and a feature output layer;

[0065] The smooth curve layer is used to analyze the voltage and current characteristic curve X i Perform smoothing preprocessing to obtain the preprocessed voltage and current characteristic curve; the smoothing curve layer smoothes the voltage and current characteristic curve to be analyzed, effectively removing noise and abnormal fluctuations in the original curve, making subsequent feature extraction more stable and accurate; especially in complex voltage and current data, noise often affects the results of subsequent analysis, and smoothing helps improve data quality;

[0066] The curve feature extraction layer is used to extract corner points from the preprocessed voltage and current characteristic curves using a variable-scale Gaussian kernel function and CSS technology to obtain a voltage and current feature sequence. In this layer, the variable-scale Gaussian kernel function and CSS technology are used to extract corner points from the preprocessed voltage and current characteristic curves, capturing key feature points in the curves. This method can help identify significant change points in the voltage and current characteristics, such as threshold voltage, conduction region, saturation region, and other key features, providing an effective basis for subsequent feature analysis.

[0067] The strong correlation feature extraction layer is used to extract the strong correlation features of the voltage and current feature sequence to obtain the voltage and current strong correlation feature Y i The strongly correlated feature extraction layer performs reinforcement learning on the extracted feature sequences, removes redundancy, and extracts features with strong correlation. This ensures that the extracted features are highly representative of the voltage and current characteristics, effectively improving the accuracy of subsequent analysis. This process can effectively reduce the dimension of the feature space and improve the efficiency of the model.

[0068] The feature output layer is used to output the voltage and current strong correlation feature Y i ;

[0069] Through the strong correlation feature extraction layer, the model can remove irrelevant and redundant parts from the original voltage and current features, retaining only the features that are strongly correlated with the voltage and current characteristics. This not only simplifies the feature set, but also improves the recognition and interpretation of the features, thereby improving the prediction and classification accuracy of the model. By strengthening feature learning and removing redundant features, the model can rely only on the most informative features in subsequent test analysis, avoiding the problem of data overfitting. By extracting and retaining features with strong correlation, it can make it easier for testers to understand the impact of each feature on the performance of the field-effect transistor. For example, key parameters such as threshold voltage, saturation region characteristics, leakage current, etc. can be well reflected in the features, thereby helping engineers better understand the physical meaning of the test data.

[0070] The specific steps of extracting strong correlation features in the strong correlation feature extraction layer include:

[0071] Based on the Bootstrap technology, K voltage and current strongly correlated individuals T are extracted from the voltage and current feature sequence. k , k=1, 2, ..., K; set the maximum number of iterations; the maximum number of iterations is set by professional technicians according to actual conditions;

[0072] Based on voltage and current to screen strong correlation individuals T k Construct a decision tree and obtain the voltage and current importance decision tree R k ; Based on the voltage and current importance decision tree R kPerform importance assessment and obtain the decision tree performance score Q k ;

[0073] When the maximum number of iterations is reached, the voltage and current strongly correlated individual to be screened corresponding to the maximum performance score of the current decision tree is output, that is, the optimal voltage and current strongly correlated individual to be screened; based on the optimal voltage and current strongly correlated individual to be screened, the voltage and current strongly correlated feature Y is output. i ;

[0074] By extracting multiple strongly correlated individuals of voltage and current to be screened from the voltage and current feature sequence through Bootstrap sampling, the model can avoid the bias caused by single sampling, making the feature selection process more stable; the Bootstrap method provides a diverse set of candidate features by resampling the data, enhancing the adaptability and robustness of the model in different test scenarios; when using decision trees for feature screening, the risk of overfitting can be reduced through multiple iterations. Each training of the decision tree is carried out on a different sampling data set, which makes the final screened features more universal and reduces the model's over-reliance on specific data sets; by constructing a decision tree based on each individual voltage and current feature to be screened, the model can evaluate the importance of each feature based on the feature's contribution to the decision-making process. The decision tree can identify and weigh the impact of each feature on the classification or regression results, thereby providing an intuitive feature importance ranking; the entire feature extraction process consists of multiple steps such as Bootstrap sampling, decision tree construction, feature importance evaluation, and iterative optimization. These steps are highly automated, which can reduce human intervention and improve the efficiency and reliability of feature selection;

[0075] For the field effect tube H under test n , collect all the FETs H n The related voltage and current strong correlation feature Y i , get the field effect tube test analysis set F n ; Based on the field effect tube test analysis set F n Analyze with the FET intelligent test analysis model to obtain the FET test result G n ; According to the field effect tube test results G n The field effect transistor H under test in the wafer to be tested n Perform subsequent operations;

[0076] The FET intelligent test analysis model includes a data preprocessing layer, a feature learning layer, and a result output layer;

[0077] The data preprocessing layer is used to analyze the field effect transistor test set F n Perform preprocessing to obtain the preprocessed field effect tube test analysis set F n ';

[0078] The feature learning layer is used to pre-process the field effect transistor test analysis set F n 'Perform feature learning and obtain the field effect tube test result G n ;

[0079] The result output layer is used to output the field effect tube test results G n ;

[0080] The specific steps for feature learning in the feature learning layer include:

[0081] The feature learning layer consists of d stacked encoder layers and decoder layers. Each encoder layer contains a multi-head attention layer and a fully connected feedforward neural network layer.

[0082] Collecting several groups of defect feature learning training samples; each group of defect feature learning training samples includes annotated test result labels and corresponding feature sets; combining several groups of defect feature learning training samples to obtain a defect feature learning training set;

[0083] Input the defect feature learning training set into the feature learning layer for model training to obtain the initial feature learning layer; perform model evaluation on the initial feature learning layer. If the initial feature learning layer passes the model evaluation, the initial feature learning layer is used as the feature learning layer in the FET intelligent test and analysis model; otherwise, continue model training using the defect feature learning training set;

[0084] The data preprocessing layer preprocesses the field effect transistor test analysis set to remove noise, fill missing values, standardize data or perform other necessary cleaning operations. This process can effectively improve the quality of input data and provide more accurate and consistent input for subsequent feature learning, thereby ensuring the accuracy and reliability of the field effect transistor test results Gn; in the feature learning layer, stacked encoder layers and decoder layers are used. Through the combination of multi-head attention mechanism and fully connected feedforward neural network, key features can be effectively extracted in high-dimensional feature space, which enhances the understanding and prediction ability of the field effect transistor characteristics to be tested, and further improves the accuracy of the test results; feature The multi-head attention layer in the learning layer adaptively focuses on the most important parts of the input data, allowing the model to focus on features that are critical to predicting test results. The attention mechanism helps the model learn input data from multiple perspectives simultaneously, capturing more potential correlations and patterns, thereby improving its ability to recognize the characteristics of different field-effect transistors. The combination of stacked encoder and decoder layers enables the model to learn abstract features layer by layer and map them to the final output space. Each encoder and decoder layer is responsible for feature representation at a specific level. The model can automatically learn complex hierarchical features in the data, making the feature extraction process highly adaptive.

[0085] During the test, the test result G n It may reveal whether the performance of the field-effect transistor to be tested meets the standard requirements; if the test results show that the performance parameters of the field-effect transistor such as the threshold voltage, gain or switching characteristics are lower or higher than the predetermined range, the system will mark it as unqualified. Suppose during the measurement process, the threshold voltage of a field-effect transistor exceeds the predetermined operating range, causing it to malfunction. In this case, the field-effect transistor will be marked as unqualified and removed from the wafer to prevent it from entering the downstream process or assembly link; if the test results show that all performance parameters are within the qualified range, the field-effect transistor will be considered qualified and enter the subsequent processes, such as packaging, inspection, etc.

[0086] Example 2, a field effect tube intelligent testing system based on electrical variables, see Figure 1 Shown, including:

[0087] The intelligent test module includes an intelligent test unit and a feature extraction unit; the intelligent test unit is used to test a total of n field effect transistors H on the wafer to be tested. n , n=1, 2, ..., N; each measurement screens M field effect transistors to be tested on the wafer to be tested for dynamic measurement, and obtains the voltage and current characteristic curve X to be analyzed i , i=1, 2, ..., I; traverse all the field effect tubes to be tested and perform dynamic measurements until all the field effect tubes to be tested H n After at least two dynamic measurements have been performed, the dynamic measurement is stopped to obtain I voltage and current characteristic curves to be analyzed; the feature extraction unit is used to extract the voltage and current characteristic curves X based on the voltage and current characteristic curves to be analyzed. i Analyze the field effect tube characteristic feature extraction model and obtain the voltage and current strong correlation feature Y i The field effect transistor characteristic feature extraction model is used to enhance feature learning of the voltage and current characteristic curves to be analyzed, remove redundant features, and obtain features with strong correlation.

[0088] The test analysis module includes a test analysis unit; the test analysis unit is used to test the field effect transistor H n , collect all the FETs H n The related voltage and current strong correlation feature Y i , get the field effect tube test analysis set F n ; Based on the field effect tube test analysis set F n Analyze with the FET intelligent test analysis model to obtain the FET test result G n ; According to the field effect tube test results G n The field effect transistor H under test in the wafer to be tested n Perform subsequent operations.

[0089] It should be understood that those skilled in the art may make improvements or modifications based on the above description, and all such improvements and modifications shall fall within the scope of protection of the appended claims. Any portion of this specification not described in detail is prior art known to those skilled in the art.

Claims

1. A field effect transistor intelligent testing method based on electrical variables, characterized in that: The following steps are involved: The wafer to be tested contains a total of n field effect transistors H to be tested. n , n=1, 2, ..., N; each measurement screens M field effect transistors to be tested on the wafer to be tested for dynamic measurement, and obtains the voltage and current characteristic curve X to be analyzed i , i=1,2,…,I; Traverse all the field effect tubes to be tested and perform dynamic measurements until all the field effect tubes to be tested H n After at least two dynamic measurements have been performed, the dynamic measurement is stopped, and I voltage-current characteristic curve to be analyzed is obtained; Based on the voltage and current characteristic curve X to be analyzed i Analyze the field effect tube characteristic feature extraction model and obtain the voltage and current strong correlation feature Y i The field effect transistor characteristic feature extraction model is used to enhance feature learning of the voltage and current characteristic curves to be analyzed, remove redundant features, and obtain features with strong correlation. For the field effect tube H under test n , collect all the FETs H n The related voltage and current strong correlation feature Y i , get the field effect tube test analysis set F n ;based on Field Effect Transistor Test Analysis Set F n Analyze with the FET intelligent test analysis model to obtain the FET test result G n ; According to the field effect tube test results G n The field effect transistor H under test in the wafer to be tested n Perform subsequent operations; The specific steps for screening the field effect transistors to be tested based on the wafers to be tested and performing dynamic measurements include: S1. Field effect tube H to be tested n The corresponding test index is Z n Among them, Z n =W1*(2-A n )+W2*B n ; A n Indicates the field effect tube H to be tested n The corresponding number of tests, W1 is A n The corresponding weight; B n Indicates the field effect tube H to be tested n The corresponding prediction test trend, W2 is B n The corresponding weight; S2. According to the test index Z n , select M field effect tubes H to be tested n Perform fixed gate voltage measurement and randomly adjust the drain voltage to obtain a fixed gate voltage drain current curve; according to the corresponding M field effect transistors H to be tested n Perform measurements with a fixed drain voltage and randomly adjust the gate voltage to obtain a fixed drain voltage-drain current curve. S3. Fusing the fixed gate voltage drain current curve and the fixed drain voltage drain current curve to obtain the voltage-current characteristic curve to be analyzed X i ; S4. After completing steps S2-S3, re-output the test index Z based on the field effect tube prediction test model n The size of the test is calculated and the next dynamic measurement is performed until all the tests have been performed for A times. n When both are greater than or equal to 2, stop dynamic prediction.

2. The method for intelligent testing of field effect transistors based on electrical variables according to claim 1, characterized in that: The FET prediction test model includes a frequency change layer, a test trend prediction layer, and an index output layer; The times change layer is used to change the field effect transistor H to be tested selected in step S2. n A n Perform an add operation to get the new number of tests A n ; The test trend prediction layer is used to obtain the H n The test data of the field effect tube to be tested is obtained n ; Based on the test data P of the field effect tube to be tested n and current prediction test trends B n Perform predictive analysis to obtain new predictive test trend B n ; Specific steps for training and testing the trend prediction layer: Constructing an initial test trend prediction layer based on a BP neural network model; collecting several groups of prediction test training samples, each group of prediction test training samples containing test data, current test trend data, and corresponding test trend labels; using several groups of prediction test training samples to perform model training on the initial test trend prediction layer to obtain a trained test trend prediction layer; The exponential output layer is used to calculate the number of times A has been tested. n and new prediction test trend B n Output the changed test index Z n .

3. The method for intelligent testing of field effect transistors based on electrical variables according to claim 2, characterized in that: The field effect tube characteristic feature extraction model includes a smooth curve layer, a curve feature extraction layer, a strong correlation feature extraction layer and a feature output layer; The smooth curve layer is used to analyze the voltage and current characteristic curve X i Perform smoothing preprocessing to obtain a preprocessing voltage-current characteristic curve; The curve feature extraction layer is used to extract corner points of the preprocessed voltage and current characteristic curve using the variable-scale Gaussian kernel function and CSS technology to obtain the voltage and current feature sequence; The strong correlation feature extraction layer is used to extract the strong correlation features of the voltage and current feature sequence to obtain the voltage and current strong correlation feature Y i ; The feature output layer is used to output the voltage and current strong correlation feature Y i .

4. The method for intelligent testing of field effect transistors based on electrical variables according to claim 3, characterized in that: The specific steps of extracting strong correlation features in the strong correlation feature extraction layer include: Based on the Bootstrap technology, K voltage and current strongly correlated individuals T are extracted from the voltage and current feature sequence. k , k=1, 2, …, K; set the maximum number of iterations; Based on the voltage and current to screen the strong correlation individual T k Construct a decision tree and obtain the voltage and current importance decision tree R k ; Based on the voltage and current importance decision tree R k Perform importance assessment and obtain the decision tree performance score Q k ; When the maximum number of iterations is reached, the voltage and current strongly correlated individual to be screened corresponding to the maximum performance score of the current decision tree is output, that is, the optimal voltage and current strongly correlated individual to be screened; based on the optimal voltage and current strongly correlated individual to be screened, the voltage and current strongly correlated feature Y is output. i .

5. The method for intelligent testing of field effect transistors based on electrical variables according to claim 4, characterized in that: The FET intelligent test analysis model includes a data preprocessing layer, a feature learning layer, and a result output layer; The data preprocessing layer is used to analyze the field effect transistor test set F n Perform preprocessing to obtain the preprocessed field effect tube test analysis set F n '; The feature learning layer is used to pre-process the field effect transistor test analysis set F n 'Perform feature learning and obtain the field effect tube test result G n ; The result output layer is used to output the field effect tube test results G n .

6. The method for intelligent testing of field effect transistors based on electrical variables according to claim 5, characterized in that: The specific steps for feature learning in the feature learning layer include: The feature learning layer consists of d stacked encoder layers and decoder layers. Each encoder layer contains a multi-head attention layer and a fully connected feedforward neural network layer. Collecting several groups of defect feature learning training samples; each group of defect feature learning training samples includes annotated test result labels and corresponding feature sets; combining several groups of defect feature learning training samples to obtain a defect feature learning training set; The defect feature learning training set is input into the feature learning layer for model training to obtain the initial feature learning layer; the initial feature learning layer is evaluated. If the initial feature learning layer passes the model evaluation, the initial feature learning layer is used as the feature learning layer in the field effect transistor intelligent test and analysis model; otherwise, the defect feature learning training set is used to continue model training.

7. An intelligent field effect transistor test system based on electrical variables, characterized in that: The system applies the field effect transistor intelligent testing method based on electrical variables as described in any one of claims 1 to 6, including: The intelligent test module includes an intelligent test unit and a feature extraction unit; the intelligent test unit is used to test a total of n field effect transistors H on the wafer to be tested. n , n=1, 2, ..., N; each measurement screens M field effect transistors to be tested on the wafer to be tested for dynamic measurement, and obtains the voltage and current characteristic curve X to be analyzed i , i=1, 2, ..., I; traverse all the field effect tubes to be tested and perform dynamic measurements until all the field effect tubes to be tested H n After at least two dynamic measurements have been performed, the dynamic measurement is stopped to obtain I voltage and current characteristic curves to be analyzed; the feature extraction unit is used to extract the voltage and current characteristic curves X based on the voltage and current characteristic curves to be analyzed. i Analyze the field effect tube characteristic feature extraction model and obtain the voltage and current strong correlation feature Y i The field effect transistor characteristic feature extraction model is used to enhance feature learning of the voltage and current characteristic curves to be analyzed, remove redundant features, and obtain features with strong correlation. The test analysis module includes a test analysis unit; the test analysis unit is used to test the field effect transistor H n , collect all the FETs H n The related voltage and current strong correlation feature Y i , get the field effect tube test analysis set F n ; Based on the field effect tube test analysis set F n Analyze with the FET intelligent test analysis model to obtain the FET test result G n ; According to the field effect tube test results G n The field effect transistor H under test in the wafer to be tested n Perform subsequent operations.

Images