Failure physical model transfer learning method of MOS field effect transistor, storage medium and equipment

Abstract

The invention discloses an MOS field effect transistor failure physical model transfer learning method, a storage medium and equipment, and belongs to the technical field of MOS field effect transistor bias temperature instability failure physical modeling. The objective of the invention is to solve the problem that an existing transfer learning method cannot carry out transfer learning on a failure mechanism development rule of an MOS field effect transistor failure physical model, so that the effect of a migrated model is not ideal. According to the method, historical data and a simulation analysis method are combined, multiple correlation models of different candidate source fields are constructed, key parameters influencing failure physical model coefficients are determined, then according to the multiple correlation models, the distances between the different candidate source fields and a target field are calculated, and a mobility judgment criterion of the MOS field effect transistor is constructed; secondly, establishing an electronic component failure physical parameter identification model in the candidate source field, and constructing a parameter identification model of a target domain and a failure physical model coefficient influenced by the target domain by adopting an instance transfer learning method based on a transferable domain discrimination criterion;

Description

Technical Field

[0001] The invention belongs to the technical field of MOS field effect tube bias temperature instability failure physical modeling, and relates to a study of a MOS field effect tube failure physical model transfer learning method. Background Art

[0002] The failure physics model is a model that quantitatively describes the intrinsic physical and chemical processes of the failure phenomena of the underlying basic units (materials, structures, processes, etc.) of electronic components under various failure inducements (electrical, thermal, vibration stress, etc.). MOS field effect transistors are widely used in various types of electronic systems to ensure the reliable operation of electronic systems. However, in actual applications, MOS field effect transistors are affected by a variety of complex task profiles, resulting in their degradation, which in turn affects the reliability of electronic systems. Relevant statistics show that electronic system failures due to MOS field effect tube degradation account for more than 40% of the total. According to the typical failure mechanism of MOS field effect tubes, the failure physics model is established, and then applied to the reliability analysis of electronic systems, which is of great significance and application value.

[0003] At present, based on the typical failure modes and failure mechanisms of MOS field effect tubes, the performance parameter degradation data obtained based on accelerated degradation tests are obtained, and the failure physical model is established in combination with the typical mission profile of MOS field effect tubes. However, there are still some problems in the current research: the failure physical model currently established is only for the specific failure mechanism of a specific MOS field effect tube, and does not have the ability to generalize to other specifications of MOS field effect tubes of the same category. Therefore, how to establish a universal model is the key issue of the reliability prediction method based on failure physics. For MOS field effect tubes with the same failure mechanism, the essence of the change in its failure physical model is caused by the changes in its underlying key materials, structures, and processes. Analyzing its material, structure, and process characteristics, and establishing a migration model that characterizes the relationship between material, structure, process parameters and failure physical model coefficients can effectively solve the problems of high cost, long cycle, and poor applicability of failure physical models based entirely on degradation tests, and provide guarantees for reliability prediction. Summary of the invention

[0004] The present invention aims to solve the problem that the existing transfer learning method cannot perform transfer learning on the failure mechanism evolution law of the MOS field effect tube failure physical model, resulting in unsatisfactory model effect after transfer.

[0005] A failure physical model transfer learning method for MOS field effect transistors, comprising:

[0006] Step S1: Combine historical data and simulation analysis methods to construct complex correlation models for different candidate source fields, that is, to reflect the set of coefficients of the MOS field effect tube failure physical model. With the key parameter set The complex correlation model between them; the key parameters include material, structure, and process parameters; the complex correlation coefficients are sorted according to the calculation results, and the first multiple parameters are selected as the key parameters affecting the failure physical model coefficients and recorded as

[0007] Step S2: According to the complex correlation model of step S1, the distances between different candidate source domains and target domains are calculated, and the migration criterion of MOS field effect transistors is constructed:

[0008]

[0009] In the formula, Z is the similarity ordinal number, and rank represents the ranking function; represents the mean distance between the kth candidate source domain and the target domain, represents the distance variance between the kth candidate source domain and the target domain;

[0010] Step S3: for the candidate source domain, the key parameters in the candidate source domain are As input, the failure physics model coefficients affected by it As output, a physical parameter identification model of electronic component failure in the candidate source field is established;

[0011] Step S4: Based on the transferable domain discrimination criteria in step S2, the target domain is constructed using the instance transfer learning method. and the failure physical model coefficients affected by it Parameter identification model.

[0012] Furthermore, the process of constructing complex correlation models for different candidate source domains includes the following steps:

[0013] Step S11: Analyze the MOS field effect tube manufacturing process to determine

[0014] Step S12: key parameters and corresponding failure physical model coefficients obtained in step S11 A complex correlation model between key parameters and failure physical model coefficients is constructed, as shown in formulas (1) and (2):

[0015]

[0016] In the formula, is a set of complex correlation model coefficients, P represents the actual value of the model coefficient, and for the source domain P corresponds to The coefficients in Represents the average value of the model coefficient; R represents the complex correlation coefficient, and its value range is [0,1]. The larger the complex correlation coefficient, the closer the correlation between the key parameters and the failure physical model coefficient.

[0017] Step S13: Sort the complex correlation coefficients according to the calculation results and select the top s (k)n The parameters are the key parameters that affect the failure physical model coefficients. The key parameters selected according to the calculation results are recorded as

[0018] Furthermore, the key parameter set in step S11 is Select [T ox ,W,L,N ch ,N sub ], where T ox is the gate oxide thickness, W is the channel width, L is the channel length, N ch is the channel doping concentration, N sub is the substrate doping concentration.

[0019] Furthermore, the actual value of the model coefficient in step S12

[0020]

[0021] Furthermore, the key parameters in step S13 are For [T ox ,W,L].

[0022] Furthermore, the specific process of constructing the criterion for determining the mobility of the MOS field effect transistor in step S2 includes the following steps:

[0023] Step S21: Assume that the target domain is Datasets in Each candidate source field has an annotated dataset. The annotated dataset of the kth candidate source field is represented as Then the MMD distance between the key parameters of the candidate source domain and the target domain is as shown in formula (3):

[0024]

[0025] In the formula, Characterizes the maximum mean difference between the candidate source domain and the target domain, X T Indicates D T All instance features in Indicates D s(k) All instance features in ; Indicates D s(k) The jth and j'th instance features in the candidate source domain are the key parameters that affect the failure physical model coefficients; i 、x i' Indicates D T The i-th and i'th instance features in the target domain are the key parameters that affect the coefficients of the failure physical model; is the number of instances in the candidate source domain dataset; n T is the number of instances in the target domain dataset; k(·,·) represents the kernel function in the RKHS space;

[0026] Step S22: Calculate the variances of different candidate source fields and target fields, and obtain the variances of the same candidate source field and target field, as shown in formula (5):

[0027]

[0028] Where k(·,·) represents the kernel function in the RKHS space;

[0029] Step S23: Use an ordinal function to sort the distance means and variances of different candidate source domains and target domains to form a transferability criterion

[0030] Furthermore, the process of establishing the electronic component failure physical parameter identification model in the candidate source field in step S3 includes the following steps:

[0031] Step S31: Use a neural network to construct a system that reflects the key parameters and the failure physics model coefficients Parameter identification model of

[0032] Step S32: using an optimization algorithm to construct a parameter identification model for adaptively searching for the best combination of parameters unit and epoch.

[0033] Furthermore, the specific process of step S4 includes the following steps:

[0034] Step S41: Select the target domain D T The closest candidate source domain dataset Building a new dataset Indicates data set fusion;

[0035] Step S42: Initialize the sample weights ω of the source domain and the target domain Ti =ω Si , a target domain failure physical parameter identification model is constructed based on the TrAdaBoost algorithm; wherein, the instance features of the new data set are key parameters, the instance space annotations are failure physical model coefficients, and the established mapping model is the target domain failure physical parameter identification model;

[0036] The failure physical parameter identification model is established as shown in formula (6):

[0037] p D =F(x D) (6)

[0038] In the formula, F represents the parameter identification model based on AdaBoost, P D represents the instance space annotation of the reconstructed dataset D, x D Represents the instance features of the reconstructed dataset;

[0039] Step S43: Calculate the model coefficient prediction result and the actual model coefficient error. The error function is defined as:

[0040]

[0041] Among them, e iD is sample i D The relative error, is the true value of the annotation space, Characterization error normalization;

[0042] Step S44: Update the source domain dataset D according to the model error of step S43 s(k) Weight, until the error of formula (6) meets the requirements, the weight update formula is as follows;

[0043] ω T(N+1) =ω TN β (7)

[0044] ω S(N+1) =ω SN ·(1-β) (8)

[0045] Where ω TN ,ω T(i+1) represents the current target domain weight and the adjusted target domain weight; β represents the adjustment coefficient; ω SN ,ω S(i+1) Represents the current source domain weight and the adjusted source domain weight, and N represents the number of iterations.

[0046] A computer storage medium stores at least one instruction, and the at least one instruction is loaded and executed by a processor to implement a failure physical model transfer learning method for a MOS field effect tube.

[0047] A MOS field effect tube failure physical model transfer learning device, characterized in that the device includes a processor and a memory, the memory stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the MOS field effect tube failure physical model transfer learning method.

[0048] Beneficial effects:

[0049] The present invention provides a method for transferring and learning the failure physical model of electronic components. The method reveals the evolution law of MOS field effect tube materials, structures, processes and their internal failure mechanisms, analyzes the interaction law between key materials, structures, processes and their performance parameters and degradation processes, and quantitatively establishes the law of key parameters and failure physical model coefficients to form a migration model for the failure physical models that have been established for several specifications of products through the method of induction and summary. Therefore, the present invention can effectively improve the model effect after migration. Furthermore, through the migration model, the failure physical model coefficients can be quickly extrapolated and calculated for any new specification product related key parameters, so as to realize the rapid establishment of the failure physical model, effectively save the resources and time required for MOS field effect tube failure physical modeling, and lay the foundation for accurately establishing the MOS field effect tube reliability prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A flow chart of a failure physical model transfer learning method for MOS field effect transistors;

[0051] Figure 2 The result diagram of the target domain failure physical model coefficient A based on the TradaBoost parameter identification method;

[0052] Figure 3 The result diagram of the target domain failure physical model coefficient B based on the TradaBoost parameter identification method;

[0053] Figure 4 The result diagram of the target domain failure physical model coefficient Ea based on the TradaBoost parameter identification method;

[0054] Figure 5 The result diagram of the target domain failure physical model coefficient C based on the TradaBoost parameter identification method;

[0055] Figure 6 It represents the comparison chart of validation specification 1 based on TradaBoost fitting results and actual degradation data;

[0056] Figure 7 It represents the comparison chart of the degradation data fitted by TradaBoost and the actual degradation data for validation specification 2;

[0057] Figure 8 It represents the comparison chart of the validation specification 3 based on TradaBoost fitting results and actual degradation data;

[0058] Fig. 9 It shows the comparison chart of validation specification 4 based on TradaBoost fitting results and actual degradation data. DETAILED DESCRIPTION

[0059] In view of the above problems existing in the prior art, the present invention provides a MOS field effect tube failure physical model transfer learning method, which first combines historical data and simulation analysis methods to construct a complex correlation coefficient model of key parameters and failure physical models, and then uses an improved maximum mean difference method to calculate the distance of key parameters of MOS field effect tubes of different materials to construct a transferable domain model of MOS field effect tubes of different materials. After that, a function model of the same material related parameter vector and the failure physical model coefficient affected by it is constructed; finally, an example transfer learning method is used to construct a parameter identification migration model of the key materials, structures, and process vectors of MOS field effect tubes of different materials and the failure physical model coefficients affected by them.

[0060] The method reveals the evolution law of MOS field effect tube materials, structures, processes and their internal failure mechanisms, analyzes the interaction between key materials, structures, processes and their performance parameters and degradation processes, and establishes a migration model of different key materials, structures, processes and failure physical model coefficients, effectively saving the resources and time required for MOS field effect tube failure physical modeling. The present invention is further described below in conjunction with specific implementation methods. Before the specific description, the parameters and definitions are first described:

[0061] (1) Key parameters: electronic component materials, structures, and process parameters that affect the failure physical model coefficients.

[0062] (2) Candidate source fields: represents the candidate source domain, that is, there are a large number of labeled samples available for learning, represents the candidate source domain feature space. s(k) represents the kth candidate source domain dataset, Represents the instance characteristics of the candidate source domain, that is, the key parameters affecting the failure physical model, which is specifically expressed as It represents the spatial annotation of the source domain instance, that is, the failure physical model coefficient (the mathematical model coefficient of the failure mechanism of the MOS field effect tube paranoid temperature instability), which is specifically expressed as is the number of instances in the candidate source domain dataset.

[0063] (3) Target areas: represents the target domain, which usually has fewer labeled samples to learn from. represents the target domain feature space, P T (x) represents the marginal probability distribution of the target domain data. T represents the target domain dataset, x i is the target domain instance feature, and the target domain instance feature is expressed as represents instance space annotation, and the target domain instance space annotation is represented as n T is the number of instances in the target domain dataset.

[0064] Specific implementation method 1: Combination Figure 1 To explain this embodiment,

[0065] This embodiment is a failure physical model transfer learning method for MOS field effect transistors, comprising the following steps:

[0066] Step S1: Combine historical data and simulation analysis methods to construct complex correlation models for different candidate source fields, that is, to reflect the set of coefficients of the MOS field effect tube failure physical model. With the key parameter set The multiple correlation model between

[0067] The specific steps are as follows:

[0068] Step S11: Analyze the MOS field effect tube manufacturing process to determine The specific form depends on the actual situation. In this implementation, the materials, structures, and process parameters related to the failure physical model are selected as follows: ox ,W,L,N ch ,N sub ], where T ox is the gate oxide thickness, W is the channel width, L is the channel length, N ch is the channel doping concentration, N sub is the substrate doping concentration.

[0069] Step S12: key parameters obtained in step S11 [T ox ,W,L,N ch ,N sub ] and the corresponding failure physical model coefficients A complex correlation model between key parameters and failure physical model coefficients is constructed, as shown in formulas (1) and (2):

[0070]

[0071] In the formula, is a set of complex correlation model coefficients, P represents the actual value of the model coefficient, and for the source domain P corresponds to The coefficients in the target domain P correspond to The coefficients in ; Represents the average value of the model coefficient; R represents the multiple correlation coefficient, and its value range is [0,1]. The larger the multiple correlation coefficient is, the closer the correlation between the key parameters and the failure physical model coefficient is.

[0072] Step S13: Sort the complex correlation coefficients according to the calculation results, select some parameters as key parameters affecting the failure physical model coefficients (select according to the complex correlation coefficient R result), and the key parameters selected according to the calculation results are recorded as (In this embodiment, specifically [T ox ,W,L]).

[0073] The same method as the candidate source field is used to establish a complex correlation model for the target field. The materials, structures, and process parameters related to the failure physical model are selected. After verification, the influence of the failure physical model coefficients of the target field is basically the same. The number of some parameters for the target field is selected to be the same as that of some parameters in the source field.

[0074] Step S2: According to the complex correlation model of step S1, based on the improved maximum mean difference method, the distance between different candidate source domains and the target domain is calculated to construct the criterion for the transferability of MOS field effect tubes; the specific steps are as follows:

[0075] Step S21: Assume that the target domain is Datasets in There are a large number of annotated datasets in each candidate source field. The annotated dataset of the kth candidate source field is represented as Then the MMD distance between the key parameters of the candidate source domain and the target domain is as shown in formula (3):

[0076]

[0077] In the formula, Characterizes the maximum mean difference between the candidate source domain and the target domain, X T Indicates D T All instance features in Indicates D s(k) All instance features in ; Indicates D s(k) The jth and j'th instance features in the candidate source domain are the key parameters that affect the failure physical model coefficients; i 、x i' Indicates D T The i-th and i'th instance features in the target domain are the key parameters that affect the coefficients of the failure physical model; is the number of instances in the candidate source domain dataset; n T is the number of instances in the target domain dataset; k(·,·) represents the kernel function in the RKHS space.

[0078] Step S22: Calculate the variance of different candidate source domains and target domains according to the statistical knowledge variance calculation formula. The calculation process is as shown in formula (5):

[0079] Variance formula:

[0080]

[0081] Among them, E(x) represents expectation, and x represents the data set to be calculated.

[0082] Substituting formula (3) into formula (4), we can obtain the variance of the candidate source domain and the target domain, as shown in formula (5):

[0083]

[0084] Where k(·,·) represents the kernel function in the RKHS space;

[0085] Step S23: Use an ordinal function to sort the distance means and variances of different candidate source domains and target domains to form a transferability judgment criterion. The specific formula is shown in (5):

[0086]

[0087] In the formula, the similarity ordinal number Z ranges from [1, M] and is an integer. represents the mean distance between the kth candidate source domain and the target domain, represents the distance variance between the kth candidate source domain and the target domain; rank represents the ranking function.

[0088] When Z takes the minimum value of 1, it means that the similarity between the candidate source domain and the target domain is the smallest; when Z takes the maximum value of M, it means that the similarity between the candidate source domain and the target domain is the largest.

[0089] Step S3: for the candidate source domain, the key parameters in the candidate source domain are As input, the failure physics model coefficients affected by it As output, a physical parameter identification model of electronic component failure in the candidate source field is established; the specific steps are as follows:

[0090] Step S31: Use a neural network to construct a system that reflects the key parameters and the failure physics model coefficients Parameter identification model.

[0091] Step S32: Use the bat optimization algorithm to construct a parameter identification model for adaptively searching for the best combination of parameters unit and epoch.

[0092] Step S4: Based on the transferable domain discrimination criteria in step S2, the target domain is constructed using the instance transfer learning method. and the failure physical model coefficients affected by it The parameter identification model of ; the specific steps are as follows:

[0093] Step S41: Select the target domain D T The closest candidate source domain dataset Building a new dataset Indicates data set fusion;

[0094] Step S42: Initialize the sample weights ω of the source domain and the target domain Ti =ω Si , based on the TrAdaBoost algorithm, a target domain failure physical parameter identification model is constructed. That is, although the data in the candidate source domain and the target domain are different, there is still a part of the candidate source domain that is similar to the target domain data. Find the instance features that are suitable for the target domain from the candidate source domain data, and reconstruct the data set by combining these instance features with the target domain features. Take the new dataset instance feature x D As input, the new dataset instance space annotation p D As the output, a TrAdaBoost-based mapping model is established, in which the instance features of the new data set are key parameters, the instance space annotations are failure physical model coefficients, and the established mapping model is the target domain failure physical parameter identification model.

[0095] The failure physical parameter identification model is established as shown in formula (6):

[0096] p D =F(x D ) (6)

[0097] In the formula, F represents the parameter identification model based on AdaBoost, P D represents the instance space annotation of the reconstructed dataset D, x D Represents the instance features of the reconstructed dataset.

[0098] Step S43: Calculate the model coefficient prediction result and the actual model coefficient error. The error function is defined as:

[0099]

[0100] Among them, e iD is sample i D The relative error, is the true value of the annotation space, Characterization error normalization;

[0101] Step S44: Update the source domain dataset D according to the model error of step S43 s(k)Weight, until the error of formula (6) meets the requirements, the weight update formula is as follows;

[0102] ω T(N+1) =ω TN β (7)

[0103] ω S(N+1) =ω SN ·(1-β) (8)

[0104] Where ω TN ,ω T(i+1) represents the current target domain weight and the adjusted target domain weight; β represents the adjustment coefficient; ω SN ,ω S(i+1) Represents the current source domain weight and the adjusted source domain weight, and N represents the number of iterations.

[0105] Example:

[0106] Step S1: Determine the MOS field effect tube failure physical model

[0107]

[0108] In the formula, V th (t) is the threshold voltage at time t, V th (0) is the initial value of the threshold voltage, A BTI is the BTI effect factor, B is the electric field acceleration factor, E a is the failure activation energy, k is the Boltzmann constant, and C is the time power exponent.

[0109] Combining historical data and simulation analysis methods, a physical model coefficient P=[P1, P2, ... P n ] and the materials, structures, and processes of MOS field effect tubes X = [x1, x2, ..., x n ] to determine the key parameters that affect the failure physical model coefficients:

[0110] First, combining historical data and simulation methods, the key parameters of MOS field effect tubes are determined as: M = [T ox ,W,L,N ch ,N sub ]; where T ox is the gate oxide thickness, W is the channel width, L is the channel length, N ch is the channel doping concentration, N sub is the substrate doping concentration, and the failure physical model coefficient is [A, B, E a ,C].

[0111] Then, according to formula (1) and formula (2), the gate oxide thickness, channel width, channel length, channel doping concentration, N sub is the complex correlation model of substrate doping concentration and failure physical model coefficient, and the complex correlation coefficient is calculated. The specific results are shown in Table 1:

[0112] Table 1 Multiple correlation coefficients of different parameter combinations

[0113]

[0114] Finally, according to the results in Table 1, T ox The complex correlation coefficient under the combination of W and L parameters is the largest, so the key parameter combination that affects the MOS failure physical model is determined to be T ox ,W,L, that is, the first three parameters are selected as the key parameters affecting the failure physical model coefficients.

[0115] Step S2: Based on the complex correlation model of step S1, an improved maximum mean difference method is proposed to calculate the distance between the target domain MOS field effect transistor and the candidate source domain MOS field effect transistor. The detailed construction parameters of the candidate source domain data set experiment are shown in Table 2, where D s(1) -D s(6) Indicates the candidate source area, and 1-6 are codes for different materials.

[0116] Table 2 Experimental settings of the double-moon artificial dataset

[0117]

[0118] The effectiveness of MMD variance divergence in measuring the distribution differences between domains: The experimental results are shown in Table 3

[0119] Table 3 MMD distance Distance variance σ 2 , MMD similarity ordinal number Z and migration accuracy Acc

[0120]

[0121] Table 3 lists the target areas and candidate source fields MMD distance between Distance variance σ 2 , MMD similarity ordinal Z, and the migration accuracy Acc obtained by using the transfer learning algorithm TrAdaBoost to migrate each candidate source domain to the target domain. As the k value increases, the candidate source domain Target Area The larger the actual distribution difference is, the greater the difference is. The following conclusions can be drawn from the analysis of the experimental results: As the distribution difference between the candidate source domain and the target domain gradually increases, the MMD distance between the target domain and the candidate source domain calculated Distance variance σ 2 , the MMD similarity index Z gradually decreases.

[0122] Step S3: construct the same material parameter vector M = [T ox ,W,L] and the failure physical model coefficient P = [A,B,E a ,C] function mapping model; the specific steps are as follows:

[0123] According to different stress combinations, the failure physical model is established, the failure physical model coefficients under different stress combinations are obtained, and the failure physical model transfer learning research based on KNN is carried out. The source domain data is converted into multivariate feature sample data through PolynomialFeatures processing, and the high-dimensional space data samples are obtained after data normalization. The high-dimensional Euclidean distance between the key parameters of the MOS field effect tube of the test set specification and the key parameters of the training specification is calculated, and sorted in ascending order of distance, and then K points with the smallest distance to the key parameters of the MOS field effect tube of the specification to be predicted are selected. Finally, the average value of the values ​​of the K points is calculated as the output of the mapping model.

[0124] Step S4: Based on the transferable domain model in step S2, an instance transfer learning method is used to construct a target domain MOS field effect tube failure physical migration model:

[0125] Select the candidate source domain dataset that is closest to the target domain of the key parameter and construct a new dataset Taking the key manufacturing parameters of the target field as input and the model coefficients as output, the TradaBoost method is used to establish the physical migration model of electronic component failure of new materials. Figure 2 The coefficient A of the target domain failure physical model is based on the result of the TradaBoost parameter identification method. Figure 3 The coefficient B of the target domain failure physical model is based on the result of the TradaBoost parameter identification method. Figure 4 The coefficient Ea of the target domain failure physical model is based on the result of the TradaBoost parameter identification method. Figure 5 The coefficient C of the target domain failure physical model is based on the result of the TradaBoost parameter identification method. Figure 2-Figure 5 It shows that the coefficients of the PoF model identified based on TradaBoost parameters are not much different from the actual ones, which is consistent with the actual change trend.

[0126] In order to further prove the effectiveness of this method, the model coefficients calculated based on TradaBoost are introduced into the failure physical model to obtain the average degradation trajectory of the verification specification, as shown in Figure 6-Figure 9 As shown, Figure 6 It indicates that the validation specification 1 is based on the comparison between the TradaBoost fitting results and the actual degradation data. Figure 7 It indicates that the validation specification 2 is based on the comparison between the TradaBoost fitted degradation data and the actual degradation data. Figure 8 It indicates that the validation specification 3 is based on the comparison between the TradaBoost fitting results and the actual degradation data. Fig. 9 It indicates the comparison between the fitting results of TradaBoost and the actual degradation data of verification specification 4. The degradation curves of the fitting data and the actual data are analyzed. Table 5 shows that the actual average life of the MOS field effect tube stress reaches 500 hours, and the error of the failure physical migration method based on TradaBoost. By comparison, it is found that the maximum value of the average life error of the method proposed in the present invention is only 0.0281, which proves the effectiveness of the method.

[0127] Table 4 Comparison of fitting RMSE of the two methods

[0128] Specific implementation method 2:

[0130] This embodiment is a computer storage medium, in which at least one instruction is stored. The at least one instruction is loaded and executed by a processor to implement the failure physical model transfer learning method of a MOS field effect transistor.

[0131] It should be understood that the instructions include computer program products, software or computerized methods corresponding to any method described in the present invention; the instructions can be used to program a computer system, or other electronic device. Computer storage media may include readable media on which instructions are stored, which may include but are not limited to magnetic storage media, optical storage media; magneto-optical storage media include read-only memory ROM, random access memory RAM, erasable programmable memory (e.g., EPROM and EEPROM) and flash memory layers, or other types of media suitable for storing electronic instructions. Specific implementation method three:

[0133] This embodiment is a MOS field effect tube failure physical model transfer learning device, the device includes a processor and a memory, it should be understood that including any device including a processor and a memory described in the present invention, the device may also include other units and modules that perform display, interaction, processing, control, etc. and other functions through signals or instructions;

[0134] At least one instruction is stored in the memory, and the at least one instruction is loaded and executed by the processor to implement the failure physical model transfer learning method of a MOS field effect transistor.

[0135] Those skilled in the art will appreciate that at least one instruction stored is a computer program product corresponding to the method or system. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes. The scheme in the embodiment of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal scripting language JavaScript, etc.

[0136] The present application is described with reference to the flowcharts and / or block diagrams of the methods, systems, and computer program products according to the embodiments of the present application, and may also be used for corresponding devices. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0137] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0138] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0139] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0140] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

[0141] The above calculation examples of the present invention are only used to explain the calculation model and calculation process of the present invention in detail, and are not intended to limit the implementation methods of the present invention. For ordinary technicians in the relevant field, other different forms of changes or modifications can be made based on the above description. It is impossible to list all the implementation methods here. All obvious changes or modifications derived from the technical solution of the present invention are still within the scope of protection of the present invention.

Claims

1. A failure physical model transfer learning method for MOS field effect transistors, characterized in that: include: Step S1: Combine historical data and simulation analysis methods to construct complex correlation models for different candidate source fields, that is, to reflect the set of coefficients of the MOS field effect tube failure physical model. With the key parameter set The complex correlation model between them; the key parameters include material, structure, and process parameters; the complex correlation coefficients are sorted according to the calculation results, and the first multiple parameters are selected as the key parameters affecting the failure physical model coefficients and recorded as Step S2: According to the complex correlation model of step S1, the distances between different candidate source domains and target domains are calculated, and the migration criterion of MOS field effect transistors is constructed: In the formula, Z is the similarity ordinal number, and rank represents the ranking function; represents the mean distance between the kth candidate source domain and the target domain, represents the distance variance between the kth candidate source domain and the target domain; Step S3: for the candidate source domain, the key parameters in the candidate source domain are As input, the failure physics model coefficients affected by it As output, a physical parameter identification model of electronic component failure in the candidate source field is established; Step S4: Based on the transferable domain discrimination criteria in step S2, the target domain is constructed using the instance transfer learning method. and the failure physical model coefficients affected by it Parameter identification model.

2. The method for transferring the failure physical model of a MOS field effect tube according to claim 1, characterized in that: The process of building complex correlation models for different candidate source domains includes the following steps: Step S11: Analyze the MOS field effect tube manufacturing process to determine Step S12: key parameters and corresponding failure physical model coefficients obtained in step S11 A complex correlation model between key parameters and failure physical model coefficients is constructed, as shown in formulas (1) and (2): In the formula, is a set of complex correlation model coefficients, P represents the actual value of the model coefficient, and for the source domain P corresponds to The coefficients in Represents the average value of the model coefficient; R represents the complex correlation coefficient, and its value range is [0,1]. The larger the complex correlation coefficient, the closer the correlation between the key parameters and the failure physical model coefficient. Step S13: Sort the complex correlation coefficients according to the calculation results and select the top s (k)n The parameters are the key parameters that affect the failure physical model coefficients. The key parameters selected according to the calculation results are recorded as 3. The method for transferring the failure physical model of a MOS field effect tube according to claim 2, characterized in that: Key parameter set in step S11 Select [T ox ,W,L,N ch ,N sub ], where T ox is the gate oxide thickness, W is the channel width, L is the channel length, N ch is the channel doping concentration, N sub is the substrate doping concentration.

4. The method for transferring the failure physical model of a MOS field effect tube according to claim 3 is characterized in that: The actual value of the model coefficient in step S12 5. The method for transferring the failure physical model of a MOS field effect tube according to claim 4 is characterized in that: Key parameters in step S13 For [T ox ,W,L].

6. A method for transferring and learning a failure physical model of a MOS field effect transistor according to any one of claims 1 to 5, characterized in that: The specific process of constructing the criterion for determining the mobility of the MOS field effect transistor in step S2 includes the following steps: Step S21: Assume that the target domain is Datasets in Each candidate source field has an annotated dataset. The annotated dataset of the kth candidate source field is represented as Then the MMD distance between the key parameters of the candidate source domain and the target domain is as shown in formula (3): In the formula, Characterizes the maximum mean difference between the candidate source domain and the target domain, X T Indicates D T All instance features in Indicates D s(k) All instance features in ; Indicates D s(k) The jth and j'th instance features in the candidate source domain are the key parameters that affect the failure physical model coefficients; i 、x i' Indicates D T The i-th and i'th instance features in the target domain are the key parameters that affect the coefficients of the failure physical model; is the number of instances in the candidate source domain dataset; n T is the number of instances in the target domain dataset; k(·,·) represents the kernel function in the RKHS space; Step S22: Calculate the variances of different candidate source fields and target fields, and obtain the variances of the same candidate source field and target field, as shown in formula (5): Where k(·,·) represents the kernel function in the RKHS space; Step S23: Use an ordinal function to sort the distance means and variances of different candidate source domains and target domains to form a transferability criterion 7. The method for transferring the failure physical model of a MOS field effect transistor according to claim 6, characterized in that: The process of establishing the electronic component failure physical parameter identification model in the candidate source field in step S3 includes the following steps: Step S31: Use a neural network to construct a system that reflects the key parameters and the failure physics model coefficients Parameter identification model of Step S32: using an optimization algorithm to construct a parameter identification model for adaptively searching for the best combination of parameters unit and epoch.

8. The method for transferring the failure physical model of a MOS field effect transistor according to claim 6, characterized in that: The specific process of step S4 includes the following steps: Step S41: Select the target domain D T The closest candidate source domain dataset Building a new dataset Indicates data set fusion; Step S42: Initialize the sample weights ω of the source domain and the target domain Ti =ω Si , a target domain failure physical parameter identification model is constructed based on the TrAdaBoost algorithm; wherein, the instance features of the new data set are key parameters, the instance space annotations are failure physical model coefficients, and the established mapping model is the target domain failure physical parameter identification model; The failure physical parameter identification model is established as shown in formula (6): p D =F(x D ) (6) In the formula, F represents the parameter identification model based on AdaBoost, P D represents the instance space annotation of the reconstructed dataset D, x D Represents the instance features of the reconstructed dataset; Step S43: Calculate the model coefficient prediction result and the actual model coefficient error. The error function is defined as: Among them, e iD is sample i D The relative error, is the true value of the annotation space, Characterization error normalization; Step S44: Update the source domain dataset D according to the model error of step S43 s(k) Weight, until the error of formula (6) meets the requirements, the weight update formula is as follows; oh T(N+1) =ω TN ·b (7) oh S(N+1) =ω SN ·(1-b) (8) Where ω TN ,ω T(i+1) represents the current target domain weight and the adjusted target domain weight; β represents the adjustment coefficient; ω SN ,ω S(i+1) Represents the current source domain weight and the adjusted source domain weight, and N represents the number of iterations.

9. A computer storage medium, characterized in that The storage medium stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the failure physical model transfer learning method of a MOS field effect transistor as described in any one of claims 1 to 8.

10. A MOS field effect tube failure physical model transfer learning device, characterized in that: The device includes a processor and a memory, wherein the memory stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement a failure physical model transfer learning method for a MOS field effect transistor as described in any one of claims 1 to 8.

Images