A wafer surface defect detection method, system, electronic device and medium

Abstract

The application discloses a wafer surface defect detection method and system, electronic equipment and medium, and relates to the field of semiconductor wafer manufacturing.The method divides a plurality of defective and non-defective pictures of each type of wafer into a support set and a query set to form a task set; a deep learning model and a MAML framework are constructed; a MAML inner loop is executed, and the support set in part of the task set is taken for model training; the learning rate in the training is updated by using an ADAM algorithm, and the inner layer parameters are updated by using a gradient descent method; then a MAML outer loop is executed, the model is continuously trained by using the query set, the loss functions of all tasks are added together as the loss function of the outer loop, and the outer layer parameters are updated by using the gradient descent method; the MAML inner loop and the outer loop are repeatedly executed until the model converges.The application can obtain a defect detection model with better robustness by using a smaller amount of training samples, and the defect detection model has strong generalization ability, high efficiency and strong practicability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of semiconductor wafer manufacturing, in particular to a wafer surface defect detection method and system, an electronic device and a medium. BACKGROUND

[0002] In recent years, the development of computer science and artificial intelligence has driven the development of the semiconductor industry, and the demand for chips in various industries has greatly increased. Therefore, it is of great economic significance to improve the yield of semiconductor wafers, and defect detection is undoubtedly an important step in improving the yield. In the manufacturing process of semiconductor wafers, each process is closely linked, and the technology is complex. Small changes in factors such as materials, environment, and process parameters often lead to defects in chips, such as contamination, scratches, and edge collapse. If these defects are not found and removed in time, it will lead to waste of production resources and more defective products entering the market, causing economic losses.

[0003] Traditional machine vision methods have the problems of slow speed, low accuracy, and poor generalization performance. Usually, a new batch of wafers needs to adjust the defect detection algorithm, which is a tedious and inefficient process. Although the current popular conventional deep learning method is powerful, it requires a large number of data samples to train a good detection model. However, in the real industrial scene, the defect samples of wafers are very few, so it is difficult for the conventional deep learning method to achieve good results. SUMMARY

[0004] The purpose of the present application is to provide a wafer surface defect detection method and system, an electronic device and a medium, which can use a small amount of training samples to obtain a defect detection model with good robustness, and has strong generalization ability, high efficiency and strong practicality.

[0005] To achieve the above purpose, the present application provides the following scheme:

[0006] In one aspect, the present application provides a wafer surface defect detection method, comprising:

[0007] Collecting wafer pictures of different types of chips to form a task set; the different types of chips include pressure sensor chips, LED chips, and infrared sensor chips; the task set includes a support set and a query set;

[0008] Constructing an MAML framework based on a plurality of deep learning models, and randomly initializing the parameters of the plurality of deep learning models to obtain the initialization parameters of the MAML inner loop;

[0009] Performing an MAML inner loop based on the MAML framework, training the initialization parameters of the MAML inner loop using the support set in part of the task set, and obtaining the updated parameter vector of the MAML inner loop;

[0010] performing an MAML outer loop based on the MAML framework, training the parameter vector updated by the MAML inner loop by using a query set in the partial task set, to obtain a parameter vector updated by the MAML outer loop;

[0011] In the process of performing the MAML outer loop, the loss functions of each deep learning model are added together as a total loss function of the outer loop, and it is determined whether the total loss function converges;

[0012] If the total loss function does not converge, the parameter vector updated by the MAML outer loop is taken as the initialization parameter of the MAML inner loop, and the step of "performing an MAML inner loop based on the MAML framework, training the initialization parameter of the MAML inner loop by using a support set in the partial task set, to obtain a parameter vector updated by the MAML inner loop" is returned;

[0013] If the total loss function converges, the parameter vector updated by the MAML outer loop is assigned to the plurality of deep learning models, to obtain a surface defect detection model for different types of chip wafers;

[0014] The surface defect detection model is used for surface defect detection on a wafer picture to be detected.

[0015] Optionally, the wafer pictures of different types of chips are collected to form a task set, specifically including:

[0016] For each type of chip wafer in T types of chip wafers, k defective wafer pictures and k non-defective wafer pictures are collected respectively, to form k pairs of wafer pictures;

[0017] q pairs of wafer pictures in the k pairs of wafer pictures are divided into a support set, and the remaining k-q pairs of wafer pictures are taken as a query set, to jointly form T task sets.

[0018] Optionally, the step of performing an MAML inner loop based on the MAML framework, training the initialization parameter of the MAML inner loop by using a support set in the partial task set, to obtain a parameter vector updated by the MAML inner loop, specifically includes:

[0019] t task sets are randomly sampled from the T task sets as a partial task set;

[0020] The support set in the partial task set is used to train the plurality of deep learning models, the learning rate in the training is updated by using the update step of the ADAM algorithm, the initialization parameter of the MAML inner loop is updated by using the gradient descent method, until all wafer pictures of the support set are traversed, to obtain a parameter vector updated by the MAML inner loop.

[0021] Optionally, the MAML outer loop is executed based on the MAML framework, and the parameter vector updated by the MAML inner loop is trained by using a query set in the partial task set to obtain a parameter vector updated by the MAML outer loop, and the method specifically comprises the following steps.

[0022] The parameter vector updated by the MAML inner loop is used to initialize a plurality of deep learning model parameters of the MAML outer loop, and the plurality of deep learning models are further trained by using the query set in the partial task set, and the plurality of deep learning model parameters are updated by using the gradient descent method until all wafer pictures in the query set are traversed, and a parameter vector updated by the MAML outer loop is obtained.

[0023] Optionally, the surface defect detection model is used for surface defect detection of the wafer picture to be detected, and the method specifically comprises the following steps.

[0024] The chip type corresponding to the wafer picture to be detected is determined as a target type.

[0025] Defective and non-defective wafer pictures of the target type chip are collected to form a training set.

[0026] A surface defect detection model corresponding to the target type chip is extracted, and parameter training is performed by using the training set until all wafer pictures in the training set are traversed, and a surface defect detection model specially for the target type chip is obtained.

[0027] The surface defect detection model specially for the target type chip is used for surface defect detection of the wafer picture to be detected.

[0028] In another aspect, the application further provides a wafer surface defect detection system, comprising:

[0029] A wafer picture acquisition module is configured to acquire wafer pictures of different types of chips to form a task set, wherein the different types of chips include pressure sensor chips, LED chips and infrared sensor chips, and the task set includes a support set and a query set.

[0030] A MAML framework construction module is configured to construct a MAML framework based on a plurality of deep learning models, and randomly initialize parameters of the plurality of deep learning models to obtain initialization parameters of the MAML inner loop.

[0031] A MAML inner loop module is configured to execute the MAML inner loop based on the MAML framework, and train the initialization parameters of the MAML inner loop by using a support set in a partial task set to obtain a parameter vector updated by the MAML inner loop.

[0032] a MAML outer loop module, configured to perform a MAML outer loop based on the MAML framework, train the updated parameter vector of the MAML inner loop based on a query set in the partial task set, and obtain an updated parameter vector of the MAML outer loop;

[0033] a total loss function convergence judgment module, configured to add the loss function of each deep learning model as a total loss function of the outer loop during the performance of the MAML outer loop, and judge whether the total loss function converges;

[0034] a parameter iterative update module, configured to, if the total loss function does not converge, take the updated parameter vector of the MAML outer loop as an initial parameter of the MAML inner loop, and return the step of performing a MAML inner loop based on the MAML framework, training the initial parameter of the MAML inner loop based on a support set in the partial task set, and obtaining an updated parameter vector of the MAML inner loop;

[0035] a surface defect detection model generation module, configured to, if the total loss function converges, assign the updated parameter vector of the MAML outer loop to the plurality of deep learning models, and obtain a surface defect detection model of different types of chip wafers;

[0036] a surface defect detection module, configured to perform surface defect detection on a wafer picture to be detected based on the surface defect detection model.

[0037] In another aspect, the present application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the wafer surface defect detection method.

[0038] In another aspect, the present application also provides a non-transitory computer readable storage medium, which stores a computer program executable to implement the wafer surface defect detection method.

[0039] According to the embodiments of the present application, the following technical effects are achieved:

[0040] The application provides a wafer surface defect detection method, system, electronic equipment and medium, the method comprises the following steps: collecting wafer pictures of different types of chips to form a task set; constructing a MAML framework based on multiple deep learning models, and randomly initializing the parameters of the multiple deep learning models to obtain the initialization parameters of the inner loop of the MAML; executing the inner loop of the MAML based on the MAML framework, training the initialization parameters of the inner loop of the MAML using the support set in part of the task set to obtain the updated parameter vector of the inner loop of the MAML; executing the outer loop of the MAML based on the MAML framework, training the updated parameter vector of the inner loop of the MAML using the query set in part of the task set to obtain the updated parameter vector of the outer loop of the MAML; in the process of executing the outer loop of the MAML, adding the loss function of each deep learning model as the total loss function of the outer loop, and judging whether the total loss function converges; if not, taking the updated parameter vector of the outer loop of the MAML as the initialization parameters of the inner loop of the MAML, and returning to the step of executing the inner loop of the MAML based on the MAML framework, training the initialization parameters of the inner loop of the MAML using the support set in part of the task set to obtain the updated parameter vector of the inner loop of the MAML; if yes, assigning the updated parameter vector of the outer loop of the MAML to the multiple deep learning models to obtain a wafer surface defect detection model for different types of chips; and using the surface defect detection model to detect the surface defects of a wafer picture to be detected. The method can obtain a defect detection model with better robustness using a smaller amount of training samples, and has strong generalization ability, high efficiency and strong practicability. BRIEF DESCRIPTION OF DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative effort.

[0042] Figure 1 The flowchart of the wafer surface defect detection method of the present application;

[0043] Figure 2 The technical roadmap of the wafer surface defect detection method of the present application;

[0044] Figure 3 The task set division schematic diagram of the present application;

[0045] Figure 4 The general MAML framework constructed based on multiple deep learning models of the present application. DETAILED DESCRIPTION

[0046] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0047] The present application aims to provide a wafer surface defect detection method, system, electronic device and medium, which can obtain a defect detection model with better robustness using a small amount of training samples, and has strong generalization ability, high efficiency and strong practicability.

[0048] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0049] Figure 1 Flow chart of a wafer surface defect detection method according to the present application; Figure 2 Technical roadmap of a wafer surface defect detection method according to the present application. Referring to Figure 1 and Figure 2 The wafer surface defect detection method according to the present application comprises:

[0050] Step 1: Collect wafer pictures of different types of chips to form a task set.

[0051] Collect wafer pictures of different types of chips, as long as the die patterns are different, they are different types, for example, different types of chips can include pressure sensor chips, LED chips, infrared sensor chips, etc. The task set includes a support set and a query set.

[0052] Figure 3 Task set division schematic diagram provided by the embodiments of the present application. Referring to Figure 3 , assume that there are T types in total, each type contains k positive samples (wafer pictures without defects) and k negative samples (wafer pictures with defects). For each type of wafer of the T types of chips, k wafer pictures with defects and k wafer pictures without defects of the wafer are collected respectively to form k pairs of wafer pictures; then q pairs of wafer pictures in the k pairs of wafer pictures are divided into a support set (Support Set), and the remaining k-q pairs of wafer pictures are used as a query set (Query Set), which together form T task sets {Task1, Task2,...TaskT}. Each task set corresponds to a different chip type (or wafer type).

[0053] Step 2: Build a MAML framework based on multiple deep learning models, and randomly initialize the parameters of multiple deep learning models to obtain the initialization parameters of the MAML inner loop.

[0054] First, build multiple deep learning models of the same type and randomly initialize their model parameters to obtain the initialization parameters θ i The deep learning model constructed here can be VGG, GOOGLENET, ResNet, etc. In a specific implementation of the present invention, the VGG16 model is preferably used.

[0055] Then, a MAML framework was built based on multiple deep learning models, such as Figure 4 As shown, the MAML framework includes an inner loop and an outer loop, corresponding to Figure 4 The upper and lower parts of the . Figure 4 All MODEL blocks in represent different individuals of the same deep learning model. In the inner loop, we first randomly sample t task sets {Task1,Task2,...Taskt} from the task set, extract the support set (Support Set) in each task set, and then randomly initialize the parameters of each model individual. We use the support set of each task to train the model individual corresponding to that task, and use the Adam-based gradient descent method (Adam-based gradient descent) to update the parameter θ i , until all support sets are traversed, an inner loop is completed, and the parameters θ of each model are obtained i1 ...θ it Then enter the outer loop and use the parameter θ of the inner loop i1 ...θ it Initialize the model individuals corresponding to each task in the outer loop respectively, then take out the query set of each task and train the corresponding model individuals respectively. Add the loss function of each model individual to form the total loss function L of the entire MAML model, and then use gradient descent to update the initialization parameter θ of the inner loop. i , until all query sets are traversed and parameter θ is obtained i+1 , complete the outer loop. One inner loop and one outer loop constitute a large loop. Then the parameter θ obtained by the outer loop is i+1 As the initialization parameter θ of the inner loop i , continue to initialize all model individuals in the inner loop, resample t tasks, continue to enter the next large loop, and repeat this process until the total loss function of the outer loop converges and the final model parameters θ are obtained.

[0056] Step 3: performing a MAML inner loop based on the MAML framework, training the initialization parameters of the MAML inner loop by using the support set in the partial task set, and obtaining a parameter vector of the MAML inner loop after updating.

[0057] The MAML framework in step 2 is trained by using the task set obtained in step 1. In the MAML inner loop, t task sets are randomly sampled from the T task sets as partial task sets. All support set samples in the partial task set are respectively put into the corresponding model, and a plurality of deep learning models are trained. The learning rate in the training is updated by using the update step of the ADAM algorithm, and the initialization parameters θ of the MAML inner loop are updated by using the gradient descent method i , until all wafer pictures of the support set are traversed, and a parameter vector θ' of the MAML inner loop after updating is obtained i .

[0058] The gradient descent method based on ADAM is used to update the model parameters of each task respectively, as shown in the following formula (1).

[0059]

[0060] Wherein θ i represents the parameter vector before updating, and is t-dimensional. The parameter values in the vector are the same during initialization, and are used as the initialization parameters of the MAML inner loop; θ' represents the parameter vector after updating by executing the MAML inner loop, and is also t-dimensional. The values in the vector are θ i , θ i1 , θ i2 , θ it respectively, and are used as the parameter vector after updating of the MAML inner loop; represents the loss function of the deep learning model trained by using the i th support set sample ; Itera (i=0, i=k) (f (i)) represents iteration of the variable i in the function f from 0 to k, as shown in the following formula (2). Figure 4 The loss functions of the t task models in the MAML inner loop are L1, L2,..., L t respectively. represents the gradient value obtained by taking the partial derivative of the loss function L with respect to each parameter θ i ; β represents the learning rate of the inner loop, and is calculated by the ADAM algorithm composed of the following formulas (2)-(6), and is updated by formula (7).

[0061]

[0062] mn = β1m n-1 + (1 - β1)g n (3)

[0063]

[0064]

[0065]

[0066]

[0067] n in the above formula represents the iteration number of the inner loop, also known as the time point, the first support set picture input training represents the 0 time point, and so on. g n in formula (2) represents the loss function , and the partial derivative of the parameter θ n is taken, where θ n represents different values in different model individuals of the inner loop. m n and m n-1 in formula (3) respectively represent the first moment estimation of the gradient in the momentum form at the n time point and the n-1 time point. v n and v n -1 in formula (4) respectively represent the second moment estimation of the gradient in the momentum form at the n time point and the n-1 time point. β in formula (5) represents the second moment estimation after the bias correction of m n . β in formula (6) represents the second moment estimation after the bias correction of v n . Formula (7) is the final update formula of β, where η is the learning rate of the ADAM algorithm, ∈ is a very small constant to prevent the denominator from being 0, both of which are hyperparameters of the model and their values are set in advance. β1 and β2 represent weight parameters, which are also hyperparameters of the model, and respectively represent the n-th power of β1 and β2. The initial values m0 and v0 of m n and v n are both 0.

[0068] Step 4: performing an MAML outer loop based on the MAML framework, training the parameter vector updated by the MAML inner loop using the query set in the partial task set, to obtain an MAML outer loop updated parameter vector.

[0069] Referring to Figure 4 , in the MAML outer loop, the parameter vector θ' updated by the MAML inner loop iAs the initial parameters of multiple models in the outer loop of MAML, the parameter vector θ' updated by the inner loop of MAML is used i Initialize the multiple deep learning model parameters of the outer loop of MAML; and respectively put all the query set samples in the t tasks into the corresponding model for training, continue to train the multiple deep learning models using the query set in the part of the task set, update the multiple deep learning model parameters using the gradient descent method until all the wafer pictures of the query set are traversed, and obtain the parameter vector θ of the multiple deep learning models updated by the outer loop of MAML i+1 .

[0070] In the process of executing the outer loop of MAML, the loss functions L1, L2,..., L t are added as the total loss function L of the whole MAML model, that is

[0071]

[0072] The total loss function is used to update the parameters using the gradient descent method, as shown in the following formula (9).

[0073]

[0074] Where θ i represents the parameter vector before updating, which is t-dimensional, and the parameter values in the vector are the same at the beginning; θ i+1 represents the parameter vector after updating, which is also t-dimensional, and is the parameter vector after updating of the outer loop of MAML; α represents the learning rate of the outer loop, which is a hyperparameter and is set in advance; k-q represents the number of samples in the i-th query set query(i). Itera(r=0, r=k-q)(f(r)) represents the iteration of the variable r in the function f from 0 to k-q, represents the total loss function of the t tasks, which can be written as L. represents the gradient value of the total loss function L with respect to each parameter θ i obtained by taking the partial derivative.

[0075] Step 5: In the process of executing the outer loop of MAML, the loss functions of each deep learning model are added as the total loss function of the outer loop, and it is judged whether the total loss function converges.

[0076] θ i+1 is used as the initialization parameter of the inner loop, and is assigned to each deep learning model in the inner loop, and it is judged whether the model converges, if yes, step 7 is entered, if not, steps 3-5 are repeatedly executed until the total loss function of the outer loop converges, and step 7 is entered.

[0077] Step 6: If the total loss function does not converge, the updated parameter vector of the MAML outer loop cycle is taken as the initialization parameter of the MAML inner loop cycle, and the step 3 is returned.

[0078] Step 7: If the total loss function converges, the updated parameter vector of the MAML outer loop cycle is assigned to the plurality of deep learning models to obtain a surface defect detection model of different types of chip wafers.

[0079] Referring to Figure 2 , if the total loss function L converges, the updated parameter vector θ i+1 is assigned to the corresponding plurality of deep learning models as the final model parameter θ, and a surface defect detection model M of different types of chip wafers is obtained.

[0080] Step 8: The surface defect detection model is used for surface defect detection of the wafer picture to be detected.

[0081] The above steps 1 to 7 mainly describe the training phase of the model, which is used to train a surface defect detection model M applicable to a plurality of different types of chip wafers. Referring to Figure 2 , in the model application phase, first, the chip type corresponding to the wafer picture to be detected is determined as the target type, and the wafer pictures with defects and without defects of the target type chip are collected to form a training set; the surface defect detection model corresponding to the target type chip is extracted, that is, the model M is initialized using the parameter θ, and the parameter of the model M is trained using the training set until all wafer pictures of the training set are traversed, obtaining the model parameter θ* specially for the target type chip, and the parameter θ* is assigned to the model M, that is, the surface defect detection model M* specially for the target type chip is obtained; the surface defect detection model specially for the target type chip can be used for surface defect detection of the wafer picture to be detected, thereby greatly improving the robustness of surface defect detection of the wafer picture to be detected.

[0082] The above steps 3 and 4 of the method of the application use the framework of the MAML algorithm, which belongs to a meta-learning algorithm and can complete training using a small sample, and can solve the problem of few wafer defect samples in actual industrial scenarios. In step 3, the gradient descent method based on ADAM is used when the MAML inner loop cycle is executed, which is equivalent to using the fusion of the second-order gradient and momentum, and can accelerate the convergence speed of the MAML inner loop cycle, thereby improving the training efficiency. In step 4, the initialization parameter of the outer loop cycle uses the training result of the inner loop cycle, which is equivalent to using the idea of model pre-training, and can improve the convergence speed of the model. Therefore, the method of the application can obtain a wafer surface defect detection model with good robustness using a small amount of training samples, and has strong generalization ability, high efficiency and strong practicability.

[0083] Based on the method provided by the application, the application further provides a wafer surface defect detection system, comprising:

[0084] A wafer picture acquisition module is configured to acquire a task set composed of wafer pictures of different types of chips; the different types of chips include pressure sensor chips, LED chips, and infrared sensor chips; the task set includes a support set and a query set;

[0085] A MAML framework construction module is configured to construct a MAML framework based on multiple deep learning models, and randomly initialize parameters of the multiple deep learning models to obtain initialization parameters of a MAML inner loop;

[0086] A MAML inner loop module is configured to execute a MAML inner loop based on the MAML framework, train the initialization parameters of the MAML inner loop using a support set in a partial task set, and obtain an updated parameter vector of the MAML inner loop;

[0087] A MAML outer loop module is configured to execute a MAML outer loop based on the MAML framework, train the updated parameter vector of the MAML inner loop using a query set in the partial task set, and obtain an updated parameter vector of the MAML outer loop;

[0088] A total loss function convergence judgment module is configured to add loss functions of each deep learning model together as a total loss function of the outer loop during execution of the MAML outer loop, and judge whether the total loss function converges;

[0089] A parameter iterative update module is configured to, if the total loss function does not converge, take the updated parameter vector of the MAML outer loop as the initialization parameters of the MAML inner loop, and return to the step of "executing a MAML inner loop based on the MAML framework, training the initialization parameters of the MAML inner loop using a support set in a partial task set, and obtaining an updated parameter vector of the MAML inner loop";

[0090] A surface defect detection model generation module is configured to, if the total loss function converges, assign the updated parameter vector of the MAML outer loop to the multiple deep learning models, and obtain a surface defect detection model for wafers of different types of chips;

[0091] A surface defect detection module is configured to perform surface defect detection on a wafer picture to be detected using the surface defect detection model.

[0092] The method and system of the present application divide a plurality of pictures of each type of wafer with defects and without defects into a support set and a query set to form a task set by collecting different types of wafer pictures; and construct a deep learning model and a MAML framework, and perform random initialization of model parameters; perform a MAML inner loop, take part of the tasks in the task set, and train the model using the support set, update the learning rate in the training using the update step of the ADAM algorithm, update the inner layer parameters using the gradient descent method until all the pictures in the support set are traversed; then perform a MAML outer loop, continue to train the model using the pictures in the query set for each task, add the loss functions of all the tasks as the loss function of the outer loop, and update the outer layer parameters using the gradient descent method; repeat the MAML inner loop and the outer loop until the model converges, and obtain the final surface defect detection model. The method and system of the present application based on the improved MAML algorithm can accurately determine whether there is a defect on the wafer grain picture, improve the generalization ability of the detection model, and have high model training efficiency and strong practicability.

[0093] Further, the present application also provides an electronic device, which can include a processor, a communication interface, a memory and a communication bus. Wherein the processor, the communication interface and the memory complete mutual communication through the communication bus. The processor can call the computer program in the memory to execute the wafer surface defect detection method.

[0094] In addition, the computer program in the memory described above is realized in the form of a software functional unit and sold or used as an independent product, which can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a plurality of instructions to make a computer device (which can be a personal computer, a server or a network device, etc.) execute all or part of the steps of the method described in the embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory, a random access memory, a magnetic disk or an optical disk, and various media that can store program codes.

[0095] Further, the present application also provides a non-transitory computer readable storage medium, which stores a computer program, and the computer program can implement the wafer surface defect detection method when executed.

[0096] The various embodiments described in this specification are presented for the purpose of illustrating the principles of the present application and its best mode of operation. Each of the embodiments described in this specification has been provided for the purpose of illustration only and the various embodiments are not intended to limit the present application in any way unless otherwise specifically indicated. The same parts and / or features of the various embodiments described in this specification can be referenced using the same reference numerals for the ease of understanding of the present application.

[0097] The principles and implementations of the present application have been described in the above embodiments, which are only used to help understand the method of the present application and its core idea. Meanwhile, for those skilled in the art, the specific implementation and application range of the present application can be changed according to the idea of the present application. In summary, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A wafer surface defect detection method, characterized in that: include: The task set is to collect wafer images of different types of chips, including pressure sensor chips, LED chips, and infrared sensor chips. The task set includes a support set and a query set; Build a MAML framework based on multiple deep learning models, and randomly initialize the parameters of multiple deep learning models to obtain the initialization parameters of the MAML inner loop; Executing a MAML inner loop based on the MAML framework, training initialization parameters of the MAML inner loop using a support set in a partial task set, and obtaining an updated parameter vector of the MAML inner loop; Executing a MAML outer loop based on the MAML framework, and using a query set in a portion of the task set to train a parameter vector updated by the MAML inner loop, thereby obtaining an updated parameter vector for the MAML outer loop; During the execution of the MAML outer loop, the loss function of each deep learning model is added as the total loss function of the outer loop, and whether the total loss function converges is determined; If the total loss function has not converged, using the updated parameter vector of the MAML outer loop as the initialization parameter of the MAML inner loop, and returning to the step of "executing the MAML inner loop based on the MAML framework, training the initialization parameter of the MAML inner loop using the support set in the partial task set, and obtaining the updated parameter vector of the MAML inner loop"; If the total loss function converges, assigning the parameter vector updated by the MAML outer loop to the multiple deep learning models to obtain surface defect detection models for different types of chip wafers; The surface defect detection model is used to perform surface defect detection on the wafer image to be inspected.

2. The wafer surface defect detection method according to claim 1, characterized in that: The task set of collecting wafer images of different types of chips specifically includes: For each type of chip in the T types of chips, k defective wafer images and k non-defective wafer images are collected to form k pairs of wafer images; The q pairs of wafer images among the k pairs of wafer images are divided into support sets, and the remaining kq pairs of wafer images are used as query sets, which together constitute T task sets.

3. The wafer surface defect detection method according to claim 2, characterized in that: The method of executing the MAML inner loop based on the MAML framework, training the initialization parameters of the MAML inner loop using the support set in the partial task set, and obtaining the updated parameter vector of the MAML inner loop specifically includes: Randomly sampling t task sets from the T task sets as partial task sets; The support set in the partial task set is taken to train the multiple deep learning models. The learning rate in the training is updated using the update step size of the ADAM algorithm, and the gradient descent method is used to update the initialization parameters of the MAML inner loop until all wafer images in the support set are traversed to obtain the parameter vector after the MAML inner loop is updated.

4. The wafer surface defect detection method according to claim 3, characterized in that: The method of executing the MAML outer loop based on the MAML framework and using the query set in the partial task set to train the parameter vector updated by the MAML inner loop to obtain the parameter vector updated by the MAML outer loop specifically includes: The parameter vector updated by the MAML inner loop is used to initialize the multiple deep learning model parameters of the MAML outer loop, and the query set in the partial task set is used to continue training the multiple deep learning models. The gradient descent method is used to update the multiple deep learning model parameters until all wafer images in the query set are traversed to obtain the parameter vector updated by the MAML outer loop.

5. The wafer surface defect detection method according to claim 4, characterized in that: Surface defect detection is performed on the wafer image to be inspected using the surface defect detection model, specifically including: Determine the chip type corresponding to the wafer image to be inspected as the target type; Collect defective and non-defective wafer images of the target type of chips to form a training set; Extracting a surface defect detection model corresponding to the target type chip, and using the training set to perform parameter training until all wafer images in the training set are traversed to obtain a surface defect detection model specifically for the target type chip; Surface defect detection is performed on the wafer image to be inspected using the surface defect detection model specifically for the target type of chip.

6. A wafer surface defect detection system, characterized in that: include: A wafer image acquisition module is used to collect wafer images of different types of chips to form a task set; the different types of chips include pressure sensor chips, LED chips, and infrared sensor chips; The task set includes a support set and a query set; The MAML framework construction module is used to build a MAML framework based on multiple deep learning models and randomly initialize the parameters of multiple deep learning models to obtain the initialization parameters of the MAML inner loop; A MAML inner loop module is used to execute the MAML inner loop based on the MAML framework, train the initialization parameters of the MAML inner loop using the support set in the partial task set, and obtain an updated parameter vector of the MAML inner loop; A MAML outer loop module is configured to execute a MAML outer loop based on the MAML framework, train the updated parameter vector of the MAML inner loop using the query set in the partial task set, and obtain the updated parameter vector of the MAML outer loop; A total loss function convergence judgment module is used to add the loss functions of each deep learning model as the total loss function of the outer loop during the execution of the MAML outer loop, and to judge whether the total loss function has converged; a parameter iterative update module configured to, if the total loss function has not converged, use the updated parameter vector of the MAML outer loop as the initialization parameter of the MAML inner loop, and return to the step of "executing the MAML inner loop based on the MAML framework, training the initialization parameter of the MAML inner loop using the support set in the partial task set, and obtaining the updated parameter vector of the MAML inner loop"; A surface defect detection model generation module is used to assign the parameter vector updated by the MAML outer loop to the multiple deep learning models if the total loss function converges, so as to obtain surface defect detection models for different types of chip wafers; The surface defect detection module is used to perform surface defect detection on the wafer image to be inspected using the surface defect detection model.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the wafer surface defect detection method according to any one of claims 1 to 5 is implemented.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed, the wafer surface defect detection method according to any one of claims 1 to 5 is implemented.

Images