A method for constructing a deep learning model and a nonlinear coordinate transformation method

Abstract

The present invention relates to a method for constructing a deep learning model and a method for nonlinear coordinate conversion. The construction method includes: obtaining a training sample set, wherein the training samples in the training sample set include a first coordinate and a second coordinate, wherein the first coordinate and the second coordinate are coordinates of any point of a standard film in the wafer coordinate system of a preceding detection device and a review detection device respectively; using the first coordinate as input and the second coordinate as output to train a deep learning model; an embodiment of the present invention provides a method for constructing a deep learning model, which realizes conversion between nonlinear coordinates based on the deep learning model, does not require multiple assumptions and attempts to find suitable coefficients, and the conversion process is simple and can be automatically implemented, and the conversion result is accurate.

Description

Technical Field

[0001] The present invention relates to the field of semiconductor process manufacturing, and in particular to a method for constructing a deep learning model and a nonlinear coordinate transformation method. Background Art

[0002] Wafer defect detection can be divided into pre-process inspection and review inspection. Pre-process inspection equipment performs a preliminary, rapid detection of the object of interest, without concern for the details of the defect. If further, more detailed information is required, review inspection equipment is used to collect information such as shape, size, and elemental composition before classifying the defect. With the rapid advancement of semiconductor manufacturing processes, critical process dimensions have reached below 10nm. Small defects have a greater impact on process quality. Therefore, the importance of extremely small defects is increasing.

[0003] Therefore, establishing a high-precision correspondence between the wafer coordinate system of the review inspection equipment and the wafer coordinate system of the previous inspection equipment becomes crucial. Traditional methods for coordinate system conversion primarily involve linear transformations, primarily affine transformations. Alternatively, nonlinear coordinate transformations are performed based on these transformations, followed by affine transformations, with the fitting parameters calculated using the least squares method. While linear transformations are simple and stable, they struggle to reduce the mean absolute error (MAE) to less than 20 μm. MAEs often exceed 20 μm, making them inadequate for detecting small defects. Furthermore, the instability of traditional nonlinear transformations often leads to even lower accuracy than linear transformations. If the error between these correspondences increases, the review inspection equipment will need to spend more time searching near a given defect coordinate, severely impacting throughput. Alternatively, increasing the search range by using a larger field of view (FOV) reduces the number of pixels occupied by the defect within the FOV, inevitably leading to inability to locate and identify small defects. The accuracy of this correspondence directly impacts the throughput of automated inspections and the minimum detectable defect size, impacting semiconductor process quality and efficiency.

[0004] Furthermore, during the pre-inspection and re-inspection processes for wafer defect detection, differences in equipment usage, errors in implementation technology, and differences between pre-inspection and re-inspection equipment determine that the correspondence between the wafer coordinate systems of the re-inspection and pre-inspection equipment is often nonlinear. Therefore, it is necessary to find a suitable nonlinear correspondence. However, the traditional nonlinear transformation method assumes a set of nonlinear functions and then finds appropriate coefficients. This approach requires manual assumptions in advance, and in practice, the final accuracy does not meet the requirements, requiring further assumptions. As a result, practical applications are complex and the final accuracy is not good enough. Summary of the Invention

[0005] In response to the technical problems existing in the prior art, the present invention provides a method for constructing a deep learning model and a nonlinear coordinate conversion method. The method realizes the conversion between nonlinear coordinates based on the deep learning model, does not require multiple assumptions and attempts to find suitable coefficients, and the conversion process is simple and can be automatically implemented, and the conversion result is accurate.

[0006] According to a first aspect of the present invention, a method for constructing a deep learning model is provided, comprising:

[0007] Step 1: Obtain a training sample set, wherein the training samples in the training sample set include a first coordinate and a second coordinate, wherein the first coordinate and the second coordinate are coordinates of any point of the standard film in the wafer coordinate system of the pre-order inspection device and the review inspection device respectively;

[0008] Step 2: Based on a deep learning algorithm, use the first coordinate as input and the second coordinate as output to train a deep learning model.

[0009] On the basis of the above technical solution, the present invention can also make the following improvements.

[0010] Optionally, the process of obtaining the training sample in step 1 includes:

[0011] Make a standard film containing multiple landmarks;

[0012] Putting the standard sheet into a pre-process inspection device, and obtaining the coordinates of the marking point in the wafer coordinate system as the first coordinates;

[0013] The standard film is placed in a review and detection device, and the coordinates of the marking point in the wafer coordinate system are obtained as the second coordinates.

[0014] Optionally, obtaining the coordinates of the marker point in the wafer coordinate system as the second coordinates includes:

[0015] Obtaining the coordinates of the wafer center and the wafer notch in the table coordinate system of the review and inspection equipment, as well as the coordinates of the wafer center and the wafer notch in the wafer coordinate system of the review and inspection equipment, and calculating a conversion matrix between the table coordinate system and the wafer coordinate system;

[0016] The coordinates of the marking point on the standard film in the table coordinate system of the review and detection equipment are detected, and the coordinates of the marking point in the wafer coordinate system of the review and detection equipment are obtained according to the conversion matrix as the second coordinates.

[0017] Optionally, the deep learning model includes an input layer, a hidden layer, and an output layer, the hidden layer has an m-layer structure, and m is an integer greater than 1;

[0018] From the input layer to each layer of the hidden layer, the number of neurons in each hidden layer gradually increases.

[0019] Optionally, the 1st to m-1th layers in the hidden layer are respectively connected to the mth layer; the layers are connected in a fully connected manner, and nonlinear transformation is performed between the layers.

[0020] Optionally, the deep learning model is optimized by gradient descent during training, and the optimization formula is:

[0021]

[0022] Among them, ∈ is the learning rate, which is a positive scalar that determines the size of the optimization step, v is a tensor composed of horizontal and vertical coordinates, and f(v) is the loss function. Represents the partial derivative in the v direction.

[0023] Optionally, the step 2 further includes: selecting a test set from the training sample set;

[0024] The deep learning model is evaluated using the test set through an evaluation function; the evaluation function is a mean square error function, a root mean square error function, a mean absolute error function, or an R-square function.

[0025] According to a second aspect of the present invention, a nonlinear coordinate transformation method is provided. Based on the deep learning model constructed by the construction method provided by the present invention, the nonlinear coordinate transformation method includes:

[0026] The coordinates to be converted are input into the deep learning model, and the deep learning model outputs the converted coordinates.

[0027] On the basis of the above technical solution, the present invention can also make the following improvements.

[0028] According to a third aspect of the present invention, a method for detecting wafer defects is provided. Based on a nonlinear coordinate transformation method provided by the present invention, the method for detecting wafer defects includes:

[0029] Extract the coordinates of the defect from the inspection result file of the previous inspection equipment, input them into the trained deep learning model, and output the converted coordinates;

[0030] Converting the converted coordinates output by the deep learning model into coordinates in the table coordinate system of the review and inspection equipment according to a conversion matrix between the table coordinate system and the wafer coordinate system of the review and inspection equipment;

[0031] The defect is detected in the review and detection equipment according to the coordinates of the defect in the table coordinate system.

[0032] On the basis of the above technical solution, the present invention can also make the following improvements.

[0033] Optionally, the training samples are periodically obtained from the preceding detection device and the review detection device, and the deep learning model is updated after training the deep learning model.

[0034] The present invention provides a nonlinear coordinate transformation method and wafer defect detection method. This method uses deep learning to fit nonlinear transformations and finds the correspondence between the wafer coordinate system of a review inspection device and the wafer coordinate system of a preceding inspection device. This method has the advantages of low implementation difficulty, high practicality, high precision, and high stability. It can significantly improve the throughput and detection rate of wafer defect review inspection equipment, enabling the review inspection equipment to identify and detect smaller defects. Furthermore, the method can promptly adapt to deformation of the defect detection equipment, resulting in high universality and improved repeatability of defect detection by the review inspection equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 A flowchart of a method for constructing a deep learning model provided by the present invention;

[0036] Figure 2 A schematic diagram of the structure of a deep learning model embodiment provided by the present invention;

[0037] Figure 3 A flow chart of a nonlinear coordinate transformation method provided by the present invention;

[0038] Figure 4 A schematic diagram of the application principle of an embodiment of a nonlinear coordinate transformation method provided by the present invention;

[0039] Figure 5 A flowchart of an embodiment of a method for detecting wafer defects provided by an embodiment of the present invention;

[0040] Figure 6 A schematic diagram of visualizing coordinate system transformation provided by an embodiment of the present invention;

[0041] Figure 7 The present invention provides Figure 6 A partial enlarged schematic diagram. DETAILED DESCRIPTION

[0042] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0043] Example 1

[0044] Example 1 provided by the present invention is an embodiment of a method for constructing a deep learning model provided by the present invention. Figure 1A flowchart of a method for constructing a deep learning model provided by the present invention is shown in FIG. Figure 1 As shown, the construction method includes:

[0045] Step 1: Obtain a training sample set. The training samples in the training sample set include a first coordinate and a second coordinate. The first coordinate is obtained by a preceding detection device, and the second coordinate is obtained by a review detection device. The first coordinate and the second coordinate correspond one to one. The first coordinate and the second coordinate are the coordinates of any point of the standard film in the wafer coordinate system of the preceding detection device and the review detection device respectively.

[0046] In a possible embodiment, the process of obtaining training samples includes:

[0047] Create a standard film containing multiple landmarks.

[0048] Place the standard wafer into the pre-process inspection equipment and obtain the coordinates of the marking point in the wafer coordinate system as the first coordinates.

[0049] Place the standard film into the review and detection equipment, and obtain the coordinates of the marking point in the wafer coordinate system as the second coordinate.

[0050] The training sample contains the coordinates of the previous detection device (X k , Y k ) and the coordinates of the inspection equipment (X gt , Y gt ), and use the ID number of the mark point on the wafer as the corresponding connection condition to make a one-to-one correspondence between the coordinates of the previous inspection equipment and the coordinates of the review inspection equipment. That is, the training sample format can be (id, X k , Y k , X gt , Y gt ). The training sample set contains coordinates of multiple landmark points that correspond one to one.

[0051] In a possible embodiment, obtaining the coordinates of the marker point in the wafer coordinate system as the second coordinate includes:

[0052] Obtain the coordinates of the wafer center and the wafer notch in the table coordinate system of the review and inspection equipment, as well as the coordinates of the wafer center and the wafer notch in the wafer coordinate system in the review and inspection equipment, and calculate the conversion matrix between the table coordinate system and the wafer coordinate system; the table coordinate system is the wafer carrier coordinate system.

[0053] The coordinates of the marking point on the standard film in the stage coordinate system are detected, and the coordinates of the marking point in the wafer coordinate system are obtained as the second coordinates according to the conversion matrix.

[0054] Step 2: Based on the deep learning algorithm, the first coordinate is used as input and the second coordinate is used as output to train a deep learning model.

[0055] In a possible embodiment, step 2 further includes:

[0056] A test set is selected from the training sample set. For example, 20% of the training sample set is selected as the test set by random sampling.

[0057] Using the test set, the deep learning model is evaluated by the evaluation function; the evaluation function is the mean square error function, the root mean square error function, the mean absolute error function or the R square function. Specifically, the formula of the evaluation function can be:

[0058] Mean square error function:

[0059] Root mean square error function:

[0060] Mean absolute error function:

[0061] R square function:

[0062] Where n is the number of training samples in the test set, y i is the true value on the test set, The predicted value on the test set, The average value on the test set.

[0063] In one possible embodiment, Figure 2 The figure shows a schematic diagram of the structure of a deep learning model embodiment provided by the present invention, combined with Figure 2 It can be seen that the deep learning model includes an input layer, a hidden layer, and an output layer. The hidden layer has an m-layer structure, where m is an integer greater than 1. Figure 2 In the given embodiment, m is 6, that is, there are 6 hidden layers, the input layer inputs a 2-dimensional tensor to the first hidden layer, the first hidden layer inputs a 32-dimensional tensor to the second hidden layer, the second hidden layer inputs a 128-dimensional tensor to the third hidden layer, the third hidden layer inputs a 256-dimensional tensor to the fourth hidden layer, the fourth hidden layer inputs a 512-dimensional tensor to the fifth hidden layer, and the first five hidden layers input a 1952-dimensional tensor to the sixth hidden layer.

[0064] From the input layer to the hidden layer, the number of neurons in each hidden layer increases layer by layer.

[0065] For example, n neurons are set in the first layer and 2n neurons are set in the second layer.

[0066] In the hidden layer, two adjacent layers are connected to each other.

[0067] The hidden layers are connected in a fully connected manner, and nonlinear transformation is performed between the layers. In a specific implementation, the nonlinear transformation between the layers is realized by a piecewise function. Furthermore, the piecewise function can be a ReLU function.

[0068] In one possible embodiment, the deep learning model is optimized by gradient descent during training, and the optimization formula is:

[0069]

[0070] Among them, ∈ is the learning rate, which is a positive scalar that determines the step size, v is a tensor consisting of horizontal and vertical coordinates, and f(v) is the loss function. Represents the partial derivative in the v direction. The increase in the number of hidden layers can easily lead to vanishing gradients during back propagation. During the training phase, it is not conducive to updating the network weights of the deep learning model, which ultimately leads to the inability to reduce losses. Therefore, layers 1 to m-1 are connected to the mth layer respectively. In the embodiment given in the present invention, the outputs of the first 5 layers are simultaneously connected to the 6th layer. The last layer is then converted to 2 neurons for output. Using the above method, the deep learning model will not experience vanishing gradients during the training process, so that the predicted values ​​obtained by the trained deep learning model are more accurate and closer to the true values.

[0071] Considering the calculation process of nonlinear coordinate transformation using nonlinear functions, it is necessary to find appropriate coefficients. This approach requires manual assumptions in advance, and when the coordinate transformation accuracy does not meet the standards, it is necessary to make assumptions again. Therefore, the actual application is complex and the final accuracy is not good enough. An embodiment of the present invention provides a method for constructing a deep learning model, which realizes the transformation between nonlinear coordinates based on the deep learning model. It does not require multiple assumptions to try to find appropriate coefficients. The conversion process is simple and can be automatically implemented, and the coordinate transformation results are accurate.

[0072] Example 2

[0073] The nonlinear coordinate conversion method provided in this embodiment 2 can be used to convert the wafer coordinate system of the review inspection equipment and the previous inspection equipment, such as Figure 3 FIG. 1 is a flow chart of a nonlinear coordinate transformation method provided by the present invention, as shown in FIG. Figure 4 The figure shows the application principle of an embodiment of a nonlinear coordinate transformation method provided by the present invention, combined with Figure 3 and Figure 4 It can be seen that the embodiment of the nonlinear coordinate transformation method includes:

[0074] Step 1: Obtain a training sample set. The training samples in the training sample set include a first coordinate and a second coordinate. The first coordinate is obtained by a preceding detection device, and the second coordinate is obtained by a review detection device. The first coordinate and the second coordinate correspond one to one. The first coordinate and the second coordinate are the coordinates of any point of the standard film in the wafer coordinate system of the preceding detection device and the review detection device respectively.

[0075] In specific implementation, in order to find the mapping relationship between the coordinate system of the previous inspection equipment and the coordinate system of the review inspection equipment, it is necessary to collect paired coordinate data in advance, that is, the wafer coordinates of a point in the previous inspection equipment and the wafer coordinates of this point in the review inspection equipment.

[0076] In a possible embodiment, the process of obtaining training samples includes:

[0077] A standard film containing multiple marking points is produced. Specifically, the multiple marking points can be defects on the wafer. A standard film is produced for calibration. Preferably, the multiple marking points are evenly distributed on the wafer.

[0078] The standard film is placed in a pre-process detection device, and the coordinates of the marking point are obtained as first coordinates. The first coordinates can be directly obtained from the output file of the pre-process detection device.

[0079] Place the standard film into the review and detection equipment, and obtain the coordinates of the marking point as the second coordinate.

[0080] In a possible embodiment, obtaining the coordinates of the marker point as the second coordinates includes:

[0081] After taking out the standard wafer and placing it in the review and inspection equipment, the coordinates of the wafer center and the wafer notch in the table coordinate system of the review and inspection equipment, as well as the coordinates of the wafer center and the wafer notch in the wafer coordinate system of the review and inspection equipment are obtained. The conversion matrix between the table coordinate system and the wafer coordinate system is calculated using formulas (6)-(10).

[0082] The coordinates of the marking point on the standard film in the stage coordinate system are detected, and the coordinates of the marking point in the wafer coordinate system are obtained as the second coordinates according to the conversion matrix.

[0083]

[0084] a1=s1×cos(θ) (7)

[0085] a2=s2×-sin(θ) (8)

[0086] a3=s3×sin(θ) (9)

[0087] a4=s4×cos(θ) (10)

[0088] Among them, t x and t y Represents the translation amount in the X-axis and Y-axis directions respectively; x is the horizontal coordinate of the wafer center or wafer notch in the table coordinate system, y is the vertical coordinate of the wafer center or wafer notch in the table coordinate system, x′ is the horizontal coordinate of the wafer center or wafer notch in the wafer coordinate system, y′ is the vertical coordinate of the wafer center or wafer notch in the wafer coordinate system; θ is the parameter of wafer rotation, s i (i=1, 2, 3 or 4) is the parameter of wafer scaling, which is determined by the review and inspection equipment.

[0089] Specifically, the conversion matrix can be:

[0090]

[0091] After obtaining the transformation matrix, the coordinates of the marker points on the standard film in the stage coordinate system are detected one by one, and then the corresponding coordinates in the wafer coordinate system are obtained by formula (12). The coordinates of the marker points in the wafer coordinate system constitute the second coordinates.

[0092] W=M cow ×S (12)

[0093] Wherein, W is the coordinate of the mark point in the wafer coordinate system, and S is the coordinate of the mark point in the stage coordinate system.

[0094] Step 2: Based on the deep learning algorithm, the first coordinate is used as input and the second coordinate is used as output to train a deep learning model.

[0095] Based on the training sample set obtained in step 1, the designed deep learning network is trained, and a deep learning model is obtained after the training is completed.

[0096] In practice, the deep learning model obtained after training is a file containing the connection weights between neurons. This file is saved and loaded into the deep learning model before use. The trained deep learning model can be considered the required correspondence between the coordinate system of the pre-inspection device and the coordinate system of the review device.

[0097] Step 3: Input the coordinates to be converted into the deep learning model, and the deep learning model outputs the converted coordinates.

[0098] The coordinates to be converted are the coordinates obtained by the previous inspection equipment, and the coordinates after conversion are the wafer coordinates corresponding to the review inspection equipment. For example, a coordinate to be converted (X before , Y before) is input into the deep learning model to obtain a predicted coordinate (X post , Y post ), the predicted coordinates are the coordinates of the defect in the wafer coordinate system in the review and inspection equipment.

[0099] Example 3

[0100] Example 3 provided by the present invention is an embodiment of a wafer defect detection method provided by the present invention. The wafer defect detection method is based on a nonlinear coordinate transformation method provided by an embodiment of the present invention.

[0101] The precursor to patternless wafer defect review using a Scanning Electron Microscope (SEM) is to generate preliminary defect inspection results using pre-process inspection equipment. This type of equipment performs rapid, preliminary inspections of processes and tools for ICs, wafers, devices, and advanced memory design nodes, obtaining basic defect information such as approximate coordinates and dimensions. A detailed review using the SEM then yields more detailed defect information, including specific shape, highly accurate size, classification, and component elements.

[0102] like Figure 5 As shown, a flowchart of an embodiment of a wafer defect detection method provided by an embodiment of the present invention is combined with Figure 1 - Figure 5 It can be seen that the embodiment of the wafer defect detection method includes:

[0103] Load the test result file of the previous test equipment to determine whether there is a defect. If yes, extract the defect coordinates; if not, end the test.

[0104] The extracted defect coordinates are input into the trained deep learning model, and forward prediction is performed through the trained deep learning model to obtain the coordinates of the defect in the wafer coordinate system of the re-inspection equipment.

[0105] The results output by the pre-process inspection equipment include the defect ID number, X-axis coordinate, Y-axis coordinate and size information.

[0106] For example, the coordinates of the defect are (X before , Y before ), input it into the deep learning model to get a predicted coordinate (X post , Y post ), the predicted coordinates are the coordinates of the defect in the wafer coordinate system of the review and detection equipment.

[0107] According to the conversion matrix between the table coordinate system of the review and inspection equipment and the wafer coordinate system, the predicted coordinates output by the deep learning model are converted into coordinates in the table coordinate system in the review and inspection equipment.

[0108] Specifically, referring to formula (13), the coordinate W of the defect in the wafer coordinate system in the review and detection equipment can be converted into the coordinate S in the stage coordinate system.

[0109]

[0110] M cow The inverse matrix of .

[0111] According to the coordinates of the defect in the table coordinate system, in the review inspection equipment, the table is moved to this coordinate to inspect the defects in the area. After completing the defect inspection of the defect area in the review inspection equipment, it is determined whether there is the next defect coordinate in the inspection result file of the previous inspection equipment. After all the defect coordinates in the inspection result file are taken out and the defect inspection in the review inspection equipment is completed, the inspection result is obtained.

[0112] Through the trained deep learning model, the high-precision coordinates of defects in the wafer coordinate system of the review and inspection equipment can be obtained. The review and inspection equipment does not need to adjust the FOV again and search for defects time-consumingly, which improves the efficiency and accuracy of defect identification of the review and inspection equipment.

[0113] In addition, this embodiment also provides a method to ensure the long-term stability of the device by periodically acquiring training sample data from the pre-processing detection device and the review detection device to train and update the deep learning model.

[0114] Since defect detection equipment may deform with increased usage, this deformation inevitably results in significant errors when using linear transformations to find coordinate correspondences. Therefore, a mechanism is needed to ensure that these correspondences can be updated in some form. Therefore, the process must employ a nonlinear fitting approach. For engineering feasibility, the correspondences must be minimal. Furthermore, a calibration process must be performed on a regular basis, such as daily, weekly, monthly, or annually. Automated processes, such as acquiring new training sample sets, deriving new deep learning models, and loading these models, allow for timely adaptation to defect detection equipment deformation.

[0115] Figure 6A schematic diagram of visualizing coordinate system transformation provided by an embodiment of the present invention. There are three types of coordinates in the figure, namely the wafer coordinates of the pre-processing inspection equipment, the actual value and the predicted value of the wafer coordinates of the review inspection equipment. By comparing the difference between the actual value and the predicted value of the wafer coordinates of the review inspection equipment, the effect of the coordinate transformation can be known. Figure 6 As shown, the wafer coordinates of the pre-process inspection equipment can be roughly matched to the true values ​​of the wafer coordinates of the review inspection equipment by a certain rotation method. While there are certain differences in different methods for achieving this conversion, the present invention can achieve the best results through deep learning.

[0116] Figure 7 The present invention provides Figure 6 In order to judge the deviation more clearly, Figure 6 The defect ID 59383 in the image is magnified. It's easy to determine the difference between the actual and predicted wafer coordinates of this defect's inspection equipment. The difference is approximately 4 μm on the X-axis and 5 μm on the Y-axis. Traditional linear transformations struggle to reduce the MAE error to less than 20 μm, and the instability of traditional nonlinear changes often exceeds the accuracy of linear transformations. The present invention, using deep learning-based nonlinear transformations, can reduce the MAE error to less than 1 μm.

[0117] The embodiments of the present invention provide a nonlinear coordinate transformation method and wafer defect detection method. This method uses deep learning to fit nonlinear transformations and find the correspondence between the wafer coordinate system of the review inspection equipment and the wafer coordinate system of the preceding inspection equipment. This method has the advantages of low implementation difficulty, high practicality, high precision, and high stability. It can significantly improve the throughput and detection rate of wafer defect review inspection equipment, enabling the review inspection equipment to identify and detect smaller defects. It can also promptly adapt to deformation of the defect detection equipment.

[0118] It should be noted that, in the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant description of other embodiments.

[0119] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0120] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, 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 computer, or other programmable data processing device to produce 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 flowcharts and / or block diagrams. 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.

[0121] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work 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 The function specified in one or more boxes.

[0122] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0123] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional 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 that fall within the scope of the present invention.

[0124] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A method for constructing a deep learning model, characterized in that: The construction method comprises: Step 1: Obtain a training sample set, wherein the training samples in the training sample set include a first coordinate and a second coordinate, wherein the first coordinate and the second coordinate are coordinates of any point of the standard film in the wafer coordinate system of the pre-order inspection device and the review inspection device respectively; Step 2: Based on a deep learning algorithm, the first coordinate is used as input and the second coordinate is used as output to train a deep learning model; The process of obtaining the training sample in step 1 includes: Make a standard film containing multiple landmarks; Putting the standard sheet into a pre-process inspection device, and obtaining the coordinates of the marking point in the wafer coordinate system as the first coordinates; Putting the standard film into a review and detection device, and obtaining the coordinates of the marking point in the wafer coordinate system as the second coordinates; The acquiring the coordinates of the marking point in the wafer coordinate system as the second coordinates includes: The coordinates of the marking point on the standard film in the table coordinate system of the review and detection equipment are detected, and the coordinates of the marking point in the wafer coordinate system of the review and detection equipment are obtained according to the conversion matrix as the second coordinates.

2. The construction method according to claim 1, characterized in that The step of obtaining the coordinates of the marker point in the wafer coordinate system as the second coordinates further includes: The coordinates of the wafer center and the wafer notch in the table coordinate system of the review and inspection equipment are obtained, as well as the coordinates of the wafer center and the wafer notch in the wafer coordinate system of the review and inspection equipment, and the conversion matrix between the table coordinate system and the wafer coordinate system is calculated.

3. The construction method according to claim 1, characterized in that The deep learning model includes an input layer, a hidden layer and an output layer, the hidden layer has an m-layer structure, and m is an integer greater than 1; From the input layer to each layer of the hidden layer, the number of neurons in each hidden layer gradually increases.

4. The construction method according to claim 3, characterized in that The 1st to m-1th layers in the hidden layer are respectively connected to the mth layer; the layers are connected in a fully connected manner, and nonlinear transformation is performed between the layers.

5. The construction method according to claim 1, characterized in that The training of the deep learning model is optimized by gradient descent, and the optimization formula is: ； in, is the learning rate, which is a positive scalar that determines the size of the optimization step, and v is a tensor consisting of the horizontal and vertical coordinates. is the loss function, Representatives request Directional derivative.

6. The construction method according to claim 1, characterized in that The step 2 further includes: selecting a test set from the training sample set; The deep learning model is evaluated using the test set through an evaluation function; the evaluation function is a mean square error function, a root mean square error function, a mean absolute error function, or an R-square function.

7. A nonlinear coordinate transformation method, based on the deep learning model constructed by the construction method according to any one of claims 1 to 6, characterized in that: The nonlinear coordinate transformation method comprises: The coordinates to be converted are input into the deep learning model, and the deep learning model outputs the converted coordinates.

8. A method for detecting wafer defects, comprising: The detection method comprises: Extract the coordinates of the defect from the inspection result file of the previous inspection equipment, input them into the trained deep learning model, and output the converted coordinates; Converting the converted coordinates output by the deep learning model into coordinates in the table coordinate system of the review and inspection equipment according to a conversion matrix between the table coordinate system and the wafer coordinate system of the review and inspection equipment; The defect is detected in the review and detection equipment according to the coordinates of the defect in the table coordinate system.

9. The detection method according to claim 8, characterized in that The training samples are periodically obtained from the preceding detection device and the review detection device, and the deep learning model is updated after training the deep learning model.

Images