WO2021129398A1

WO2021129398A1 - Wafer grinding device and method

Info

Publication number: WO2021129398A1
Application number: PCT/CN2020/135042
Authority: WO
Inventors: 陈兴伟; 曾斌; 王海升
Original assignee: 青岛歌尔微电子研究院有限公司
Priority date: 2019-12-23
Filing date: 2020-12-09
Publication date: 2021-07-01
Also published as: CN111037455A; CN111037455B

Abstract

A wafer grinding device and method. The wafer grinding device comprises: a grinding table, the grinding table being used for placing a wafer (10); a camera device (30), the camera device (30) being provided above the grinding table and used for photographing a wafer image on the grinding table; a controller, the controller being electrically connected to the camera device (30) and being used for receiving the wafer image photographed by the camera device (30); and an alarm, the controller being electrically connected to the alarm, and the controller sending prompt information by means of the alarm according to the wafer image.

Description

Wafer grinding equipment and method

This application claims the priority of the Chinese patent application filed on December 23, 2019 with the application number 201911338861.X, the entire content of which is incorporated into this application by reference.

Technical field

This application relates to the field of electronic machinery technology, and in particular to a wafer polishing equipment and method.

Background technique

This section provides only background information related to the present disclosure, which is not necessarily prior art.

At present, the grinders on the market have the function of identifying the insertion weight and the insertion angle of the wafer, but there is no monitoring function for the product reversal phenomenon. If the upward reversal phenomenon occurs in the loading, not only the impact on the product is irreparable, but also on the equipment. There is also a certain degree of damage. At present, in order to solve the problem of loading and unloading, only manual procedures can be added: manual inspection before loading, double check and signature. This method relies too much on manpower, has low reliability, and consumes manpower.

For the wafer polishing process, the main problems during loading are: 1. Heavy insertion; 2. Inclined insertion; 3. Upside-down loading. For heavy and slanted insertion, it is easy to cause the machine to hit the wafer and cause damage to the machine and the wafer. However, for this problem, the wafer polishing equipment generally has a scanning function, which can effectively solve the above problems, but the phenomenon of loading is reversed. At present, there is no effective solution on the market. The main dangers of reverse loading are as follows: 1. The product is scrapped. After the wafer is reversed, the equipment grinding wheel will directly grind the chip surface, which will completely grind the chip away, causing the product to be completely scrapped. 2. The grinding wheel is scrapped. Since the front of the product will be glued, the teeth of the grinding wheel and the glue film will damage the teeth of the grinding wheel during the grinding process, causing the grinding wheel to be scrapped; 3. The grinding table is damaged. When the wafer is ground to a very thin degree, Since the protective film on the front side will be worn off if it is placed upside down, the product will be severely split, and the silicon slag from the split will cause scratches on the grinding table.

Technical solutions

The purpose of this application is to solve at least one of the above-mentioned problems in the prior art, and this purpose is achieved through the following technical solutions:

The first aspect of the present application provides a wafer polishing equipment. The wafer polishing equipment includes: a polishing table, the polishing table is used to place the wafer; a camera device, the camera device is arranged above the polishing table, used to photograph the polishing table The controller, the controller is electrically connected with the camera device, and is used to receive the wafer image taken by the camera device; the alarm, the controller is electrically connected with the alarm, and the controller sends out prompt information through the alarm according to the wafer image .

In an embodiment, the wafer polishing equipment further includes a polishing head disposed facing the polishing table, and a rotatable polishing pad is provided on the polishing head, and the controller operates the polishing pad to polish the surface of the wafer according to the wafer image.

In one embodiment, the wafer polishing equipment further includes a flipping device arranged at the bottom of the polishing table and connected to the polishing table, and the controller adjusts the tilt angle of the polishing table and the wafer through the flipping device according to the wafer image.

In an embodiment, the wafer polishing equipment further includes an installation groove provided on the polishing table, and an inner wall of the installation groove is provided with a retractable clamping block for clamping the wafer.

In an embodiment, the alarm includes an audible and visual alarm.

The second aspect of the present application provides a wafer polishing method. The wafer polishing method is executed according to the wafer polishing equipment of the first aspect of the present application. The wafer polishing method includes: controlling the guidance in the wafer polishing equipment The program module starts after the wafer is loaded, and guides the operation module in the wafer polishing equipment to start through the boot program module; controls the operation module to start the camera device in the wafer polishing equipment, and imports the image captured by the camera device into the wafer image Recognize the model, determine whether there is a wafer in the screen through the wafer image recognition model; if there is a wafer in the screen, determine the current state of the wafer according to the pre-stored wafer state in the wafer image recognition model; according to the current state of the wafer Control the alarm to send out prompt information.

In one embodiment, controlling the boot program module in the wafer polishing equipment to start after the wafer is loaded, and guiding the operation module in the wafer polishing equipment through the boot program module to start before starting includes: using the camera in the wafer polishing equipment The device shoots multiple training images containing pre-stored wafer states, and trains a wafer image recognition model through the convolutional neural network and multiple training images.

In one embodiment, the imaging device in the wafer polishing equipment is used to capture multiple training images containing pre-stored wafer states, and training the wafer image recognition model through the convolutional neural network and multiple training images specifically includes: extracting multiple images. Multiple feature values of multiple pre-stored wafer states in a training screen; multiple feature values are converted into numerical values through convolution calculation and pooling calculation and stored in the convolution kernel; the convolution kernel is passed through the fully connected layer to generate predictions Value and store; use the gradient descent method to measure the error between the predicted value and the true value through the loss function layer, and optimize and save the predicted value; repeatedly shoot multiple training images, through the convolutional neural network and repeatedly shoot multiple training images Train a wafer image recognition model.

In an embodiment, extracting multiple feature values of multiple pre-stored wafer states in multiple training screens specifically includes: adding the upper edge, lower edge, left edge, right edge, upper left corner, and upper left corner of the pre-stored wafer state in the training screen. The upper right corner, the lower left corner and the lower right corner are used as feature values and stored in the convolution kernel.

In an embodiment, converting multiple feature values into numerical values through convolution calculation and pooling calculation and storing them in the convolution kernel specifically includes: performing the first convolution calculation and the first pooling calculation on the multiple feature values And stored in the convolution kernel; the second convolution calculation and the second pooling calculation are performed on the multiple feature values in the convolution kernel after the first convolution calculation and the first pooling calculation.

Those skilled in the art can understand that by adding an imaging device to the polishing table inside the wafer polishing equipment, the normally loaded wafers are photographed through the imaging device, and then the information including the state of the pre-stored wafer is recorded at this time. For multiple training screens, during subsequent loading, the camera device will take pictures of the wafer and generate a wafer image, and then compare the wafer image with the pre-stored wafer status in multiple training screens. If the wafer is normally loaded, The wafer image is consistent with the state of the pre-stored wafer in multiple training screens, and the wafer polishing equipment is working normally; if the wafer is not loaded normally (such as the wafer position is reversed), the wafer image and the multiple training screens The pre-stored wafer state has a large error value. When the error value exceeds the preset error range, the wafer polishing equipment will send a warning message through an alarm, and the wafer polishing equipment will stop working until the wafer is adjusted to a normal state.

Description of the drawings

By reading the detailed description of the preferred embodiments below, various other advantages and benefits will become clear to those of ordinary skill in the art. The drawings are only used for the purpose of illustrating the preferred embodiments, and are not considered as a limitation to the application. Also, throughout the drawings, the same reference symbols are used to denote the same components. In the attached picture:

FIG. 1 is a schematic diagram of a partial structure of a wafer polishing equipment according to an embodiment of the application.

FIG. 2 is a schematic flowchart of a wafer polishing method according to an embodiment of the application.

Among them, the reference signs are as follows:

10. Wafer;

20. Workbench;

30. Camera device.

Embodiments of the present invention

Hereinafter, exemplary embodiments of the present disclosure will be described in more detail with reference to the accompanying drawings. Although the drawings show exemplary embodiments of the present disclosure, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

It should be understood that the terms used in the text are only for the purpose of describing specific example embodiments, and are not intended to be limiting. Unless the context clearly indicates otherwise, the singular forms "a", "an" and "said" as used in the text may also mean that the plural forms are included. The terms "including", "including" and "having" are inclusive, and therefore indicate the existence of the stated features, elements, and/or components, but do not exclude the existence or addition of one or more other features, elements, and components , And/or their combination.

In addition, in the description of this application, unless otherwise clearly specified and limited, the terms "set" and "connection" should be understood in a broad sense, for example, it may be a fixed connection, a detachable connection, or an integral connection; It can be directly connected or indirectly connected through an intermediary. For those skilled in the art, the specific meaning of the above-mentioned terms in this application can be understood according to specific circumstances.

For ease of description, spatial relative relation terms may be used in the text to describe the relation of one element or feature relative to another element or feature as shown in the figure. These relative relation terms are, for example, "upper", "facing", "bottom". ", "内", "下", etc. This spatial relative relationship term is intended to include different positions of the mechanism in use or operation other than those depicted in the figure. For example, if the mechanism in the figure is turned over, then elements described as "below other elements or features" or "below other elements or features" will then be oriented as "above other elements or features" or "below other elements or features" Above features". Thus, the example term "below" can include an orientation of above and below. The mechanism can be otherwise oriented (rotated by 90 degrees or in other directions) and the spatial relationship descriptors used in the text are explained accordingly.

As shown in Figures 1 and 2, according to the embodiments of the present application, the first aspect of the present application provides a wafer polishing equipment. The wafer polishing equipment includes a polishing table, a camera device, a controller, and an alarm. Used to place wafers, the camera device is set above the polishing table, used to capture the wafer image on the polishing table, the controller is electrically connected to the camera device, used to receive the wafer image taken by the camera device, the controller and the alarm Electrically connected, the controller sends out prompt information through an alarm according to the wafer image.

In this embodiment, by adding an imaging device to the polishing table inside the wafer polishing equipment, the normally loaded wafers are photographed through the imaging device, and then multiple training sessions including pre-stored wafer states are recorded at this time. During subsequent loading, the camera device will take a picture of the wafer and generate a wafer image, and then compare the wafer image with the pre-stored wafer status in multiple training screens. If the wafer is normally loaded, the wafer image Consistent with the pre-stored wafer status in multiple training screens, the wafer polishing equipment is working normally; if the wafer is not loaded normally (such as the wafer position is reversed), the wafer image and the pre-stored wafers in multiple training screens The state error value is relatively large. When the error value exceeds the preset error range, the wafer polishing equipment sends out a warning message through an alarm, and the wafer polishing equipment stops working until the wafer is adjusted to a normal state.

Continuing to refer to FIG. 1, according to the first embodiment of the present application, the wafer polishing equipment further includes a polishing head disposed facing the polishing table, and a rotatable polishing pad is provided on the polishing head. The controller operates the polishing pad to grind the crystal according to the wafer image. Round surface. Further, the wafer polishing equipment further includes a flipping device arranged at the bottom of the polishing table and connected to the polishing table, and the controller adjusts the tilt angle of the polishing table and the wafer through the flipping device according to the wafer image.

In this embodiment, the polishing pad can operably contact the surface of the wafer according to the wafer image. At the same time, the turning device adjusts the tilt angle of the polishing table and the wafer according to the wafer image, so that the polishing pad can align the wafer at a suitable position. The circle is polished at an appropriate angle, thereby reducing the misoperation of the polishing pad when polishing the wafer, and improving the polishing accuracy of the wafer by the wafer polishing equipment.

Continuing to refer to FIG. 1, according to the first embodiment of the present application, the wafer polishing equipment further includes an installation groove provided on the polishing table, and the inner wall of the installation groove is provided with a retractable clamp block for clamping the wafer.

In this embodiment, the wafer is installed in the mounting groove and clamped by the retractable clamping block, so as to reduce the phenomenon of wafer displacement when the workbench adjusts the state of the wafer.

Continuing to refer to Fig. 1, according to the first embodiment of the present application, the alarm includes an audible and visual alarm.

In this embodiment, when the state of the wafer needs to be adjusted by the flipping device, the acousto-optic alarm will emit a flashing light during the flipping process of the wafer to indicate that the wafer is in the flipping process. When adjusting the state of the wafer, the audible and visual alarm emits a flashing light and a voice message to inform the user that the state of the wafer needs to be adjusted.

As shown in FIG. 2, the second aspect of the present application provides a wafer polishing method. The wafer polishing method is executed according to the wafer polishing apparatus of the first aspect of the present application. The wafer polishing method includes: The boot program module in the circular polishing equipment is activated after the wafer is loaded, and the boot program module guides the operation module in the wafer polishing equipment to start; the control operation module activates the camera device in the wafer polishing equipment, and captures the image of the camera device. Import the wafer image recognition model into the wafer image recognition model, and determine whether there is a wafer in the image according to the wafer image recognition model; if there is a wafer in the image, determine the current state of the wafer according to the pre-stored wafer status in the wafer image recognition model; According to the current state of the wafer, the alarm is controlled to send out prompt information.

According to an embodiment of the present application, controlling the boot program module in the wafer polishing equipment to start after the wafer is loaded, and guiding the operation module in the wafer polishing equipment through the boot program module to start before starting includes: The camera device of ”captures multiple training images containing pre-stored wafer states, and trains a wafer image recognition model through convolutional neural networks and multiple training images.

According to an embodiment of the present application, shooting multiple training images containing pre-stored wafer states through the camera device in the wafer polishing equipment, and training the wafer image recognition model through the convolutional neural network and multiple training images specifically includes: Extract multiple feature values of multiple pre-stored wafer states in multiple training screens; convert multiple feature values into numerical values through convolution calculation and pooling calculation and store them in the convolution kernel; pass the convolution kernel through the fully connected layer Generate predicted values and store them; use the gradient descent method to measure the error between the predicted value and the true value through the loss function layer, and optimize and save the predicted value; repeatedly shoot multiple training images, through the convolutional neural network and multiple repeated shots The training screen trains the wafer image recognition model.

According to an embodiment of the present application, extracting multiple feature values of multiple pre-stored wafer states in multiple training screens specifically includes: adding the upper edge, lower edge, left edge, right edge, and upper left edge of the pre-stored wafer state in the training screen. The corner, upper right corner, lower left corner, and lower right corner are used as feature values and stored in the convolution kernel.

According to an embodiment of the present application, converting multiple feature values into numerical values through convolution calculation and pooling calculation and storing them in the convolution kernel specifically includes: performing the first convolution calculation and the first pooling on the multiple feature values The convolution calculation is stored in the convolution kernel; the second convolution calculation and the second pooling calculation are performed on the multiple eigenvalues in the convolution kernel after the first convolution calculation and the first pooling calculation.

This application is based on the principle of the wafer image recognition model trained by the deep neural network model through training data as follows:

Step 1: Read the pre-stored wafer status data set and predefine the data;

Read the pre-stored wafer status data set captured at the official installation location using a monochrome camera with a 1080p resolution of 1920*1080 pixels, to ensure that the trained wafer image recognition model and the wafer image angle captured by the camera after the camera are installed are used to read the pre-stored wafer status data set. ?????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????? The positions are the same, and the total pixels in the model are defined as 1920*1080=2088960;

Step 2: Set the weight and offset value function;

Generate random variables: The generated value obeys the normal distribution with the specified average and standard deviation. If the generated value is greater than 2 standard deviations from the average, it will be discarded and reselected;

Output a random value from the truncated normal distribution, and select a random value near the mean value of the normal distribution = 0.1;

Step 3: Define the convolution function and pooling function

Enter the image information matrix, the value of the convolution kernel, the step length of the convolution kernel to move to the right and down, the right and down steps of the convolution calculation are set to 1, the right and downward of the pooling calculation The step size is set to 2.

The size of the convolution kernel is set to 320*180 pixels, and the image features of the pre-stored wafer state are converted into numerical values and placed in the convolution kernel.

Step 4: The first convolution + pooling

Convolutional layer 1 network structure definition

Convolution kernel 1: Since the size of the convolution kernel here is 320*180, the number of input channels is 1, and the number of output channels is 32.

The size of the output picture after the first convolution is 1920*1080*32

In order to reduce the calculation, the picture is pooled, the size of the output picture after the first pooling is 960*540, and the non-linear processing is performed through the activation function ReLU.

Linear rectification function (Rectified Linear Unit, ReLU), also known as modified linear unit, is a commonly used activation function (activation function) in artificial neural networks, usually refers to the non-linear function represented by the ramp function and its variants. ?????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????

In the usual sense, linear rectification function refers to the ramp function in mathematics, that is, f(x)=max(0,x). In neural networks, linear rectification is used as the activation function of the neuron, which defines the linearity of the neuron. The non-linear output result after transforming w ^{T x+b.} For the input vector x from the upper layer of neural network into the neuron, the neuron using the linear rectification activation function will output max(0, w ^T x+b). To the next layer of neurons or as the output of the entire neural network (It depends on where the current neuron is in the network structure).

Take the average value in the area of the matrix 2*2, and the pooling step size is 2.

The value of the convolution kernel here is equivalent to the weight value, which is obtained by generating a random number sequence

Since the picture size of the pre-stored wafer state data set is 1920*1080, and it is black and white, the accurate picture size is 1920*1080*1 (1 means that the picture has only one color layer, and the color pictures are all 3 color layers. ——RGB), so after the first convolution, the number of output channels changes from 1 to 32, and the picture size becomes: 1920*1080*32 (equivalent to stretched height)

After the first pooling (pooling step size is 2) and activation, the picture size is 960*540*32

Step 5: The second convolution + pooling

Convolutional layer 2 network structure definition

Convolution kernel 2: The size of the second convolution kernel is also 320*180, the number of input channels is 32, and the number of output channels is 64.

The size of the output image after the second convolution is 960*540*64

In order to further reduce the amount of calculation, the second pooling activation (pooling step size is 2), the size of the output picture after the pooling activation is 480*270*64.

Step 6: Set up fully connected layer 1, fully connected layer 2

Fully connected layer 1

The input of the fully connected layer 1 is the output after the second pooling, the size is 480*270*64, and the fully connected layer 1 has 1024 neurons.

Calculate the other shape attribute values of the array according to the existing dimensions. For example, a three-dimensional array is [[[0], [1]], [[2], [3]], [[4], [5]]] , Then its shape is (3,2,1).

In order to reduce the over-fitting phenomenon. Each time only some neurons are involved in the work to adjust the weights.

Fully connected layer 2

The fully connected layer 2 has 10 neurons, which is equivalent to the generated classifier.

After fully connected layers 1 and 2, the predicted value obtained after the previous convolution pooling is saved.

Step 7: Choose the gradient descent method to optimize the loss function layer and find the accuracy rate

The loss function uses a quadratic cost function to measure the error between the predicted value and the true value.

Because the data set is too large, the gradient descent method is used for learning, and the learning rate is 1e-4. The optimizer used here is the AdamOptimizer optimizer.

Store the result in a boolean list.

Returns the predicted label value for the input.

In order to calculate the accuracy of the classification, the returned Boolean array is converted into a floating point number to represent right and wrong, and then the average value is taken.

Step 8: Set other parameters, save parameters

Set the original image data package to come from the DangCheQiang data set, one batch contains 50 data

Save model parameters

Setting the neuron participation rate to 0.5, only half of the neurons are involved in the work.

Step 9: Repeat the operation 10,000 times to obtain a more accurate wafer image recognition model;

Every time it runs, the difference between the result of the operation and the actual image is evaluated by the loss function. After running 10,000 times, the recognition rate of the wafer image recognition model can reach more than 95%.

The above are only preferred specific implementations of this application, but the protection scope of this application is not limited to this. Any person skilled in the art can easily think of changes or changes within the technical scope disclosed in this application. Replacement shall be covered within the scope of protection of this application. Therefore, the protection scope of this application should be subject to the protection scope of the claims.

Claims

A wafer polishing equipment, wherein the wafer polishing equipment includes:

A polishing table, which is used to place wafers;

An imaging device, the imaging device is arranged above the polishing table, and is used to capture an image of the wafer on the polishing table;

A controller, which is electrically connected to the camera device, and is configured to receive wafer images taken by the camera device;

An alarm, the controller is electrically connected to the alarm, and the controller sends out prompt information through the alarm according to the wafer image.
The wafer polishing equipment according to claim 1, wherein the wafer polishing equipment further comprises a polishing head disposed facing the polishing table, a rotatable polishing pad is disposed on the polishing head, and the controller is configured according to The wafer image operates the polishing pad to polish the surface of the wafer.
The wafer polishing equipment according to claim 1, wherein the wafer polishing equipment further comprises a flip device arranged at the bottom of the polishing table and connected to the polishing table, and the controller is based on the wafer The image is used to adjust the inclination angle of the polishing table and the wafer through the turning device.
The wafer polishing equipment according to claim 1, wherein the wafer polishing equipment further comprises a mounting groove provided on the polishing table, and an inner wall of the mounting groove is provided with a retractable structure for clamping the wafer. Stuck.
The wafer polishing apparatus according to claim 1, wherein the alarm includes an acousto-optic alarm.
A wafer polishing method, wherein the wafer polishing method is performed according to the wafer polishing equipment of any one of claims 1 to 5, and the wafer polishing method comprises:

Controlling the boot program module in the wafer polishing equipment to start after the wafer is loaded, and guiding the operation module in the wafer polishing equipment to start through the boot program module;

Controlling the operation module to activate the camera device in the wafer polishing equipment, import the picture taken by the camera device into a wafer image recognition model, and determine whether there is a wafer in the picture according to the wafer image recognition model;

According to the wafer in the screen, the current state of the wafer is determined according to the pre-stored wafer state in the wafer image recognition model;

The alarm is controlled to send out prompt information according to the current state of the wafer.
The wafer polishing method according to claim 6, wherein the boot program module in the control wafer polishing equipment is activated after the wafer is loaded, and the boot program module guides the boot program module in the wafer polishing equipment Before the operation module starts, it includes:

A plurality of training images including pre-stored wafer states are captured by the camera device in the wafer polishing equipment, and a wafer image recognition model is trained through a convolutional neural network and a plurality of training images.
The wafer polishing method according to claim 7, wherein the imaging device in the wafer polishing equipment captures a plurality of training images including a pre-stored wafer state, and a convolutional neural network and a plurality of training images The trained wafer image recognition model specifically includes:

Extract multiple feature values of multiple pre-stored wafer states in multiple training screens;

Through convolution calculation and pooling calculation, multiple feature values are converted into numerical values and stored in the convolution kernel;

Pass the convolution kernel through the fully connected layer to generate and store the predicted value;

Use the gradient descent method to measure the error between the predicted value and the true value through the loss function layer, and optimize and save the predicted value;

Multiple training images are repeatedly shot, and the wafer image recognition model is trained through a convolutional neural network and multiple training images repeatedly shot.
8. The wafer polishing method according to claim 8, wherein said extracting a plurality of characteristic values of a plurality of pre-stored wafer states in a plurality of training images specifically comprises:

The upper edge, lower edge, left edge, right edge, upper left corner, upper right corner, lower left corner and lower right corner of the pre-stored wafer status in the training screen are used as feature values and stored in the convolution kernel.
The method for detecting a vehicle retaining wall based on a camera according to claim 9, wherein said converting multiple feature values into numerical values through convolution calculation and pooling calculation and storing them in the convolution kernel specifically comprises:

Perform the first convolution calculation and the first pooling calculation for multiple eigenvalues and store them in the convolution kernel;

Perform the second convolution calculation and the second pooling calculation on the multiple feature values in the convolution kernel after the first convolution calculation and the first pooling calculation.