Wafer test long die detection and elimination method based on YOLO target detection

Abstract

The invention relates to the technical field of semiconductor testing, and particularly discloses a Wafer test long die detection and elimination method based on YOLO target detection, and the method comprises the steps: obtaining and preprocessing a wafermap image; carrying out information labeling on the Wafermap image after the preprocessing is completed; randomly distributing the prepared wafermap data set according to a preset proportion; detecting rules of the isolated chips are defined on the basis of the specified lonely die type, automatic identification of the isolated chips is carried out by combining the detection result output by the YOLO, the chips meeting the conditions of the isolated chips are marked and removed, a report containing the coordinates and the number of the isolated chips is generated, and visualization is carried out on a wafermap image. Compared with a manual screening mode, the detection efficiency is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of semiconductor testing technology, and more specifically to a method for detecting and eliminating lonely dies in Wafer testing based on YOLO target detection. Background Technology

[0002] In semiconductor wafer testing, to ensure chip quality and reliability, it is typically necessary to perform bin-by-bin testing on the chips on the wafer and visualize the chip distribution using a wafer map. In actual production, a common quality issue is isolated chips, where the chip at the center of the 3x3 grid is in the "Pass" state (passes the relevant test and meets expectations), but the surrounding chips are all in the "Fail" state (fails the relevant test and does not meet expectations). These isolated chips often pose a high quality risk and need to be rejected in subsequent processes.

[0003] Currently, the detection and removal of such isolated chips usually relies on manual judgment or rule-based algorithms, but these methods have problems such as low efficiency and easy to miss detection.

[0004] The main purpose of object detection is to identify all objects in an image that meet specified criteria based on human interest, and to analyze the category and location of the detected objects. Currently, there are two main categories of deep learning-based object detection algorithms: the first category is two-stage object detection algorithms, such as the R-CNN (Region-based Convolutional Neural Networks) series. These algorithms have high detection accuracy but poor real-time performance. The other category is one-stage object detection algorithms, such as the SSD (Single Shot MultiBox Detector) and YOLO (You Only Look Once) series. These algorithms often have faster detection speeds, sacrificing some detection accuracy, but achieving real-time detection, thus better meeting practical requirements. Therefore, these algorithms are more widely used.

[0005] To address the shortcomings of traditional isolated die detection and rejection methods, this invention proposes a Wafer test method for isolated die detection and rejection based on YOLO target detection. Summary of the Invention

[0006] The purpose of this invention is to provide a method for detecting and eliminating lonely dies in Wafer testing based on YOLO object detection, thereby solving the following technical problems.

[0007] The objective of this invention can be achieved through the following technical solutions: The Wafer test lonely die detection and rejection method based on YOLO object detection includes the following steps: The customer provides the wafer and generates the original wafer map image using ATE testing equipment; The wafermap image is acquired and preprocessed, including image segmentation, data augmentation, binarization, and normalization. The preprocessed Wafermap image is annotated with information, indicating the state (pass / fail) of each die and its position in the wafer, and then converted into YOLO dataset format; The prepared wafermap dataset is randomly allocated in a 7:2:1 ratio, that is, 70% is used as the training set, 20% as the validation set, and 10% as the test set. The dataset is used to train the YOLO model. The goal of the model training is to identify the state of each die in the wafermap image and its spatial distribution. The trained YOLO model is used to detect the wafer map under test. The detection rules for isolated chips are defined based on the specified lonely die type. The specific rules are as follows: a. All eight chips around the perimeter are in a fail state; b. The chip located in the center of the 3x3 grid (lonely die); c. For chips at the edge of the wafer, if their adjacent die is a fail, they are also judged as lonely dies (without the need for 8 surrounding fail dies). d. "Lonely die" can refer to a collection, not just individual deaths. e. Lonely die, a type of failure that appears as several rows or columns of failed chips in the wafer map; Based on the detection results output by YOLO, and combined with the above rules, isolated chips are automatically identified. Chips that meet the isolated chip criteria are marked and removed, and a report containing the coordinates and number of isolated chips is generated. The information of isolated chips is transmitted to the subsequent production process to achieve automatic removal, and the data is visualized on the wafer map image.

[0008] The beneficial effects of this invention are as follows: Based on the YOLO target detection algorithm, isolated chips can be identified and rejected without manual intervention, significantly improving detection efficiency compared to manual screening. Through deep learning models and strict isolated chip definition rules, missed detections and false detections can be effectively avoided. It is widely applicable to different types of wafer testing equipment and production environments, and can be further extended to other types of abnormal chip detection and rejection scenarios. Attached Figure Description

[0009] The invention will now be further described with reference to the accompanying drawings.

[0010] Figure 1 This is a schematic diagram illustrating the basic principle structure of the YOLO algorithm in this embodiment of the invention; Figure 2 This is the lone die wafermap image with no pass in this embodiment of the invention (left image), and the lone die wafermap image with a nine-grid rule failure type (right image). Figure 3 In this embodiment of the invention, there is a lonely die wafermap with set failure type (left) and a lonely die wafermap with edge failure type (right). Figure 4 This is a lone die wafermap diagram showing a strip-shaped failure type in this embodiment of the invention. Figure 5 This is a diagram illustrating the process of YOLO algorithm detecting Lonely Die targets in an embodiment of the present invention; Figure 6 This is an overall flowchart of the detection of Lonely Die based on the YOLO algorithm in an embodiment of the present invention; Figure 7 This is the original WaferMap generated after testing by the ATE testing equipment in this embodiment of the invention. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] This invention is a method for detecting and removing lonely dies in wafer testing based on YOLO target detection. By performing automated target detection and analysis on wafer map images, it can accurately identify and remove dies that meet the criteria for isolated chips, thereby improving detection efficiency and accuracy.

[0013] like Figure 7 The original Wafermap image shown is generated using the following principle: In Wafer testing, the generation of Wafermap images is a visualization process that spatially maps physical locations to electrical / functional test results. Its core principle is as follows: 1. Fundamentals of Wafer Coordinate Systems: a. Die grid layout: The chips (dies) on the wafer are arranged strictly in a row and column matrix (e.g., Row 1-100, Column 1-100). b. Coordinate positioning: The probe card of the ATE tester ensures that the test point (Pad) of each die can be accurately contacted through precise mechanical control; c. Location recording: When testing each die, the ATE test system automatically records its coordinate position (X, Y).

[0014] 2. Binding test data to location: a. Test execution: ATE applies test vectors (such as functional tests and parameter tests) to each die. b. Result classification: Pass / Fail (Basic Category); Bin Code (Fine-level classification: e.g., Bin1 = All Passed, Bin2 = Connectivity Failed, Bin3 = Failed for a Specific Function, etc.). c. Data storage: The test result (Bin Code) of each Die is stored in association with its coordinates (X, Y); The data format is typically structured text (such as STDF files) or a database.

[0015] 3. Wafermap generation algorithm: a. Data mapping: Convert the stored (X, Y, Bin) data into a two-dimensional matrix; b. Color mapping: Assign a unique color to each Bin Code; c. Image rendering: Render each square containing a die with its mapped color based on its coordinates, leaving blank / untested areas as white or gray.

[0016] 4. Key technology processing: a. Wafer edge recognition: Determines boundaries through wafer notch or contour scanning, and automatically removes invalid edge points; b. Normalization process: Coordinate alignment: compensating for the offset between the testing machine and the photolithography coordinate system; Rotation correction: Addresses wafer placement angle deviations.

[0017] The basic principle of the YOLO algorithm: The core idea of ​​the YOLO algorithm is to transform the object detection task into a regression problem. It divides the input image into several grids, and each grid is responsible for predicting a fixed number of bounding boxes and whether an object exists in those bounding boxes.

[0018] Specifically, the YOLO algorithm mainly includes the following steps: 1. Image segmentation: The input image is divided into an s*s grid, with each grid cell responsible for predicting b bounding boxes and their confidence scores. The confidence score indicates the probability of an object being present within the bounding box and the accuracy of the bounding box.

[0019] 2. Feature extraction: Convolutional neural networks are used to extract image features. These features typically include a variety of information such as color, texture, and shape, and are crucial for subsequent object detection.

[0020] 3. Bounding box prediction: Each grid cell predicts the coordinates (x, y, w, h) and confidence scores of b bounding boxes based on the extracted features. Here, (x, y) represents the center coordinates of the bounding box, and (w, h) represents the width and height of the bounding box.

[0021] 4. Category prediction: Each grid cell predicts the probability of c categories. These probabilities represent the likelihood of an object belonging to a different category within that grid cell.

[0022] 5. Nonmaximum suppression: After obtaining the prediction results for all grids, redundant bounding boxes are removed using a non-maximum suppression algorithm, retaining the best detection results.

[0023] like Figure 1 As shown, the CBL module is a fundamental component of the YOLOv3 model. The Darknet network's convolutional layers are followed by the BatchNormalization (BN) and Leakyrelu modules. Except for the last convolutional layer, BN and Leakyrelu are already the smallest inseparable parts of the convolutional layers, together forming the minimum components.

[0024] The YOLOv3 backbone network uses a total of 5 Resx structures, where x represents the quantity. For example, Res1, Res2, Res4, and Res8 indicate that the Res_block consists of x Res_units. An explanation of the Res_block module can also be found at [link to documentation]. Figure 1 As seen in the bottom right corner of the result, its basic component is also CBL.

[0025] The prediction branch includes a tensor concat operation. This is implemented by concatenating the upsampled data from an intermediate layer in Darknet with the data from a layer immediately following it. It's important to note that tensor concatenation differs from the add operation in the Res_unit structure. Tensor concatenation expands the tensor's dimensions, while add simply adds the data without changing the tensor's dimensions.

[0026] Method and steps: 1. Generating raw data: The customer provides the wafer raw materials, and the fab testing plant creates a corresponding Device Map based on the raw materials to ensure that the Device Map matches the initial pattern of the wafer. The wafer is subjected to relevant tests on the ATE test equipment, and the initial wafer map image is generated using the self-designed and developed WaferMap tool.

[0027] 2. Data preparation: A large amount of wafermap image data was collected and preprocessed, including image segmentation, data augmentation, binarization and normalization. The preprocessed wafermap image data is labeled, that is, the state (pass / fail) of each die and its position in the wafer are labeled; The wafermap image data with labeled target information is converted into YOLO dataset format to facilitate subsequent model training.

[0028] 3. Model Training: The prepared wafermap dataset is randomly distributed in a 7:2:1 ratio, that is, 70% of the images in the dataset are randomly selected as the training set, 20% as the validation set, and 10% as the test set. The goal of model training is to identify the state of each die in the wafermap image and its spatial distribution.

[0029] 4. Isolated chip detection rules: Detection rules for isolated chips are defined using algorithms: a. All eight chips around the perimeter are in a fail state; b. The chip located in the center of the 3x3 grid (lonely die); c. For chips at the edge of the wafer, if their adjacent die is a fail, they are also judged as lonely dies (without the need for 8 surrounding fail dies). d. "Lonely die" can refer to a collection, not just individual deaths. e. Lonely die, a type of failure that appears as several rows or columns of failed chips in the wafer map; Based on the detection results output by YOLO, and combined with the above rules, isolated chips are automatically identified.

[0030] 5. Result Removal and Output: Generate a report containing the coordinates and number of isolated chips; The information of isolated chips is transmitted to subsequent production processes to achieve automatic rejection; The detected isolated chips are labeled and visualized on wafermap images.

[0031] like Figure 5 As shown, when using the YOLO algorithm to detect Lonely Dies, the preprocessed wafer map image is first input into the deep learning network, dividing the image into an S*S grid. Next, it checks whether the center of the Lonely Die in the image lies within any grid cell, selecting these grid cells that may contain the target as Regions of Interest (ROIs). The size and position of B bounding boxes, as well as their confidence scores, are predicted. Then, a maximum suppression method is used to remove redundant bounding boxes. Finally, the target's category is determined based on its features, thus completing the image feature extraction. After feature extraction, the detection and removal of Lonely Dies can be achieved by labeling the center point of the feature target in each grid cell.

[0032] The overall process of detecting Lonely Die based on the YOLO algorithm is as follows: Figure 6As shown. First, a wafer map is acquired (this image can be directly obtained from the CP test) and preprocessed accordingly—image segmentation (removing unnecessary information and retaining only the wafer map itself), data augmentation (expanding the diversity of the wafer map to lay the foundation for subsequent generalization), binarization (processing the color markers in the wafer map to make their features more prominent), and normalization are performed. Next, the preprocessed wafer map is labeled, with each die marked as pass / fail. Then, the wafer map image data containing target information is converted to YOLO dataset format for easy import into the algorithm model for processing. During training, the dataset is randomly allocated in a 7:1:2 ratio, that is, 70% of the images in the dataset are randomly selected as the training set, 10% as the test set, and 20% as the validation set. The model is trained for 100 rounds. After the model training is completed, its algorithm can accurately identify and remove dies that meet the isolated chip condition when detecting Lonely Dies, improving detection efficiency and accuracy.

[0033] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A method for detecting and eliminating lonely dies in Wafer testing based on YOLO object detection, characterized in that, Includes the following steps: Step 1: Acquire and preprocess the wafermap image. Preprocessing includes image segmentation, data augmentation, binarization, and normalization. Step 2: Annotate the preprocessed Wafermap image, labeling the state of each die and its position in the wafer, and convert it into YOLO dataset format; Step 3: Randomly distribute the prepared wafermap dataset according to a preset ratio, and use the dataset to train the YOLO model. The goal of the model training is to identify the state of each die in the wafermap image and its spatial distribution. Step 4: Use the trained YOLO model to detect the wafer map under test; Step 5: Define the detection rules for isolated chips based on the specified lonely die type, combine the detection results output by YOLO to automatically identify isolated chips, mark and remove chips that meet the isolated chip conditions, generate a report containing the coordinates and quantity of isolated chips, pass the information of isolated chips to the subsequent production process to achieve automatic removal, and visualize it on the wafer map image.

2. The Wafer test lonely die detection and rejection method based on YOLO target detection according to claim 1, characterized in that, In step 1, image segmentation removes unnecessary information from the wafer map image, retaining only the wafer map itself; data augmentation enhances the diversity of the wafer map; and binarization processes the colors in the wafer map to make its features more prominent. Normalization is a process of standardizing image data.

3. The Wafer test lonely die detection and rejection method based on YOLO target detection according to claim 1, characterized in that, In step 2, when marking the position of each die in the wafer, a coordinate positioning method is used to accurately record the position information of each die.

4. The Wafer test lonely die detection and rejection method based on YOLO target detection according to claim 1, characterized in that, In step 3, during the YOLO model training process, a multi-round training method is adopted, and the prepared wafermap dataset is randomly distributed in a 7:2:1 ratio, where 70% of the images are used as the training set, 20% of the images are used as the validation set, and 10% of the images are used as the test set.

5. The Wafer test lonely die detection and rejection method based on YOLO target detection according to claim 1, characterized in that, In step 5, the detection rules for isolated chips are as follows: a. In a 3x3 grid structure, the chip located in the center of the grid and surrounded by 8 chips that are all in a fail state is an isolated chip. b. For chips at the edge of the wafer, if their adjacent die is a fail, they are also considered isolated chips. d. An isolated chip can be a collection of several adjacent dies, not a single die; e. Lonely die, a type of failure that appears as several rows or columns of failed chips in the wafer map.

6. The Wafer test lonely die detection and rejection method based on YOLO target detection according to claim 1, characterized in that, In step 5, when visualizing the wafermap image, the detected isolated chips are specially marked to distinguish them from other chips.

7. The Wafer test lonely die detection and rejection method based on YOLO target detection according to claim 1, characterized in that, In step 3, the YOLO model uses the YOLOv3 model, whose backbone network uses 5 Resx structures and has tensor splicing operations on the prediction branches.

8. The Wafer test lonely die detection and rejection method based on YOLO target detection according to claim 1, characterized in that, In step 4, when the trained YOLO model is used to detect the wafer map to be tested, the input image is divided into several grids, and each grid is responsible for predicting a fixed number of bounding boxes and whether there is a target in these bounding boxes, thereby realizing the detection of lonely die.

Images