High-speed prediction method for die bonding shift
Patent Information
- Application Number
- TW114106930
- Authority / Receiving Office
- TW · TW
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2025-02-25
- Publication Date
- 2026-09-01
- Estimated Expiration
- 2045-02-24
AI Technical Summary
Conventional die-bonding offset detection methods require post-package testing, leading to reduced product yield and increased scrapping due to poor electrical connections, and lack a high-speed, accurate prediction method for die-bonding offset.
A high-speed prediction method using a deep learning model with a YOLOv8 architecture combined with a RepGhost module for real-time die-bonding offset detection, which includes data collection, training, identification, prediction, and adjustment of die-bonding machine parameters to improve yield and accuracy.
Enables rapid training, real-time detection, and prediction of die-bonding offset, enhancing production efficiency and yield by automatically adjusting die-bonding parameters, thereby reducing defects and costs.
Smart Images
Figure TWG2TA001073963_001 
Figure TWG2TA001073963_002 
Figure TWG2TA001073963_003
Abstract
Description
Technical Field
[0001] The present invention relates to a method for improving a semiconductor packaging process, and more particularly to a high-speed prediction method for die-bonding offset for improving die-bonding yield. Prior Art
[0002] During the semiconductor packaging process, the multiple dies on the finished wafer are first tested for electrical properties and appearance. The wafer is then diced to separate the dies. The selected high-quality dies are then bonded one by one to a carrier through a die bonding process. Wire bonding is then used to electrically connect the die to the carrier's circuits, and the die and circuitry are then covered with a sealant to protect them.
[0003] The die bonding process includes die removal, adhesive / flux application, alignment, heating, pressurization, cooling, curing, and testing. Die offset can lead to poor electrical connections and affect the functionality of the chip product. Conventional automated die bonding equipment not only improves production efficiency but also eliminates die offset by adjusting process parameters. However, conventional die offset detection methods require post-package testing to screen out defective products and identify the cause of the offset before corrective adjustments can be made. This results in reduced product yield and significant costly scrapping.
[0004] In view of this, the conventional die-bond shift detection method still needs to be improved. Summary of the Invention
[0005] To solve the above problems, the present invention aims to provide a high-speed prediction method for die-bonding offset, which can improve product yield and reduce production costs.
[0006] A second object of the present invention is to provide a high-speed prediction method for die-bonding offset, which can improve the training speed of the prediction model and production efficiency.
[0007] Another object of the present invention is to provide a high-speed prediction method for die-bonding offset, which can improve the prediction accuracy.
[0008] The use of the quantifiers "a" or "an" in the elements and components described throughout the present invention is merely for convenience and to provide a general meaning of the scope of the present invention; in the present invention, they should be interpreted as including one or at least one, and the single concept also includes the plural case, unless it is obvious that it means otherwise.
[0009] The high-speed die-bonding offset prediction method of the present invention includes: a data collection step, using multiple die-bonding images and annotating each of the die-bonding images, wherein the annotated content includes multiple die-bonding alignment results, as a training data set; a training step, inputting the data set into an object detection model for image recognition training of the die-bonding results, wherein the object detection model combines the C2f module in the YOLOv8 architecture with the RepGhost module; an identification step, using the trained object detection model to real-time identify a die-bonding image after die-bonding is completed by a machine, and outputting corresponding multiple edge alignment results and multiple corner alignment results; a prediction step, calculating a cumulative value based on the multiple edge alignment results and the multiple corner alignment results, and determining whether the die-bonding result is a good or bad product based on the cumulative value, thereby generating a predicted classification result; and an adjustment step, real-time adjusting a die-bonding control parameter of the die-bonding machine based on the predicted classification result to reduce die-bonding offset.
[0010] The high-speed die-stick offset prediction method of the present invention is based on real-time detection and prediction technology using a deep learning model. It uses a training dataset to train the object detection model in the prediction module. The RepGhost module implements convolution operations with lower computational effort, improving inference speed and compensating for feature loss. It also automatically identifies die-stick images and outputs multiple edge and corner offset results. It then calculates cumulative values and generates a predicted classification result, which is used to determine the quality of the die-stick image and adjust the machine's die-stick control parameters accordingly. This high-speed die-stick offset prediction method allows for rapid training, real-time detection and prediction of die-stick offset, and real-time adjustment of the machine's die-stick control parameters. Furthermore, the object detection model can be calibrated based on actual classification results to improve the accuracy of the prediction module.
[0011] The die-bonding alignment results labeled in the data collection step include four labels: good edge, bad edge, good corner, and bad corner. Thus, the die-bonding alignment results can be used for training, identification, and prediction, effectively operating a training model.
[0012] The dataset includes a true classification result. A confusion matrix is generated based on the true classification result and the predicted classification result, and the accuracy of the object detection model is calculated. This provides an objective assessment of the model's performance and ensures the accuracy of the prediction results.
[0013] In the C2f module of the object detection model, data is input into a convolutional unit, split, and then passed through n RepGhost bottleneck units to generate residuals. This residuals are then concatenated and passed through another convolutional unit before being output. This reduces the amount of convolutional computation and compensates for feature loss, effectively improving inference speed.
[0014] Each RepGhost bottleneck unit has a compression and excitation block located between two RepGhost layers, and has two path modes: true shortcut and false shortcut. This compensates for feature loss after increasing speed, improving prediction accuracy.
[0015] The convolutional unit sequentially comprises a convolutional layer, a batch normalization layer, and a SiLu activation function layer. The convolutional layer downsamples to reduce feature size and increase the number of channels. The batch normalization layer and the SiLu activation function layer are then combined to enhance the model's nonlinear representation capabilities, reducing parameters and improving feature extraction capabilities.
[0016] The object detection model can be of version YOLOv8n, YOLOv8s, or YOLOv8n-ghost. The YOLOv8n-ghost version offers faster training speeds, while the YOLOv8s version offers higher accuracy, allowing users to select the appropriate training model based on their object detection needs.
[0017] The alignment results output by the identification step include good or bad right edge alignment, good or bad left edge alignment, good or bad top edge alignment, good or bad bottom edge alignment, good or bad top right corner alignment, good or bad top left corner alignment, good or bad bottom right corner alignment, and good or bad bottom left corner alignment. Thus, the identification step can determine the die bonding offset conditions of the four sides and four pins, effectively classifying die bonding anomalies.
[0018] If the cumulative value calculated in the prediction step is four sets of good edge alignment or four sets of good corner alignment, the corresponding die-bonding image is determined to be a good die-bonding product. If the cumulative value is three sets of good edge alignment and three sets of good corner alignment, the corresponding die-bonding image is also determined to be a good die-bonding product. If the cumulative value is any other than three sets of good edge alignment, the corresponding die-bonding image is determined to be a defective die-bonding product. In this way, the prediction step can integrate the results of object detection model training and recognition, effectively classifying good and defective die-bonding products.
[0019] The die bonding control parameters include dispensing parameters, placement accuracy, placement speed, downforce, and heating temperature. This adjustment step can adjust the die bonding machine based on the predicted classification results to reduce or eliminate die bonding offset factors, thereby improving die bonding yield and process efficiency. Simple diagram description
[0020] [Figure 1] A flow chart of the steps of a preferred embodiment of the present invention. [Figure 2] A block diagram of the object detection model according to a preferred embodiment of the present invention. [Figure 3] A block diagram of the RepGhost bottleneck unit according to a preferred embodiment of the present invention. [Figure 4] Block diagram of a convolution unit according to a preferred embodiment of the present invention. Implementation Method
[0021] In order to make the above and other objects, features and advantages of the present invention more obvious and easy to understand, the preferred embodiments of the present invention are given below and described in detail with reference to the accompanying drawings.
[0022] Please refer to FIG. 1 , which is a flow chart of a preferred embodiment of the high-speed prediction method for die-bonding offset of the present invention, comprising a data collection step S1 , a training step S2 , an identification step S3 , a prediction step S4 , and an adjustment step S5 .
[0023] The data collection step S1 uses several die bonding images. These can be images of actual die bonding operations captured from the die bonding machine, preferably images of die bonding alignment results captured from the die bonding direction. Each die bonding image is annotated to generate a corresponding XML tag file. This file can include four tags: good edge, bad edge, good corner, and bad corner. This XML tag file can also be converted to VOC format to serve as a training dataset. In this embodiment, the dataset can be split into 70% training data, 15% validation data, and 15% test data, respectively, for training, identification, and prediction, but the present invention is not limited to this.
[0024] The training step S2 involves inputting the data set into an object detection model for image recognition training of die bonding results. In this embodiment, the object detection model used for training is a YOLO (You Only Look Once) model based on the Python framework, such as YOLOv8n, YOLOv8s, or YOLOv8n-ghost, to accelerate die bonding object detection and image recognition.
[0025] Please refer to Figures 2 and 3, which are preferred embodiments of the module relationship diagram of the object detection model. They combine the C2f module in the YOLOv8 architecture with the RepGhost module, replacing the bottleneck of the C2f module with the RepGhost bottleneck for feature extraction. The RepGhost bottleneck achieves convolution operations with less computational effort, thereby increasing inference speed and compensating for feature loss. Reparameterization is used to improve the speed of the object detection model. As shown in Figure 2, in the C2f module, data is input into a convolutional unit 1, undergoes a splitting (P) process, and then passes through n RepGhost bottleneck units 2 to generate residual values. This residual value is then processed through a splicing (C) process and then passes through another convolutional unit 1 for output. As shown in Figure 3, each RepGhost bottleneck unit 2 includes a Squeeze & Excitation (SE) block 21 located between two RepGhost layers 22. This SE block 21 has two path modes: Shortcut = True and Shortcut = False. This uses convolutions with less computational effort to compensate for feature loss after increasing speed.
[0026] As shown in FIG. 4 , the convolution unit 1 may further include a convolution layer 11, a batch normalization layer 12, and a Sigmoid Linear Unit Activation Function (CBS) layer 13. The parameters of the convolution layer 11 may be k=1, s=1, p=0, and c=cout, for downsampling to reduce feature size and increase the number of channels. Furthermore, the batch normalization layer 12 and the Sigmoid Linear Unit Activation Function layer 13 are connected after the convolution layer 11 to enhance the nonlinear representation capability of the model.
[0027] In one embodiment, the training step S2 uses the object detection model to train and infer 366 die-stack images, setting the image size to 320x320, the batch size to 32, and the iteration count to 300. The training and inference results are recorded. For example, the object detection model version YOLOv8n has a training time of 1.1 hours, an inference time of 1.5 milliseconds, uses 3,006,428 parameters, and performs at 8.1 billion floating-point operations per second; the object detection model version YOLOv8s has a training time of 2.6 hours, an inference time of 3 milliseconds, uses 11,127,132 parameters, and performs at 28.4 billion floating-point operations per second; and the object detection model version YOLOv8n-ghost has a training time of 0.8 hours, an inference time of 1.4 milliseconds, uses 1,714,856 parameters, and performs at 5 billion floating-point operations per second.
[0028] The recognition step S3 utilizes the trained object detection model to instantly recognize a die-bonding image after die bonding is completed, and outputs corresponding edge alignment results and corner alignment results. For example, for a square wafer, the recognition results include good / bad right edge alignment, good / bad left edge alignment, good / bad top edge alignment, good / bad bottom edge alignment, good / bad top right corner alignment, good / bad top left corner alignment, good / bad bottom right corner alignment, and good / bad bottom left corner alignment. However, the present invention is not limited to this.
[0029] The prediction step S4 calculates a cumulative value based on the plurality of edge alignment results and the plurality of corner alignment results. Based on this cumulative value, the die bonding result of the die bonding machine is determined to be good or defective, thereby generating a predicted classification result. For example, taking the eight sets of identification results as an example, if four sets of good edge alignment results or four sets of good corner alignment results are accumulated, the corresponding die bonding image is determined to be good. Furthermore, if three sets of good edge alignment results and three sets of good corner alignment results are accumulated simultaneously, the corresponding die bonding image is also determined to be good. For all other accumulated results, the corresponding die bonding image is determined to be defective.
[0030] The adjustment step S5 adjusts a die bonding control parameter of the die bonding machine in real time based on the predicted classification result to reduce die bonding offset. The die bonding control parameter may include dispensing parameters, placement accuracy, placement speed, downforce, heating temperature, and other machine operating conditions, thereby improving die bonding yield and process efficiency.
[0031] Different versions of the object detection model were integrated into the machine's control system to identify and predict die-bond offset. Performance data such as frame per second (FPS), accuracy, recall, and precision were recorded. For example, the YOLOv8n object detection model achieved a frame rate of 370 frames per second, 98.66% accuracy, 95.8% recall, and 99.7% precision. The YOLOv8s object detection model achieved a frame rate of 243.9 frames per second, 99.17% accuracy, 95.9% recall, and 99.8% precision. The YOLOv8n-ghost object detection model achieved a frame rate of 454.55 frames per second, 98.5% accuracy, 95.6% recall, and 99.5% precision.
[0032] Based on the aforementioned object detection model training and prediction results, the YOLOv8n-ghost object detection model is the preferred choice, as it offers significant advantages in training speed while maintaining high accuracy. For the highest accuracy, the YOLOv8s object detection model is recommended.
[0033] In this embodiment, the training data set used in the training step S2 preferably includes post-bonding images captured by the die bonding machine. This allows for training and testing using the actual image data set captured by the die bonding machine, improving the model's adaptability to actual working components and enhancing the accuracy and practicality of predicting die-bonding offset in the identification step S3 and the prediction step S4.
[0034] In this embodiment, the training step S2 can further utilize the actual classification results and the predicted classification results to generate a confusion matrix to calculate the accuracy of the object detection model. Furthermore, these predicted classification results are stored in a separate folder, and the actual classification results and predicted classification results are aggregated to generate the confusion matrix. This allows for rapid evaluation of the object detection model's accuracy and continuous model optimization.
[0035] In this embodiment, the training step 2 can further adjust the hyperparameters of the object detection model based on the performance data to improve its accuracy and performance, including adjusting the number of training epochs and batch size, so that the model can be calibrated according to the actual classification results, thereby improving prediction accuracy and overall model performance.
[0036] In summary, the high-speed die-stick offset prediction method of the present invention is based on real-time detection and prediction technology using a deep learning model. It uses a training dataset to train the object detection model in the prediction module. The RepGhost module implements convolution operations with reduced computational effort, improving inference speed and compensating for feature loss. It also automatically identifies die-stick images and outputs multiple edge and corner offset results. It then calculates cumulative values and generates a predicted classification result, which is used to determine the quality of the die-stick image and adjust the machine's die-stick control parameters accordingly. This high-speed die-stick offset prediction method allows for rapid training, real-time detection and prediction of die-stick offset, and real-time adjustment of the machine's die-stick control parameters. Furthermore, the object detection model can be calibrated based on actual classification results to improve the accuracy of the prediction module.
[0037] Although the present invention has been disclosed using the preferred embodiments described above, they are not intended to limit the present invention. Any person skilled in the art may make various changes and modifications to the above embodiments without departing from the spirit and scope of the present invention. These changes and modifications are still within the technical scope protected by the present invention. Therefore, the scope of protection of the present invention shall include all changes within the meaning and equivalent scope of the appended patent claims.
[0038] S1: Data collection steps S2: Training step S3: Identification step S4: Prediction step S5: Adjustment steps 1: Convolutional unit 11: Convolutional layer 12: Batch Normalization Layer 13: SiLu activation function layer 2: RepGhost bottleneck unit 21: Compression and incentive blocks 22: RepGhost layer P: Split C: Splicing
Claims
1. A high-speed die-bonding shift prediction method, comprising: a data collection step, using multiple die-bonding images and annotating each die-bonding image, wherein the annotations include multiple die-bonding alignment results, as a training dataset; a training step, inputting the dataset into an object detection model for image recognition training of die-bonding results, wherein the object detection model combines the C2f module in the YOLOv8 architecture with the RepGhost module; a recognition step, using the trained object detection model to recognize a die-bonding image of a machine after die-bonding is completed, and outputting corresponding edge alignment results and corner alignment results; a prediction step, calculating a cumulative value based on the edge alignment results and the corner alignment results, and determining whether the die-bonding result is good or bad based on the cumulative value, thereby generating a predicted classification result; and an adjustment step, adjusting a die-bonding control parameter of the die-bonding machine based on the predicted classification result in real time to reduce die-bonding shift.
2. The high-speed prediction method for die-bond shift of claim 1, wherein: The die bonding alignment results marked in the data collection step include four labels: good edge, bad edge, good corner, and bad corner.
3. The high-speed prediction method for die-bond shift of claim 1, wherein: The data set includes a true classification result. A confusion matrix is generated according to the true classification result and the predicted classification result, and the accuracy of the object detection model is calculated.
4. The high-speed prediction method for die-bond shift of claim 1, wherein: In the C2f module of the object detection model, data is input into a convolutional unit, split, and then passed through n RepGhost bottleneck units to generate residual values. This is then concatenated and output through another convolutional unit.
5. The high-speed prediction method for die-bond shift according to claim 4, wherein: Each of the RepGhost bottleneck units has a compression and excitation block and is located between two RepGhost layers, and has two path modes: shortcut true and shortcut false.
6. The high-speed prediction method for die-bond shift according to claim 4, wherein: The convolution unit sequentially includes a convolution layer, a batch normalization layer and a SiLu activation function layer.
7. The high-speed prediction method for die-bond shift according to claim 1, wherein: The object detection model version is YOLOv8n, YOLOv8s, or YOLOv8n-ghost.
8. The high-speed prediction method for die-bond shift according to claim 1, wherein: The alignment results output by the identification step include right edge alignment good or bad, left edge alignment good or bad, upper edge alignment good or bad, lower edge alignment good or bad, upper right corner alignment good or bad, upper left corner alignment good or bad, lower right corner alignment good or bad, and lower left corner alignment good or bad.
9. The high-speed prediction method for die-bond shift according to claim 8, wherein: If the cumulative value calculated in the prediction step is four sets of good edge alignments or four sets of good corner alignments, the corresponding die-bonding image is determined to be a good die-bonding product; if the cumulative value is three sets of good edge alignments and three sets of good corner alignments, the corresponding die-bonding image is also determined to be a good die-bonding product; if the other cumulative results are obtained, the corresponding die-bonding image is determined to be a defective die-bonding product.
10. The high-speed prediction method for die-bond shift according to claim 1, wherein: The die bonding control parameters include dispensing parameters, placement accuracy, placement speed, down force channel, and heating temperature.