Semiconductor Package Inspection with Predictive Model for Wirebond High Frequency Performance

Abstract

Methods, systems, and apparatuses are provided for use in semiconductor package manufacturing of a package having a die and a die substrate. The method for use in semiconductor package manufacturing includes forming one or more wirebond interconnects between the die and the die substrate, capturing input data representative of characteristics of the wirebond interconnects during inspection of the formed wirebond interconnects, and passing the captured input data to a machine learning (ML) engine. The method further includes processing the captured input data by the machine learning engine using a trained model to obtain an output sequence of data, evaluating the output sequence of data to determine a predicted radio frequency (RF) performance rating, and outputting the predicted RF performance rating. Processing the training dataset may include applying the image data and parameters to a multi-layer neural network to obtain a set of candidate ML models and selecting an optimal ML model.

Description

[Technical field]

[0001] This field relates to semiconductor package inspection for use in manufacturing and wire bond assembly. [Background technology]

[0002] Related Technology Wire bonding is a semiconductor package manufacturing process that uses wire bonds to connect an integrated circuit (IC) or other semiconductor device to a substrate. In some applications, wire bonding is used to connect a monolithic microwave integrated circuit (MMIC) on a die to a substrate. The wire bonds essentially act as an antenna or lumped element, affecting the radio frequency (RF) response of the circuitry in the MMIC. The resulting performance impact may render the resulting package unusable.

[0003] To detect negative performance impacts, manufacturers test packages with wire bond interconnects. Testing occurs after wire bond assembly. This testing can include time consuming and expensive testing and probing of the package. Testing at this late stage often requires rework to correct the wire bond interconnects. Additional testing of the reworked package is also required. This rework and retest can add significant cost or delay to the semiconductor package manufacturing process. Summary of the Invention

[0004] In embodiments, a method, system, and apparatus are provided for semiconductor package inspection with RF performance prediction. Machine learning based RF performance prediction is used to evaluate packages having wire bond interconnects.

[0005] In one embodiment, an inspection method for use in semiconductor package manufacturing includes forming one or more wirebond interconnects between a die and a die substrate, capturing input data representative of characteristics of the wirebond interconnects during inspection of the formed wirebond interconnects, and passing the captured input data to a machine learning engine. The method further includes processing the captured input data by the machine learning engine using a trained model to obtain an output sequence of data, evaluating the output sequence of data to determine a predicted RF performance rating, and outputting the predicted RF performance rating.

[0006] In another embodiment, the inspection method also includes rejecting or passing the package production according to the output predicted RF performance rating. In one aspect, rejecting or passing the package production occurs during wire bonding assembly of the semiconductor package production.

[0007] In a further embodiment, the method includes generating an alert based on the predicted RF performance rating.

[0008] In yet another embodiment, the method includes storing a training dataset in a computer readable memory and processing the training dataset with a machine learning engine to obtain a trained model. In one aspect, processing the training dataset includes applying the image data and parameters to a multi-layer neural network with feature extraction and classification to obtain a set of candidate ML models and selecting an optimal ML model from the set of candidate ML models for use as the trained model.

[0009] In another embodiment, the trained model is a trained neural network model, the ML engine has an inference stage, and processing the captured input data includes classifying the input data in the inference stage with the trained neural network model to obtain an output array of data having a predicted value for an RF performance rating corresponding to the package having the formed wire bond interconnects. In another aspect, the RF performance rating identifies an RF performance level from among a plurality of performance levels.

[0010] In a further embodiment, the inspection system includes a monitoring device and an RF performance predictor tool. The monitoring device is configured to capture input data characteristic of a formed wirebond interconnect during inspection of the wirebond interconnect. The RF performance predictor tool is configured to process the captured input data with a machine learning engine having a trained model to obtain an output array of data, evaluate the output array of data to determine a predicted radio frequency (RF) performance rating, and output the predicted RF performance rating.

[0011] In another embodiment, the RF performance prediction tool is further configured to reject or pass manufacturing of the package according to the output predicted RF performance rating. An alert generator is configured to generate an alert based on the predicted RF performance rating.

[0012] In a further embodiment, an inspection system for a semiconductor package having a die, a die substrate, and one or more wire bond interconnects formed between the die and the die substrate includes monitoring equipment, a wire bond assembly supervision system, and an RF performance prediction equipment tool.

[0013] In yet another embodiment, a system for predicting RF performance of a wirebond interconnect formed between a die and a die substrate in a package during wirebond assembly includes a computer readable memory configured to store a trained model; and at least one processor configured to: process captured input data with a machine learning engine using the trained model to obtain an output array of data; evaluate the output array of data to determine a predicted RF performance rating of the formed wirebond interconnect; and output the predicted RF performance rating.

[0014] Another embodiment relates to a device having a computer-readable storage medium having instructions stored thereon, the instructions configured to cause at least one processor to perform operations to predict RF performance of wirebond interconnects formed between a die and a die substrate in a package.

[0015] Further embodiments, features, and advantages of the present invention, as well as the structure and operation of the various embodiments of the present invention, are described in detail below with reference to the accompanying drawings.

[0016] DETAILED DESCRIPTION OF THE EMBODIMENTS The embodiments are described with reference to the accompanying drawings, in which like reference numbers may indicate identical or functionally similar elements, and the drawing in which an element first appears is typically indicated by the leftmost digit(s) in the corresponding reference number. [Brief description of the drawings]

[0017] [Figure 1] FIG. 1 is a diagram of an inspection system for use in semiconductor package manufacturing with predicted RF performance for wirebond interconnects formed in a wirebond assembly, according to an embodiment. [Diagram 2]2A and 2B are diagrams illustrating one view of the package and monitoring equipment and parameters for wirebond interconnects according to an embodiment. FIG. 2C is a diagram illustrating a wirebond assembly having wirebond interconnects forming inputs and outputs to a die according to an embodiment. FIG. 2D is a diagram illustrating a wirebond assembly having wirebond interconnects forming inputs and outputs to a die according to another embodiment. [Diagram 3] FIG. 3 is a flowchart of an illustration of a method for semiconductor package inspection with predicted RF performance for wirebond interconnects, according to an embodiment. [Figure 4] FIG. 4 is a flow chart diagram of a control action process according to an embodiment. [Diagram 5] FIG. 5 is a flowchart diagram of a training process for training an ML engine, according to an embodiment. [Figure 6] FIG. 6 is a flow chart diagram illustrating the training data set processing steps of FIG. 5 in more detail, according to an embodiment. [Figure 7] FIG. 7 is a flow chart diagram illustrating in more detail the input data processing steps of FIG. 3 with the trained model, according to an embodiment. [Figure 8] FIG. 8 is a block diagram of an ML engine according to an embodiment. [Figure 9] FIG. 9 is a diagram of a convolutional neural network (CNN) according to an embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0018] This disclosure describes predicted RF performance for semiconductor package manufacturing. Methods, systems, and apparatus for inspection with RF performance prediction are described. In embodiments, machine learning based RF performance prediction is used to inspect packages with wire bond interconnects and generate an RF performance rating.

[0019] In the embodiments, reference is made to the description set forth herein with reference to specific applications. It is to be understood that the invention is not limited to the embodiments. Those skilled in the art, having access to the teachings provided herein, will recognize further modifications, applications, and embodiments that are within the scope thereof, as well as further fields in which the embodiments would have significant utility.

[0020] In the detailed description of the embodiments herein, references to "one embodiment," "embodiment," "exemplary embodiment," and the like indicate that the described embodiment may include a particular feature, structure, or characteristic, but not all embodiments may necessarily include that particular feature, structure, or characteristic. Moreover, such phrases do not necessarily refer to the same embodiment. Furthermore, when a particular feature, structure, or characteristic is described in connection with an embodiment, it is submitted that it is within the knowledge of one of ordinary skill in the art to implement such feature, structure, or characteristic in connection with other embodiments, whether or not explicitly described.

[0021] The terms "wirebond" and "wirebond interconnect" are used interchangeably herein.

[0022] System with predicted RF performance 1 is a diagram of a semiconductor package inspection system 100 with predicted RF performance for wirebond interconnects formed in a wirebond assembly, according to an embodiment. System 100 includes monitoring equipment 110 and a wirebond assembly supervision system (WASS) 120. A semiconductor package 105 is supported by a stage 107. In an example, semiconductor package 105 may have a die, a die substrate, and one or more wirebond interconnects formed between the die and the die substrate.

[0023] WASS 120 is also coupled to a machine learning (ML) engine 130, a trained ML model 132, a training dataset 135, and a database 140. The ML engine 130 and its operation with respect to the trained ML model 132 and the training dataset 135 are described in more detail below with respect to Figures 8 and 9. The database 140 may be one or more databases, for example, a relational database used by a database management system.

[0024] A monitoring device 110 is positioned relative to the package 105 to capture input data during inspection of the wirebond interconnects. The inspection can be performed during or after wirebond assembly as part of quality control or other manufacturing or assembly processes. The input data can be image data or other sensor data representative of the characteristics of the wirebond interconnects. In one embodiment, the monitoring device 110 includes one or more optical or infrared camera devices and a sensor system. The camera devices capture digital images and are positioned relative to the package 105 such that the package area having the formed wirebond connections is within the field of view of the camera devices. The sensor system is positioned relative to the package and configured to capture distance data. The distance data can include a set of data points representing distance values ​​between the package area having the formed wirebond connections and the sensor system. The data captured by the monitoring device 110 is output to the WASS 120.

[0025] The monitoring equipment 110 may be coupled to the WASS 120 to pass data and control signals over communications links. Wired or wireless communications links may be used. These links may also include communications links capable of transmitting data or control signals over one or more data networks, such as a local area network, a medium area network, or a large area network (e.g., the Internet).

[0026] WASS 120 includes a controller 122, a data manager 124, an RF performance predictor tool 126, and an alert generator 128. The controller 122 controls the initiation and operation between the components of the system 120 (i.e., the data manager 124, the RF performance predictor tool 126, and the alert generator 128). The controller 122 may also control the authorization of administrators or other users, and the sending of communication messages (data and control) to and from authorized users. The data manager 124 manages the storage and retrieval of captured input data received from the monitoring devices 110.

[0027] During operation, the RF performance prediction device tool 126 processes the captured input data by the ML engine 130 using the trained ML model 132 to obtain an output array of data. The RF performance prediction device tool 126 evaluates the output array of data to determine a predicted RF performance rating. The RF performance prediction device tool 126 may then output the predicted RF performance rating for storage, transmission, or display. Upon obtaining the RF performance rating, the RF performance prediction device tool 126 may reject or pass the production of the package 105 according to the output predicted RF performance rating.

[0028] The warning generator 128 generates warnings based on the predicted RF performance ratings output by the RF performance prediction tool 126. In one aspect, the warning generator 128 generates warnings based on the predicted RF performance ratings during wire bond assembly of semiconductor package manufacture. In this manner, performance failures are predicted during wire bond assembly prior to testing, and costly rework and testing is avoided.

[0029] In one embodiment, WASS 120, including its components (controller 122, data manager 124, RF performance predictor tool 126, and alert generator 128), may be implemented in software, firmware, or hardware, or any combination thereof, on one or more computing devices at the same or different locations. A computing device, as used herein, may be any type of computing device, such as, without limitation, a smartphone, laptop, desktop, tablet, workstation, kiosk, embedded device, or other computing device having at least one processor and computable readable memory. For example, an embedded device may have an embedded processor and memory, where the processor is configured to execute software or firmware stored in the memory. In further embodiments, WASS 120, including some or all of its components, may be implemented using one or more application programming interfaces (APIs) as necessary to access different services to perform operations on one or more remote servers, as described herein.

[0030] The ML engine 130 may be implemented in software, firmware, or hardware, or any combination thereof, on one or more computing devices at the same or different locations. The ML engine 130 may also be implemented using one or more application programming interfaces (APIs). For example, the RF performance predictor tool 126 may pass requests to the API of a remote cloud computing service that implements the ML engine 130, as described herein.

[0031] The trained ML model 132 and the training dataset 135 may be stored in computer readable memory in the same or different locations. WASS 120 is coupled to the computer readable memory to access and retrieve data in the trained ML model 132 and the training dataset 135.

[0032] The operation of system 100 is described in more detail below with respect to a method for manufacturing a semiconductor package using predicted RF performance as shown in Figures 3-7. For simplicity, the method is described with reference to the exemplary monitoring equipment 110 and wire bond connections on package 105 shown in Figures 2A-2B. Routines illustrating the use of training data set 135 to train ML engine 130 to obtain trained ML model 132 are described in more detail with respect to Figures 5-6.

[0033] Package surface monitoring and parameters Figures 2A and 2B are diagrams illustrating two views of package 105 and a monitoring instrument 110 for wirebond interconnects and parameters, according to an embodiment. Figure 2A shows a top view of package 105 as seen by monitoring instrument 110. Package 105 has an MMIC die 206 mounted on a die substrate 207. Figure 2B shows a cross-sectional view of Figure 2A.

[0034] Eight wirebond interconnects (WB1-WB8) are formed between the die 206 and the substrate 207. Eight traces (Tr1-Tr8) exit from the wirebond interconnects WB1-WB8. Die pads (DP1-DP8) may be used to support connecting one end of each wirebond interconnect (WB1-WB8) to the die 206. Packaging pads (PP1-PP8) may be used to support connecting the other end of each wirebond interconnect (WB1-WB8) to the substrate 207. The traces (Tr1-Tr8) are coupled to respective packaging pads (PP1-PP8). Each trace (Tr1-Tr8) has a respective length and width. The eight wirebond interconnects (WB1-WB8), die pads (DP1-DP8), and traces (Tr1-Tr8) are shown for illustrative purposes and are not intended to be limiting. Fewer or more wirebond interconnects, die pads, or traces may be used in different configurations. For example, the wirebond interconnects (e.g., WB1-WB8) may be located on different sides of a die, substrate, printed circuit board, or package surface. The wirebond interconnects (e.g., WB1-WB8) may each be spaced or angled at the same spacing or angle, or at different spacings or angles, relative to one another or to another surface.

[0035] 2A, the monitoring device 110 may include an infrared camera device 212, an optical camera device 214, and / or a light detection and ranging (LIDAR) sensor system 216. The infrared camera device 212 and the optical camera device 214 capture infrared and optical digital images, respectively, and are positioned relative to the package 105 such that package areas on the top surface are within the field of view of the monitoring device 110. These package areas on the top surface may include a LIDAR sensor system 216 positioned relative to the package 105 to capture distance data. The distance data may include a set of data points representing distance values ​​from the LIDAR sensor system 216 to points on different areas of the top surface of the package 105. The package areas on the top surface of the package 105 monitored by the infrared camera device 212, the optical camera device 214, and the LIDAR sensor system 216 may include any surfaces on the package 105 and its components formed during wire bond assembly. These surfaces include surfaces on the MMIC 206, the wire bond interconnects (WB1-WB8), the substrate 207, the die pads (DP1-DP-8), the packaging pads (PP1-PP8), and the traces (TR1-TR8).

[0036] The camera devices 212, 214 may be positioned directly above the surface of the package 105 at a 90 degree angle, or at an angle to the normal. The camera devices 212, 214 may be fixed at the same angle or at different angles. The camera devices 212, 214 or the stage 107 may move relative to one another to scan the surface of the package 105.

[0037] As shown in Figures 2A and 2B, several parameters 220 may be obtained from the monitoring. Wirebond interconnects (e.g., WB1-WB8) may be formed on the surface of the die 206 as shown. Adjacent wirebond interconnects (e.g., WB1-WB8) may be separated by spacing. The spacing between adjacent wirebond interconnects (e.g., WB1-WB8) may be the same or different from one another depending on the particular layout and the shape and length of the wirebonds. For example, a first spacing S1 between a first wirebond interconnect WB1 and a second wirebond interconnect WB2 is shown in Figure 2A. Additionally, a second spacing S2 between a second wirebond interconnect WB2 and a third wirebond interconnect WB3 is shown in Figure 2A. Adjacent traces (e.g., Tr1-Tr8) may also be separated by spacing. The spacing between adjacent traces (e.g., Tr1-Tr8) may be the same or different from one another depending on the particular layout. For example, a third spacing S3 between the fifth trace Tr5 and the sixth trace Tr6 is shown in FIG. 2A. Additionally, a fourth spacing S4 between the sixth trace Tr6 and the seventh trace Tr7 is shown in FIG. 2A.

[0038] Each wirebond interconnect (e.g., WB1-WB8) may have one or more parameters 220 associated with it, including, for example and without limitation, a wirebond planar length (e.g., represented by L1 in FIG. 2A), a curved length (e.g., represented by L2 in FIG. 2B), or a loop height (e.g., illustrated by H1 in FIG. 2B). The wirebond planar length L1 is the length of the wirebond along the plane between the ends of the wirebond. The curved length L2 is the length from one end of the wirebond itself to the other end. The loop height H1 is the height of the wirebond measured from the top end of the wirebond to the top surface of the substrate 207 or die 206. The wirebond angle (e.g., illustrated by θ1 and θ2 in FIG. 2B) may also be a monitored parameter 220. These parameters 220 are examples and are not intended to be limiting. Other parameters 220 and combinations of parameters 220 may be used, as would be apparent to one of ordinary skill in the art in view of this description.

[0039] 2A and 2B show in more detail exemplary parameters 220 that may be monitored and captured for a pair of wirebond interconnects (WB2, WB6) on a surface of the package 105. FIG. 2A shows the wirebond planar length of WB2 extending between the two ends of WB2. A first spacing S1 between WB2 and nearest neighbor 1 (WB1) and a second spacing S2 between WB2 and nearest neighbor 2 (WB3) are shown. Also shown are respective traces Tr2 having a length and a width. The substrate 207 can also have a spacing between Tr6 and nearest neighbor 1 (Tr5) (e.g., a third spacing S3) and a spacing between Tr5 and nearest neighbor 2 (Tr7) (e.g., a fourth spacing S4).

[0040] FIG. 2B shows two exemplary wire bond angles θ1 and θ2 for wire bond (WB6). Wire bond angle θ1 is the angle of WB6 at one end on a packaging pad (e.g., PP6) near trace (Tr6). A loop height H1 is shown from the top end of wire bond WB2 to the top surface of substrate 207. Wire bond angle θ2 is the angle of WB6 at the other end on a die pad (e.g., DP6) on die 206. A thickness (e.g., represented by T1 in FIG. 2B) for trace Tr2 is also shown. Exemplary parameters 220 are shown in detail for WB2, WB6, but are illustrative and not intended to be limiting. Similar parameters 220 may be monitored for each of WB1-WB8.

[0041] The examples of Figures 2A and 2B are exemplary, and other package and wirebond configurations can be monitored with the RF predicted performance described herein. For example, Figure 2C illustrates a wirebond assembly structure 230 having one or more first exemplary wirebond interconnects 233 forming an input to (or output from) an MMIC die 206 on a die pedestal 232, according to an embodiment. Each of the one or more first exemplary wirebond interconnects 233 is coupled at one end to an exemplary die pad 231 (which may be an input or output in the example). The other end of the one or more first exemplary wirebond interconnects 233 is coupled to a PCB pad 241 on a printed circuit board (PCB) 242. Additionally, one or more second exemplary wirebond interconnects 240 may also be coupled to a PCB trace 245 or to other elements on the PCB 242.

[0042] 2D illustrates a wirebond assembly 250 having one or more first exemplary wirebond interconnects 233 forming outputs from (or inputs to) an MMIC die 206 on a die pedestal 232, according to an embodiment. Each of the one or more first exemplary wirebond interconnects 233 is coupled at one end to a die pad 231 (which may be an input or output in the examples). The other end of the one or more first exemplary wirebond interconnects 233 is coupled to a PCB pad and trace 251 on a PCB 242.

[0043] Behavior with respect to predicted RF performance 3 is a flow chart diagram of a semiconductor package manufacturing method 300 with predicted RF performance for wirebond interconnects, according to an embodiment (steps 310-360). For simplicity, the method 300 is described with reference to the exemplary wirebond interconnects (e.g., WB1-WB8) of FIGS. 2A and 2B. However, this is not intended to be limiting, as the semiconductor package manufacturing method 300 with predicted RF performance can be used for other wirebond interconnects (e.g., without limitation, the wirebond assemblies 230, 250 shown in FIGS. 2C and 2D).

[0044] In step 310, one or more wirebond interconnects are formed between the die and the die substrate during wirebond assembly. In the example of FIG. 2A, multiple wirebond interconnects (e.g., WB1-WB8) are formed. Although eight wirebond interconnects (WB1-WB8) are shown in FIG. 2A, the structure of the wirebond assembly is not so limited. For example, embodiments are contemplated in which fewer than eight wirebond interconnects or more than eight wirebond interconnects are formed. Any wirebond formation technique may be used. For example, ball bonding or wedge bonding techniques may be used to form the wirebonds. Heat, pressure, and / or sonic energy may be used as thermosonic, thermocompression, and ultrasonic bonding techniques.

[0045] At step 320, input data representative of the characteristics of the wirebond interconnects is captured during inspection of the formed wirebond interconnects. For example, stage 107 may position package 105 relative to monitoring equipment 110 for inspection. The monitoring equipment 110 then captures input data representative of the characteristics of the wirebond interconnects during inspection of the formed wirebond interconnects. For example, monitoring equipment 110 may capture image data and sensor data and output data to WAAS 120. Data manager 124 may receive and store the input data in a record in database 140 along with inspection event information (e.g., a package identifier associated with package 105 and a timestamp).

[0046] At step 330, the retrieved input data is then sent to the ML engine 130. In an embodiment, the RF performance predictor tool 126 retrieves the input data from the database 140 and outputs the retrieved input data to the ML engine 130. For example, the RF performance predictor tool 126 sends an API request to the ML engine 130 requesting classification of the input data according to the trained ML model 132. The API request may be one or more messages passing the input data (e.g., image data and sensor data) and identifying the trained ML model 132 to be used to classify the input data.

[0047] In step 340, the ML engine 130 processes the captured input data using the trained model 132 to obtain an output array of data. The ML engine 130 outputs the array of data to the RF performance prediction device tool 126. For example, the output array may be sent to fulfill an initiation API request sent by the RF performance prediction device tool 126. The processing of step 340 is further described below with respect to FIG.

[0048] At step 350, the output array of data is evaluated to determine a predicted radio frequency (RF) performance rating. For example, the RF performance predictor tool 126 may evaluate one or more values ​​in the output array of data to determine an RF performance rating. The RF performance rating is a prediction of the RF performance of the package 105 having wire bond interconnects. In one example, the RF performance rating identifies an RF performance level from among a number of performance levels. These different performance levels may correspond to pass / fail levels, pass / warning / fail levels, high / medium / low, or other ranges of performance levels. The RF performance rating may be a numeric value (e.g., 0-100), a label (pass / fail), or other identifier associated with the predicted RF performance.

[0049] In step 360, the RF performance prediction device tool 126 outputs the predicted RF performance rating. For example, the rating may be output for storage or display, or for transmission to an administrator.

[0050] Further control actions may be taken in response to the predicted RF performance rating, as shown in Figure 4. Figure 4 is a flow chart diagram of a control action process 400 according to an embodiment (steps 410-420). In step 410, the RF performance prediction tool 126 rejects or passes the manufacturing of the package 105 according to the output predicted RF performance rating. Depending on the performance level used, the RF performance prediction tool 126 may instruct the package 105 for further testing or rework for a warning rating.

[0051] At step 420, one or more alerts may be generated based on the predicted RF performance rating. Alert generator 128 generates an alert based on the predicted RF performance rating. The alert may be, for example, a visual, audio, or tactile or other message that can inform an administrator or other authorized user of an alert condition.

[0052] Machine learning, including training and inference, is described in more detail below. Training, performed in the training phase of ML engine 130, is described with reference to routine 500 shown in Figures 5-6. Inference, performed in the inference phase of ML engine 130, is described in more detail with reference to routine 500 in Figure 7. An embodiment of ML engine 130 having training and inference phases is described with reference to Figures 8-9. For simplicity, these routines 500 and steps 340 are described with reference to ML engine 130, but are not necessarily limited thereto.

[0053] Machine Learning FIG. 8 is a block diagram of an ML engine 130 according to an embodiment. The ML engine 130 includes an inference stage 810 and a training stage 820. The training stage 820 further includes a multi-layer neural network (NN) 830. As shown in FIG. 9, in some embodiments, the NN 830 can be a convolutional neural network (CNN) having a feature extraction layer 910 and a classification layer 920. The feature extraction layer 910 may include an input layer, a convolution layer, and a pooling layer. The input layer sets the size of the input image data and can resize it as needed. The input image data is convolved by the convolution layer with multiple learning kernels using shared weights 809. The pooling layer attempts to reduce the image size while retaining the information required for objection detection and feature extraction and form a feature map. The feature map is then output to the classification layer 920. The classification layer 920 combines the extracted features in a fully connected layer with multiple nodes. The output layer has an output neuron for each object category. The output layer is connected to receive and classify the output from the classification layer to obtain the classification result (label 840) and the probability that the result is correct. See, for example, the description of feature extraction and classification of image data described by Van Hiep Phung and Eun Joo Rhee, “A High-Accuracy Model Average Ensemble of Convolutional Neural Networks for Classification of Cloud Image Patches on Small Datasets,” J. of Appl. Sci. 2019, 9, 4500, 16 pages.

[0054] training During training, the training stage 820 applies the training dataset 135 to the NN 830 to obtain a trained ML model 132. The training in the training stage 820 may be supervised learning, unsupervised learning, or reinforcement learning. The training stage 820 processes the data in the training dataset 135 by the NN 830, including applying parameters or features 807 and weights 809, to obtain a set of candidate models. The training stage 820 further selects a trained ML model 132 from the set of candidate models. For example, the trained ML model 132 may be selected to minimize a loss function or meet other design criteria. Hyperparameters may also be tuned to further aid in the selection of the trained ML model 132.

[0055] In an embodiment, training in the training stage 820 involves modifying weights associated with nodes in a layer over many iterations until an expected output is obtained for the particular training input data. One or more learning algorithms may be used to train the layers of the NN 830. For example, if the NN 830 is a deep multi-layer neural network, gradient descent and backpropagation algorithms may be used in parallel. Supervised or unsupervised learning may be used to modify weights to minimize a loss function. Reinforcement learning may be used to modify weights to maximize a reward function. In a further example, activation functions (e.g., sigmoid functions) may also be used, especially after layers with weights. Data fitting or regularization techniques to achieve a balanced CNN and avoid undesirable overfitting or underfitting may also be used. Further optimizations may be employed to improve the training, such as expanding the training dataset by augmentation, increasing the training time or depth (or width) of the model, adding regularization, or increasing hyperparameter tuning, as will be apparent to one of ordinary skill in the art in light of this description.

[0056] FIG. 5 is a flow chart diagram of a training process 500 for training the ML engine 130 according to an embodiment (steps 510-530). First, in step 530, the training stage 820 preprocesses raw input data from test images or other predefined images, sensor data or other synthetic data to obtain a training data set 135. For example, the test images may be cropped or oriented to a common size and orientation. Test images with excessive blur or highlights may be removed. Images and sensor data acquired by monitoring equipment 110 on a sample package with known performance ratings may be used. Alternatively, the training stage 820 may simply receive the training data set 135, omitting the need for preprocessing.

[0057] In step 520, the training stage 820 stores the training dataset 135 in a computer-readable memory coupled to the ML engine 130. For example, the training dataset 135 may be stored in a memory at the same location as the ML engine 130 or at a different location. The memory that stores the training dataset 135 may be local to the ML engine 130 on the same computing device or may be stored at a remote network location accessible over a network. This may include storing the training dataset 135 on a cloud-based drive or storage service accessible by the ML engine 130.

[0058] In step 530, the training stage 820 processes the training dataset 135 to obtain a trained ML model 132. As shown in more detail in FIG. 6, step 530 first involves applying the image data and parameters to a multi-layer neural network with feature extraction and classification to obtain a set of candidate ML models (step 610). For example, the training stage 820 applies the image data and parameters or features 807 from the training dataset 135 to a multi-layer neural network with feature extraction and classification 830 to obtain a set of candidate ML models. In step 620, the best ML model is then selected from the set of candidate ML models to use as the trained model 132. For example, the training stage 820 can select the best ML model from the set of candidate ML models that minimizes a loss function in supervised or unsupervised learning. The training stage 820 can also select the best ML model from the set of candidate ML models that maximizes a reward function in reinforcement learning, depending on the particular application.

[0059] One or more parameters 807 (also referred to as features) related to wirebond assembly or RF performance are used by the training stage 820. The parameters 807 may be used by the training stage 820 in obtaining a set of candidate models, evaluating a loss or reward function, and selecting a trained model. The parameters 807 are shown separately in FIG. 8 for clarity, but may be included as part of the training dataset 135. Labels 840 may also be included in the training dataset 135 to facilitate training to obtain an optimal trained ML model 132. The training stage 820 may also use hyperparameters to further tune the trained model 132.

[0060] In an embodiment, parameters 807 may be a default set of desired parameters or may be manually set by an administrator or operator of ML engine 130. Parameters 807 may also include more parameters that are automatically identified during training. Hyperparameters may be a default set or may be manually set by an administrator or operator of ML engine 130.

[0061] As will be apparent to one of ordinary skill in the art in view of this description, other training techniques can be used to obtain a trained ML model, for example by augmenting the image data or other data used in the training dataset.

[0062] In addition to image data, the training dataset 135 may also include sensor data (e.g., distance data points or range data acquired in a LIDAR). The training stage 820 may then be trained using the sensor data from the training dataset 135 along with the image data. The sensor data is applied to train the NN 830 as described herein with respect to the image data.

[0063] The training data set 135 may also include test equipment data collected during the design, simulation, and manufacturing of the product or similar products. The test equipment data may include product performance and RF performance characteristics (e.g., voltage, current, scattering parameters, impedance, and spectral data).

[0064] Training data set 135 can also include synthetic data related to RF performance (e.g., data obtained from EM modeling software). Examples of EM modeling software tools that can provide RF performance parameters are the SONNET SUITES tools available from Sonnet Software, Inc. and the ANSYS HFSS design tools available from Ansys, Inc.

[0065] The training stage 820 may also be trained using synthetic data in the training dataset 135. The synthetic data is applied to further train the NN 830 as described herein with respect to image data and sensor data.

[0066] The above description of the ML engine 130 with respect to the CNN 830 is exemplary and not intended to be limiting. In further embodiments, the ML engine 130 may use other multi-layer deep learning neural networks and architectures.

[0067] Example Wirebond and RF Performance Prediction Parameters and Labels As previously mentioned, the training data set 135 may include a number of parameters and labels to facilitate training. Such parameters may include wire bond parameters and RF performance parameters. The labels may include known image data or sensor data, or RF rating labels generated from testing of previous packages with wire bond assemblies associated with known parameters or characteristics.

[0068] In one embodiment, the wire bond parameters include one or more of the following parameters for each wire bond: wire bond planar length (e.g., L1), wire bond curvilinear (3D) length (e.g., L2), wire bond loop height (e.g., H1), wire bond diameter, or angle of the wire bond from a bird's eye view relative to a plane.

[0069] In another embodiment, the package has traces coupled to wire bonds in the corresponding sections, and the wire bond parameters include one or more of the following parameters: wire bond planar length (e.g., L1), wire bond curvilinear (3D) length (e.g., L2), wire bond loop height (e.g., H1), wire bond diameter, number of wire bonds, rough spacing between adjacent wire bonds, angle of wire bonds from a bird's eye view relative to a plane, trace width of the corresponding sections, or trace length of the corresponding sections.

[0070] MMIC production line In an embodiment, as wirebond interconnects and other structures and physical features (also referred to herein as "elements") are created during fabrication of the MMIC on a die, system 100 determines an RF rating for each such element and / or combination of elements. A passing rating may include multiple performance levels (e.g., high, medium, low, or a selection of one of multiple RF performance ranges). A failing rating may include instructions to redo the element(s).

[0071] training The training data set 135 in these examples for a MMIC manufacturing line may include raw training data drawn from one or more sources. The raw training data used may be parameter values ​​for size and positioning of elements on the MMIC, RF performance of each element of a combination of elements, and RF rating labels for each vector or other output of machine learning in the ML engine 130. Sources for training data include: one or more test MMICs, one or more optical, IR, and / or LIDAR devices that capture parameter values ​​for each element or combination of elements, a tester that measures the RF performance of each element or combination of elements on the test MMICs. A user or other administrator may provide RF labels that correspond to the RF performance data. Additional sources of training data may be used, such as historical data (e.g., records of previously manufactured MMICs), or composite values ​​of parameters generated by EM modeling software. Any combination of the above types and sources of training data may be used.

[0072] Also in training, model 132 can be designed to focus on one or more structures individually (e.g., the RF response of each wire bond on an MMIC die). Alternatively, model 132 can be designed to determine the RF response of a set of structures that include conductive paths (e.g., the RF response of a conductive path that includes the combination of trace 2 (Tr2), packaging pad 2 (PP2), wire bond 2 (WB2), and die pad 2 (DP2)).

[0073] Live Action In live operation, the inference stage 810 may receive input data 805 and access the trained ML model 132. The inference stage 810 processes the input data 805 with the trained ML model 132 to obtain an output array of data having predicted values ​​for RF performance ratings corresponding to packages 105 having formed wirebond interconnects. The output array of data may include labels 840 associated with the predicted values ​​for the RF performance ratings.

[0074] 7 is a flow chart diagram illustrating in more detail the input data processing step 340 by the trained model 132 according to an embodiment (steps 710-720).

[0075] In step 710, an inference stage 810 applies the input data 805 and parameters 807 to the trained ML model 132. In step 720, the trained ML model 132 extracts features and classifies the input data and parameters to obtain an output array of data having predicted values ​​for an RF performance rating corresponding to a package having wire bond interconnects formed therein.

[0076] For example, the input data 805 may be live image data of the package 105 captured by the monitoring device 110 as described above. The parameters 807 may be live sensor data captured by the monitoring device 110. The parameters 807 may also be default or known parameters associated with the live input data that are used by the inference stage 810 to classify the input data 805.

[0077] The inference stage 810 may further generate a label 840 associated with a predicted value for an RF performance rating corresponding to the package having the formed wire bond interconnects. For example, if the RF performance rating identifies an RF performance level from among multiple performance levels (e.g., pass / fail, pass / warning / fail, or high / medium / low), the label 840 may identify a particular performance level that is predicted.

[0078] Instructions may be output, for example a fail rating or label 840 indicating failure may be provided along with instructions to redo an element or elements of the package 105, such as any of the wirebond interconnects (WB1-WB8).

[0079] In one technical advantage, performing inference using a model trained using the aforementioned parameters increases the likelihood that manufactured parts will meet their performance criteria, improving manufacturing yields. For example, the trained model 135 can be deployed on an MMIC manufacturing line. As wire bonds and / or other elements are formed on the MMIC, an inference stage 810 of the ML engine 130 receives input data 805 to rank the predicted RF performance of individual elements or combinations of elements. For example, a vector of data may be pulled from the input data 805 for classification by the model 132 to obtain a label 840 indicative of predicted RF performance. In this manner, the WASS 120 can identify structures on an MMIC die (or an assembly including one or more MMIC dies) that may have defective RF responses.

[0080] Further predicted RF performance applications and configurations In further examples, a wire bond is a conductive piece of metal that connects two points and is defined by the material (conductive, often gold), thickness (mils), 3D curve length (mils), planar length (mils, bird's-eye view), number of wires (number), spacing and angle of the wires (relative shape of the group). Pads (e.g., I / O pads) are conductive metal pads, often gold plated, that form the inputs or outputs of a custom integrated circuit. For RF circuits, pads can have a GSG (ground-signal-ground) or GSSG (ground-signal-signal-ground) configuration, with single or multiple wires per pad.

[0081] The conductive traces may be transmission lines on a printed wiring assembly consisting of the electrical properties of the dielectric (height, dissipation factor, dielectric constant, isotropic or anisotropic), trace height, plating, trace length, and trace width. These electrical properties may be used as one or more parameters in parameters 807. The location of the conductive traces on the printed wire assembly (PWA) can help infer the parameter values ​​relative to a specified nominal value (given manufacturing tolerances).

[0082] In one simulation performed by the inventors, RF performance was simulated using the simulation design tool ANSYS HFSS available from Ansys, Inc. The ANSYS HFSS model produced optimal results for a set of parameters. This set of parameters from the simulation can be used as synthetic data in the training data set 135, as previously described.

[0083] Further embodiments and examples Various embodiments (including WASS 120 and its components 122-128) may be implemented on one or more computing devices. The computing devices may be in the same location or in different locations. The computing devices may be any type of device having one or more processors and memory. For example, a computing device may be a workstation, a mobile device (e.g., a mobile phone, a personal digital assistant, a tablet, or a laptop), a computer, a server, a computer cluster, a server farm, a game console, a set-top box, a kiosk, an embedded device or system, or other device having at least one processor and computer-readable memory. In addition to at least one processor and memory, such computing devices may include software, firmware, hardware, or combinations thereof. The software may include one or more applications and an operating system. The hardware may include, without limitation, a processor, memory, and a user interface display or other input / output devices.

[0084] Aspects of the computing embodiments may also include client-side and server-side (including remote users on remote computing devices coupled to WASS120) and may be implemented electronically using hardware, software modules, firmware, tangible computer-readable or computer-usable storage media having instructions stored thereon, or combinations thereof, and may be embodied in one or more computer systems or other processing systems.

[0085] The embodiments are also directed to computer program products including software stored on any computer usable medium. Such software, when executed in one or more data processing devices (e.g., processors), operates the data processing device(s) as described herein or enables the synthesis and / or manufacture of electronic devices (e.g., ASICs, or processors) that execute the embodiments described herein as described above. The embodiments use any computer usable or readable medium, and any computer usable or readable storage medium known now or in the future. Examples of computer usable or computer readable media include, without limitation, primary storage (e.g., any type of random access memory), secondary storage (e.g., hard drives, floppy disks, CD-ROMs, ZIP disks, tapes, magnetic storage, optical storage, MEMS, nanotechnology storage, etc.), and communication media (e.g., wired and wireless communication networks, local area networks, wide area networks, intranets, etc.). Computer usable or computer readable media can include any form of transitory media (including signals) or non-transitory media (excluding signals). Non-transitory media include, by way of non-limiting example, the physical storage devices mentioned above (eg, primary storage and secondary storage).

[0086] Further embodiments The present disclosure is also directed to the following exemplary embodiments.

[0087]

[0021] Embodiment 1. An inspection system for use in semiconductor package manufacturing of a semiconductor package having a die, a die substrate, and one or more wirebond interconnects formed between the die and the die substrate, the inspection system comprising: a monitoring device positioned relative to a stage supporting the semiconductor package and configured to capture input data characteristic of the wirebond interconnects during inspection of the formed wirebond interconnects; a wirebond assembly oversight system comprising: a controller; a data manager configured to manage data including the captured input data, a training data set, and a trained machine learning model; an RF performance predictor tool configured to: process the captured input data with a machine learning engine using the trained machine learning model to obtain an output sequence of data; evaluate the output sequence of data to determine a predicted radio frequency (RF) performance rating; and output the predicted RF performance rating; and an alert generator configured to generate an alert based on the predicted RF performance rating.

[0088]

[0023] Embodiment 2. The system of embodiment 1, wherein the semiconductor package includes a substrate having a monolithic microwave integrated circuit (MMIC) coupled to the one or more wirebond connections.

[0089]

[0023] Embodiment 3. The system of embodiment 1 or 2, wherein the semiconductor package further includes a die input pad or a die output pad coupled to the one or more wire bond connections.

[0090]

[0023] Embodiment 4. A tool for predicting RF performance of a wirebond interconnect formed between a die and a die substrate in a package, the tool including: a computer readable memory configured to store a trained model; and at least one processor configured to: process captured input data with a machine learning engine using the trained model to obtain an output array of data; evaluate the output array of data to determine a predicted radio frequency (RF) performance rating of the formed wirebond interconnect; and output the predicted RF performance rating.

[0091]

[0014] Embodiment 5. The tool of embodiment 4, wherein the machine learning engine comprises a multi-layer neural network.

[0092]

[0021] Embodiment 6. The tool of embodiment 5, wherein the multi-layer neural network comprises a convolutional neural network (CNN).

[0093] Embodiment 7. A system for predicting RF performance of a wirebond interconnect formed between a die and a die substrate in a package, the system including: means for processing captured input data by a machine learning engine using a trained model to obtain an output array of data, means for evaluating the output array of data to determine a predicted radio frequency (RF) performance rating, and means for outputting the predicted RF performance rating.

[0094]

[0023] Embodiment 8. An apparatus having a computer-readable storage medium having instructions stored thereon, the instructions configured to cause at least one processor to perform operations for predicting radio frequency (RF) performance of a wirebond interconnect formed between a die and a die substrate in a package, the operations including receiving input data representative of a characteristic of the wirebond interconnect captured during inspection of the formed wirebond interconnect, processing the captured input data by a machine learning engine using a trained model to obtain an output array of data, evaluating the output array of data to determine a predicted RF performance rating, and outputting the predicted RF performance rating.

[0095] The embodiments of the present invention have been described above using functional building blocks illustrating implementation of certain functions and relationships thereof. The boundaries of these functional building blocks have been arbitrarily defined herein for convenience of description. Alternative boundaries can be defined so long as the specified functions and relationships thereof are appropriately performed. The breadth and scope of the present invention should not be limited by any of the exemplary embodiments described above.

[0096] The foregoing description of the specific embodiments fully discloses the general nature of the present invention, so that others, by applying knowledge within the skill of the art, can readily modify and / or adapt such specific embodiments for various applications without departing from the general concept of the present invention without undue experimentation. Such adaptations and modifications are therefore intended to be within the meaning and range of equivalents of the disclosed embodiments, based on the teaching and guidance presented herein. It should be understood that the phrases or terms used herein are for the purpose of description, not of limitation, as the terms or terms used herein would be interpreted by those skilled in the art in light of the teaching and guidance.

Claims

1. 1. A method for testing a semiconductor package having a die and a die substrate, comprising: forming one or more wirebond interconnects between the die and the die substrate; capturing input data characteristic of a wirebond interconnect during inspection of the formed wirebond interconnect; Passing the captured input data to a machine learning engine; processing the captured input data by the machine learning engine using a trained model to obtain an output array of data; evaluating the output array of data to determine a predicted radio frequency (RF) performance rating; and outputting the predicted RF performance rating.

2. The method of claim 1 , further comprising: rejecting or accepting production of the package according to the output predicted RF performance rating.

3. 3. The method of claim 2, wherein the failing or passing of the manufacturing of the package according to the output predicted RF performance rating occurs during or after wire bonding assembly of the semiconductor package manufacturing.

4. The method of any of claims 1 to 3, further comprising generating an alert based on the predicted RF performance rating.

5. The method of claim 4 , wherein the generating of the warning occurs during a wire bonding assembly in the semiconductor package manufacturing.

6. The method of any of claims 1 to 5, wherein the capturing of the input data includes capturing one or more digital images with an optical or infrared camera device positioned relative to the package such that a package area having the formed wire bond connections is within a field of view of the optical or infrared camera device.

7. 7. The method of claim 6, wherein the capturing of the input data includes capturing distance data by a sensor system positioned relative to the package, the distance data including a set of data points representing distance values ​​between a package area having the formed wire bond connections and the sensor system.

8. storing a training data set in a computer readable memory; The method of any of claims 1 to 7, further comprising: processing the training data set with the machine learning engine to obtain the trained model.

9. The processing of the training data set includes: applying the image data and parameters to a multi-layer neural network with feature extraction and classification to obtain a set of candidate machine learning models; and selecting an optimal machine learning model from the set of candidate machine learning models for use as the trained model.

10. 10. The method of claim 9, wherein the training data set includes wirebond parameters, RF performance parameters, and RF rating labels generated from previous wirebond assemblies.

11. 11. The method of claim 10, wherein the wire bond parameters include, for each wire bond, parameters of wire bond planar length, wire bond curvilinear (3D) length, wire bond loop height, wire bond diameter, and angle of the wire bond from a bird's-eye view relative to a plane.

12. 11. The method of claim 10, wherein traces are coupled to the wire bonds at corresponding sections, and the wire bond parameters include one or more of the following parameters: wire bond planar length, wire bond curvilinear (3D) length, wire bond loop height, wire bond diameter, number of wire bonds, rough spacing between adjacent wire bonds, angle of wire bonds from a bird's-eye view relative to a plane, trace width of the corresponding section, or trace length of the corresponding section.

13. 11. The method of claim 10, wherein traces are coupled to the wire bonds at corresponding sections, and the wire bond parameters include a wire bond planar length, a wire bond curvilinear (3D) length, a wire bond loop height, a wire bond diameter, a number of wire bonds, a rough spacing between adjacent wire bonds, an angle of the wire bonds from a bird's-eye view relative to a plane, a trace width of the corresponding section, or a trace length of the corresponding section.

14. The method of claim 10 , wherein the training data set further comprises synthetic data regarding RF performance obtained from modeling software.

15. 11. The method of claim 10, wherein the trained model includes a trained neural network model, the machine learning engine includes an inference stage, and the processing of the captured input data includes classifying the input data in the inference stage using the trained neural network model to obtain an output array of data having a predicted value for an RF performance rating corresponding to the package having the formed wire bond interconnects.

16. The method of claim 15 , wherein the RF performance rating identifies an RF performance level from among a plurality of performance levels.

17. 1. An inspection system for use in the manufacture of a semiconductor package having a die, a die substrate, and one or more wire bond interconnects formed between the die and the die substrate, comprising: a monitoring device configured to capture input data characteristic of the wirebond interconnects during inspection of the formed wirebond interconnects; 1. A radio frequency (RF) performance prediction device tool comprising: processing the captured input data with a machine learning engine having a trained model to obtain an output array of data; evaluating the output array of data to determine a predicted RF performance rating; and outputting the predicted RF performance rating.

18. 20. The system of claim 17, wherein the RF performance predictor tool is further configured to reject or accept the manufacturing of the semiconductor package according to the output predicted RF performance rating.

19. 19. The system of claim 17 or 18, further comprising an alert generator configured to generate an alert based on the predicted RF performance rating.

20. 20. The system of claim 19, wherein the alert generator is configured to generate an alert based on the predicted RF performance rating during a wire bonding assembly of the semiconductor package manufacture.

21. The system of any of claims 17 to 20, wherein the monitoring equipment includes one or more optical or infrared camera devices for capturing digital images, the optical or infrared camera devices positioned relative to the semiconductor package such that the package area having the formed wire bond connections is within a field of view of the optical or infrared camera device.

22. 22. The system of claim 21, wherein the monitoring equipment further comprises a sensor system positioned relative to the semiconductor package and configured to capture distance data including a set of data points representing distance values ​​between a package area having the formed wire bond connections and the sensor system.

23. a computer readable memory configured to store a training data set; The system of any of claims 17 to 22, further comprising at least one processor configured to process the training data set with the machine learning engine to obtain the trained model.

24. The at least one processor Applying the image data and parameters to a multi-layer neural network with feature extraction and classification to obtain a set of candidate machine learning models; and selecting an optimal machine learning model from the set of candidate machine learning models for use as the trained model.

25. 24. The system of claim 23, wherein the training data set includes wire bond parameters, RF performance parameters, and RF rating labels generated from previous wire bond assemblies.

26. 26. The system of claim 25, wherein the wire bond parameters include, for each wire bond, parameters of wire bond planar length, wire bond curvilinear (3D) length, wire bond loop height, wire bond diameter, and angle of the wire bond from a bird's eye view relative to a plane.

27. 26. The system of claim 25, wherein traces are coupled to the wire bonds at corresponding sections, and the wire bond parameters include one or more of the following parameters: wire bond planar length, wire bond curvilinear (3D) length, wire bond loop height, wire bond diameter, number of wire bonds, rough spacing between adjacent wire bonds, angle of wire bonds from a bird's eye view relative to a plane, trace width of the corresponding section, or trace length of the corresponding section.

28. 26. The system of claim 25, wherein traces are coupled to the wire bonds at corresponding sections, and the wire bond parameters include parameters of wire bond planar length, wire bond curvilinear (3D) length, wire bond loop height, wire bond diameter, number of wire bonds, rough spacing between adjacent wire bonds, angle of wire bonds from a bird's eye view to a plane, trace width of the corresponding section, or trace length of the corresponding section.

29. 26. The system of claim 25, wherein the training data set further comprises synthetic data regarding RF performance obtained from modeling software.

30. 26. The system of claim 25, wherein the trained model includes a trained neural network model, the machine learning engine includes an inference stage, and the at least one processor is configured to process the captured input data including classifying the input data in the inference stage using the trained neural network model to obtain an output array of data having predicted values ​​for an RF performance rating corresponding to the semiconductor package having the formed wire bond interconnects.

31. The system of any of claims 17 to 30, wherein the RF performance rating identifies an RF performance level from among a plurality of performance levels.

32. The system of any of claims 17 to 31, further comprising a machine learning engine coupled to the RF performance predictor.

33. 33. The system of claim 32, wherein the machine learning engine includes an inference stage, a training stage, and a multi-layer neural network.

34. 34. The system of claim 33, wherein the multi-layer neural network comprises a convolutional neural network (CNN).

Images