Microvoid analysis system and method in crystal growth using a continuous Czochralski pulling apparatus

Abstract

The computer device includes at least one processor that communicates with at least one memory device. The at least one processor is programmed to a) receive at least one image of the silicon molten crystal in a crucible, b) run a model that has been trained to segment the at least one image into different classes, c) analyze the segmentation to determine the quality of the crystal, and / or d) approve or reject the crystal based on the analysis.

Description

Cross - reference to related applications

[0001] This application claims the priority of U.S. Provisional Patent Application No. 63 / 496,830, filed on April 18, 2023, the entire disclosure of which is incorporated herein by reference.

Technical Field

[0002] This field generally relates to the analysis of microvoids in crystals grown by a continuous Czochralski pulling apparatus, and more specifically, to analyzing images of silicon melts in a triple crucible with artificial intelligence to detect and / or predict microvoids.

Background Art

[0003] The demand for 300 - mm - diameter silicon wafers doped with high concentrations of arsenic and phosphorus has been steadily increasing. This requirement can be very effectively met by using crystals grown by the continuous Czochralski method (CCz). However, the demand for 300 - mm - diameter silicon wafers also requires a low number of voids. Over time, the requirements have become more stringent, and the number of defects per wafer has decreased from less than 50 to less than 10. To meet this requirement, it is necessary to repeatedly improve the technology for controlling the number of voids (local light scattering - LLS). Therefore, a system for improving the void control ability of 300 - mm - diameter crystals grown by the CCz method is desired.

[0004] This background section is intended to introduce the reader to various aspects of technologies that may be related to various aspects of the present disclosure described and / or claimed below. This discussion is considered useful in providing the reader with background information for a better understanding of the various aspects of the present disclosure. Therefore, these descriptions should be read from this perspective and should not be construed as an admission of prior art.

Summary of the Invention

[0005] In one embodiment, the computer system includes a computing device which includes at least one processor that communicates with at least one memory device. The at least one processor may be configured to perform: a) receiving at least one image of the silicon molten crystal in a crucible; b) running a model that has been trained to segment the at least one image into different classes; c) analyzing the segmentation to determine the quality of the crystal; and / or d) approving or rejecting the crystal based on the analysis.

[0006] In another embodiment, the computerized method may be performed by a continuous Czochralski image analysis (CCZIA) computerized apparatus including at least one processor communicating with at least one memory device. The method may include a) receiving at least one image of the silicon molten crystal in a crucible, b) running a model trained to segment the at least one image into different classes, c) analyzing the segmentation to determine the quality of the crystal, and / or d) approving or rejecting the crystal based on the analysis.

[0007] In another embodiment, when executed by a computing device having at least one non-volatile computer-readable medium having computer-executable instructions embodied thereon, and having at least one processor and at least one memory device communicately connected, the computer-executable instructions may cause at least one processor to perform the following operations: a) receiving at least one image of silicon molten crystal in a crucible; b) running a model trained to segment at least one image into different classes; c) analyzing the segmentation to determine the quality of the crystal; and / or d) approving or rejecting the crystal based on the analysis.

[0008] The advantages will become more apparent to those skilled in the art from the following description of preferred embodiments illustrated and described by drawings. As will be understood, other different embodiments are possible, and their details can be modified in various ways. Accordingly, the drawings and description are to be considered illustrative and not restrictive.

[0009] The figures described below illustrate various aspects of the disclosed system and method. Each figure illustrates an embodiment of a particular aspect of the disclosed system and method, and each figure is intended to correspond to a possible embodiment. Furthermore, the following description refers to the reference figures included in the following figures whenever possible, and features depicted in multiple figures are designated by consistent reference figures. [Brief explanation of the drawing]

[0010] [Figure 1A] Figure 1A is a schematic diagram of a triple crucible without a homogeneous cullet distribution. [Figure 1B] Figure 1B is a schematic diagram of a triple crucible with a homogeneous cullet distribution. [Figure 2A] Figure 2A shows a diagram of the silicon molten ring inside the crucible. [Figure 2B] Figure 2B shows a diagram of the silicon molten ring inside the crucible. [Figure 3A] Figure 3A shows an example of how a crucible is used. [Figure 3B] Figure 3B shows an example of the labeling scheme for the crucible in use shown in Figure 3A. [Figure 4A] Figure 4A shows an example of a crucible in use before the image is segmented using the labeling scheme shown in Figure 3B. [Figure 4B] Figure 4B shows an example of a diagram of a crucible in use after the image has been segmented using the labeling scheme shown in Figure 3B. [Figure 5A] Figure 5A shows an example of the stages of crystal growth. [Figure 5B]Figure 5B shows an example of the stages of crystal growth. [Figure 5C] Figure 5C shows an example of the stages of crystal growth. [Figure 5D] Figure 5D shows an example of the stages of crystal growth. [Figure 6A] Figure 6A shows an example of a further stage in crystal growth. [Figure 6B] Figure 6B shows an example of a further stage in crystal growth. [Figure 6C] Figure 6C shows an example of a further stage in crystal growth. [Figure 6D] Figure 6D shows an example of a further stage in crystal growth. [Figure 7] Figure 7 shows an example of the process for analyzing the continuous Czochralski method. [Figure 8] Figure 8 shows an example system for executing the process shown in Figure 7. [Figure 9] Figure 9 shows an example of the configuration of a user computer device. [Figure 10] Figure 10 shows an example of a server computer system configuration.

[0011] Similar reference numerals in various drawings indicate the same elements. [Modes for carrying out the invention]

[0012] This field generally relates to the microvoid analysis of crystals grown by a continuous Chochralski (CCz) pulling apparatus, and more specifically, to analyzing images of silicon melt in a triple crucible with artificial intelligence to detect and / or predict microvoids. In the CCz process, during crystal growth, by continuously adding polysilicon and dopants to the annular portion between the outer crucible and the middle crucible, it is possible to achieve quasi-steady-state heat and mass transfer conditions. In the continuous supply mode, in the steady state, the mass of the silicon crystal growing at regular time intervals is equal to the amount of polysilicon added to the melt, the mass balance is maintained, and the volume of the silicon melt is kept constant.

[0013] The generation of voids within the crystal, and thus within the wafer, is inherently related to the CCz methodology. In fact, when continuously adding polysilicon raw materials during crystal growth, there is a possibility of generating argon bubbles in the melt. The bubbles usually have a size of 10 μm or less and are carried by the silicon melt flowing between adjacent melt regions. The bubbles are incorporated into the crystal and form voids. This phenomenon is well known in the CCz method, and historically, the problem of voids seen on the wafer has imposed significant limitations on the use of the CCz process.

[0014] In the case of 200-mm crystals pulled by CCz, a method has been used to control the melting characteristics and avoid the incorporation of voids into the crystal by using a small quartz piece called a "caulet". The caulet functions as a barrier and forms a meandering path for the melt to follow.

[0015] The process of using a cullet in the silicon melt is carried out to prevent the formation of microvoids. The purpose is to create a meandering path for the silicon melt to move from the outer region to the inner region in order to reduce the microvoids in a 200 mm crystal grown by the CCz method. The quartz body is cylindrical, but can be of any size and shape. The cullet can be placed in the outer crucible region between the outer crucible and the middle crucible. In other configurations, the cullet is placed in the transition region between the inner crucible and the middle crucible. During the CCz process, the quartz body is consumed by the melt. Thereafter, it is necessary to replenish the new quartz body.

[0016] By using a cullet, the generation of microvoids in the crystal grown by the CCz method can be suppressed, and by replenishing the quartz body at the end of crystal growth, the generation of further microvoids in the crystal due to the consumption of the quartz body can be kept low. However, this solution has been found to be ineffective for crystals with a diameter of 300 mm.

[0017] One solution to the problem of 300 mm diameter crystals is based on controlling the ratio of the weight of the cullet (grams) added before the start of crystal growth and the elapsed time (hours) from the addition of the cullet by the supply system to the start of crystal growth. At an appropriate ratio, the void content at the crystal / silicon melt interface is less than 30, which is the specification limit for 300 mm diameter crystals. Specifically, the appropriate ratio is 60 g / hr, and for melts with a ratio exceeding 60 g / hr, the void content at the crystal / silicon surface decreases.

[0018] The utilization efficiency of the cullet depends on the distribution of the underwater cullet and the floating cullet in the outer silicon melt annular part. The underwater cullet mainly comes from the amount of cullet placed on the initial polysilicon solid charge, and the floating cullet mainly comes from the amount dumped onto the melt by the feeder. A uniform distribution may lead to an improvement in efficiency by reducing the amount of microbubbles formed by its cushion effect and increasing the filter effect by forming a wide path in the melt.

[0019] This disclosure describes using a camera and artificial intelligence to determine whether the cullet distribution within the outer ring portion is uniform or non-uniform.

[0020] This disclosure is based on the case where the optimal ratio is just under 60 g / hr, in which case the void number was never below 30 (@0.2um), and the specification upper limit has been updated to 10 (@0.2um).

[0021] Knowing the distribution of cullet within the outer ring is fundamental to proper crystal growth, as it verifies whether the wafer produced from the grown crystal falls within the microvoid number specifications. This solution can be applied in addition to controlling the ratio of the cullet weight (grams) added before crystal growth begins to the elapsed time (hours) from cullet addition by the feeder system to the start of crystal growth.

[0022] Furthermore, this solution includes a check backup in case the timer for cullet ejection fails, ejection is not performed properly, or an anomaly occurs (e.g., if growth parameters such as crystal diameter become uncontrollable). By installing a camera to check the cullet distribution, it is possible to confirm that everything is working correctly and to know if the wafer void count is 10@0.2um or less.

[0023] This system describes a preventative control system for ensuring uniformity of cullet distribution in the external silicon molten annular portion. As described herein, uniform distribution leads to increased efficiency by reducing the amount of microbubble formation through its cushioning effect and increasing the filtering effect through meandering paths in the molten material. Without the process described herein, feedback cannot be obtained before laser inspection performed on the silicon wafer, and microvoid defects are detected downstream of the crystal growth process.

[0024] The cullet is placed in the outer annular portion of the CCz crucible, where polysilicon is continuously discharged and the melting level is kept constant. The cullet covers the entire surface of the polysilicon annular portion to avoid voids where the discharged polysilicon comes into direct contact with the molten material (forming microbubbles in the silicon molten material).

[0025] Figures 1A and 1B are schematic diagrams of a triple crucible without a homogeneous cullet distribution and a triple crucible with a homogeneous cullet distribution. Figures 1A and 1B are schematic diagrams of triple crucible 100 viewed from above. More specifically, Figure 1A schematically shows the suboptimal state before crystal growth due to the non-uniform distribution of cullet. Figure 1B, on the other hand, shows the optimal state before crystal growth due to the uniform cullet distribution.

[0026] To ensure good void counting capability in silicon single crystal growth using CCz technology, it is particularly important to guarantee that the cullet is homogeneously distributed across the outer annular portion of the crucible. The cullet needs to cover the entire surface of the polysilicon annular portion before and during crystal growth. This can be achieved by installing a camera combined with artificial intelligence. This code helps to distinguish between good conditions (schematically shown in Figure 1B) and potentially hazardous conditions (schematically shown in Figure 1A).

[0027] Crystals with a ratio higher than the optimal 60g / hr have excellent void counting capabilities. Since this ratio is fundamental to obtaining 300mm diameter crystals without microvoid contamination, a timer is installed on the pulling device to ensure the correct ratio is maintained during crystal transport. The ratio is checked before shipping the quartz, and any crystals with an unfavorable ratio for any reason are not shipped. While this ratio is the most important parameter for controlling the void count, it is not the only parameter used to judge the safety of a crystal in terms of microvoid contamination. In fact, depending on the crystal segment, the void count may exceed 10 counts (the current stricter limit) even when the appropriate ratio is achieved. Situations can occur during the crystal stage where the homogeneous cullet layer is stressed, altering its structure. For example, this can happen if the polysilicon supply is inconsistent, causing a sudden feeder problem that disrupts and eliminates the homogeneity of the cullet layer. Furthermore, even with child crystals of the appropriate ratio, improper cullet addition to the parent crystal can result in a cullet distribution different from that of a standard process of record (POR). In such cases, considering only the ratio is insufficient to guarantee void counting capability, as it does not necessarily reflect the actual cullet distribution. In another situation, the crystal ratio may be quite close to the critical limit of the optimal ratio. In such cases, the ratio may fall below the critical value, potentially leading to the formation of microvoids. Since this ratio is below the critical threshold, microvoids may be present, and the cullet flow layer needs to be analyzed with a camera to determine the risk of microvoid presence. This improves yield and reduces the number of crystals that are not shipped.

[0028] Prior to laser inspection of the wafer, additional control measures that can be taken during the crystallization process to ensure that there are no voids in the crystal are particularly important, especially in the case described above. A monitoring camera perpendicular to the annular portion of the external silicon molten material is applied to a 4 cm diameter circular area, and one or more models provided by artificial intelligence are used to evaluate the cullet distribution at the crystalline stage. Images of the external silicon molten material are captured by camera 825 (shown in Figure 8).

[0029] In the exemplary embodiments, the image creation (elaboration) process using the trained model requires considerable computational resources in terms of hardware. In these embodiments, it may not be possible to acquire an image every second in order to improve the processing performance of system 800 (shown in Figure 8). To simplify the process, an image may be acquired after each rotation of crucible 100 is completed, with a slight delay added between the previous and next images. This method allows for a complete scan of the entire external silicon molten ring after several rotations of crucible 100.

[0030] The captured and saved images are scanned by the CCZIA (Continuous Czochralski Image Analysis) server 810 (shown in Figure 8). The CCZIA server 810 runs one or more models. One or more models are trained to qualitatively predict the quality of the caret layer.

[0031] Figures 2A and 2B are two images showing a silicon molten ring in a crucible 100 (shown in Figure 1). Figures 2A and 2B show two images obtained by running one or more models and using the images as input. Figure 2A is associated with image 200, which was recognized as not being in an optimal state and displayed a “Warning” message 205. Figure 2B is associated with image 210, which was recognized as having a good cullet distribution with no liquid regions and displayed a “Pass” message 215. The objective is to prevent argon microbubbles from polysilicon discharge in the liquid region from being incorporated into the solid crystal growth region, as in 200 in Figure 2A.

[0032] Figures 2A and 2B represent two extremely different cases. Figure 2A shows islands of cullet near transparent silicon molten material. Figure 2B shows a very dense cullet structure. In many cases, the actual distribution of cullet can be far more complex than the two cases above. For example, there may be numerous small islands of cullet around a region of liquid silicon, or a dense cullet layer with a small liquid region on top. A quantitative approach is needed to calculate the proportion of liquid region in images captured by a camera and processed by an AI model.

[0033] Figure 3A shows an example of the usage state of crucible 100 (shown in Figure 1). The diagram shows an example of a labeling scheme for crucible 100 in use, as shown in Figure 3A.

[0034] This system employs semantic segmentation, a deep learning algorithm that associates a label or category with each pixel in an image. Segmentation is the process of dividing an image into separate regions that share common features. The result of the segmentation process is an image with the same dimensions as the initial image and composed of monochromatic regions. The first step is labeling, which aims to obtain an equivalent image with each pixel masked according to its associated class.

[0035] In this segmentation, four different classes were selected with four different colors: 1) background (black), 2) silicon melt (blue), 3) cullet (gray), and 4) crucible (yellow). These colors were chosen for training and user purposes, but other colors could also be used for the mask. Figure 3A shows the image, and Figure 3B shows the image after labeling.

[0036] Multiple training images may be generated to train one or more models to identify classes within the images. The training process for a semantic segmentation network for classifying images includes the following steps: 1) analyzing a collection of pixel-labeled images, 2) creating a semantic segmentation network and training the network to classify images into pixel categories, and 3) evaluating the network's accuracy. One or more models described herein are trained for prediction and are trained using either supervised or unsupervised learning.

[0037] The inference phase uses the trained model to perform predictions using a segmentation algorithm.

[0038] Figures 4A and 4B show examples of the crucible 100 (shown in Figure 1) in use, before and after segmenting the image using the labeling scheme shown in Figure 3B.

[0039] Figure 4A is the real or unprocessed image, and Figure 4B is a relative mask image in which four different classes are separated by four different colors.

[0040] The number of pixels in each class is counted, and by dividing this by the total surface area, the proportion of each class can be calculated. For example, the proportion of molten silicon can be calculated, and this proportion correlates with the inclusion rate of microvoids in the crystal.

[0041] Figures 5A–5D show examples of crystal growth stages. More specifically, Figures 5A–5D show the correlation between 0.12 μm microvoids in each crystal segment and images collected by a camera during crystal growth. A slight increase in microvoids is observed from Figure 5A to Figure 5D. This example is associated with a ratio of subcrystals exceeding the critical threshold, where cullet was not ejected in the previous parent crystal.

[0042] Figure 5A shows the crystalline region between 10 cm and 45 cm, Figure 5B shows the crystalline region between 45 cm and 80 cm, Figure 5C shows the crystalline region between 80 cm and 114 cm, and Figure 5D shows the crystalline region between 114 cm and 148 cm. In this example, the crystal has a zero-loss defect at 179 cm, and it is not possible to measure microvoids beyond Figure 5D.

[0043] Figure 5D shows the 75th percentile approaching 10 counts. The correlation between the number of microvoids and the liquid area on the image is clear. If the cullet distribution is compact (as shown in Figures 5A and 5B), no microvoids will be present in the crystal. When liquid areas begin to form (as shown in Figure 5C), the number of microvoids tends to be considerably higher than in Figures 5A and 5B. If the liquid area is consistent (as shown in Figure 5D), the number of microvoids increases and tends to approach 10 counts. From one or more models, it can be determined that to avoid an increase in microvoid inclusion (above 10 counts), the liquid area in the image captured by the camera must be lower than 1 / 4 of the total surface inspected by the camera. One or more of the described models make it possible to quantify the liquid area in each image during the crystalline phase, and a warning message is displayed if the liquid area exceeds 1 / 4 of the total surface.

[0044] Figures 6A–6D illustrate examples of further stages in crystal growth. Figures 6A–6D show another example where a proportion of child crystals above the critical threshold report high microvoid numbers. Similar to the previous example shown in Figures 5A–5D, this originates from child crystals where cullet from the parent crystal was not ejected before the crystalline stage. In this case, an increase in microvoids was observed from Figure A to Figure D. While the uniformity of the cullet distribution is acceptable up to Figure C, in Figure D the proportion of liquid region exceeds 1 / 4 of the entire imaged surface, and the 75th percentile exceeds 10 counts in Figure D.

[0045] Figure 7 shows an example of process 700 for analyzing the Czochralski method. In this exemplary embodiment, process 700 is executed by the CCZIA server 810 (shown in Figure 8).

[0046] In the exemplary embodiment, the CCZIA server 810 trains a model using multiple pixel-labeled images 705 so that the model can classify the pixels of an image into pixel categories. In some embodiments, training the model 705 includes the steps of 1) analyzing a collection of pixel-labeled images, 2) creating a semantic segmentation network and training the network to classify images into pixel categories, and 3) evaluating the accuracy of the network. In some embodiments, the CCZIA server 810 is further programmed to train the model on multiple pixel-labeled images so that the model can classify the pixels of an image into pixel categories. In at least one embodiment, the model is trained 705 to segment at least one image into multiple segments including, but not limited to, background, silicon molten, cullet, and crucible. In at least one embodiment, the model is a semantic segmentation network.

[0047] In the exemplary embodiment, the CCZIA server 810 receives at least one image of the crystalline silicon molten in the crucible 100 (shown in Figure 1). In the exemplary embodiment, at least one image is received by 710 from a camera 825 (shown in Figure 8) positioned perpendicular to the annular portion of the outer silicon molten. In the exemplary embodiment, the silicon molten is associated with a continuous Czochralski process. In the exemplary embodiment, the crucible 100 is a triple crucible 100, as shown in Figure 1.

[0048] In the exemplary embodiment, the CCZIA server 810 runs a model 715 that has been trained to segment at least one image into different classes. In the exemplary embodiment, the model is trained to segment at least one image pixel by pixel. In at least one embodiment, the model produces an output of segmented images.

[0049] In the exemplary embodiment, the CCZIA server 810 analyzes the segmentation of at least one image to determine the quality of the crystal 720. In at least one embodiment, the CCZIA server 810 masks the segmented image to determine the proportion of regions associated with one or more classifications 725.

[0050] In exemplary embodiments, the CCZIA server 810 approves or rejects the crystal based on the analysis 730. In some of these embodiments, the CCZIA server 810 approves or rejects the crystal based on the number of microvoids predicted to occur in the crystal based on the analysis 735. In some further embodiments, the CCZIA server 810 rejects the crystal if the area of ​​the silicon molten exceeds one-quarter of the total surface area in the crucible 100.

[0051] In some embodiments, the CCZIA server 810 performs one or more actions based on the approval and / or rejection 735 of the crystal, instructs other computer devices and / or systems to perform one or more actions, and / or instructs one or more users to perform one or more actions. In response to the rejection 735 of the crystal, one or more actions include, but are not limited to, stopping crystal growth, removing the crystal, draining a new amount of cullet into the crucible to restore a homogeneous distribution, and / or restarting the growth of a new crystal.

[0052] In some further embodiments, the CCZIA server 810 retrains the model with subsequent pixel-labeled images.

[0053] While the above describes the use of the systems and processes described herein to analyze silicon molten material in the CCz process, those skilled in the art will understand that these systems and methods can also be used to classify the conditions of other liquid molten materials.

[0054] Figure 8 shows an exemplary system 800 for performing process 700 (shown in Figure 7). In the exemplary embodiment, system 800 is used to analyze image data for voids in the continuous Czochralski method. Furthermore, system 800 is a continuous Czochralski image analysis (CCZIA) computer system configured to analyze crucible images.

[0055] As will be described in more detail later, the CCZIA server 810 is programmed to analyze crucible images to identify and count potential voids. Furthermore, the CCZIA server 810 is programmed to train one or more models to be used for analyzing crucible images. The CCZIA server 810 is programmed to a) receive at least one image of the silicon molten crystal in crucible 100, b) run a trained model to segment at least one image into different classes, c) analyze the segmentation to determine the quality of the crystal, and / or d) approve or reject the crystal based on the analysis.

[0056] In the exemplary embodiments, client device 805 is a computer including a web browser or software application, thereby enabling client device 805 to communicate with CCZIA server 810 using the Internet, a local area network (LAN), or a wide area network (WAN). In some embodiments, client device 805 is connected to the Internet via many interfaces, including but not limited to at least one of the following: the Internet, LAN, WAN, or a network such as an Integrated Services Digital Network (ISDN), a dial-up connection, a digital subscriber line (DSL), a cellular connection, a satellite connection, and a cable modem. Client device 805 is a desktop computer, a laptop computer, a PDA (Personal Digital Assistant), a mobile phone, a smartphone, a tablet, a phablet, a wearable electronic device, a smartwatch, a virtual headset or glasses (which could be any device that can access a network such as the Internet, e.g., an AR (Augmented Reality), VR (Virtual Reality), or XR (Augmented Reality) headset or glasses), a chatbot, a voice bot, a ChatGPT bot or ChatGPT-based bot, or other web-based connectable device or mobile device.

[0057] In the exemplary embodiment, the CCZIA computer device 810 (also known as the CCZIA server 810) is a computer including a web browser or software application that enables the CCZIA server 810 to communicate with client devices 805 and cameras / sensors 825 using the Internet, a local area network (LAN), or a wide area network (WAN). In some embodiments, the CCZIA server 810 is connected to the Internet via many interfaces, including, but not limited to, the Internet, LAN, WAN, or a network such as ISDN (Integrated Services Digital Network), dial-up connection, DSL (Digital Subscriber Line), cellular connection, satellite connection, and cable modem. CCZIA Server 810 is a desktop computer, laptop computer, PDA (Personal Digital Assistant), mobile phone, smartphone, tablet, phablet, wearable electronic device, smartwatch, virtual headset or glasses (e.g., AR (Augmented Reality), VR (Virtual Reality), or XR (Augmented Reality) headset or glasses, etc.), chatbot, voice bot, ChatGPT bot or ChatGPT-based bot, or other web-based connectable device or mobile device.

[0058] The database server 815 is communicatively coupled to the database 820 that stores data. In one embodiment, the database 820 is a database containing one or more analytical models and / or analytical information. In one embodiment, the database 820 is stored remotely from the CCZIA server 810. In some embodiments, the database 820 is distributed. In the exemplary embodiment, a person can access the database 820 via the client device 805 by logging into the CCZIA server 810.

[0059] The camera / sensor 825 can be any camera and / or sensor that transmits images to the CCZIA server 810 and with which the CCZIA server 810 can communicate. In exemplary embodiments, there is a camera / sensor 825 that communicates with the CCZIA server 810 using the Internet, a local area network (LAN), or a wide area network (WAN). In some embodiments, the camera / sensor(s) 825 is coupled to the Internet via many interfaces, including but not limited to at least one of the following: the Internet, LAN, WAN, or a network such as an Integrated Services Digital Network (ISDN), a dial-up connection, a digital subscriber line (DSL), a cellular connection, a satellite connection, and a cable modem.

[0060] Figure 9 shows an example configuration 900 of the user computer device 902. In this exemplary embodiment, the user computer device 902 may be similar to or the same as the client device 805 (shown in Figure 8). The user computer device 902 may be operated by user 901.

[0061] The user computer device 902 may include a processor 905 for executing instructions. In some embodiments, executable instructions may be stored in a memory area 910. The processor 905 may include one or more processing units (e.g., a multi-core configuration). The memory area 910 may be any device that stores and allows retrieval of information such as executable instructions and / or transaction data. The memory area 910 may include one or more computer-readable media.

[0062] The user computer device 902 may also include at least one media output component 915 for presenting information to the user 901. The media output component 915 may be any component capable of conveying information to the user 901. In some embodiments, the media output component 915 may include an output adapter (not shown), such as a video adapter and / or an audio adapter. The output adapter may be operably coupled to the processor 905 and operably coupled to an output device such as a display device (e.g., a cathode ray tube (CRT), liquid crystal display (LCD), light-emitting diode (LED) display, or "electronic ink" display) or an audio output device (e.g., speakers or headphones).

[0063] In some embodiments, the media output component 915 may be configured to present a graphical user interface (e.g., a web browser and / or a client application) to the user 901. The graphical user interface may include, for example, an interface for viewing information items provided by the CCZIA server 810 (shown in Figure 8). In some embodiments, the user computer device 902 may include an input device 920 for receiving input from the user 901. The user 901 may use the input device 920 to transmit information by voice or typing, but is not limited to these methods.

[0064] The input device 920 may include, for example, a keyboard, pointing device, mouse, stylus, touch-sensitive panel (e.g., touchpad or touchscreen), gyroscope, accelerometer, position detector, biometric input device, and / or audio input device. A single component, such as a touchscreen, may also function as both an output device and an input device 920 for the media output component 915.

[0065] The user computer device 902 may also include a communication interface 925 coupled to a remote device such as a CCZIA server 810 for communication. The communication interface 925 may include, for example, a wired or wireless network adapter and / or a wireless data transceiver for use in a mobile communication network.

[0066] Storing in the memory area 910 are, for example, computer-readable instructions for providing a user interface to user 901 via the media output component 915 and, optionally, for receiving and processing input from input device 920. The user interface may also include a web browser and / or a client application. The web browser allows users like user 901 to view and interact with media and other information typically embedded in web pages or websites from the CCZIA server 810. A client application allows user 901 to interact with the CCZIA server 810, for example. For instance, instructions are stored by a cloud service, and the execution output of the instructions is sent to the media output component 915.

[0067] Figure 10 shows an example configuration 1000 of the server computer device 1002. In this exemplary embodiment, the server computer device 1002 is similar to, or may be the same as, the CCZIA server 810 and the database server 815 (both shown in Figure 8). The server computer device 1002 may also include a processor 1005 for executing instructions. Instructions may be stored in a memory area 1010. The processor 1005 may include one or more processing units (e.g., a multi-core configuration).

[0068] The processor 1005 can be operably coupled to the communication interface 1015 so that the server computer unit 1002 can communicate with remote devices such as another server computer unit 1002, the CCZIA server 810, the camera / sensor 825, and the client device 805 (shown in Figure 8) (for example, using wireless communication or data transmission over one or more wireless links or digital communication channels). For example, the communication interface 1015 can receive input from the client device 805 over the internet, as shown in Figure 8.

[0069] The processor 1005 may also be operably coupled to a storage device 1025. The storage device 1025 may be, but is not limited to, any computer operating hardware suitable for storing and / or retrieving data, such as data related to one or more models. In some embodiments, the storage device 1025 may be integrated into a server computer unit 1002. For example, the server computer unit 1002 may include one or more hard disk drives as the storage device 1025.

[0070] In other embodiments, the storage device 1025 is located outside the server computer unit 1002 and may be accessed by multiple server computer units 1002. For example, the storage device 1025 may include multiple storage units such as a storage area network (SAN), a network-attached storage (NAS) system, and / or hard disks and / or solid-state disks in an inexpensive disk redundancy array (RAID) configuration.

[0071] In some embodiments, the processor 1005 is operably coupled to the storage device 1025 via a storage interface 1020. The storage interface 1020 can be any component that can provide the processor 1005 with access to the storage device 1025. The storage interface 1020 may include, for example, an ATA (Advanced Technology Attachment) adapter, a SATA (Serial ATA) adapter, a SCSI (Small Computer System Interface) adapter, a RAID controller, a SAN adapter, a network adapter, and / or any component that provides the processor 1005 with access to the storage device 1025.

[0072] The processor 1005 can execute computer-executable instructions for carrying out embodiments of the present disclosure. In some embodiments, the processor 1005 is converted into a special-purpose microprocessor by executing computer-executable instructions or by being programmed in other ways. For example, the processor 1005 can be programmed with instructions as shown in Figure 7.

[0073] At least one of the technical problems addressed by this system may include: (i) improving wafer analysis, (ii) reducing material loss due to malfunction, (iii) determining wafer quality early, (iv) improving the accuracy of wafer analysis, and / or (v) improving the accuracy of wafer analysis.

[0074] The technical effects of the systems and processes described herein can be achieved by performing at least one of the following steps: (i) receiving at least one image of the silicon molten crystal in a crucible; (ii) running a model trained to segment at least one image into different classes; (iii) analyzing the segmentation to determine the quality of the crystal; and / or (iv) approving or rejecting the crystal based on the analysis. [Machine Learning and Others]

[0075] The computer implementations discussed herein may include additional, fewer, or alternative operations, including those discussed elsewhere herein. These implementations may be carried out via one or more local or remote processors, transceivers, servers, and / or sensors (such as processors, transceivers, servers, and / or sensors mounted in a vehicle or mobile device, or processors, transceivers, servers, and / or sensors associated with a smart infrastructure or remote server), and / or via non-transient computer-readable media or computer-executable instructions stored on media.

[0076] In some embodiments, the HMIA server 410 is configured to perform machine learning so that, without explicit programming, the HMIA server 410 “learns” to analyze, organize, and / or process data. Machine learning can be performed through machine learning methods and algorithms ("ML methods and algorithms"). In exemplary embodiments, a machine learning module ("ML module") is configured to perform ML methods and algorithms. In some embodiments, ML methods and algorithms are applied to a data input to produce a machine learning output ("ML output"). The data input includes, but is not limited to, images. The ML output includes, but is not limited to, identified objects, item classifications, and / or other data extracted from the images. In some embodiments, the data input may include a specific ML output.

[0077] In some embodiments, at least one of several ML methods and algorithms can be applied, including, but not limited to, linear or logistic regression, instance-based algorithms, regularization algorithms, decision trees, Bayesian networks, cluster analysis, association rule learning, artificial neural networks, deep learning, complex learning, reinforcement learning, dimensionality reduction, and support vector machines. In various embodiments, the ML methods and algorithms implemented are directed towards at least one of several classifications of machine learning, such as supervised learning, unsupervised learning, and reinforcement learning.

[0078] In one embodiment, the ML module employs supervised learning. Supervised learning involves identifying patterns from existing data and then making predictions about subsequently received data. Specifically, the ML module is "trained" using training data that includes input examples and associated output examples. Based on the training data, the ML module generates a prediction function that maps outputs to inputs, and can use the prediction function to generate ML outputs based on data inputs. The input and output examples in the training data may include either the data inputs or ML outputs described above. In an exemplary embodiment, the processing element may be trained by providing a large sample of images having known properties or features. Such information may include, for example, information related to multiple images of multiple different objects, articles, and / or melts.

[0079] In another embodiment, the ML module may employ unsupervised learning. Unsupervised learning involves finding meaningful relationships from unorganized data. Unlike supervised learning, unsupervised learning does not involve user-driven learning based on input examples and their associated outputs. Rather, in unsupervised learning, the ML module can organize unlabeled data according to relationships determined by at least one ML method / algorithm employed by the ML module. The unorganized data can include any combination of the data inputs and / or ML outputs described above.

[0080] In yet another embodiment, the ML module can employ reinforcement learning to optimize its output based on feedback from a reward signal. Specifically, the ML module can receive a user-defined reward signal definition, receive data input, utilize a decision model to generate an ML output based on the data input, receive a reward signal based on the reward signal definition and the ML output, and then modify the decision model to receive a stronger reward signal for the subsequently generated ML output. Other types of machine learning, such as deep learning or hybrid learning, can also be employed.

[0081] Based on these analyses, the processing elements can learn how to identify features and patterns that can be applied to the analysis and classification of objects. This information is used to determine which classification model to use and which classification to provide. [Other considerations]

[0082] As understood in accordance with the preceding specification, the embodiments described above of this disclosure can be implemented using computer programming or engineering techniques, including computer software, firmware, hardware, or any combination or subset thereof. Any resulting program having computer-readable code means may be embodied or provided in one or more computer-readable media, thereby creating a computer program product, in other words, a product, in accordance with the embodiments discussed in this disclosure. Computer-readable media include, but are not limited to, fixed (hard) drives, diskettes, optical discs, magnetic tapes, semiconductor memory such as read-only memory (ROM), and / or transmission / reception media such as communication networks and links, such as the Internet. Products containing computer code can be manufactured and / or used by executing the code directly from one medium, by copying the code from one medium to another, or by transmitting the code over a network.

[0083] These computer programs (also known as programs, software, software applications, “apps,” or code) contain machine instructions for programmable processors and can be implemented in high-level procedural and / or object-oriented programming languages ​​and / or assembly / machine languages. As used herein, the terms “machine-readable medium” and “computer-readable medium” refer to any computer program product, apparatus, and / or device (e.g., magnetic disks, optical disks, memory, programmable logic devices (PLDs)) used to provide machine instructions and / or data to a programmable processor, including machine-readable medium that receives machine instructions as machine-readable signals. However, “machine-readable medium” and “computer-readable medium” do not include transient signals. “Machine-readable signals” refer to any signals used to provide machine instructions and / or data to a programmable processor.

[0084] As used herein, the term “database” may refer to the data itself, a relational database management system (RDBMS), or both. As used herein, a database may include any collection of data, including hierarchical databases, relational databases, flat-file databases, object-relational databases, object-oriented databases, and any other structured collection of records or data stored in a computer system. The above examples are illustrative and do not limit the definition or meaning of the term database. Examples of RDBMS include, but are not limited to, Oracle Database, MySQL, IBM DB2, Microsoft SQL Server, Sybase, and PostgreSQL. However, any database that enables the systems and methods described herein may be used. (Oracle is a registered trademark of Oracle Corporation (Redwood Shores, California), IBM is a registered trademark of International Business Machines Corporation (Armonk, New York), Microsoft is a registered trademark of Microsoft Corporation (Redmond, Washington), and Sybase is a registered trademark of Sybase (Dublin, California).)

[0085] As used herein, a processor can include any programmable system, including a microcontroller, a reduced instruction set circuit (RISC), an application-specific integrated circuit (ASIC), a logic circuit, and a system that uses other circuitry or processors capable of performing the functions described herein. The above examples are illustrative and do not limit the definition or meaning of the term “processor.”

[0086] As used herein, the terms “software” and “firmware” are interchangeable and include any computer program stored in memory for execution by a processor, including RAM memory, ROM memory, EPROM memory, EEPROM memory, and non-volatile RAM (NVRAM) memory. The above types of memory are examples and do not limit the types of memory that can be used to store computer programs.

[0087] In another example, a computer program is provided, and that program is embodied on a computer-readable medium. For example, this system runs on a single computer system without requiring a connection to a server computer. Further examples include a system running in a Windows environment (Windows is a registered trademark of Microsoft Corporation, Redmond, Washington). Still another example is a system running in a mainframe and UNIX server environment (UNIX is a registered trademark of X / Open Company Limited, Reading, Berkshire, UK). Further examples include a system running on an iOS environment (iOS is a registered trademark of Cisco Systems, Inc., San Jose, California). Still another example is a system running on a Mac OS environment (Mac OS is a registered trademark of Apple Inc., Cupertino, California). Still yet another example is a system running on Android OS (Android is a registered trademark of Google, Inc., Mountain View, California). Finally, a system running on Linux OS (Linux is a registered trademark of Linus Torvalds, Boston, Massachusetts). This application is flexible and designed to work in a variety of different environments without compromising its core functionality.

[0088] In some embodiments, the system includes multiple components distributed across multiple computing devices. One or more components may take the form of computer-executable instructions embodied in a computer-readable medium. The systems and processes are not limited to the specific embodiments described herein. In addition, each system and each process component can be implemented independently and separately from other components and processes described herein. Each component and process can also be used in combination with other assembly packages and processes.

[0089] In this specification, elements or steps described in the singular and referred to with the words “a” or “an” should be understood not to exclude multiple elements or steps unless such exclusion is expressly stated. Furthermore, references to “examples” or “one example” in this disclosure are not intended to be construed as excluding the existence of additional examples that also incorporate the mentioned features. Moreover, to the extent that the terms “includes,” “including,” “has,” “contains,” and their variations are used herein, such terms are intended to be comprehensive in the same manner as the open conjunction “comprises,” without excluding additional or other elements.

[0090] Furthermore, as used herein, the term “real time” refers to at least one of the time the relevant event occurs, the time the given data is measured and collected, the time the data is processed, and the time the system responds to the event and the environment. In the examples described herein, these activities and events occur substantially instantaneously.

[0091] The claims at the end of this document are not intended to be construed under 35 U.S.SC Section 112(f) unless the conventional means-plus-function terminology, such as “means for” or “steps for,” is explicitly stated in the claims.

[0092] This specification uses examples to disclose the disclosure, including the best mode, and to enable a person skilled in the art to practice the disclosure, including the manufacture and use of any device or system, and the execution of any method incorporated therein. The patentable scope of the disclosure is defined by the claims and may include other examples that a person skilled in the art can conceive. Such other examples are intended to be included in the claims if they have structural elements that are not different from the language of the claims, or if they include equivalent structural elements that are substantially not different from the language of the claims.

Claims

1. A computer device including at least one processor that communicates with at least one memory device, The aforementioned at least one processor is To receive at least one image of the crystalline silicon molten in the crucible, Running a model trained to segment at least one of the aforementioned images into different classes, Analyzing segmentation to determine crystal quality, and Approving or rejecting the crystal based on the above analysis, It is programmed to do, Computer device.

2. The aforementioned silicon melt is associated with the continuous Czochralski process. The computer device according to claim 1.

3. The aforementioned crucible is a triple crucible. The computer device according to claim 2.

4. The model is trained to segment the at least one image pixel by pixel. The computer device according to claim 1.

5. The aforementioned model generates a segmented image output. The computer device according to claim 4.

6. The at least one processor is further programmed to apply a mask to the segmented image in order to determine the proportion of regions associated with one or more classifications. The computer device according to claim 5.

7. The model is trained to segment the at least one image into multiple segments, including but not limited to a background, molten silicon, cullet, and a crucible. The computer device according to claim 4.

8. The at least one processor is programmed to reject the crystal if the area of ​​the molten silicon exceeds one-quarter of the total surface area in the crucible. The computer device according to claim 1.

9. The aforementioned model is a semantic segmentation network, The at least one processor is further programmed to train the model on multiple pixel-labeled images so that the model can classify the pixels of an image into pixel categories. The computer device according to claim 1.

10. The aforementioned at least one processor is further programmed to retrain the model on a subsequent set of pixel-labeled images. The computer device according to claim 9.

11. The at least one processor is further programmed to approve or reject the crystal based on the number of microvoids predicted to occur in the crystal based on the analysis. The computer device according to claim 1.

12. The aforementioned at least one image is received from a camera positioned perpendicular to the annular portion of the external molten silicon. The computer device according to claim 1.

13. A computer implementation method performed by a computer system including at least one processor that communicates with a chatbot and at least one memory device, The aforementioned method, To receive at least one image of the crystalline silicon molten in the crucible, Running a model trained to segment at least one of the aforementioned images into different classes, Analyzing segmentation to determine crystal quality, and Approving or rejecting the crystal based on the above analysis, A computer implementation method, including

14. The aforementioned silicon melt is associated with the continuous Czochralski process. The computer implementation method according to claim 13.

15. The aforementioned crucible is a triple crucible. The computer implementation method according to claim 14.

16. The aforementioned model is trained to segment at least one image pixel by pixel. The computer implementation method according to claim 13.

17. The aforementioned model generates a segmented image output. The computer implementation method according to claim 16.

18. This further includes masking the segmented image in order to determine the proportion of regions related to the 1-tertiary classification. The computer implementation method according to claim 17.

19. The model is trained to segment the at least one image into multiple segments, including but not limited to a background, molten silicon, cullet, and crucible. The computer implementation method according to claim 16.

20. The further includes rejecting crystals if the area of ​​the molten silicon exceeds one-quarter of the total surface area in the crucible. The computer implementation method according to claim 13.

21. The aforementioned model is a semantic segmentation network, The method further impairs the ability of the model to train using multiple pixel-labeled images in order to classify image pixels into pixel categories. The computer implementation method according to claim 13.

22. This further includes retraining the model with subsequent pixel-labeled images. The computer implementation method according to claim 21.

23. Further including approving or rejecting a crystal based on the number of microvoids predicted to occur in the crystal based on the analysis, The computer implementation method according to claim 13.

24. The aforementioned at least one image is received from a camera positioned perpendicular to the annular portion of the external molten silicon. The computer implementation method according to claim 13.

Images