KR102585478B1 - System for inspecting photoresist dispensing condition of spin coater - Google Patents

Abstract

The present invention includes a PR image acquisition unit installed in the spin coater to acquire an image of the photoresist supplied from the supply nozzle of the spin coater; a horizontal wafer image acquisition unit installed in the spin coater to obtain an image of the photoresist supplied from the supply nozzle to the upper surface of the wafer on the spin chuck and spreading by centrifugal force; A deep learning defect determination unit that determines whether the photoresist is defective by a deep learning algorithm from the images acquired from the first to vertical wafer image acquisition units; A control unit configured to output a control signal for generating an alarm when a photoresist defect occurs according to the determination result of the deep learning defect determination unit; and a defect notification unit that generates an alarm according to a control signal from the control unit. It relates to a system for inspecting the photoresist application state of a spin coater using deep learning. According to the present invention, the state of the nozzle used in photo resist (PR) application, which is one of the semiconductor manufacturing processes, and the application area of the photo resist are used to accurately detect defects by using deep learning technology, and the photo resist application is applied. By improving the detection and processing ability of the video status of the process, the nozzle status is monitored in real time, and when an abnormal defect condition is detected, an alarm is sounded in the relevant unit, thereby maximizing production efficiency by preventing process defects and enhancing yield.

Description

System for inspecting photoresist dispensing condition of spin coater using deep learning {System for inspecting photoresist dispensing condition of spin coater}

The present invention relates to a system for inspecting the photoresist application state of a spin coater, and more specifically, to accurately detect the abnormal state of photoresist (PR) application, which is one of the semiconductor manufacturing processes, using deep learning technology. This relates to a system for inspecting the photoresist application state of a spin coater using deep learning.

In general, the photo-lithography process is a process in which a certain pattern of the photo mask is transferred to the photoresist film by applying photoresist to the wafer, exposing it to light using a photomask, and developing it in a developer. At this time, a process of applying photoresist to the wafer is performed, which is mainly performed using a spin coater.

In this spin coater, when an externally supplied wafer is placed on a rotary chuck that fixes and rotates the wafer, the rotary chuck fixes the wafer and starts rotating. Then, the photoresist supply nozzle moves from the standby position to the spraying position and sprays a certain amount of photoresist on the center of the rotating wafer. This liquid photoresist spreads widely by centrifugal force on the rotating wafer and is affected by its own viscosity. It is distributed in an almost even thickness across the wafer. Afterwards, the liquid photoresist goes through a series of curing processes such as baking to form a solid state and is then inputted into the exposure process.

Quality control of photoresist application on wafers by such spin coaters is not only closely related to the wafer defect rate, but is also becoming more important in realizing fine line widths in semiconductors.

As a technology for controlling the quality of application of conventional photoresist, Korean Patent Publication No. 10-2003-0085692, “Photoresist sensing device and sensing method of spin coater,” has been proposed, which is a rotating device that can be rotated by a driving means. Chuck and; a photoresist supply nozzle equipped with a third sensing unit capable of detecting the amount of photoresist and applying a certain amount of photoresist to the wafer located on the upper part of the rotary chuck; a photoresist container provided with a first sensing unit capable of detecting the amount of photoresist and containing photoresist to be dispersed on the wafer; a pumping unit provided with a second sensing unit capable of detecting the amount of photoresist and pumping the photoresist contained in the photoresist container so that it can be supplied to the photoresist supply nozzle; A controller is provided to control the driving means so that the rotary chuck can rotate, and to control the pumping unit to pump the photoresist contained in the photoresist container and move it to the nozzle for supplying the photoresist.

However, as shown in Figure 1, the prior art has limitations in accurately determining defects that cause abnormal defects, unlike the normal photoresist application quality, and as a result, the degree to which it contributes to reducing the defect rate of wafers is not significant. There was a problem that it didn't work. This caused photoresist thickness non-uniformity on the wafer, seriously affecting the formation of circuit patterns on the wafer, and adversely affecting yield.

In order to solve the problems of the prior art as described above, the present invention utilizes deep learning technology to determine the state of the nozzle used in photo resist (PR) application, which is one of the semiconductor manufacturing processes, and the application area of the photo resist. It enables accurate defect detection and improves the accuracy of photoresist defect detection by intensively monitoring the edge of the wafer as well as the area around the supply nozzle or the plane of the wafer, where there is a high possibility of photoresist defects occurring, especially in semiconductor lithography. The purpose is to satisfy the demands for high reliability and increased yield of semiconductor devices required by the semiconductor industry by detecting the photoresist supply status of the nozzle for photoresist application, which is a key process equipment in the (lithography) process.

Other objects of the present invention may be easily understood through the description of the examples below.

In order to achieve the above object, according to one aspect of the present invention, it is a system for inspecting the application state of photoresist of a spin coater used in a photolithography process for manufacturing semiconductor devices, wherein the photoresist is supplied from the supply nozzle of the spin coater. a PR image acquisition unit that acquires an image of the photoresist; A defect occurs when the photoresist is applied from the supply nozzle twisted to the left or right by a deep learning algorithm from the image acquired from the PR image acquisition unit, and the photoresist hangs in the form of a water drop from the supply nozzle and then falls. A deep learning defect determination unit that determines defects that occur and defects that occur due to residues generated on the outer surface of the supply nozzle; A control unit configured to output a control signal for generating an alarm when a photoresist defect occurs according to the determination result of the deep learning defect determination unit; And a defect notification unit that provides to the outside the fact that a defect has occurred according to a control signal from the control unit, wherein the deep learning defect determination unit generates a feature map by max pooling from the image acquired from the PR image acquisition unit. Extract, distinguish object area (foreground) and non-object area (background) from the feature map, pass the feature map through CNN (Conv.3), and then again through CNN (Conv.1), This is connected to a fully connected layer to classify the object area and the non-object area, and a bounding-box circumscribing the object area and the non-object area is obtained. The location of the bounding box is located in the object area. is the center coordinate of the rectangle surrounding , the size of the bounding box is displayed in width and height, and the area is expressed using the display result, anchor box, and ground-truth box, The detection process involves the detection of a ROI (region of interesting) pooling layer, two fully connected layers, areas that are shifted to the left or right of the photoresist, droplet-shaped areas of the photoresist, and residue areas. and the output of the ROI pooling layer goes through two fully connected layers and is divided into the left or right side area of the photoresist, the droplet-shaped area of the photoresist, the residue area, and the background, using deep learning. A system for inspecting the photoresist application state of the spin coater used is provided.

The PR image acquisition unit consists of a pair and is arranged at an angle of 90 degrees around the supply nozzle, and the control unit calculates the image from the cross section of the photoresist normally supplied from the supply nozzle and the image obtained from the PR image acquisition unit. By displaying the cross-sections of the actual photoresist so that they overlap, supply defects of the photoresist can be displayed intuitively.

The control unit may control the spin coater to stop operation when the defect occurs in the photoresist according to the determination result of the deep learning defect determination unit.

a horizontal wafer image acquisition unit installed in the spin coater to obtain an image of the photoresist supplied from the supply nozzle to the upper surface of the wafer on the spin chuck and spreading by centrifugal force from the edge of the wafer in a direction horizontal to the wafer; And a vertical wafer image acquisition unit installed in the spin coater to obtain an image of the photoresist supplied from the supply nozzle to the upper surface of the wafer on the spin chuck and spreading by centrifugal force from the upper side of the wafer, and the deep learning The defect determination unit uses the image of the photoresist acquired by the PR image acquisition unit to determine the photo supplied downward from the bottom of the supply nozzle with respect to a virtual vertical line extending vertically downward from the bottom of the supply nozzle. Defects in the photoresist are determined according to the spacing amount of the resist, and defects in the convex or concave portion of the photoresist are determined using the image of the edge in the horizontal direction of the wafer obtained by the horizontal wafer image acquisition unit. Using the image of the upper surface of the wafer obtained by the vertical wafer image acquisition unit, defects of the photoresist are identified according to the ratio of the area having a color or shape that is different from the center on the wafer on which the photoresist is applied. It can be determined.

After the photoresist is applied to the wafer, a defect rate receiving unit is provided to receive the defect rate from a wafer yield management system that calculates the defect rate for each die of the wafer through probe test data inspected for each die of the wafer in a subsequent process. Included, the deep learning defect determination unit may modify the weight learned by the deep learning algorithm to be proportional to the defect rate and use it when determining defects in the photoresist.

According to the photoresist application state inspection system of a spin coater using deep learning according to the present invention, the state of the nozzle used for photoresist (Photo Resist, PR) application, which is one of the semiconductor manufacturing processes, and the application area of the photoresist are deep learned. It utilizes technology to enable accurate defect detection, improves the detection processing ability of the image state of the photoresist application process, monitors the nozzle state in real time, and sounds an alarm in the relevant unit when an abnormal defect state is detected to improve the process. Production efficiency can be maximized by preventing defects and enhancing yield, and the accuracy of photoresist defect detection can be improved by intensively monitoring the edge of the wafer as well as the area around the supply nozzle or the plane of the wafer, where there is a high possibility of photoresist defects occurring. In particular, by detecting the photoresist supply status of the nozzle for photoresist application, which is a key process equipment in the semiconductor lithography process, the demand for high reliability and increased yield of semiconductor devices required by the semiconductor industry can be met. It has a satisfying effect.

Figure 1 is an image showing a normal state and an abnormal state for conventional photoresist application, respectively.
Figure 2 is a schematic diagram of a state detection solution for a spin coater photoresist application state inspection system using deep learning according to an embodiment of the present invention.
Figure 3 is a front view showing the interior of a spin coater in which a system for inspecting the photoresist application state of the spin coater using deep learning according to an embodiment of the present invention is provided.
Figure 4 is a configuration diagram showing a system for inspecting the photoresist application state of a spin coater using deep learning according to an embodiment of the present invention.
Figure 5 is an image showing four types of photoresist coating state defects detected by a spin coater photoresist coating state inspection system using deep learning according to an embodiment of the present invention.
Figure 6 is a plan view showing the arrangement of the PR image acquisition unit in the photoresist application state inspection system of the spin coater using deep learning according to an embodiment of the present invention.
Figure 7 is a side view for explaining the defect determination criteria of the photoresist application state inspection system of the spin coater using deep learning according to an embodiment of the present invention.
Figure 8 is a partial enlarged view for explaining defect measurement through images acquired by the horizontal wafer image acquisition unit in the photoresist application state inspection system of the spin coater using deep learning according to an embodiment of the present invention.
Figure 9 is a partial enlarged view for explaining defect measurement through images acquired by the vertical wafer image acquisition unit in the photoresist application state inspection system of the spin coater using deep learning according to an embodiment of the present invention.
Figure 10 shows an anchor box and a reference box in a system for inspecting the photoresist application state of a spin coater using deep learning according to an embodiment of the present invention.
Figure 11 is a diagram illustrating the implementation of Faster R-CNN-based detection of the deep learning defect determination unit of the photoresist application state inspection system of the spin coater using deep learning according to an embodiment of the present invention.
Figure 12 shows the front-back and left-right spacing values based on the cross-section of a normal photoresist in a planar view in a system for inspecting the photoresist application state of a spin coater using deep learning according to an embodiment of the present invention.
Figure 13 is an image showing an exemplary monitoring screen in a spin coater photoresist application state inspection system using deep learning according to an embodiment of the present invention.

Since the present invention can be subject to various changes and can have various embodiments, specific embodiments will be illustrated in the drawings and described in detail. However, this is not intended to limit the present invention to specific embodiments, but should be understood to include all changes, equivalents, and substitutes included in the technical idea and scope of the present invention, and may be modified into various other forms. The scope of the present invention is not limited to the following examples.

Hereinafter, embodiments according to the present invention will be described in detail with reference to the attached drawings, and identical or corresponding components will be assigned the same reference numerals regardless of reference numerals, and redundant description thereof will be omitted.

Figure 2 is a schematic diagram of the state detection solution of the photoresist application state inspection system of the spin coater using deep learning according to an embodiment of the present invention, and Figure 3 is a photo of the spin coater using deep learning according to an embodiment of the present invention. It is a front view showing the inside of a spin coater in which a resist application state inspection system is provided, and Figure 4 is a configuration diagram showing a photoresist application state inspection system of a spin coater using deep learning according to an embodiment of the present invention.

Referring to Figures 2 to 4, the photoresist application state inspection system 100 of a spin coater using deep learning according to an embodiment of the present invention includes a PR image acquisition unit 110 and a deep learning defect determination unit 120. , It may include a control unit 130 and a defect notification unit 140, and is a system that inspects the photoresist application state of the spin coater 10 used in the photolithography process for manufacturing semiconductor devices.

The PR image acquisition unit 110 acquires an image of the photoresist supplied from the supply nozzle 210 of the spin coater 10, and is installed in the spin coater 10 for this purpose. The PR image acquisition unit 110, like the horizontal wafer image acquisition unit 160 and the vertical wafer image acquisition unit 170, which will be described later, may be used with various photographing devices, including, for example, a camera.

The PR image acquisition unit 110 needs to acquire images from various directions in order to accurately detect whether the photoresist supplied downward from the supply nozzle 210 has a uniform thickness and is properly supplied vertically. In order to accurately detect defects with the smallest number, as shown in FIG. 6, a pair may be arranged at an angle of 90 degrees around the supply nozzle 210. Each of the PR image acquisition units 110 may be installed in the casing 230 of the spin coater 10 to face the photoresist supplied downward from the supply nozzle 210.

The deep learning defect determination unit 120 uses a deep learning algorithm from the image acquired from the PR image acquisition unit 110 to turn the photoresist from the supply nozzle to the left (①) or right (②), as shown in FIG. 5. Distinguish between defects caused by spraying, defects caused by photoresist hanging from the supply nozzle in the form of water droplets (③) and then falling, and defects caused by residues generated on the outer surface of the supply nozzle (④). The deep learning defect determination unit 120 is located in the control box 231 provided in the casing 230 of the spin coater 10 along with the control unit 130, defect notification unit 140, and defect rate receiver 150, which will be described later. It can be accepted. Referring to FIG. 5, ① and ② represent defective elements that cause the photoresist to be applied incorrectly in the right or left direction due to residues on the nozzle and other reasons. ③ After application, the photoresist should be sucked back into the nozzle by about 2~3mm by the suck back technology, but as it cannot be sucked back in normally, it hangs in the form of water droplets at the end of the nozzle and falls off, or is applied together during the next application. If this occurs, it represents a defective element that causes a change in the viscosity of the photoresist and causes serious defects in film thickness. ④ represents a defective element that causes defects due to residue on the outer surface of the nozzle due to chemical vaporization of the photoresist.

The photoresist application state inspection system 100 of a spin coater using deep learning according to an embodiment of the present invention may further include a horizontal wafer image acquisition unit 160 and a vertical wafer image acquisition unit 170.

The horizontal wafer image acquisition unit 160 is a spin coater to obtain an image of the photoresist supplied from the supply nozzle 210 to the upper surface of the wafer on the spin chuck 220 and spreading by centrifugal force from the edge of the wafer in a direction horizontal to the wafer. It is installed at (10). The horizontal wafer image acquisition unit 160 acquires an image (B) of the edge of the wafer (W) in a direction horizontal to the wafer, as shown in FIG. 8, and captures the convex portion (b1) or This is advantageous for observing defects in the photoresist (PR), such as the concave portion (b2). In other words, there is a high risk of defects occurring in the photoresist (PR) at the edge of the wafer (W). When an image of the edge of the wafer (W) is acquired in the horizontal direction, a convex portion that impairs the flatness of the photoresist (PR) Observation of (b1) becomes easier and more accurate, and since the concave part (b2) shows a smudge-like shape compared to other parts when the wafer (W) is rotated, this also becomes easier and more accurate to observe.

The vertical wafer image acquisition unit 170 is installed on the spin coater 10 to obtain an image of the photoresist supplied from the supply nozzle 210 to the upper surface of the wafer on the spin chuck 220 and spreading by centrifugal force from above the wafer. . For example, the vertical wafer image acquisition unit 170 may be installed to face downward on the nozzle arm 211 on which the supply nozzle 210 is fixed.

For example, as shown in FIG. 7, the deep learning defect determination unit 120 uses the image of the photoresist acquired by the PR image acquisition unit 110, and extends vertically downward from the bottom of the supply nozzle 210. Defects in the photoresist can be determined according to the distance between the photoresist P supplied downward from the bottom of the supply nozzle 210, for example, the maximum distance d, with respect to the virtual vertical line H. This spacing amount can be calculated based on a predetermined thickness of the photoresist (P) through image processing. As another example, the deep learning defect determination unit 120 uses the image B of the edge in the horizontal direction of the wafer W acquired by the horizontal wafer image acquisition unit 160, as shown in FIG. It is possible to determine defects in the convex portion (b1) or concave portion (b2) of the resist PR. As another example, the deep learning defect determination unit 120 uses the image of the upper surface of the wafer acquired by the vertical wafer image acquisition unit 170, as shown in FIG. 9, to determine the wafer (W) on which the photoresist is applied. Defects in the photoresist can be determined based on the ratio of the area (A2) that has a color or shape that is different from the center (A1) in the image. This also uses an area calculation algorithm to calculate the difference in color or shape for each pixel through an image process. Therefore, the area ratio of the area A2 having a different color or shape from the center A1 may be calculated and used to determine defects in the photoresist.

The CNN hardware accelerator used in the deep learning defect determination unit 120 can be implemented with Verilog-based PL (Programmable Logic) hardware, data transmission can use the AMBA AXI4 bus, and Convolution is used for real-time processing of CNN in an embedded environment. It can be implemented as a hardware structure that improves the speed of CNN by processing calculations and other operations in parallel, and can be composed of a Convolution module, Max Pooling module, and Fully Connected module. Their operation process stores the weight, bias, and input image required for CNN calculation in Block Memory (weight memory, bias memory, input image) through the Bus Interface, and sequentially implements the operation starting from the Convolution1 module, using Max Pooling (Max Pooling is performed simultaneously on all feature maps, and the operation results of each module are stored in block memory. Here, the hardware structure can be implemented so that a multiplier can be used to minimize the execution time of the convolution layer, the convolution operation can be performed in 1 clk, and calculations can be performed continuously using a double buffer. The deep learning defect determination unit 120 can be manufactured as an FPGA board for neural network system development, AXI4 simulation environment and logic design, Channel write input arbiter, address/data fifo design, Channel write input arbiter, address/data fifo design , Boundary write split design, AXI I/F write wrapper design, AXI Master wrapper I/F, Slave I/F design, FPGA top module design, simulation environment construction and modeling for parallel processors, object recognition algorithm mapping on parallel processors, etc. This is implemented, and in addition to storing the learned weights in memory, the camera interface design is also implemented.

In order to implement a deep learning algorithm, the deep learning defect determination unit 120 may have a configuration for detecting four types of defects in the photoresist application state based on Faster R-CNN as shown in Figure 11, using Faster R-CNN. An algorithm is implemented to detect defective areas such as left and right distribution of photoresist and drops and residues in the photoresist supply nozzle 210, and the region proposal includes the extraction method in CNN. By allowing the entire process from input to output to be operated as a single network, it is implemented to reduce processing time and ensure learning consistency.

The deep learning defect determination unit 120 extracts a feature map from the image acquired from the PR image acquisition unit 110 by Max pooling, and detects the object area (foreground) where the object to be detected exists in the feature map. Distinguish the non-object area (background) where the object you want to target does not exist, pass the feature map through CNN (Conv.3), pass it again through CNN (Conv.1), and then transfer it to a fully connected layer. , classify the object area and non-object area, and obtain a bounding box that circumscribes the object area and non-object area. The location of the bounding box is the center coordinate (x, y), and the size of the bounding box is displayed as width (w) and height (h) (x,y,w,h), and the display result (object area, non-object area, x,y,w,h The region is expressed using (6 values of) an anchor box (see Figure 10) and a ground-truth box (see Figure 10), and the detection process is performed using a ROI (region of interesting) pooling layer. layer), two fully connected layers, the area that is shifted to the left or right of the photoresist, the water drop-shaped area of the photoresist, and the residue area are detected all at once, and the output of the ROI pooling layer Through these two fully connected layers, the area that is turned to the left or right of the photoresist, the droplet-shaped area of the photoresist, and the residue area can be distinguished from the background. Each object area has a real number between 0 and 1. This value is obtained using an anchor box and a reference box, and the anchor box is a window for searching the detection area. The value to distinguish between the object area and the non-object area uses Equation 1 below.

Here, A is the anchor box area, B is the reference box area, and S is the area where the two boxes overlap. Area values for the object area and non-object area can be set.

The detection process by the deep learning defect determination unit 120 includes a ROI (region of interesting) pooling layer, two fully connected layers, left and right dispensing of photoresist, and water droplets. It may consist of a detection process of the (drop) shape area and the residues area. The output of the ROI pooling layer passes through two fully connected layers (FC) and is divided into left & right dispensing, drop, residues areas, and background. The process of fine-tuning the position of the bounding box surrounding the object area is to confirm the detected area and is also the process of finding a rectangle circumscribed to the detected object.

In the case of anchor box and reference box detection, a simulator for operation verification for RTL can be developed, which is necessary for verification to confirm hardware and correct operation. This can be implemented with a C-language-based simulator, and network compression such as the insertion of a 1 It can be implemented using techniques. In addition, the amount of computation can be improved by reducing existing network layers and introducing new network techniques such as route (skip connection) and Reorg to minimize recognition performance degradation. The process of implementing the profiling function for verification, the process of creating a log file for the input/output information of each layer, and the process of generating the weight and parameter information of each layer as a log file can be performed.

The data to be learned has many difficulties in accessing data collection due to security issues due to the nature of the semiconductor industry, making it very insufficient for learning defect data images. Therefore, in order to secure sufficient data for learning, it is possible to secure supply nozzles and photoresists for photoresist application used in industrial sites by utilizing related maintenance companies, and then secure them through similar environmental experiments in other locations. It can also be secured by using transformation (parallel movement, angle adjustment, etc.) processing on the original image. Learning can be done through four steps: the first step is learning the RPN using training data, the second step is learning the detector using the proposal region obtained from the learned RPN, and the third step is convolution. After fixing the weights of the (CNN) and Max pooling steps, the RPN can be fine-tuned. The fourth step is the process of fine-tuning the detector using the proposed area obtained from the third step. Here, in the third and fourth steps, the weights of the convolution and Max pooling processes are shared. Additionally, by integrating RPNs that exist separately outside into the network, the convolution and max pooling parts can be shared, greatly shortening the object detection time. To develop a platform-based imaging system, apply a deep learning video surveillance platform, develop a UI-based defect inspection monitoring program specialized for semiconductor equipment (see Figure 13), video input of 8 or more network channels, screen split (1 split, 4 split, 8 split) Split), continuous or event video storage modes can be implemented, H.265, H.264, MJPEG compression methods, HDMI, VGA monitor output, and storage space of 1TB or more can be implemented.

Resolution expansion for high image quality is required for the image acquisition units (110, 160, 170). To this end, it can be configured to expand to a resolution of 128*128 or higher to obtain a high level of impact, and maintain performance to maximize integration. However, it is necessary to reduce the pixel area to secure larger resolution in the existing overall chip size.

High sensitivity of pixels is required to be able to detect objects even in low illumination. For this purpose, a high gain of the differential amplifier is required, so there is a problem that the fill factor of the photodiode that collects light is reduced. To solve this problem, a differential amplifier is used. (Differencing amp) can be replaced with a pulse generator.

The control unit 130 outputs a control signal for generating an alarm when a photoresist defect occurs according to the determination result of the deep learning defect determination unit 120. The control unit 130 can control the operation of the spin coater 10 to stop when a defect in the photoresist occurs according to the determination result of the deep learning defect determination unit 120. For this purpose, a device provided in the spin coater 10 is provided. The defect stopping unit 250 may be configured to receive a control signal from the control unit 130 and stop the operation of the spin coater 10.

The control unit 130 consists of a pair of cross sections of photoresist when normally supplied from the supply nozzle 210, and images acquired from the PR image acquisition unit 110 arranged at an angle of 90 degrees around the supply nozzle 210. By displaying the cross-section of the actual photoresist calculated from the display unit 180 so as to overlap, supply defects of the photoresist can be displayed intuitively. Accordingly, the operator can not only check the current status of the supply nozzle 210 through these display results, but also take prompt follow-up measures regarding it. The control unit 130 is a pair of PR image acquisition units 110 that acquire images so as to be orthogonal to each other. The photoresist supplied from each acquired image matches to some extent the vertical line (see FIG. 7), which is a standard for determining whether the photoresist is normal. spaced apart from each other is obtained numerically through image processing, and based on this, the cross-sectional shape of the photoresist supplied from the supply nozzle 210 is obtained through experiment, and the cross-sectional shape is measured to some extent from the vertical line that serves as the standard. By applying the numerically obtained value of the distance and arranging it from the reference, as shown in FIG. 12, the cross-section of the photoresist actually supplied reflects the front-to-back and left-right spacing values based on the cross-section (reference) of the normal photoresist in a planar plane ( By indicating the actual photoresist defect deviation, it is possible to quickly and intuitively identify the defect deviation of the actual photoresist.

The defect notification unit 140 provides an alert to the outside, for example, a designated terminal, or generates an alarm externally of the fact that a defect has occurred, in accordance with a control signal from the control unit 130. For this purpose, the defect notification unit 140 communicates with the designated terminal through wired or wireless communication. provides an alarm, displays an alarm message through the display unit 180 provided on the outside of the casing 230, consists of a speaker to output an alarm sound, or provides a warning by emitting light or flashing. It can be made up of units.

The photoresist application state inspection system 100 of a spin coater using deep learning according to an embodiment of the present invention includes an input unit 190 for inputting data or commands, and a power supply unit (not shown) for distributing power required for operation. ), a memory unit (not shown) for storing various data or programs, etc. may be provided, and after application of photoresist to the wafer, a probe test is performed for each die of the wafer in the subsequent process. A defect rate receiver 150 may be provided to receive the defect rate as data using wired or wireless communication from the wafer yield management system 300, which calculates the defect rate for each wafer die through data.

At this time, the deep learning defect determination unit 120 may modify the weight learned by the deep learning algorithm to be proportional to the defect rate and use it when determining defects in the photoresist. Here, the proportionality constants range from 0.1 to 10 and can be obtained experimentally to minimize the defect rate. Here, the wafer yield management system 300 receives probe test data inspected for each wafer die, and uses the input probe test data to calculate the yield for each wafer according to a set yield calculation method. , the identification information for each wafer and the corresponding yield are provided to the defect rate receiver 150.

The spin coater 10 may be provided with a nozzle opening/closing unit 212 for opening and closing the supply of photoresist through the supply nozzle 210, and has a nozzle arm 211 on which the supply nozzle 210 is fixed within the casing 230. A nozzle transfer unit 213 may be provided to load the onto the upper side of the spin chuck 220 or unload it from the upper side of the spin chuck 220, and the spin coater 10 may be provided with a nozzle opening and closing unit to perform a predetermined process. A main control unit 240 may be provided to control (212), the nozzle transfer unit 213, and the spin chuck 220.

According to the spin coater's photoresist application state inspection system using deep learning according to the present invention, type data on the state of the spin coater's supply nozzle using deep learning technology is quantified, and the photoresist application process is quantified. By improving the detection processing ability of the image status, the supply nozzle status can be monitored in real time, and if an abnormal defect condition is detected, an alarm is sounded in the relevant unit, thereby maximizing production efficiency by preventing process defects and enhancing yield.

In addition, according to the present invention, deep learning-based image analysis combines the sophistication and flexibility of visual inspection with the reliability, consistency, and speed of computer vision, enabling precise, repeatable performance of difficult vision applications that are nearly impossible to maintain with existing machine vision. It can distinguish unacceptable defects while tolerating natural variations in complex patterns, and has the advantage of being able to immediately adjust to new examples without reprogramming the core algorithm.

In addition, according to the present invention, deep learning-based technology classifies images based on the defect criteria of the photoresist supply nozzle state, and defect area classification and object detection use a convolutional neural network (CNN) to classify images. It is different from the machine vision method in that it can improve defect image processing capabilities and increase inspection reliability by using an image processing algorithm that creates a feature map to classify defective areas and detect objects in image data. will have

In this way, damage caused by thickness abnormalities and unevenness of the resist film due to supply defects from the photoresist nozzle of the photo-lithography process, which is the initial stage of the wafer manufacturing process, is prevented, resulting in economical savings through import substitution and cost reduction. effect can be achieved.

Although the present invention has been described with reference to the accompanying drawings, it goes without saying that various modifications and variations can be made without departing from the technical spirit of the present invention. Therefore, the scope of the present invention should not be limited to the described embodiments, but should be determined by the claims described below as well as equivalents to these claims.

110: PR image acquisition unit 120: Deep learning defect determination unit
130: Control unit 140: Defect notification unit
150: Defect rate receiver 160: Horizontal wafer image acquisition unit
170: vertical wafer image acquisition unit 180: display unit
190: input unit 210: supply nozzle
211: nozzle arm 212: nozzle opening and closing part
213: nozzle transfer unit 220: spin chuck
230: Casing 231: Control box
240: main control unit 250: defect termination unit
300: Wafer yield management system

Claims (5)

A system for inspecting the photoresist application state of a spin coater used in a photolithography process for manufacturing semiconductor devices,
a PR image acquisition unit that acquires an image of the photoresist supplied from the supply nozzle of the spin coater;
A defect occurs when the photoresist is applied from the supply nozzle twisted to the left or right by a deep learning algorithm from the image acquired from the PR image acquisition unit, and the photoresist hangs in the form of a water drop from the supply nozzle and then falls. A deep learning defect determination unit that determines defects that occur and defects that occur due to residues generated on the outer surface of the supply nozzle;
A control unit configured to output a control signal for generating an alarm when a photoresist defect occurs according to the determination result of the deep learning defect determination unit; and
It includes a defect notification unit that provides the fact that a defect has occurred to the outside according to a control signal from the control unit,
The deep learning defect determination unit,
A feature map is extracted from the image acquired from the PR image acquisition unit by Max pooling, an object area (foreground) and a non-object area (background) are distinguished from the feature map, and the feature map is converted into a CNN (Conv. .3) and again through CNN (Conv.1), connect it to a fully connected layer to classify the object area and non-object area, and circumscribe the object area and non-object area. Obtain a bounding-box, where the location of the bounding box is the center coordinate of a rectangle surrounding the object area, the size of the bounding box is displayed in width and height, and the display result and anchor box (anchor box) are calculated. The region is expressed using a ground-truth box and a ground-truth box, and the detection process involves an ROI (region of interesting) pooling layer, two fully connected layers, and the left side of the photoresist. Alternatively, the area shifted to the right, the droplet-shaped area of the photoresist, and the residue area are detected all at once, and the output of the ROI pooling layer passes through two fully connected layers to the left or right side of the photoresist. A system for inspecting the photoresist application state of a spin coater using deep learning, which is divided into the area that is turned to the right, the water droplet-shaped area of the photoresist, the residue area, and the background. In claim 1,
The PR image acquisition department,
It consists of a pair and is arranged at an angle of 90 degrees around the supply nozzle,
The control unit,
Deep learning is used to intuitively display photoresist supply defects by displaying the cross-section of the photoresist normally supplied from the supply nozzle and the cross-section of the actual photoresist calculated from the image acquired from the PR image acquisition unit to overlap. Photoresist application inspection system using a spin coater. In claim 1,
The control unit,
A system for inspecting the photoresist application state of a spin coater using deep learning, which controls to stop the operation of the spin coater when the defect in the photoresist occurs according to the determination result of the deep learning defect determination unit. In claim 2,
a horizontal wafer image acquisition unit installed in the spin coater to obtain an image of the photoresist supplied from the supply nozzle to the upper surface of the wafer on the spin chuck and spreading by centrifugal force from the edge of the wafer in a direction horizontal to the wafer; and
It further includes a vertical wafer image acquisition unit installed in the spin coater to obtain an image of the photoresist supplied from the supply nozzle to the upper surface of the wafer on the spin chuck and spreading by centrifugal force from above the wafer,
The deep learning defect determination unit,
Using the image of the photoresist acquired by the PR image acquisition unit, the separation amount of the photoresist supplied downward from the bottom of the supply nozzle with respect to an imaginary vertical line extending vertically downward from the bottom of the supply nozzle. Accordingly, defects in the photoresist are determined, and defects in the convex or concave part of the photoresist are determined using the image of the edge in the horizontal direction of the wafer obtained by the horizontal wafer image acquisition unit, Using the image of the upper surface of the wafer acquired by the vertical wafer image acquisition unit, defects in the photoresist are determined according to the ratio of the area having a color or shape that is different from the center on the wafer on which the photoresist is applied, A spin coater photoresist application condition inspection system using deep learning. The method according to any one of claims 1 to 4,
After the photoresist is applied to the wafer, a defect rate receiving unit is provided to receive the defect rate from a wafer yield management system that calculates the defect rate for each die of the wafer through probe test data inspected for each die of the wafer in a subsequent process. A system for inspecting the photoresist application state of a spin coater using deep learning, including:

Images