Film thickness estimation from machine learning based processing of substrate images

Abstract

To planarize a substrate surface to improve planarity for photolithography during fabrication of an integrated circuit.SOLUTION: In a method, a neural network is trained for use in a substrate thickness measurement system by obtaining ground truth thickness measurements of a top layer of a calibration substrate at multiple locations, each location being a defined position for a die being fabricated on the substrate. Multiple color images of the calibration substrate are obtained, each color image corresponding to a region for a die being fabricated on the substrate. A neural network is trained to convert color images of die regions from an in-line substrate imager into thickness measurements for the top layer in the die region. The training is performed using training data that includes the multiple color images and ground truth thickness measurements, where each color image is paired with a ground truth thickness measurement for the die region associated with the respective color image.SELECTED DRAWING: None

Description

[Technical Field]

[0001] The present disclosure relates to optical metrology, for example, to detecting layer thicknesses on substrates using machine learning approaches. [Background technology]

[0002] Integrated circuits are typically formed on substrates by sequentially depositing conductive, semiconductive, or insulating layers on a silicon wafer. Planarization of the substrate surface may be required to remove filler layers or to improve flatness for photolithography during integrated circuit fabrication.

[0003] Chemical mechanical polishing (CMP) is an accepted method of planarization. This planarization method typically requires the substrate to be mounted on a carrier or polishing head. The exposed surface of the substrate is typically pressed against a rotating polishing pad. The carrier head applies a controllable load to the substrate, forcing the substrate against the polishing pad. A polishing slurry is typically applied to the surface of the polishing pad.

[0004] Various optical metrology systems, such as spectroscopy or ellipsometry, can be used, for example, in-line or in a stand-alone metrology station, to measure the thickness of the substrate layer before and after polishing.

[0005] In parallel, advances in hardware resources such as graphics processing units (GPUs) and tensor processing units (TPUs) have led to dramatic improvements in deep learning algorithms and their applications. One area where deep learning is developing is computer vision and image recognition. Such computer vision algorithms are primarily designed for image classification or segmentation. Summary of the Invention

[0006] In one aspect, a method for training a neural network for use in a substrate thickness measurement system includes acquiring ground truth thickness measurements of a top layer of a calibration substrate at multiple locations, each location being a defined position of a die fabricated on the substrate. Multiple color images of the calibration substrate are acquired, each color image corresponding to an area of ​​a die fabricated on the substrate. The neural network is trained to convert the color images of the die areas from an in-line substrate imager into thickness measurements of the top layer at the die areas. The training is performed using training data including the multiple color images and ground truth thickness measurements, each respective color image paired with a ground truth thickness measurement for the die area associated with the respective color image.

[0007] In another aspect, a method for controlling polishing includes, in an in-line monitor station of a polishing system, acquiring a first color image of a first substrate, dividing the first color image into multiple second color images using a die mask, each second color image corresponding to a region of a die fabricated on the first substrate, generating thickness measurements for one or more locations, and determining polishing parameters for the first substrate or a subsequent second substrate based on the thickness measurements. Each of the one or more locations corresponds to a respective region of a die fabricated on the first substrate. To generate thickness measurements for the regions, the second color image corresponding to the regions is processed through a neural network trained with training data including multiple third color images of dies of a calibration substrate and ground truth thickness measurements for the calibration substrate, and each of the third color images is paired with a ground truth thickness measurement for the die region associated with the respective third color image.

[0008] Implementations may include one or more of the following potential advantages: The thickness of multiple dies on a substrate can be measured quickly. For example, an in-line metrology system can determine the thickness of a substrate based on a color image of the substrate without impacting throughput. The estimated thickness can be directly used in a multivariate run-to-run control scheme.

[0009] Using the described approach, a model can be trained to generate thickness measurements with an error of less than 5% of the actual film thickness. While thickness measurements can be extracted from color images with three color channels, a hyperspectral camera can be added to the substrate imager system to provide higher-dimensional feature input to the model. This can facilitate the training of more complex models to understand more physical properties of the film stack.

[0010] Deep learning in metrology systems can achieve high inference speed and high-resolution measurements of thickness profiles on substrates, making the metrology systems a fast and low-cost pre- and post-metrology measurement tool with high thickness accuracy for memory applications.

[0011] The details of one or more implementations are set forth in the accompanying drawings and the description below. Other aspects, features, and advantages will become apparent from the description and drawings, and from the claims. [Brief explanation of the drawings]

[0012] [Figure 1] FIG. 1 illustrates an example of an in-line light measurement system. [Figure 2A] FIG. 1 shows an example of an exemplary image of a substrate used to train the model. [Figure 2B] FIG. 1 is a schematic diagram of a computer data storage system. [Figure 3] FIG. 1 illustrates a neural network used as part of a controller for a polishing apparatus. [Figure 4]FIG. 1 is a flow diagram illustrating a method for detecting layer thicknesses on a substrate using a deep learning approach. DETAILED DESCRIPTION OF THE INVENTION

[0013] Like numbers in the drawings refer to like elements.

[0014] Dry metrology systems are used to measure thin film thickness in CMP processes due to variations in polishing rates during CMP. These dry metrology techniques often use spectroscopy or ellipsometry, which fit variables from an optical model of the film stack to collected measurements. These metrology techniques typically require precise alignment of the sensor to the measurement spot on the substrate so that the model can be applied to the collected measurements. As a result, measuring a large number of points on the substrate can be time-consuming, and collecting high-resolution thickness profiles can be impractical.

[0015] However, machine learning may enable rapid measurement of film thickness on substrates. A deep neural network can be trained using color images of the die on the substrate and associated thickness measurements from other reliable metrology systems, and the die film thickness can be measured by applying the input image to the neural network. This system can be used as a high-throughput, economical solution for, for example, low-cost memory applications. In addition to film thickness estimation, this technique can also be used to classify residue levels on substrates using image segmentation.

[0016] 1, a polishing apparatus 100 includes one or more carrier heads 126, each configured to carry a substrate 10, one or more polishing stations 106, and a transfer station for loading and unloading substrates from the carrier heads. Each polishing station 106 includes a polishing pad 130 supported on a platen 120. The polishing pad 130 may be a two-layer polishing pad having an outer polishing layer and a softer backing layer.

[0017] The carrier heads 126 are suspended from a support 128 and are movable between the polishing stations. In some implementations, the support 128 is an overhead track, and each carrier head 126 is coupled to a carriage 108 mounted on the track such that each carriage 108 can selectively move between the polishing stations 124 and the transfer station. Alternatively, in some implementations, the support 128 is a rotatable carousel, and rotation of the carousel simultaneously moves the carrier heads 126 along a circular path.

[0018] Each polishing station 106 of the polishing apparatus 100 may include a port, for example, at the end of the arm 134, for dispensing a polishing fluid 136, such as a polishing slurry, onto the polishing pad 130. Each polishing station 106 of the polishing apparatus 100 may also include a pad conditioning device for polishing the polishing pad 130 and maintaining the polishing pad 130 in a consistent polishing condition.

[0019] Each carrier head 126 is operable to hold a substrate 10 against a polishing pad 130. Each carrier head 126 may have independent control of polishing parameters, such as pressure, associated with each respective substrate. In particular, each carrier head 126 may include a retaining ring 142 for holding the substrate 10 beneath a flexible membrane 144. Each carrier head 126 may include multiple independently controllable pressurizable chambers defined by the membrane, e.g., three chambers 146a-146c, which may apply independently controllable pressures to associated zones on the flexible membrane 144 and, therefore, on the substrate 10. While only three chambers are shown in FIG. 1 for ease of illustration, there may be one or two chambers, or four or more chambers, e.g., five chambers.

[0020] Each carrier head 126 is suspended from a support 128 and connected by a drive shaft 154 to a carrier head rotation motor 156 so that the carrier head can rotate about axis 127. Optionally, each carrier head 126 can be oscillated laterally, for example, by driving carriage 108 on a track or by rotational oscillation of the carousel itself. In operation, the platen rotates about its central axis, and each carrier head rotates about its central axis 127 and translates laterally across the upper surface of the polishing pad.

[0021] A controller 190, such as a programmable computer, is connected to each motor and independently controls the rotational speed of the platen 120 and carrier head 126. The controller 190 may include a central processing unit (CPU) 192, memory 194, and support circuits 196, such as input / output circuits, power supplies, clock circuits, cache, etc. The memory is connected to the CPU 192. The memory is a non-transitory, computable, readable medium and may be one or more readily available memories, such as random access memory (RAM), read-only memory (ROM), a floppy disk, a hard disk, or other form of digital storage. Furthermore, although illustrated as a single computer, the controller 190 may be a distributed system, including, for example, multiple processors and memories operating independently.

[0022] Polishing apparatus 100 also includes an in-line (also called in-sequence) optical metrology system 160. The color imaging system of in-line optical metrology system 160 is positioned within polishing apparatus 100, but does not take measurements during the polishing process; rather, measurements are collected between polishing processes, e.g., while a substrate is being moved from one polishing station to another, or before or after polishing, e.g., while a substrate is being moved from a transfer station to a polishing station or vice versa. Furthermore, in-line optical metrology system 160 may be positioned in a fab interface unit or a module accessible from the fab interface unit to measure a substrate after it has been removed from a cassette and before it is moved to a polishing unit, or after it has been cleaned and before it is returned to the cassette.

[0023] The in-line optical metrology system 160 includes a sensor assembly 161 that provides color imaging of the substrate 10. The sensor assembly 161 may include a light source 162, a light detector 164, and circuitry 166 for sending and receiving signals between the controller 190, the light source 162, and the light detector 164.

[0024] The light source 162 may be operable to emit white light. In one implementation, the emitted white light comprises light having a wavelength between 200 and 800 nanometers. A suitable light source is an array of white light emitting diodes (LEDs), or a xenon or xenon mercury lamp. The light source 162 is oriented to direct light 168 onto the exposed surface of the substrate 10 at a non-zero angle of incidence α. The angle of incidence α may be, for example, between about 30° and 75°, e.g., 50°.

[0025] The light source can illuminate a substantially linear, elongated region spanning the width of the substrate 10. For example, the light source 162 can include optics, such as a beam expander, to spread the light from the light source into the elongated region. Alternatively or additionally, the light source 162 can include a linear array of light sources. The light source 162 itself, and the illuminated region on the substrate, can be elongated and have a longitudinal axis parallel to the surface of the substrate.

[0026] Placing a diffuser 170 in the path of the light 168 or including a diffuser in the light source 162 can diffuse the light before it reaches the substrate 10 .

[0027] Detector 164 is a color camera sensitive to light from light source 162. The camera includes an array of detector elements. For example, the camera may include a CCD array. In some implementations, the array is a row of detector elements. For example, the camera may be a line scan camera. The row of detector elements may extend parallel to the longitudinal axis of the elongated region illuminated by light source 162. If light source 162 includes a row of light emitting elements, the row of detector elements may extend along a first axis parallel to the longitudinal axis of light source 162. The row of detector elements may include 1024 or more elements.

[0028] The camera 164, configured with appropriate focusing optics 172, projects a field of view of the substrate onto an array of detector elements. The field of view may be long enough to view the entire width of the substrate 10, e.g., 150-300 mm long. The camera 164, including its associated optics 172, may be configured so that individual pixels correspond to an area having a length of approximately 0.5 mm or less. For example, assuming the field of view is approximately 200 mm long and the detector 164 contains 1024 elements, an image produced by a line scan camera may have pixels having a length of approximately 0.5 mm. To determine the linear resolution of an image, the length of the field of view (FOV) can be divided by the number of pixels the FOV is imaged to obtain the linear resolution.

[0029] The camera 164 can also be configured so that the pixel width is comparable to the pixel length. For example, an advantage of a line scan camera is its very fast frame rate. The frame rate can be at least 5 kHz. The frame rate can be set at a frequency such that the pixel width is comparable to the pixel length as the imaged area is scanned across the substrate 10, which can be, for example, about 0.3 mm or less.

[0030] The light source 162 and the light detector 164 may be supported on a stage 180. If the light detector 164 is a line scan camera, the light source 162 and the camera 164 may be movable relative to the substrate 10 so that the imaged area can be scanned across the length of the substrate. In particular, the relative motion may be in a direction parallel to the surface of the substrate 10 and perpendicular to the rows of detector elements of the line scan camera 164.

[0031] In some implementations, the stage 182 is stationary and the substrate support moves. For example, the carrier head 126 can move, e.g., by movement of the carriage 108 or rotational oscillation of a carousel, or a robotic arm holding the substrate at a factory interface unit can move the substrate 10 past the line scan camera 182. In some implementations, the stage 180 is movable, but the carrier head or robotic arm remains stationary for image acquisition. For example, the stage 180 can be movable along rails 184 by a linear actuator 182. In either case, this allows the light source 162 and camera 164 to remain in fixed positions relative to each other as the area to be scanned moves across the substrate 10.

[0032] A possible advantage of having a line scan camera and light source that move together across the substrate is that the relative angle between the light source and camera remains constant at different positions across the wafer, compared to, for example, a traditional 2D camera. As a result, artifacts due to variations in the viewing angle can be reduced or eliminated. Furthermore, perspective distortion, inherent in traditional 2D cameras and which must be corrected by image transformation, is eliminated with a line scan camera.

[0033] The sensor assembly 161 may include a mechanism for adjusting the vertical distance between the substrate 10, the light source 162, and the detector 164. For example, the sensor assembly 161 may include an actuator for adjusting the vertical position of the stage 180.

[0034] Optionally, a polarizing filter 174 may be positioned in the path of the light, for example, between the substrate 10 and the detector 164. The polarizing filter 174 may be a circular polarizer (CPL). A typical CPL is a combination of a linear polarizer and a quarter-wave plate. Proper orientation of the polarization axis of the polarizing filter 174 can reduce haze in the image and sharpen or emphasize desired visual features.

[0035] Assuming the outermost layer on the substrate is a semi-transparent layer, e.g., a dielectric layer, the color of the light detected by detector 164 depends, for example, on the composition of the substrate surface, the smoothness of the substrate surface, and / or the amount of interference between light reflecting from different interfaces of one or more layers (e.g., dielectric layers) on the substrate. As mentioned above, light source 162 and light detector 164 may be connected to a computing device, e.g., controller 190, operable to control their operation and receive their signals. The computing device that performs the various functions for converting the color image into a thickness measurement may be considered part of metrology system 160.

[0036] 2A, an example of an image 202 of a substrate 10 collected by an in-line optical metrology system 160 is shown. The in-line optical metrology system 160 generates a high-resolution color image 202, e.g., an image of at least 720x1080 pixels, e.g., an image of at least 2048x2048 pixels, having at least three color channels (e.g., RGB channels). The color at any particular pixel depends on the thickness of one or more layers, including the top layer, in the area of ​​the substrate corresponding to that pixel.

[0037] The image 202 is divided into one or more regions 208, each corresponding to a die 206 being fabricated on the substrate. The portion of the image providing the region 208 may be a predetermined area in the image, or the portion providing the region 208 may be automatically determined by an algorithm based on the image.

[0038] As an example of a predetermined area in an image, the controller 190 can store a die mask that identifies the location and area in the image for each region 208. For example, for a rectangular region, the area may be defined by a top right coordinate and a bottom left coordinate in the image. Thus, the mask may be a data file that includes pairs of top right and bottom left coordinates for each rectangular region. For other, non-rectangular regions, more complex functions may be used.

[0039] In some implementations, the orientation and position of the substrate can be determined, and the die mask can be aligned with respect to the image. The orientation of the substrate can be determined by a notch finder or by image processing of the color image 202, for example, by determining the angle of the scribe line in the image. The substrate position can also be determined by image processing of the color image 202, for example, by detecting a circular substrate edge and then determining the center of the circle.

[0040] As an example of automatically determining region 208, an image processing algorithm can analyze image 202 and detect scribe lines. Image 202 can then be segmented into regions between the identified scribe lines.

[0041] By segmenting the initial color image, multiple color images 204 of individual regions 208 can be collected from the substrate 10. As described above, each color image 204 corresponds to a die 206 being fabricated on the substrate. The collected color images can be exported as PNG images, although many other formats are possible, such as JPEG.

[0042] The color image 204 can be fed to an image processing algorithm to generate thickness measurements of the die shown in the color image 204. The image can be used as input data to an image processing algorithm trained, for example, by a supervised deep learning approach, to estimate layer thicknesses based on the color image. The supervised deep learning-based algorithm establishes a model between the color image and the thickness measurements. The image processing algorithm, as a deep learning-based algorithm, can include a neural network.

[0043] The intensity values ​​of each color channel of each pixel in the color image 204 are input to an image processing algorithm, e.g., the input neurons of a neural network. Based on this input data, a layer thickness measurement for the color image is calculated. Thus, inputting the color image 204 into the image processing algorithm results in an output of an estimated thickness. This system can be used as a high-throughput, economical solution, for example, in low-cost memory applications. Aside from thickness inference, this technique can also be used to classify the level of residue on a substrate using image segmentation.

[0044] Using a supervised deep learning approach, calibration images of dies may be acquired from one or more calibration substrates, as described above, to train an image processing algorithm, e.g., a neural network. That is, each calibration substrate may be scanned by a line scan camera of inline optical metrology system 160 to generate an initial calibration image, which may be segmented into multiple color images of individual regions on the calibration substrate.

[0045] Either before or after the initial color calibration image is collected, ground truth thickness measurements are collected at multiple locations on the calibration substrate using a high-precision metrology system, for example, an in-line or stand-alone metrology system. The high-precision metrology system may be a dry optical metrology system. The ground truth measurements may be obtained from offline reflectometry, ellipsometry, scatterometry, or more advanced TEM measurements, although other techniques may be suitable. Such systems are available from Nova Measuring Instruments or Nanometrics. Each location corresponds to one of the die, i.e., one of the individual regions, being manufactured.

[0046] 2B , for each individual region on each calibration substrate, a color calibration image 212 is collected with an in-line sensor in optical metrology system 160. Each color calibration image is associated with a ground truth thickness measurement 214 for a corresponding die on the calibration substrate. The images 212 and the associated ground truth thickness measurements 214 may be stored in database 220. For example, the data may be stored as records 210, with each record including a calibration image 212 and a ground truth thickness measurement 214.

[0047] A deep learning-based algorithm, e.g., a neural network, is then trained using the combined data set 218. Thickness measurements corresponding to the die centers measured from the dry metrology tool are used as labels for input images while training the model. For example, the model may be trained on approximately 50,000 images collected from five dies on substrates with a wide range of post thicknesses.

[0048] 3 illustrates a neural network 320 used as part of the controller 190 of the polishing apparatus 100. The neural network 320 is a deep neural network developed for regression analysis of RGB intensity values ​​of input images from calibration substrates with ground truth thickness measurements, and is capable of generating a model that predicts the layer thickness of an area of ​​a substrate based on a color image of that area.

[0049] Neural network 320 includes a plurality of input nodes 322. Neural network 320 may include an input node for each color channel associated with each pixel of the input color image, a plurality of hidden nodes 324 (hereinafter also referred to as "middle nodes"), and an output node 326 that generates layer thickness measurements. In a neural network with a single layer of hidden nodes, each hidden node 324 may be coupled to a respective input node 322, and output node 326 may be coupled to a respective hidden node 320. However, in practice, neural networks for image processing are likely to have many layers of hidden nodes 324.

[0050] In general, hidden nodes 324 output values ​​that are nonlinear functions of weighted sums of values ​​from the previous layer of input nodes 322 or hidden nodes to which hidden node 324 is connected.

[0051] For example, the output of a first layer hidden node 324 designated node k may be expressed as follows: tanh(0.5*a k1 (I1)+a k2 (I2)+...+a kM (I M )+b k ) In the above formula, tanh is the hyperbolic tangent, a kx is the weight for the connection between the kth hidden node and the xth input node (out of M input nodes), I M is the value at the Mth input node. However, instead of tanh, other nonlinear functions such as the rectified linear unit (ReLU) function or its variants can also be used.

[0052] Thus, neural network 320 includes an input node 322 for each color channel associated with each pixel in the input color image; for example, if there are J pixels and K color channels, where L=J*K is the number of intensity values ​​in the input color image, neural network 320 includes at least input nodes N1, N2...N L This will include:

[0053] Thus, if the number of input nodes corresponds to the number of intensity values ​​in the color image, then the output H of hidden node 324 designated as node k k can be expressed as follows: H k =tanh(0.5*a k1 (I1)+a k2 (I2)+...+a kL (I L )+b k )

[0054] The measured color image S is a column matrix (i1,i2,...,i L ), the output of an intermediate node 324, designated node k, may be expressed as: H k =tanh(0.5*a k1 (V1 S)+a k2 (V2·S)+...+a kL (V L S)+b k ) In the above formula, V is the weight value (v1, v2,...,v L ) and V x is the weight for the xth intensity value among the L intensity values ​​from the color image).

[0055] The output node 326 may generate a characteristic value CV, e.g., thickness, that is a weighted sum of the outputs of the hidden nodes. For example, this may be expressed as: CV=C1*H1+C2*H2+...+C L *H L In the above formula, C kis the weight for the output of the kth hidden node.

[0056] However, the neural network 320 may optionally include one or more other input nodes, such as node 322a, for receiving other data. This other data may be from a previous measurement of the substrate by an in-situ monitoring system, such as pixel intensity values ​​collected from an earlier stage in the processing of the substrate; from a measurement of a previous substrate, such as pixel intensity values ​​collected during the processing of another substrate; from another sensor in the polishing system, such as a measurement of the pad or substrate temperature by a temperature sensor; from a polishing recipe stored by a controller used to control the polishing system, such as a polishing parameter, such as the carrier head pressure or platen rotation speed, used to polish the substrate; from a variable tracked by the controller, such as the number of substrates since the pad was changed; or from a sensor not part of the polishing system, such as a measurement of the thickness of an underlying film by a metrology station. This allows the neural network 320 to take other process or environmental variables into account in calculating the layer thickness measurement.

[0057] The thickness measurements generated at output node 326 are provided to a process control module 330. The process control module may adjust process parameters, such as carrier head pressure, platen rotation speed, etc., based on the thickness measurements of one or more regions. Adjustments may be made for a polishing process performed on the substrate or a subsequent substrate.

[0058] Before it can be used, for example, to measure a substrate, the neural network 320 must be configured.

[0059] As part of the configuration procedure, the controller 190 may receive multiple calibration images. Each calibration image includes multiple intensity values, e.g., intensity values ​​for each color channel, for each pixel in the calibration image. The controller also receives characteristic values, e.g., thickness, for each calibration image. For example, color calibration images may be measured at specific dies fabricated on one or more calibration or test substrates. Additionally, ground truth thickness measurements at specific die locations may be performed using a dry metrology instrument, e.g., a contact profilometer or ellipsometer. Thus, ground truth thickness measurements may be correlated with color images from the same die locations on the substrate. Multiple color calibration images may be generated from, for example, five to ten calibration substrates by segmenting the calibration substrate images as described above. As part of the configuration procedure for the neural network 320, the neural network 320 is trained using the color images and characteristic values ​​of each die fabricated on the calibration substrate.

[0060] V corresponds to one of the color images and is therefore associated with a characteristic value. While the neural network 320 is operating in a training mode, such as backpropagation mode, the values ​​(v1, v2, ..., v L ) for each input node N1,N2...N L and the characteristic value CV is provided to output node 326. This can be repeated for each row. This process allows a k1 Values ​​such as:

[0061] The system is now ready for processing. Using the in-line monitor system 160, a color image is measured from the substrate. The measured color image is represented by a column matrix S=(i1,i2,...,i L ), and i j represents the intensity value at the jth intensity value among the L intensity values, where L=3n when the image contains a total of n pixels and each pixel contains three color channels.

[0062] While the neural network 320 is being used in inference mode, these values ​​(S1, S2, ..., S L ) for each input node N1,N2,...N L As a result, neural network 320 generates a property value, e.g., layer thickness, at output node 326.

[0063] The architecture of neural network 320 can vary in depth and width. For example, while neural network 320 is shown as having a single row of hidden nodes 324, it may include multiple rows. The number of hidden nodes 324 may be equal to or greater than the number of input nodes 322.

[0064] As described above, the controller 190 can associate various color images with different dies on the substrate (see FIG. 2). The output of each neural network 320 can be classified as belonging to one of the dies based on the position of the sensor on the substrate when the image was collected. This allows the controller 190 to generate a separate sequence of measurements for each die.

[0065] In some implementations, the controller 190 may be configured to have a neural network model structure composed of multiple different types of building blocks. For example, the neural network may be a residual neural network that includes a res-block function in its architecture. The residual neural network may utilize skip connections, or shortcuts, to skip over some layers. The residual neural network may be implemented, for example, with a ResNet model. In the context of residual neural networks, non-residual networks are sometimes referred to as plain networks.

[0066] In some implementations, the neural network can be trained to consider the thickness of the underlying layer from the stack during calculation, which can improve errors due to variations in the underlying layer in the thickness measurement. The effect of variations in the underlying layer thickness in the film stack can be mitigated by providing the intensity values ​​of the color image of the underlying layer thickness as an additional input to the model, improving the model's performance.

[0067] The reliability of the calculated thickness measurements can be assessed by comparing them to measured values ​​and then determining the difference between the calculated and original measurements. This deep learning model can then be used to predict thickness in inference mode immediately after scanning a new test substrate. This new approach increases overall system throughput, enabling thickness measurements to be performed on every substrate in a lot.

[0068] Referring to Figure 4, an image processing algorithm generated by machine learning techniques for use in a substrate thickness measurement system is shown. Such an image processing algorithm can receive RGB images collected from an integrated line-scan camera inspection system and enable much faster film thickness estimation. The inference time for approximately 2000 measurement points is approximately a few seconds, compared to two hours for dry measurements.

[0069] The method includes a controller assembling (500) individual image lines from the photodetectors 164 into a two-dimensional color image. The controller may apply offset and / or gain adjustments to the intensity values ​​of the image in each color channel (510). Each color channel may have a different offset and / or gain. Optionally, the image may be normalized (515). For example, the difference between the measured image and a standard predefined image may be calculated. For example, the controller may store a background image for each red, green, and blue color channel, and the background image may be subtracted from the measured image for each color channel. Alternatively, the measured image may be divided by the standard predefined image. The image may be filtered (530) to remove low-frequency spatial variations. In some implementations, a filter is generated using the luminance channel, and this filter is applied to the red, green, and blue images.

[0070] The image is transformed into a standard image coordinate frame, e.g., scaled and / or rotated and / or translated 540. For example, the image can be translated so that the die center is at the center point of the image, and / or the image can be scaled so that the edge of the substrate is at the edge of the image, and / or the image can be rotated so that there is a 0° angle between the X axis of the image and the radial section connecting the substrate center and the substrate orientation feature.

[0071] One or more regions on the substrate are selected and an image is generated for each selected region 550. This can be done using the techniques described above, for example the regions can be predetermined regions or the portions providing region 208 can be automatically determined by an algorithm.

[0072] The intensity values ​​provided by each color channel for each pixel of the image may be used as input to an image processing algorithm trained with supervised deep learning, which outputs layer thickness measurements for specific regions (560).

[0073] To reduce measurement error, we trained and validated various deep model architectures using small die test pattern boards. Models that took into account the characteristics of the underlying layers had smaller errors. Furthermore, we performed cross-tool matching validation in advance by training a model on data collected by one tool and using it for inference on data from another tool. The results were comparable to those obtained when training and inferring on data from the same tool.

[0074] Typically, the data can be used to control one or more operating parameters of the CMP tool, including, for example, platen rotation speed, substrate rotation speed, substrate polishing path, substrate velocity across the plate, pressure acting on the substrate, slurry composition, slurry flow rate, and substrate surface temperature. The operating parameters can be controlled in real time and can be automatically adjusted without human intervention.

[0075] The term substrate as used herein may include, for example, a product substrate (e.g., containing multiple memory or processor dies), a test substrate, a bare substrate, and a gating substrate. The substrate may be at various stages of integrated circuit manufacturing; for example, the substrate may be a bare wafer or may include one or more deposited and / or patterned layers. The term substrate may include a circular disk and a rectangular sheet.

[0076] However, the color imaging techniques described above may be particularly useful in the context of 3D vertical NAND (VNAND) flash memory. In particular, the layer stack used in VNAND fabrication is so complex that current metrology methods (e.g., Nova spectral analysis) may not perform reliably enough to detect areas of improper thickness. In contrast, color imaging techniques may have superior reliability in this application.

[0077] Embodiments of the invention and all functional operations described herein can be implemented in digital electronic circuitry, or computer software, firmware, or hardware, including the structural means disclosed herein and their structural equivalents, or in combinations of them.Embodiments of the invention can also be implemented as one or more computer program products, i.e., one or more computer programs tangibly embodied in a non-transitory machine-readable storage medium, for execution by or to control the operation of a data processing apparatus, such as a programmable processor, a computer, or multiple processors or computers.

[0078] It should be understood that the term relative positioning is used to indicate the positioning of the system components relative to one another, not necessarily with respect to gravity, and that the polishing surface and substrate may be held in a vertical orientation or some other orientation.

[0079] A number of implementations have been described. Nevertheless, it will be understood that various modifications may be made. For example, Instead of a line scan camera, a camera that captures the entire board can be used, eliminating the need to move the camera relative to the board. The camera can cover an area less than the full width of the board, in which case the camera must be supported on an XY stage and moved in two perpendicular directions to scan the entire board. The light source can illuminate the entire substrate, in which case the light source does not need to move relative to the substrate. The photodetector can be a spectrometer instead of a color camera, and its spectral data can be reduced to the RGB color space. The sensor assembly need not be an in-line system positioned between polishing stations or between a polishing station and a transfer station. For example, the sensor assembly can be positioned within the transfer station, in the cassette interface unit, or can be a stand-alone system. The uniformity analysis step is optional. For example, the image generated by applying the threshold transform can be fed into a feed-forward process to adjust subsequent processing steps for the substrate, or into a feedback process to adjust processing steps for subsequent substrates.

[0080] Accordingly, other implementations are within the scope of the following claims.

Claims

1. 1. A non-transitory computer-readable medium encoded with a computer program product, the computer program product causing one or more processors to: obtaining ground truth thickness measurements of a top layer of a calibration substrate at a plurality of locations, each location being a defined location of a die being fabricated on the substrate; acquiring a plurality of color images of the calibration substrate, each color image corresponding to an area of ​​a die fabricated on the substrate; training a neural network to convert a color image of a die area from an in-line substrate imager into a thickness measurement of the top layer in the die area; including instructions to: A non-transitory computer-readable medium, the instructions for training the neural network comprising instructions for using training data including the plurality of color images and ground truth thickness measurements, each respective color image paired with a ground truth thickness measurement of the die region associated with the respective color image.

2. The computer-readable medium of claim 1 , wherein the instructions for acquiring the plurality of color images include instructions for receiving a can of the calibration substrate from the in-line substrate imager.

3. 2. The computer-readable medium of claim 1, wherein the instructions for acquiring the plurality of color images comprise instructions for receiving a color image of the calibration substrate and instructions for dividing the color image into the plurality of color images based on a die mask.

4. The computer-readable medium of claim 1 , comprising instructions for obtaining ground truth thickness measurements of a top layer of a plurality of calibration substrates and instructions for obtaining a plurality of color images of each of the calibration substrates.

5. The computer-readable medium of claim 1 , wherein the instructions for obtaining ground truth thickness measurements comprise instructions for receiving thickness measurements at each of the plurality of locations from an optical profilometer.

6. obtaining ground truth thickness measurements of an underlying layer of the calibration substrate at the plurality of locations; training the neural network to convert color images of die areas from the in-line substrate imager into thickness measurements of the underlying layers; The computer-readable medium of claim 1 comprising instructions to:

7. The computer-readable medium of claim 1 , wherein the defined location is a center of the die.

8. The computer-readable medium of claim 1 comprising instructions for obtaining measurements from all dies of a substrate.

9. The computer-readable medium of claim 1 comprising instructions for obtaining measurements from all substrates of a lot both before and after chemical mechanical planarization.

10. 1. A method of training a neural network for use in a substrate thickness measurement system, comprising: obtaining ground truth thickness measurements of a top layer of a calibration substrate at a plurality of locations, each location being a defined location of a die being fabricated on the substrate; acquiring a plurality of color images of the calibration substrate, each color image corresponding to an area of ​​a die fabricated on the substrate; training a neural network to convert color images of die regions from an in-line substrate imager to thickness measurements of the top layer at the die regions, the training being performed using training data including the plurality of color images and ground truth thickness measurements, each respective color image being paired with a ground truth thickness measurement of the die region associated with the respective color image; A method comprising:

11. 1. A computer program product comprising: receiving a first color image of a first substrate from an in-line monitor station of the polishing system; dividing the first color image into a plurality of second color images using a die mask, each second color image corresponding to an area of ​​a die being fabricated on the first substrate; generating thickness measurements for one or more locations, each respective location of the one or more locations corresponding to a respective region of a die fabricated on the first substrate, wherein the instructions to generate the thickness measurements for the regions include instructions to process second color images corresponding to the regions through a neural network trained with training data including a plurality of third color images of dies of a calibration substrate and ground truth thickness measurements for the calibration substrate, generating thickness measurements for the one or more locations, each respective third color image paired with a ground truth thickness measurement for a die region associated with the respective third color image; determining values ​​of polishing parameters for the first substrate or a subsequent second substrate based on the thickness measurements; 1. A computer program product comprising a non-transitory computer-readable medium encoded with instructions to cause a

12. 12. The computer program product of claim 11, comprising instructions for receiving the first color image of the first substrate after polishing the first substrate at a polishing station.

13. 13. The computer program product of claim 12, comprising instructions for determining the polishing parameters of the polishing station for the subsequent second substrate based on the thickness measurements.

14. 12. The computer program product of claim 11, comprising instructions for receiving the first color image of the first substrate before polishing the first substrate at a polishing station.

15. 15. The computer program product of claim 14, comprising instructions for determining the polishing parameters of the polishing station for the first substrate based on the thickness measurements.

16. The computer program product of claim 11 , wherein the polishing parameters include a chamber pressure of a carrier head.

17. A polishing apparatus comprising: a polishing station including a platen supporting a polishing pad and a carrier head holding a first substrate against the polishing pad; an in-line metrology station having a color camera that generates a color image of the first substrate; 1. A control system comprising: receiving a first color image of the first substrate from an in-line monitor station of a polishing system; dividing the first color image into a plurality of second color images using a die mask, each second color image corresponding to an area of ​​a die being fabricated on the first substrate; generating thickness measurements for one or more locations, each respective location of the one or more locations corresponding to a respective region of a die fabricated on the first substrate, wherein the instructions to generate the thickness measurements for the regions include instructions to process second color images corresponding to the regions through a neural network trained with training data including a plurality of third color images of dies of a calibration substrate and ground truth thickness measurements for the calibration substrate, wherein each respective third color image generates thickness measurements for one or more locations paired with a ground truth thickness measurement for a die region associated with the respective third color image; determining values ​​of polishing parameters for the first substrate or a subsequent second substrate based on the thickness measurements; causing the polishing station to polish the first substrate or the subsequent second substrate using the determined polishing parameters; The control system is configured as follows: An apparatus comprising:

18. The apparatus of claim 17 , wherein the polishing parameters include a chamber pressure of the carrier head.

19. 20. The apparatus of claim 17, wherein the control system is configured to receive the first color image of the first substrate after polishing of the first substrate at a polishing station.

20. 1. A method for controlling polishing, comprising: acquiring a first color image of a first substrate at an in-line monitor station of the polishing system; dividing the first color image into a plurality of second color images using a die mask, each second color image corresponding to an area of ​​a die being fabricated on the first substrate; generating thickness measurements for one or more locations, each respective location of the one or more locations corresponding to a respective region of a die fabricated on the first substrate, wherein generating thickness measurements for the regions includes processing second color images corresponding to the regions through a neural network trained with training data including a plurality of third color images of dies of a calibration substrate and ground truth thickness measurements for the calibration substrate, wherein each respective third color image is paired with a ground truth thickness measurement for a die region associated with the respective third color image; determining values ​​of polishing parameters for the first substrate or a subsequent second substrate based on the thickness measurements; A method comprising:

Images