Film thickness estimation from machine learning-based substrate image processing

Abstract

A neural network is trained for a substrate thickness measurement system by obtaining surface live thickness measurements of a top layer of a calibration substrate at a plurality of locations, each location being a defined location of a die fabricated on the substrate. A plurality of color images of a calibration substrate are acquired, each color image corresponding to a region of a die fabricated on the substrate. A neural network is trained to convert a color image from a die region of an in-line substrate imager into a thickness measurement of a top layer in the die region. Training is performed using training data, the training data comprising a plurality of color images and surface live thickness measurements, each respective color image being paired with a surface live thickness measurement for a die region associated with the respective color image.

Description

[0001] This application is a divisional application of application number 202180014115.4, filed on June 25, 2021, entitled “Film Thickness Estimation from Machine Learning Based Substrate Image Processing”. TECHNICAL FIELD

[0002] The present disclosure relates to optical measurements, for example to detecting thickness of a layer on a substrate using machine learning methods. BACKGROUND

[0003] Integrated circuits are typically formed on a substrate by sequentially depositing conductive, semiconductive, or insulating layers on a silicon wafer. During integrated circuit fabrication, planarization of the substrate surface can be required to remove a fill layer or to improve lithographic planarity.

[0004] Chemical mechanical polishing (CMP) is an accepted planarization method. This planarization method typically requires mounting the substrate on a carrier or polishing head. The exposed surface of the substrate is typically placed against a rotating polishing pad. The carrier head provides a controllable load to the substrate to push the substrate against the polishing pad. A polishing slurry is typically supplied to the surface of the polishing pad.

[0005] A variety of optical metrology systems, such as spectroscopic or ellipsometric, can be used to measure the thickness of pre-polish and post-polish substrate layers, for example, in an in-line or standalone metrology station.

[0006] As a parallel problem, the development of hardware resources such as graphics processing units (GPUs) and tensor processing units (TPUs) has led to tremendous progress in deep learning algorithms and their applications. One of the growing areas of deep learning is computer vision and image recognition. Such computer vision algorithms are primarily designed for image classification or segmentation. SUMMARY

[0007] In one aspect, a method of training a neural network for a substrate thickness measurement system includes 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 fabricated on the substrate. A plurality of color images of the calibration substrate is obtained, each color image corresponding to a region of a die fabricated on the substrate. The neural network is trained to convert color images of a die region from an in-line substrate imager to thickness measurements of the top layer in the die region. The training is performed using training data, the training data including the plurality of color images and the 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.

[0008] In another aspect, a method of controlling polishing includes acquiring a first color image of a first substrate at an in-line metrology station of a polishing system; segmenting the first color image into a plurality of second color images using a die mask such that each second color image corresponds to a region of a die fabricated on the first substrate; generating a thickness measurement for one or more locations; and determining a polishing parameter for the first substrate or a subsequent second substrate based on the thickness measurement. Each respective location of the one or more locations corresponds to a respective region of a die fabricated on the first substrate. To generate a thickness measurement for a region, a second color image corresponding to the region is processed by a neural network trained using training data, the training data including a plurality of third color images of dies of a calibration substrate and in-situ thickness measurements of a surface of the calibration substrate, each respective third color image paired with an in-situ thickness measurement of a surface of the die region associated with the respective color image.

[0009] Implementations can include one or more of the following potential advantages. Thicknesses of multiple dies on a substrate can be measured quickly. For example, an in-line metrology system can determine thicknesses of a substrate based on color images of the substrate without impacting throughput. The estimated thicknesses can be used directly in a multivariate run-to-run control scheme.

[0010] A model can be trained using the described methods to produce thickness measurements with less than 5% error from actual film thicknesses. While thickness measurements can be derived from color images having three color channels, a hyperspectral camera can be added to a substrate imager system to provide higher dimensional feature inputs to the model. This can facilitate training of more complex models to understand more physical properties of the film stack.

[0011] Deep learning in a metrology system can have high inference speed and still be able to achieve high resolution measurements of thickness profiles on a substrate. It makes the metrology system a fast, low-cost pre- and post-metrology measurement tool for memory applications with higher thickness accuracy.

[0012] The accompanying drawings and below description detail one or more implementations. Other aspects, features, and advantages will be apparent from the description, drawings, and claims. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 A view showing an example of an in-line optical measurement system.

[0014] Figure 2A An example of an example image of a substrate for model training.

[0015] Figure 2B A schematic illustration of a computer data storage system.

[0016] Figure 3 A neural network is shown that is used as part of a controller for a polishing apparatus.

[0017] Figure 4 A flowchart of a method of detecting thickness of a layer on a substrate using a deep learning method is shown.

[0018] Like reference numerals in the various figures indicate like elements. DETAILED DESCRIPTION

[0019] Due to variations in polishing rates that occur in CMP processes, film thickness measurements using dry metrology systems are used in CMP processing. Such dry metrology measurement techniques often use either spectroscopic or ellipsometric methods, where variables in an optical model of the film stack are fit to the collected measurements. Such measurement techniques often require precise alignment of the sensor to the measurement point on the substrate to ensure that the model is applicable to the collected measurements. As a result, measuring a large number of points on a substrate can be time consuming, and collecting high resolution thickness profiles is not feasible.

[0020] However, using machine learning, it is possible to measure the thickness of films on a substrate with reduced time. By training a deep neural network using color images from dies on a substrate and associated thickness measurements from other reliable metrology systems, the film thickness of a die can be measured by applying an input image to the neural network. For low cost memory applications, for example, this system can be used as a high yield and economical solution. In addition to thickness inference, this technique can be used to grade residue levels on a substrate using image segmentation.

[0021] REFERENCE Figure 1 The polishing apparatus 100 includes one or more carrier heads 126, each carrier head 126 configured to carry a substrate 10, one or more polishing stations 106, and a transfer station that loads and unloads substrates from the carrier heads. Each polishing station 106 includes a polishing pad 130 supported on a platen 120. The polishing pad 130 can be a two-layer polishing pad having an outer polishing layer and a softer backing layer.

[0022] The carrier heads 126 can be suspended from a support 128 and can be moved 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 that is mounted to the track so that each carriage 108 can be selectively moved between the polishing stations 124 and the transfer station. Alternatively, in some implementations, the support 128 is a rotatable turntable, and rotation of the turntable simultaneously moves the carrier heads 126 along a circular path.

[0023] Each polishing station 106 of the polishing apparatus 100 can include a port, for example at the end of the arm 134, to dispense a polishing liquid 136, such as a polishing slurry, onto the polishing pad 130. Each polishing station 106 of the polishing apparatus 100 can also include a pad conditioning device to dress the polishing pad 130 to maintain the polishing pad 130 in a consistent polishing state.

[0024] Each carrier head 126 can be operable to hold the substrate 10 against the polishing pad 130. Each carrier head 126 can independently control a polishing parameter, such as a pressure associated with each respective substrate. In particular, each carrier head 126 can include a retaining ring 142 that holds the substrate 10 under a flexible membrane 144. Each carrier head 126 can also include a plurality of independently controllable plenum chambers (e.g., three chambers 146a-c) defined by the membrane that can apply independently controllable pressures to an associated region on the flexible membrane 144, and thereby to the substrate 10. Although three chambers are illustrated in FIG. 1 for ease of illustration, one or two chambers, or four or more chambers, such as five chambers, can be present. Figure 1

[0025] Each carrier head 126 is suspended from the support 128 and connected by a drive shaft 154 to a carrier head rotation motor 156 so that the carrier head can rotate about the shaft 127. Optionally, each carrier head 126 can oscillate laterally, for example by driving the carrier 108 on a track or by rotational oscillation of the turntable itself. In operation, the platform rotates about its central axis, each carrier head rotates about its central axis 127 and moves laterally over the top surface of the polishing pad.

[0026] A controller 190, such as a programmable computer, is connected to each motor to independently control the rate of rotation of the platform 120 and the carrier heads 126. The controller 190 can include a central processing unit (CPU) 192, memory 194, and support circuits 196, such as input / output circuits, power supplies, clock circuits, cache and the like. The memory is connected to the CPU 192. The memory is a non-transitory computer-readable medium, and can be one or more of readily available memory such as random access memory (RAM), read only memory (ROM), floppy disk, hard disk, or another form of digital storage. In addition, although shown as a single computer, the controller 190 can be a distributed system, for example including multiple processors and memories operating in concert.

[0027] ​The polishing apparatus 100 also includes an in-line (also referred to as in- sequence) optical metrology system 160. The color imaging system of the in-line optical metrology system 160 is located within the polishing apparatus 100, but does not perform measurements during polishing operations, but rather collects measurements between polishing operations, such as when moving a substrate from one polishing station to another, or before or after polishing, such as when moving a substrate from a transfer station to a polishing station or vice versa. Additionally, the in-line optical metrology system 160 can be located in a wafer fab interface unit or a module accessible from the wafer fab interface unit to measure a substrate after it is removed from a cassette but before it is moved to a polishing unit, or after it is cleaned but before it is returned to a cassette.

[0028] The in-line optical metrology system 160 includes a sensor assembly 161 that provides color imaging of the substrate 10. The sensor assembly 161 can include a light source 162, a light detector 164, and circuitry 166 for sending and receiving signals between a controller 190 and the light source 162 and light detector 164.

[0029] The light source 162 is operable to emit white light. In one implementation, the emitted white light includes light having a wavelength of 200 to 800 nanometers. Suitable light sources are arrays of white light light emitting diodes (LEDs), xenon lamps, or xenon-mercury lamps. The light source 162 is directed to direct light 168 at a non-zero angle of incidence a to the exposed surface of the substrate 10. The angle of incidence a can be, for example, about 30° to 75°, such as 50°.

[0030] The light source can illuminate a substantially linear elongated area that spans the width of the substrate 10. For example, the light source 162 can include optics, such as a beam expander, that spread the light from the light source to the elongated area. Alternatively or additionally, the light source 162 can include a linear array of light sources. The light source 162 itself and the area illuminated on the substrate can be elongated and have a longitudinal axis that is parallel to the surface of the substrate.

[0031] A diffuser 170 can be placed in the path of the light 168, or the light source 162 can include a diffuser that diffuses the light before it reaches the substrate 10.

[0032] The detector 164 is a color camera that is sensitive to light from the light source 162. The camera includes an array of detector elements. For example, the camera can include a CCD array. In some implementations, the array is a single column of detector elements. For example, the camera can be a line scan camera. A row of detector elements can extend parallel to the longitudinal axis of the elongated area illuminated by the light source 162. In the case where the light source 162 includes a row of light emitting components, the row of detector elements can extend along a first axis that is parallel to the longitudinal axis of the light source 162. A row of detector elements can include 1024 or more elements.

[0033] The camera 164 is configured with appropriate focusing optics 172 to project the field of view of the substrate onto an array of detector elements. The field of view can be long enough to observe the entire width of the substrate 10, e.g., 150 to 300 mm long. The camera 164 including the associated optics 172 can be configured so that individual pixels correspond to a region having a length equal to or less than about 0.5 mm. For example, assuming a field of view of about 200 mm long and a detector 164 including 1024 elements, the image produced by the line scan camera can have pixels of about 0.5 mm in length. To determine the length resolution of the image, the length of the field of view (FOV) can be divided by the number of pixels over which the FOV is imaged to yield the length resolution.

[0034] The camera 164 can also be configured so that the pixel width is comparable to the pixel length. For example, a line scan camera has the advantage of very high frame rates. The frame rate can be at least 5 kHz. The frame rate can be set to a frequency so that the pixel width is comparable to the pixel length, e.g., equal to or less than about 0.3 mm, as the imaging region scans the substrate 10.

[0035] The light source 162 and light detector 164 can be supported on a platform 180. In the case of a line scan camera for the light detector 164, the light source 162 and camera 164 can be moved relative to the substrate 10 so that the imaging region can scan the length of the substrate. In particular, the relative motion can be in a direction parallel to the surface of the substrate 10 and perpendicular to a row of detector elements of the line scan camera 164.

[0036] In some embodiments, the platform 182 is stationary and the support for the substrate is moved. For example, the carrier head 126 can be moved, e.g., by motion of the carriage 108 or rotational oscillation of a turntable, or a robot arm holding the substrate in a factory interface unit can move the substrate 10 past the line scan camera 182. In some embodiments, the platform 180 is movable while the carrier head or robot arm is held stationary for image acquisition. For example, the platform 180 can be moved by a linear actuator 182 along a track 184. In either case, this allows the light source 162 and camera 164 to be held in fixed positions relative to each other as the scanned region is moved past the substrate 10.

[0037] A possible advantage of moving the line scan camera and light source together over the substrate is that the relative angle between the light source and camera remains constant for different locations on the wafer, e.g., as compared to a conventional 2D camera. Thus, artifacts due to changes in viewing angle can be reduced or eliminated. In addition, a line scan camera can eliminate perspective distortion, whereas a conventional 2D camera exhibits inherent perspective distortion that then needs to be corrected by image transformation.

[0038] The sensor assembly 161 can include a mechanism to adjust the vertical distance between the substrate 10 and the light source 162 and the detector 164. For example, the sensor assembly 161 can include an actuator to adjust the vertical position of the platform 180.

[0039] Optionally, a polarization filter 174 can be located in the path of the light, for example between the substrate 10 and the detector 164. The polarization filter 174 can 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 polarization filter 174 can reduce haze in the image, sharpen or enhance desired visual features.

[0040] Assuming that the outermost layer on the substrate is a translucent layer (e.g., a dielectric layer), the color of the light detected at the detector 164 depends on, for example, the composition of the substrate surface, the smoothness of the substrate surface, and / or the amount of interference between light reflected by different interfaces of one or more layers (dielectric layers) on the substrate. As described above, the light source 162 and the light detector 164 can be connected to a computing device, such as the controller 190, which can be operable to control the operation of the light source 162 and the light detector 164 and to receive their signals. The computing device that performs the various functions to convert the color image to a thickness measurement can be considered part of the metrology system 160.

[0041] Referring to Figure 2A , an example of an image 202 of a substrate 10 collected with the inline optical metrology system 160 is shown. The inline optical metrology system 160 produces a high resolution color image 202, such as an image having at least three color channels (e.g., RGB channels) of at least 720 x 1080 pixels, such as an image of at least 2048 x 2048 pixels. The color at any particular pixel depends on the thickness of one or more layers (including the top layer) in the region of the substrate corresponding to that pixel.

[0042] The image 202 is segmented into one or more regions 208, each region corresponding to a die 206 fabricated on the substrate. The portion of the image that provides the region 208 can be a predetermined region in the image, or can be determined automatically based on the image using an algorithm.

[0043] As an example of a predetermined region in the image, the controller can store a die mask that identifies, for each region 208, a location and a region in the image. For example, for a rectangular region, the region can be defined by the upper right and lower left coordinates in the image. Thus, the mask can be a data file that includes a pair of upper right and lower left coordinates for each rectangular region. In other cases, where the regions are non-rectangular, more complex functions can be used.

[0044] 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 substrate orientation can be determined with a notch detector or through image processing of the color image 202, for example to determine the angle of the scribe line in the image. The substrate position can also be determined through image processing of the color image 202, for example by detecting the circular substrate edge and then determining the center of the circle.

[0045] As an example of automatically determining the regions 208, an image processing algorithm can analyze the image 202 and detect the scribe lines. The image 202 can then be segmented into regions between the identified scribe lines.

[0046] By segmenting the initial color image, a plurality of color images 204 of the individual regions 208 can be collected from the substrate 10. As described above, each color image 204 corresponds to a die 206 fabricated on the substrate. The collected color images can be output as PNG images, although many other formats can be used, such as JPEG, etc.

[0047] The color images 204 can be provided to an image processing algorithm to produce a thickness measurement of the die shown in the color image 204. The image is used as input data for an image processing algorithm that has been trained (e.g., through a supervised deep learning method) to estimate layer thickness based on color images. A supervised deep learning-based algorithm establishes a model between the color image and the thickness measurement. The image processing algorithm can include a neural network as the deep learning-based algorithm.

[0048] The intensity value of each color channel for each pixel of the color image 204 is keyed into the image processing algorithm, such as the input neurons of a neural network. Based on this input data, a layer thickness measurement of the color image can be computed. Thus, inputting the color image 204 to the image processing algorithm results in an output of an estimated thickness. For low-cost memory applications, for example, this system can be used as a high-yield and economical solution. In addition to thickness inference, this technique can be used to grade residue levels on a substrate using image segmentation.

[0049] To train the image processing algorithm (e.g., a neural network) using a supervised deep learning method, calibration images of the dies of one or more calibration substrates can be acquired as discussed above. That is, each calibration substrate can be scanned with the line scan camera of the inline optical metrology system 160 to produce an initial calibration image, which can be segmented into a plurality of color images of the individual regions on the calibration substrate.

[0050] Prior to or after collecting the initial color calibration images, ground truth thickness measurements are collected at multiple locations on the calibration substrate using a high accuracy metrology system, such as an inline or standalone metrology system. The high accuracy metrology system can be a dry optical metrology system. The ground truth measurements can be from offline reflectometry, ellipsometry, scatterometry, or more advanced TEM measurements, although other techniques can be suitable. Such systems can be obtained from Nova Measuring Instruments or Nanometrics. Each location corresponds to one of the dies manufactured, i.e., one of the individual regions.

[0051] For example, with reference to Figure 2B For each individual region on each calibration substrate, a color calibration image 212 is collected using an inline sensor of the optical metrology system 160. Each color calibration image is associated with a ground truth thickness measurement 214 for the corresponding die on the calibration substrate. The images 212 and associated ground truth thickness measurements 214 can be stored in a database 220. For example, the data can be stored as records 210, where each record includes a calibration image 212 and a ground truth thickness measurement 214.

[0052] The deep learning based algorithm, such as a neural network, is then trained using the combined dataset 218. When training the model, the thickness measurements corresponding to the center of the die measured with the dry metrology tool are used as labels for the input images. For example, the model can be trained on approximately 50,000 images collected from five dies on a substrate with a wide range of back thicknesses.

[0053] Figure 3 A neural network 320 used as part of the controller 190 of the polishing apparatus 100 is shown. The neural network 320 can be a deep neural network that is used to perform a regression analysis on the RGB intensity values of an input image of a calibration substrate and the ground truth thickness measurements to produce a model to predict the layer thickness of a region of a substrate based on a color image of the region.

[0054] The neural network 320 includes a plurality of input nodes 322. The neural network 320 can include an input node for each color channel associated with each pixel of an input color image, a plurality of hidden nodes 324 (also referred to as “intermediate nodes” below), and an output node 326 that will produce a layer thickness measurement value. In a neural network with a single layer of hidden nodes, each hidden node 324 can be coupled to each input node 322, while the output node 326 can be coupled to each hidden node 320. However, as a practical matter, neural networks for image processing can have many layers of hidden nodes 324.

[0055] Typically, the hidden node 324 outputs a value that is a nonlinear function of a weighted sum of values from the input nodes 322 or hidden nodes of a previous layer to which the hidden node 324 is connected.

[0056] For example, the output of a hidden node 324 in the first layer (labeled node k) can be expressed as: tanh(0.5 * a k1 (I1) + a k2 (I2) +... + a kM (I M ) + b k ) where tanh is the hyperbolic tangent, a kx is the weight of the connection between the kth hidden node and the xth input node (of M input nodes), and I M is the value at the Mth input node. However, other nonlinear functions can be used instead of tanh, such as a rectified linear unit (ReLU) function and variants thereof.

[0057] The neural network 320 thus includes an input node 322 for each color channel associated with each pixel of an input color image, e.g., where there are J pixels and K color channels, then L = J*K is the number of intensity values in the input color image, and the neural network 320 will include at least input nodes N1, N2,..., N L . Thus, where the number of input nodes corresponds to the number of intensity values in a color image, the output H k of a hidden node 324 (labeled node k) can be expressed as: H k = tanh(0.5 * a k1 (I1) + a k2 (I2) +... + a kL (I L ) + b k )

[0058] Assuming that the measured color image S is represented by a row matrix (i1, i2,..., i L ), the output of a hidden node 324 (labeled node k) can be expressed as: H k = tanh(0.5 * a k1 (V1 · S) + ak2(V2 · S) +... + a kL (V L · S) + b k ) where V is the weighted value (v1, v2, v L ), V xa weighting for the xth intensity value of L intensity values of a color image.

[0059] The output node 326 can produce a characteristic value CV, such as thickness, that is a weighted sum of the outputs of the hidden nodes. This can be represented, for example, as CV = C1*H1 + C2*H2 +... + C l *H l where Ckis a weighting for the output of the kth hidden node. k

[0060] However, the neural network 320 can optionally include one or more other input nodes (e.g., node 322a) to receive other data. This other data can come from: a previous measurement of the substrate with the in-situ monitoring system, such as pixel intensity values collected early in processing the substrate; a measurement of a previous substrate, such as pixel intensity values collected during processing of another substrate; another sensor in the polishing system, such as a measurement of the temperature of the pad or substrate with a temperature sensor; a polishing recipe stored by a controller used to control the polishing system, such as polishing parameters such as carrier head pressure and platen rotation rate used to polish the substrate; a variable tracked by the controller, such as the number of substrates after a change of pad; or a sensor that is not part of the polishing system, such as a measurement of the thickness of an underlying film with a metrology station. This allows the neural network 320 to take into account other processing or environmental variables when calculating the thickness measurement.

[0061] The thickness measurement produced at the output node 326 is provided to a process control module 330. The process control module can adjust process parameters, such as carrier head pressure, platen rotation rate, etc., based on the thickness measurement of one or more areas. The adjustment can be performed on a polishing process to be performed on the substrate or a next substrate.

[0062] The neural network 320 needs to be configured before it is used, for example, to measure a substrate.

[0063] ​As part of the configuration procedure, the controller 190 can receive a plurality of calibration images. Each calibration image has a plurality of intensity values, e.g., an intensity value for each color channel, each pixel, of the calibration image. The controller also receives a characteristic value for each calibration image, e.g., a thickness. For example, the color calibration images can be measured at a particular die fabricated on one or more calibration or test substrates. In addition, ground truth thickness measurements can be performed at the particular die location using a dry measurement device, e.g., a contact surface profiler or ellipsometer. The ground truth thickness measurements can thus be correlated with the color images of the same die location on the substrate. A plurality of color calibration images can be generated from, e.g., five to ten calibration substrates by segmenting the images of the calibration substrates as discussed above. With respect to the configuration procedure for the neural network 320, the color images and characteristic values for each die fabricated on the calibration substrates are used to train the neural network 320.

[0064] V corresponds to one of the color images, and thus is associated with a characteristic value. When the neural network 320 is operating in a training mode, e.g., a backpropagation mode, values (v1, v2,..., v L ) are provided to the respective input nodes N1, N2,... N l , while the characteristic value CV is provided to the output node 326. This process can be repeated for each column. This process sets the values of a k1 in Equations 1 or 2 above.

[0065] The system is now ready for operation. Color images are measured from a substrate using the inline monitoring system 160. The measured color images can be represented with a row matrix S = (i1, i2,..., i L ), where i j represents the intensity value of the jth intensity value of L intensity values, where L = 3n when the image includes a total of n pixels and each pixel includes three color channels.

[0066] When the neural network 320 is used in an inference mode, these values (S1, S2,..., S L ) are provided as input to the respective input nodes N1, N2,... N L . From this, the neural network 320 generates a characteristic value, e.g., a layer thickness, at the output node 326.

[0067] The architecture of the neural network 320 can vary in depth and width. For example, while the illustrated neural network 320 has a single row of intermediate nodes 324, it can include multiple rows. The number of intermediate nodes 324 can be equal to or greater than the number of input nodes 322.

[0068] As described above, the controller 190 can associate individual 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 location of the sensor on the substrate when the image was collected. This allows the controller 190 to produce a separate sequence of measurement values for each die.

[0069] In some implementations, the controller 190 can be configured to have a neural network model structure that is composed of multiple different types of building blocks. For example, the neural network can be a residual neural network that includes residual block features in its architecture. A residual neural network can utilize skip connections or shortcuts to skip over some layers. A residual neural network can be implemented with, for example, a ResNet model. In the context of a residual neural network, a non-residual neural network can be described as a plain network.

[0070] In some implementations, the neural network can be trained to take into account the underlying layer thickness of the stack during computation, which can improve errors due to variations in the thickness measurement of the underlying layer. The impact of the underlying thickness variation of the film stack can be mitigated to improve the performance of the model by providing the intensity values of the color image of the thickness of the underlying layer as an additional input to the model.

[0071] The reliability of the computed thickness measurements can be assessed by comparing the computed thickness measurements to the measurements, and then determining the difference between the computed values and the 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 method improves the throughput of the overall system, and enables thickness measurements to be performed on all substrates in a batch.

[0072] Reference Figure 4 An image processing algorithm generated by machine learning techniques for a substrate thickness measurement system. This image processing algorithm can receive RGB images collected by an integrated line scan camera inspection system and enables film thickness estimation at a much faster speed. The inference time for about 2000 measurement points is about a few seconds, in contrast to 2 hours using dry metrology.

[0073] The method includes the controller combining the individual image lines from the light detector 164 into a two-dimensional color image (500). The controller can apply an offset and / or a gain adjustment to the intensity values of the image in each color channel (510). Each color channel can have a different offset and / or gain. The image can optionally be normalized (515). For example, a difference between the measured image and a standard pre-defined image can be calculated. For example, the controller can store a background image for each of the red, green, and blue channels, and can subtract the background image from the measured image for each color channel. Alternatively, the standard pre-defined image can divide the measured image. The image can be filtered to remove low frequency spatial variations (530). In some implementations, a filter is generated using the luminance channel, and then the filter is applied to the red, green, and blue images.

[0074] The image is converted (e.g., resized and / or rotated and / or translated) to a standard image coordinate system (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 resized so that the edges of the substrate are at the edges of the image, and / or the image can be rotated so that the x-axis of the image is at a 0° angle between the radial segment connecting the substrate center and the substrate orientation feature.

[0075] One or more regions on the substrate are selected, and an image is generated for each selected region (550). This step can be performed using the techniques described above, e.g., the regions can be predetermined regions, or the regions can be determined automatically by an algorithm that provides the regions 208.

[0076] The intensity values provided by each color channel for each pixel of the image are treated as input to an image processing algorithm that is trained on the monitored data. The image processing algorithm outputs a layer thickness measurement for the particular region (560).

[0077] Various deep model architectures are trained and validated on small die test patterned substrates, with the goal of reducing error in the measurements. Models that take into account the properties of the underlying layer have less error. Additionally, a preliminary tool-to-tool matching validation is performed by training a model on data collected at one tool and using it to infer data from other tools. The results can be compared to training data from the same tool and inferring with it.

[0078] Overall, one or more operating parameters of the CMP apparatus can be controlled using the data. The operating parameters include, for example, the platform rotation speed, the substrate rotation speed, the substrate polishing path, the substrate speed on the platform, the pressure applied to the substrate, the slurry composition, the slurry flow rate, and the temperature of the substrate surface. The operating parameters can be controlled in real time, and can be adjusted automatically without further human intervention.

[0079] As used in this specification, the term substrate can include, for example, product substrates (e.g., including multiple memory or processor dies), test substrates, bare substrates, and gated substrates. The substrate can be at various stages of integrated circuit manufacturing, e.g., the substrate can be a bare die, or it can include one or more deposited and / or patterned layers. The term substrate can include discs and rectangular wafers.

[0080] However, the color image processing techniques described above can be particularly useful in the context of 3D vertical NAND (VNAND) flash memory. In particular, the layer stacks used in VNAND manufacturing are so complex that current metrology methods (e.g., Nova spectroscopy) cannot perform detection of regions with inadequate thickness with sufficient reliability. In contrast, the color image processing techniques can have higher reliability in this application.

[0081] All of the owner of the embodiments of the invention and the functional operations described in this specification can be embodied in digital electronic circuitry, or in computer software, firmware, or hardware, including the structural means disclosed in this specification and structural equivalents thereof, or in combinations of them. The embodiments of the invention can be embodied in one or more computer program products, i.e., one or more computer programs tangibly embodied in a non-transitory machine-readable storage media having instructions

[0082] The term relative position indicates the relative position of system components to each other (not necessarily relative to gravity); it should be understood that the polishing surface and substrate can be held in a vertical orientation or some other orientation.

[0083] Several implementations have been described. However, it will be understood that various modifications can be made. For example Rather than a line-scan camera, a camera that images the entire substrate can be used. In this case, no motion of the camera relative to the substrate is needed. The camera can cover less than the entire width of the substrate. In this case, the camera will need to undergo two perpendicular motions (e.g., supported on an X-Y stage) to scan the entire substrate. The light source can illuminate the entire substrate. In this case, the light source need not move relative to the substrate. The light detector can be a spectrometer rather than a color camera; the spectral data can then be reduced to the RGB color space. The sensing assembly need not be located in an in-line system between polishing stations or between a polishing station and a transfer station. For example, the sensing assembly can be located within a transfer station, in a cassette interface unit, or be a stand-alone system. The uniformity analysis step is optionally used. For example, the image produced by applying the threshold conversion can be provided to a feed forward process, to adjust a subsequent processing step of the substrate, or to a feedback process to adjust a processing step of a subsequent substrate.

[0084] Thus, other embodiments are within the scope of the following claims.

Claims

1. A non-transient computer-readable medium encoded with a computer program product, the computer program product comprising instructions to cause one or more processors to: The surface residue measurement results of the top layer of the calibration substrate are obtained at multiple locations, each location being a defined location of a die manufactured on the substrate; Acquire multiple color images of the calibration substrate, each color image corresponding to a region of a die manufactured on the substrate; and A neural network is trained to convert color images from a die region of an inline substrate imager into residue measurements of the top layer in the die region, wherein the instructions for training the neural network include instructions for using training data, the training data including the plurality of color images and ground-level residue measurements, each corresponding color image being paired with a ground-level residue measurement of the die region associated with the corresponding 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 scan of the calibration substrate from the inline substrate imager.

3. The computer-readable medium of claim 1, wherein the instructions for acquiring the plurality of color images include instructions for receiving a color image of the calibration substrate and segmenting 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 acquiring measurements of surface residue on the top layer of a plurality of calibration substrates and acquiring a plurality of color images of each of the calibration substrates.

5. The computer-readable medium of claim 1, wherein the instructions for acquiring the measured results of the surface residue include instructions for receiving the measured results of the residue at each of the plurality of locations.

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

7. The computer-readable medium of claim 1, comprising instructions for obtaining measurement results from all dies in the substrate.

8. The computer-readable medium of claim 1, comprising instructions for obtaining measurement results from all substrates in a batch before and after chemimechanical planarization.

9. A method for training a neural network for use in a substrate residue measurement system, comprising the following steps: The surface residue measurement results of the top layer of the calibration substrate are obtained at multiple locations, each location being a defined location of a die manufactured on the substrate; Acquire multiple color images of the calibration substrate, each color image corresponding to a region of a die manufactured on the substrate; as well as A neural network is trained to convert color images from the die region of an inline substrate imager into residue measurements of the top layer in the die region. Training is performed using training data, which includes the plurality of color images and ground-level residue measurements, with each corresponding color image paired with a ground-level residue measurement of the die region associated with that corresponding color image.

10. A computer program product comprising a non-transitory computer-readable medium encoded with instructions to cause one or more processors to: Receive a first color image of the first substrate from the inline monitoring system of the polishing system; The first color image is divided into a plurality of second color images using a die mask, such that each second color image corresponds to a region of a die manufactured on the first substrate. Residue measurement results are generated at one or more locations, each of the one or more locations corresponding to a corresponding region of a die manufactured on the first substrate, wherein the instruction for generating the residue measurement result for a region includes instructions for processing a second color image corresponding to the region using a neural network trained with training data, the training data including a plurality of third color images of the die of the calibration substrate and ground-based residue measurement results of the calibration substrate, each corresponding third color image being paired with a ground-based residue measurement result of the die region associated with the corresponding third color image; and The grinding parameters for the first substrate or the subsequent second substrate are determined based on the residue measurement results.

11. The computer program product of claim 10, comprising an instruction to receive the first color image of the first substrate after the first substrate has been ground in a grinding station.

12. The computer program product of claim 11, comprising instructions for determining the polishing parameters of the polishing station for the subsequent second substrate based on the residue measurement results.

13. The computer program product of claim 10, comprising an instruction to receive the first color image of the first substrate before the grinding station grinds the first substrate.

14. The computer program product of claim 13, comprising instructions for determining the polishing parameters of the polishing station of the first substrate based on the residue measurement results.

15. The computer program product of claim 10, wherein the grinding parameters include the pressure in the chamber of the carrier head.

16. A grinding apparatus comprising: A polishing station includes a platform for supporting a polishing pad and a carrier head to hold a first substrate against the polishing pad; A linear measuring station equipped with a color camera to generate a color image of the first substrate; and The control system, which is configured to Receive a first color image of the first substrate from the inline monitoring station of the polishing system. The first color image is divided into a plurality of second color images using a die mask, such that each second color image corresponds to a region of a die manufactured on the first substrate. Generate residue measurement results for one or more locations, each of the one or more locations corresponding to a corresponding region of a die manufactured on the first substrate, wherein the instruction for generating the residue measurement result for a region includes instructions for processing a second color image corresponding to the region using a neural network trained with training data, the training data including multiple third color images of the die of the calibration substrate and ground-based residue measurement results of the calibration substrate, each corresponding third color image being paired with a ground-based residue measurement result of the die region associated with the corresponding third color image, and The grinding parameters of the first substrate or the subsequent second substrate are determined based on the residue measurement results. The grinding station grinds the first substrate or the subsequent second substrate using the determined grinding parameters.

17. The apparatus of claim 16, wherein the grinding parameters include the pressure in the chamber of the carrier head.

18. The apparatus of claim 16, wherein the control system is configured to receive the first color image of the first substrate after the first substrate has been ground at the grinding station.

19. A method for controlling grinding, comprising the following steps: A first color image of the first substrate is acquired using the inline monitoring system of the polishing system; The first color image is divided into a plurality of second color images using a die mask, such that each second color image corresponds to a region of a die manufactured on the first substrate. Residue measurement results are generated at one or more locations, each of the one or more locations corresponding to a corresponding region of a die manufactured on the first substrate, wherein generating the residue measurement result for a region includes processing a second color image corresponding to the region using a neural network trained with training data, the training data including multiple third color images of the die of the calibration substrate and ground-based residue measurement results of the calibration substrate, each corresponding third color image being paired with and associated with ground-based residue measurement results of the die region; and The grinding parameters for the first substrate or the subsequent second substrate are determined based on the residue measurement results.

Images