Semiconductor substrate yield prediction based on spectral data from multiple substrate dies

Abstract

A system and method for improving the fabrication of a substrate is provided. Subsets of the dies of a substrate can be inspected at various points in the fabrication process to generate spectral data. The spectral data can be used to generate data that is input into a machine learning model to predict the yield of the substrate.

Description

Technical Field

[0001] (Claim of Priority) This application was filed as a PCT international patent application on Nov. 21, 2022, and claims the benefit and priority of U.S. Non-Provisional Patent Application No. 17 / 532,700, filed on Nov. 22, 2021, the entire disclosure of which is incorporated herein by reference.

[0002] (Field of the Invention) This disclosure relates to the manufacture and measurement of semiconductor components.

Background Art

[0003] As part of the formation of semiconductor chips or other types of integrated circuits (ICs), semiconductor substrates are manufactured or fabricated. Final IC components can be incorporated into the substrate by a series of fabrication processes. The fabrication processes may include deposition processes where thin film layers are added onto the substrate. The substrate can then be coated with photoresist and the circuit pattern of a reticle can be projected onto the substrate using lithography techniques. An etching process may then be performed. In each fabrication process, the tools performing the fabrication process may introduce defects or imperfections into the substrate.

[0004] In some applications, the semiconductor substrate consists of dies. A die is a block of semiconductor material (e.g., silicon) on which a given functional circuit is fabricated. For example, the functional circuit can take the form of a central processing unit. The functional circuits may be batch manufactured on a single substrate. The substrate is then diced into a plurality of dies, each die containing one copy of the circuit. Each die may be rectangular. A thin non-functional gap can be provided between the dies, enabling the individual dies to be separated from the substrate (e.g., by sawing) without damaging the circuits.

Summary of the Invention

[0005] Generally, this disclosure relates to the inspection of semiconductor substrates during the manufacture of the substrates. An example of a semiconductor substrate is a semiconductor wafer.

[0006] According to some aspects, the present disclosure relates to inspecting a subset of dies of a semiconductor substrate during manufacture of a substrate.

[0007] According to some other aspects, the present disclosure relates to inspecting a subset of dies of a semiconductor substrate after a plurality of manufacturing steps during manufacture of the substrate.

[0008] According to some other aspects, the present disclosure relates to inspecting a subset of dies of a semiconductor substrate by generating spectral data for each of the dies in the subset.

[0009] According to some other aspects, the present disclosure relates to generating combined data that includes spectral data of dies of a semiconductor substrate to be inspected. In some examples, the combined data can include an array of distinct data subsets, where each subset represents spectral data from one of the dies to be inspected. In some examples, the combined data can be represented as a matrix.

[0010] According to some other aspects, the present disclosure relates to generating combined data that includes spectral data of dies of a semiconductor substrate to be inspected after each of a plurality of manufacturing steps of the substrate. As used herein, terms such as manufacturing and fabrication or fabricating are used interchangeably.

[0011] According to some other aspects, the present disclosure relates to predicting the yield of a semiconductor substrate based on spectral data of dies of the substrate to be inspected.

[0012] According to some other aspects, the present disclosure relates to predicting the yield of a semiconductor substrate by inputting spectral data of dies of the substrate to be inspected into a trained machine learning model.

[0013] According to some other aspects, the present disclosure relates to predicting the yield of a semiconductor substrate by processing combined data including spectral data of dies of interest on the substrate using a convolutional neural network.

[0014] According to some other aspects, the present disclosure relates to predicting the yield of a semiconductor substrate by processing combined data including spectral data of dies of interest on the substrate inspected after each of a plurality of manufacturing processes of the substrate using a convolutional neural network.

[0015] According to some other aspects, the present disclosure relates to identifying a tool or process that causes defects in a semiconductor substrate.

[0016] According to some other aspects, the present disclosure relates to any combination of the foregoing features.

[0017] According to some specific aspects, the present disclosure relates to a method for predicting the yield of a substrate, the method including combining first spectral data for each of a plurality of dies on the substrate and second spectral data for each of the plurality of dies on the substrate to form combined data, where the first data is generated from a first inspection of the plurality of dies at a first time point and the second data is generated from a second inspection of the plurality of dies at a second time point, and providing the combined data as an input to a trained machine learning model to generate a predicted yield of the substrate.

[0018] According to some other specific aspects, the present disclosure causes a computing device to combine first spectral data for each of a plurality of dies on a substrate with second spectral data for each of the plurality of dies on the substrate to obtain combined data, where the first data is generated from a first inspection of the plurality of dies at a first point in time, the second data is generated from a second inspection of the plurality of dies at a second point in time, and to provide the combined data as an input to a trained machine learning model, thereby generating a predicted yield of the substrate. The present disclosure relates to a non-transitory computer-readable medium including instructions for performing the foregoing operations.

[0019] This summary is provided to introduce a selection of concepts in a simplified form that are further described below in the detailed description. This summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter. Additional aspects, features, and / or advantages of examples will be set forth in part in the description that follows, and in part will be apparent from the description, or may be learned by practice of the disclosure.

[0020] Non-limiting and non-exhaustive examples are described with reference to the following figures.

Brief Description of the Drawings

[0021]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

DETAILED DESCRIPTION OF THE INVENTION

[0022] Examples of the present disclosure describe systems and methods for improving substrate fabrication. An example of such a substrate is a semiconductor wafer. After the substrate is processed by at least one fabrication tool, an inspection device inspects a subset of the dies of the substrate to identify defects in the substrate. Based on the inspection data generated from the inspection process, the yield of the substrate is predicted. The yield of the substrate can refer to the percentage of dies of the substrate that meet one or more defined performance criteria, quality criteria, or other acceptability criteria.

[0023] In some examples, the predicted yield can be used to identify tools or processes used in fabricating the substrate that may be causing manufacturing defects in the substrate.

[0024] In some examples, the predicted yield can be used to improve the fabrication of the semiconductor substrate (e.g., by adjusting or replacing a tool that causes defects) to compensate for defects based on the predicted yield.

[0025] One exemplary inspection method for inspecting a substrate, particularly for inspecting dies of the substrate, includes generating spectral data for each of a subset of the dies of the substrate. The spectral data can be generated using an inspection device or tool such as a reflectometer or spectrometer that detects and measures the wavelength spectrum of light or other radiation reflected by the die. Thus, in some examples, the inspection tool is configured to generate multi-spectral imaging of the substrate.

[0026] For example, a substrate may have multiple dies, and those dies may be inspected at different stages of the manufacturing process, such as when a new layer is added to the substrate. It is not necessary to inspect each die on the substrate as part of this process. An exemplary substrate map 10 is schematically shown in FIG. 1. The substrate map 10 represents a semiconductor substrate during fabrication. The exemplary substrate map 10 shows the positions of a subset of the dies 12 of the substrate that can be inspected for the purpose of predicting the yield of the substrate in accordance with the present disclosure. Dies of the substrate map other than the dies 12, such as die 14, are not inspected in this example.

[0027] The number and positions of the dies to be inspected for predicting the yield of the substrate can be selected to optimize one or more factors. For example, inspecting fewer dies may be more cost and / or time efficient but may provide a larger margin for yield prediction error. Similarly, inspecting dies that are relatively dense on the substrate may be more cost and / or time efficient than inspecting dies to be inspected that are scattered across the entire substrate, but again, it provides a larger margin for yield prediction error across the entire substrate. Thus, the number and positions of the dies to be inspected can be selected using a convolutional neural network (CNN) or other type of machine learning model to minimize cost and time and maximize the accuracy and precision of yield prediction. In some examples, the number of dies selected can be a predetermined ratio to the total number of dies of the substrate, such as within the range of about 1 die per 50,000 dies to about 1 die per 1,000 dies. Ratios to total numbers outside this range can also be selected.

[0028] A CNN or other machine learning model can be trained (e.g., using supervised training) using spectral data from dies of substrates with known yields to predict the unknown yield of a target substrate based on spectral data from inspection of dies of the target substrate.

[0029] In some examples, the number of dies inspected on a target substrate for the purpose of predicting the yield of the target substrate is equal to the number of dies of each substrate used to train a CNN or other machine learning model.

[0030] In some examples, the positions of the dies inspected on a target substrate for the purpose of predicting the yield of the target substrate are the same as the positions of the dies of each substrate used to train a CNN or other machine learning model.

[0031] In some examples, the number and / or positions of the dies inspected on a target substrate may be different from the number and / or positions of the dies of the training substrates.

[0032] Referring further to FIG. 1, for each of the dies 12 to be inspected, spectral data is generated. The spectral data for each of the dies 12 to be inspected can be one-dimensional serial data from electromagnetic radiation of different wavelengths (e.g., light of different wavelengths). For example, for each wavelength, the signal intensity is represented by the data.

[0033] Other measurement tools such as an acoustic measurement tool or an ultrasonic measurement tool can be used to inspect the dies of the training substrate and generate the spectral data of the dies 12 of the target substrate.

[0034] In some examples, the spectral data may be obtained as a function of another spectral attribute such as frequency instead of wavelength.

[0035] FIG. 2 shows an exemplary plot 18 of visible light spectrum data of a single die 12 of a substrate represented by the map 10 of FIG. 1, where the vertical axis represents the signal intensity of the signal reflected by the die 12 and the horizontal axis represents the wavelength. The plot 18 can include a plurality of measurements made by an inspection apparatus (e.g., a reflectometer or a spectrometer) at each measured wavelength. Each such signal measurement at a given wavelength and at a given time point of a given die after a given fabrication process (i.e., at a given inspection time) can be regarded as a channel. The plot 18 can include the reflected signal intensity for each of a plurality of channels at each of a plurality of reflected wavelengths (e.g., more than 1000 wavelengths). The plurality of channels includes signal measurements made at different time points at the same wavelength for the same die after the same substrate fabrication process. Thus, the “inspection time point” is used herein to refer to the inspection of a die at a single inspection time point after a given fabrication process, but data for a plurality of channels can be acquired to improve the reliability of the signal measurement, and each channel corresponds to a different measurement time point but the same inspection time point. That is, at each inspection time point, the spectral data can be measured multiple times, and each time corresponds to an individual channel.

[0036] In some examples, one or more statistical operations can be applied to the signal intensities detected over a plurality of channels at a given wavelength to arrive at a single signal intensity value per wavelength, such as the median or average of the signal intensities per die per wavelength per inspection time point over all channels. In some examples, if it is determined that the variance of the signal intensities over a plurality of channels at a given wavelength exceeds a maximum threshold variance, the spectral data or a portion thereof is not used and / or the inspection is performed again.

[0037] Spectral data is generated for each die to be inspected on the substrate. For example, if spectral data from 12 dies 12 of a target substrate is used by a machine learning model to predict the yield of a model, a set of 12 sets of spectral data is generated for each die 12 to be inspected.

[0038] The number of dies, which is 12, is used merely as an exemplary number, and other numbers of dies may be used. For example, the number of dies selected for inspection can be based on the optimization of different factors. For example, inspecting more dies can improve the substrate yield prediction, but at the same time, it requires more time for spectrum detection and more computing power for the training and application of the machine learning model that provides the prediction, thus reducing the efficiency of the inspection process. Therefore, for example, the number of dies selected for inspection to generate a predicted substrate yield can be based on the optimization of the error margin of the prediction as a function of the number or proportion of dies to be inspected. For example, the number of dies selected for inspection of a given substrate can be the minimum number of dies that generates a substrate yield prediction with an error margin below a predetermined threshold error margin.

[0039] The position of the dies selected for inspection on the substrate can be selected based on any number of factors. For example, the position can be selected to provide the minimum number of dies required for inspection so that the substrate yield prediction falls within the threshold margin of error. In some examples, the position of the dies selected for inspection is randomly selected, for example, using a random number generator or other randomization tools.

[0040] The spectral data of the dies to be inspected is combined to obtain combined data that can be provided as an input to a machine learning model such as a CNN for predicting the yield of the substrate. An example of such combined data is a two-dimensional matrix. The yield of the substrate may be the proportion of dies on the substrate that pass the final test (i.e., the proportion of dies on the substrate that meet a predetermined acceptable operating or quality standard). More specifically, the spectral data correlates with the presence or absence of die defects that correlate with the substrate yield. Non-limiting examples of such defects can include non-uniform thickness of the deposition layer, cracks, inappropriate dimensions, contamination, etc.

[0041] A substrate can be fabricated using multiple processes and tools. The fabrication tools can include, for example, among various types of tools, an oxidation system, an epitaxial reactor, a diffusion system, an ion implantation device, a physical vapor deposition system, a chemical vapor deposition system, a photolithography device, and an etching device. The number of manufacturing steps employed to fabricate the substrate can vary from a small number (e.g., less than 10) to several hundred in some examples. Typically, the substrate can be fabricated by progressing through a fabrication line, and each fabrication tool within the fabrication line performs a process step on the substrate. The same tool can be used multiple times in the fabrication line. For example, multiple deposition steps, lithography steps, and / or etching steps can be performed to fabricate the substrate.

[0042] In some examples, the spectral data can also be correlated with the types of die defects that correlate with substrate yield. For example, some defects may have more or less impact on the acceptability of the substrate than other types of defects. If the type of defect is known, the types of defects can be weighted according to their impact on the acceptability of the substrate when using a machine learning model to predict the substrate yield. In some examples, the type of defect can be determined based on the fabrication process performed immediately prior to the substrate inspection. For example, if spectral data is acquired after a deposition process, the defect can be a deposition type defect and / or a defect related to the deposition tool, as long as the spectral data indicates a defect in the substrate die. Then, when using the spectral data to predict the substrate yield, an appropriate weighting factor for the detected deposition type defect can be applied by the machine learning model.

[0043] To predict the yield of a target substrate based on the spectral data of the dies of the substrate, a machine learning model such as a CNN can be trained using correlations. For example, the machine learning model can be trained using a training substrate with a known yield, and the spectral data of the dies of the target substrate (such as the plot shown in FIG. 2) can be compared with the corresponding spectral data of the training substrate with a known yield (for example, of the corresponding dies and / or after the same manufacturing process) to predict the yield of the target substrate. In this case, the machine learning model can use the compared spectral data of the training substrate with a known yield to classify the yield of the target substrate and provide an output yield prediction for the target substrate.

[0044] The spectral data can be combined into any suitable combined number of arrays, and the arrays can be provided as input to a machine learning model such as a CNN. For example, the combined data can include, for each die, a data subset representing the spectral data corresponding to that die. The data subsets within the combined data array may be arranged such that each data subset is distinguishable from any other data subset within the array. The value of each data within each data subset is associated with the wavelength of the corresponding spectral data of the corresponding die and represents the signal intensity at the associated wavelength of the corresponding spectral data of the corresponding die.

[0045] Non-limiting examples of combined data arrays thus arranged with data subsets include matrices. FIG. 3 shows a portion of an exemplary matrix 30 of spectral data that can be input into a machine learning model to generate a target substrate yield prediction according to the present disclosure. Matrix 30 is an exemplary array or structure of distinguishable combined data subsets corresponding to the spectral data.

[0046] Matrix 30 includes spectral data of the dies of the target substrate after the manufacturing process of the target substrate. All of the spectral data within matrix 30 is acquired at time point 1 or inspection time point 1, which means that all of the data within matrix 30 is acquired after a given manufacturing process and before the next manufacturing process or the next other process that physically changes the substrate.

[0047] The completed matrix 30 (or other array of combined spectral data) can be provided as an input to a machine learning model to predict the yield of the substrate. In some examples, one or more attributes that can be used by the machine learning model to more accurately predict the substrate yield are assigned to matrix 30 (or other array of combined spectral data). For example, a manufacturing process attribute can be assigned to matrix 30 (or other array of combined spectral data) that identifies the type of manufacturing process that was performed immediately before the spectral data included in matrix 30 (or other array of combined spectral data) was acquired.

[0048] Each of the horizontal rows of matrix 30 represents a different die of the target substrate after the manufacturing process. For example, if 12 dies are inspected, matrix 30 has 12 rows and M is 12. Thus, in one example, the first row 32 of matrix 30 can represent all of the spectral data of the first die 12 to be inspected (FIG. 1), the second row 34 of matrix 30 can represent all of the spectral data of the second die 12 to be inspected (FIG. 1), and so on. Each die to be inspected has a unique position on the substrate, i.e., a die position, that is specific to that die.

[0049] Each of the vertical columns of matrix 30 represents a different spectral wavelength. For example, the first column 36 can represent a first wavelength, the second column 38 can represent a second wavelength different from the first wavelength, and so on. In other examples, each column represents a different spectral frequency.

[0050] Each cell of the matrix 30 includes the reflected signal intensity of the corresponding die measured at the corresponding wavelength at the corresponding inspection time point. Thus, for example, the value x in the cell 40 corresponding to row 32 and column 36 of the matrix 30 corresponds to the reflected signal intensity of the first die 12 to be inspected (FIG. 1) measured at the first wavelength at inspection time point 1. The value y in the cell 42 corresponding to row 32 and column 38 of the matrix 30 corresponds to the reflected signal intensity of the first die 12 to be inspected measured at the second wavelength at time point 1. The value z in the cell 44 corresponding to row 34 and column 36 of the matrix 30 corresponds to, for each of the dies 12 to be inspected on the substrate (FIG. 1) and for each of the wavelengths to be inspected in the spectrum, the reflected signal intensity of the second die to be inspected measured at the first wavelength at time point 1, etc.

[0051] In some examples, the matrix generated in accordance with the present disclosure can be assembled into a three-dimensional structure having a third dimension corresponding to an inspection channel such that each cell of each matrix includes a value corresponding to the spectral signal intensity at a given wavelength of a given channel of a given die at a given inspection time point. In some examples, the three-dimensional structure is converted into a two-dimensional matrix, for example, as described above, by performing one or more statistical operations on data over a plurality of channels for a given die, a given wavelength, and a given inspection time point.

[0052] The number of wavelengths (or, in some examples, frequencies) to be inspected can be, for example, about 10, about 100, or about 1,000, or more. In some examples, there are about 1,000 wavelengths to be inspected for each spectrum (N is about 1,000), each die at each inspection time point is associated with about 1,000 spectral data points, the matrix 30 includes about 1,000 columns, and each data point is the signal intensity measured by a reflectance measurement device or a spectroscopic measurement device at a given wavelength.

[0053] A completed matrix 30 (e.g., where each cell contains signal strength data), or another array of combined spectral data, may be provided as an input to a machine learning model to generate a prediction of the substrate yield. In some examples, the machine learning model may be trained via a supervised training method based on previous matrices (or other arrays of combined spectral data) from substrates with known yields. In some examples, the machine learning model is a convolutional neural network. In some examples, the training matrix is the same size as matrix 30.

[0054] The matrix (or other array of combined spectral data) input to the machine learning model may be augmented at different manufacturing steps. That is, after one or more other manufacturing steps, another set or sets of spectral data can be generated by inspecting the same die of the substrate at a second inspection time point and optionally one or more other subsequent inspection time points. For example, after the first layer of the substrate is manufactured, 12 dies can be inspected. Then, using the spectral data generated from the inspection at time point 1, a 12-row matrix 30 (FIG. 3) can be generated as described above.

[0055] After another layer of the substrate is manufactured, or after another manufacturing step (e.g., etching) is performed and before the next manufacturing step or the next other step that physically changes the substrate, a second inspection of the 12 dies may be performed at inspection time point 2 (or time point 2). The matrix 30 (FIG. 3) generated from the first inspection can be augmented with the spectral data from the second inspection to generate an augmented matrix 60 (FIG. 4). For example, another 12 rows, i.e., one additional row for each die being inspected, may be added to matrix 30 to generate matrix 60. Thus, the first row 62 of matrix 60 represents the spectral data of the first die after the first manufacturing step (i.e., at inspection time point 1), and the thirteenth row 64 of matrix 60 represents the spectral data of the first die after the second manufacturing step (i.e., at inspection time point 2).

[0056] In some examples, the same die is inspected at a first inspection time and a second inspection time. In some examples, the same number of dice are inspected at the first inspection time and the second inspection time, but the dice inspected at the second inspection time do not all have to be the same as the dice inspected at the first inspection time.

[0057] The extended matrix 60 (or other extended array of the distinguished combined data subset) may be provided as an input to a machine learning model (e.g., a CNN) to predict the yield of the substrate after a second manufacturing process. The process can continue to extend the matrix performed after each of any desired number of manufacturing processes and corresponding substrate inspections. Thus, an updated yield prediction can be made after each manufacturing process using the further extended data. When each extension is supplied to the machine learning model, the machine learning model compares the input to a training matrix (or other data array) having a corresponding extension level after the same manufacturing process as the substrate in question.

[0058] As an alternative to extending the matrix 30 of FIG. 3 to include spectral data after another manufacturing process, separate matrices can be generated at each inspection time and input to the machine learning model as separate matrices to be compared with corresponding training matrices.

[0059] As a further alternative to extending the matrix 30 of FIG. 3 to include spectral data after another manufacturing process, only the matrix 60 may be generated after both manufacturing processes without first generating the matrix 30.

[0060] In some examples, a first predicted yield of a substrate is generated by inputting matrix 30 (or another array of combined spectral data) into a machine learning model. Next, another manufacturing process is performed on the substrate and another inspection is performed. Next, spectral data from the another inspection is added to matrix 30 (or another array of combined spectral data) to generate an extended matrix 60 (or another extended array of combined spectral data). Next, the extended matrix (or another extended array of combined spectral data) is input into the machine learning model to modify the first predicted yield and output a second predicted yield of the substrate that is a modification of the first predicted yield based on the extension of the matrix (or another array of combined spectral data).

[0061] A matrix (e.g., matrix 30) (or another array of combined spectral data) or an extended matrix (e.g., matrix 60) (or another extended array of combined spectral data) having spectral data of dies of a target substrate is input into a machine learning model, and the machine learning model performs one or more operations (e.g., matrix operations) on the input to classify the input. The one or more operations can generate one or more values that can output a predicted yield of the target substrate by comparing with corresponding values of training substrates having known yields.

[0062] A matrix (e.g., matrix 30) or an extended matrix (e.g., matrix 60) can be constructed within a data file using delimiters for separating columns and rows.

[0063] By predicting a substrate yield during the manufacture of a substrate using the systems and methods of the present disclosure, advantageous efficiencies can be obtained. For example, a manufacturing tool or process that causes defects can be identified and corrected or resolved earlier, e.g., before completing the manufacture of a substrate or a batch of substrates. Additionally, an approximate number of substrates that may be affected by the defects can be determined at a relatively early stage of the manufacturing process.

[0064] In addition, by re-evaluating the substrate yield prediction after each of a plurality of manufacturing steps of a substrate, the accuracy of the substrate yield prediction can be improved, which can further result in improved efficiency, minimization of unnecessary discard of acceptable substrates, and increased efficient discard of substrates that may have been given an improper pass score based on less inspection data. For example, the effect of defects after a given manufacturing step may be affected (e.g., reduced or enlarged) after subsequent manufacturing steps. Embodiments of the present disclosure are adapted to capture and evaluate such effects and refine the substrate yield prediction using an extended spectral data matrix (or other data array) as described herein.

[0065] FIG. 5 shows an exemplary three-dimensional structure 70 of spectral data that can be used to generate a substrate yield prediction according to the present disclosure. The three-dimensional structure 70 includes a series of two-dimensional matrices, each two-dimensional matrix including spectral data from dies of a substrate acquired at different times after the same fabrication process. That is, each two-dimensional matrix 30 represents data from different channels. Referring to FIG. 5, the (x)-axis corresponds to the spectral wavelength, similar to the rows of each matrix 30 (FIG. 3). The (y)-axis is a spatial axis in that it corresponds to different dies on the substrate, similar to the columns of each matrix 30 (FIG. 3). The (z)-axis is the channel axis, and each channel corresponds to measurements taken at different times.

[0066] In some examples, the two-dimensional matrices of the three-dimensional structure 70 correspond to a plurality of matrices 30 (FIG. 3), and for each of such matrices 30, signal intensity measurements acquired at different times, i.e., different channels, for a single inspection time are input.

[0067] The completed three-dimensional structure 70 (e.g., having all cells of all two-dimensional matrices 30 including signal strength data) may be provided as an input to a machine learning model to generate a prediction of the substrate yield. In some examples, the machine learning model may be trained via a supervised training method based on previous three-dimensional structures from substrates having known yields. In some examples, the machine learning model is a convolutional neural network. In some examples, the training structure is the same size as structure 70 (e.g., the same number and size of two-dimensional matrices).

[0068] FIG. 6 schematically illustrates an exemplary system 100 for predicting the yield of a semiconductor substrate according to the present disclosure.

[0069] System 100 includes a computing device 202. The computing device 202 may be a server and / or other computing device that executes operations described herein, such as the substrate yield prediction operation described herein. The computing device 202 may include computing components 206. The computing components 206 include at least one processor 208 and a memory 204. The memory 204 can include a non-transitory computer-readable medium. Depending on the exact configuration, the memory 204 (especially, storing substrate yield prediction instructions and instructions for performing other operations disclosed herein) may be volatile (such as RAM), non-volatile (such as ROM, flash memory, etc.), or some combination of the two. Further, the server 202 can also include storage (removable storage 210 and / or non-removable storage 212) including, but not limited to, solid state devices, magnetic disks or optical disks, or tapes. Further, the computing device 202 can also have an input device 216 such as a touch screen, keyboard, mouse, pen, voice input, etc., and / or an output device 214 such as a display, speaker, printer, etc. One or more communication connections 218 such as a local area network (LAN), wide area network (WAN), point-to-point, Bluetooth®, RF, etc. can be incorporated into the computing device 202.

[0070] System 100 can include one or more die inspection devices 102 that are operably communicating with the computing device 202, such as being connected via a network. Non-limiting examples of the die inspection device 102 are reflectometers or spectrometers that measure the intensity of light or other waves (such as sound waves) reflected from the substrate at different wavelengths and generate spectral data such as the spectral data shown in FIG. 2.

[0071] In an alternative configuration, one or more components of the computing device 202 are present locally on one or more die inspection devices 102. For example, one or more die inspection devices 102 may be configured to perform one or more of the substrate yield operations described herein themselves. That is, the substrate yield prediction instructions can be executed directly on one or more die inspection devices 102.

[0072] FIG. 7 shows an example of using a machine learning model to predict substrate yield according to an example of the present disclosure using the system of FIG. 6. In some examples, one or more of the operations performed in FIG. 7 may be performed by the computing device 202 of FIG. 6 and / or one or more die inspection devices 102.

[0073] Referring generally to FIG. 7, an input 302 is provided to a machine learning model 308, such as a CNN. The CNN is configured to analyze the input to generate an output 310 that can be provided to a technician using the computing device 202 of FIG. 6 (e.g., can be displayed on a display device).

[0074] The input 302 includes training data (such as a training matrix) and known yields 304, as well as test data or target data (e.g., a test matrix or target matrix) 306. The training data can include an extended combination and / or an unextended combination of spectral data of training substrates (in the form of a matrix as described above) having known substrate yields corresponding to each training data input.

[0075] The test data (e.g., a test matrix) 306 can be an extended combination or an unextended combination of spectral data of a target substrate (in the form of a matrix as described above) having an unknown yield.

[0076] The machine learning model 308 is trained using training data and a known yield 304. The machine learning model 308 operates on test data 306 using a training matrix and the known yield 304 to generate an output 310.

[0077] The output 310 includes a predicted yield 312 of the target substrate corresponding to the test data 306. In some examples, the predicted yield can also include an indication as to whether the predicted yield passes or fails one or more predetermined operation acceptabilities, quality acceptabilities, or other acceptabilities.

[0078] In some examples, the output 310 can include one or more corrective advisories 314. For example, if the predicted yield 312 correlates with a reject score of the substrate, one or more corrective actions can be proposed. The recommended corrections can be based on, for example, the extent to which the predicted yield does not reach a minimum acceptable threshold yield, the most recent manufacturing process performed, and / or other factors. For example, the corrective advisory 314 can be to recalibrate the substrate manufacturing tool used immediately prior to the generation of the test data 306. In another example, the corrective advisory 314 can be to discard the target substrate or a partially formed target substrate. Other corrective advisories 314 are possible.

[0079] FIG. 8 shows a method 400 for predicting the yield of a semiconductor substrate using the system of FIG. 6. It will be understood that different embodiments of the present disclosure can include different combinations of subsets of the steps of method 400, non-limiting examples of which are described herein.

[0080] In step 402 of method 400, the machine learning model is trained using spectral data (e.g., a spectral data matrix) generated by inspecting dies of training substrates having a known yield (e.g., using die inspection apparatus 102 (FIG. 6)).

[0081] In step 404 of method 400, a manufacturing process of the target substrate is performed.

[0082] In operation 406, a subset of the dies of the target substrate is inspected, for example, using die inspection apparatus 102 (FIG. 6).

[0083] In operation 408, based on the inspection performed in operation 406, spectral data is combined (e.g., in the form of a matrix such as spectral data matrix 30 (FIG. 3)). In some examples, operation 408 includes accessing the spectral data generated in operation 406 and generating combined spectral data using the accessed spectral data.

[0084] In some examples, following operation 408, method 400 proceeds to operation 416, where a machine learning model uses the spectral data generated in operation 408 and the training data obtained in operation 402 to predict the yield of the target substrate. In some examples, following operation 408, method 400 proceeds to operation 410.

[0085] In operation 410, another manufacturing operation is performed on the target substrate.

[0086] In operation 412, a subset of the dies of the same target substrate as in operation 406 is inspected, for example, using die inspection apparatus 102 (FIG. 6).

[0087] In operation 414, extended combined spectral data (e.g., an extended spectral data matrix such as matrix 60 (FIG. 4) or matrix 70 (FIG. 5)) is generated based on the inspection performed in operation 412. In some examples, operation 414 includes accessing the spectral data generated in operation 412 and generating extended spectral data using the accessed spectral data generated in operation 412.

[0088] In some examples, following operation 414, the method sequentially repeats operations 410, 412, and 414 one or more times to further extend the spectral data after further manufacturing operations on the target substrate.

[0089] Following operation 414 or the last instance of operation 414, in operation 416, the machine learning model uses the spectral data generated in the most recent iteration of operation 414 and the training data obtained in operation 402 to predict the yield of the target substrate. In some examples, operations 410, 412, 414, and 416 are repeated to predict the yield of the target substrate after each fabrication operation or after a specified number of fabrication operations.

[0090] In other examples, a predicted yield can be generated after operation 408 and, in some examples, that predicted yield can be refined following one or more instances of operation 414.

[0091] The embodiments described herein may be employed using software, hardware, or a combination of software and hardware to implement and execute the systems and methods disclosed herein. Throughout this disclosure, specific devices are recited as performing particular functions, but those of ordinary skill in the art will understand that these devices are provided for illustrative purposes and that other devices may be used to perform the functions disclosed herein without departing from the scope of this disclosure. Additionally, some aspects of this disclosure have been described above with reference to block diagrams and / or operational diagrams of systems and methods according to aspects of this disclosure. The functions, operations, and / or acts described in the blocks may be performed in an order different from the order shown in any respective flowchart. For example, two blocks shown in sequence may actually be performed or implemented substantially simultaneously or in the reverse order, depending on the related functions and implementations.

[0092] This disclosure describes some embodiments of the technology with reference to the accompanying drawings, which show only some of the possible embodiments. However, other aspects may be embodied in many different forms and should not be construed as limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of possible embodiments to those skilled in the art. Further, as used in this specification and the claims, the phrase "at least one of element A, element B, or element C" is intended to convey any of element A, element B, element C, element A and B, element A and C, element B and C, and element A, B, and C. Further, those skilled in the art will understand the degree conveyed by terms such as "about" or "substantially" in light of the measurement techniques utilized herein. Unless such terms can be clearly defined or understood by those skilled in the art, the term "about" shall mean plus or minus 10%.

[0093] Although specific embodiments are described herein, the scope of the technology is not limited to those specific embodiments. Further, different examples and embodiments can be described separately, but such embodiments and examples can be combined with each other when implementing the technology described herein. Those skilled in the art will recognize other embodiments or improvements within the scope and spirit of the technology. Accordingly, a particular structure, operation, or medium is disclosed only as an exemplary embodiment. The scope of the technology is defined by the following claims and any equivalents thereof.

Claims

1. 1. A method for predicting substrate yield, comprising: combining first spectral data for a first die of a plurality of dies on the substrate and second spectral data for the first die of the plurality of dies of the substrate into combined data, the first spectral data for the first die being generated from a first inspection of only the first die of the plurality of dies at a first time point and the second spectral data for the first die being generated from a second inspection of only the first die of the plurality of dies at a second time point, the second time point being different from the first time point; generating a predicted yield for the substrate based on the combined data, including providing the combined data as an input to a trained machine learning model; A method comprising:

2. The method of claim 1 , wherein the combined data includes a first data subset representing the first spectral data for the first die of the plurality of dies.

3. 3. The method of claim 2, wherein the combined data is organized such that the first data subset is distinct from a second data subset representing second spectral data for a second die of the plurality of dies.

4. The method of claim 3 , wherein each data value of the first data subset is associated with a wavelength of the first spectral data.

5. 5. The method of claim 4, wherein a data value in the first data subset represents a signal intensity of the first spectral data at the wavelength corresponding to the data value for the first die of the plurality of dies.

6. the first time occurs after a first manufacturing step is performed on the substrate and before a second manufacturing step is performed on the substrate; the second point in time occurs after the second manufacturing step is performed on the substrate. The method of claim 1.

7. extending the combined data by adding third spectral data for the first die of the plurality of dies, the third spectral data being generated from a third inspection of the first die of the plurality of dies at a third time to the combined data to form extended combined data; generating another predicted yield for the substrate, comprising providing the augmented combined data as an input to the trained machine learning model; Further comprising: the first time occurs after a first manufacturing step is performed on the substrate and before a second manufacturing step is performed on the substrate; the second point in time occurs after the second manufacturing step is performed on the substrate and before a third manufacturing step is performed on the substrate; the third point in time occurs after the third manufacturing step is performed on the substrate. The method of claim 1.

8. 2. The method of claim 1, wherein each of the first spectral data and the second spectral data includes additional spectral data for each of the plurality of dies, the plurality of dies being less than all dies on the substrate.

9. the combination data is extended combination data, the projected yield is a revised projected yield; The method further comprises, prior to the second test, generating unenhanced combined data including the first spectral data; providing the unaugmented combined data as an input to the trained machine learning model to generate a first predicted yield for the substrate; Further comprising: The method further comprises, after the second inspection, modifying the first predicted yield to generate the modified predicted yield. Further comprising: The method of claim 1.

10. A computing device, combining first spectral data for a first die of a plurality of dies of a substrate with second spectral data for the first die of the plurality of dies on the substrate into combined data, the first spectral data for the first die being generated from a first inspection of only the first die of the plurality of dies at a first time point and the second spectral data for the first die being generated from a second inspection of only the first die of the plurality of dies at a second time point, the second time point being different from the first time point; generating a predicted yield for the substrate based on the combined data, including providing the combined data as an input to a trained machine learning model; A non-transitory computer-readable medium comprising instructions to cause

11. the combined data includes a first data subset representing the first spectral data for the first die of the plurality of dies; the combined data is organized such that the first data subset is distinct from a second data subset representing second spectral data for a second die of the plurality of dies; each data value of the first data subset is associated with a wavelength of the first spectral data; a data value in the first data subset representing a signal intensity of the first spectral data at the wavelength corresponding to the data value for the first die of the plurality of dies; The non-transitory computer-readable medium of claim 10.

12. the first time occurs after a first manufacturing step is performed on the substrate and before a second manufacturing step is performed on the substrate; the second point in time occurs after the second manufacturing step is performed on the substrate. The non-transitory computer-readable medium of claim 10.

13. The computing device includes: extending the combined data by adding third spectral data for the first die of the plurality of dies, the third spectral data being generated from a third inspection of the first die of the plurality of dies at a third time to the combined data to form extended combined data; generating another predicted yield for the substrate, comprising providing the augmented combined data as an input to the trained machine learning model; further instructions configured to cause the the first time occurs after a first manufacturing step is performed on the substrate and before a second manufacturing step is performed on the substrate; the second point in time occurs after the second manufacturing step is performed on the substrate and before a third manufacturing step is performed on the substrate; the third point in time occurs after the third manufacturing step is performed on the substrate. The non-transitory computer-readable medium of claim 10.

14. 11. The non-transitory computer-readable medium of claim 10, wherein each of the first spectral data and the second spectral data includes additional spectral data for each of the plurality of dies, the plurality of dies being less than all dies on the substrate.

15. the combination data is extended combination data, the projected yield is a revised projected yield; The non-transitory computer readable medium may further include, prior to the second test, generating unenhanced combined data including the first spectral data; providing the unaugmented combined data as an input to the trained machine learning model to generate a first predicted yield for the substrate; further instructions configured to cause the The non-transitory computer readable medium may further include, after the second testing, modifying the first predicted yield to generate the modified predicted yield. and further instructions configured to cause the The non-transitory computer-readable medium of claim 10.

Images