Method and system for managing overlay alignment and computing system

Abstract

A method and system for managing overlay alignment and computing systems. This disclosure describes techniques for managing vertical alignment or overlay in semiconductor manufacturing using machine learning. Alignment of interconnect features in fan-out wafer level packaging processes is assessed and managed by the disclosed techniques. Big data and machine learning are used to train a classification that relates overlay error source factors to overlay metrology categories. The overlay error source factors include tool signals. Training the classification includes base classification and meta classification.

Description

Technical Field

[0001] The techniques described in the embodiments of the present invention generally relate to alignment or overlap in semiconductor manufacturing, and more specifically, to methods and systems for managing overlap alignments and computing systems. Background Technology

[0002] As semiconductor technology evolves, semiconductor dies are becoming smaller and smaller, while more and more functions are being integrated into a single die. Therefore, there is a need to include an increasingly larger number of I / O pads and a smaller area of ​​the die surface in integrated circuit (IC) packages. Fan-out wafer-level packaging (WLP) has emerged as a promising packaging technology to address this challenging situation. In fan-out WLP, the die is cleaved from the initial front-end die before being positioned on the carrier die, allowing for packaging using interconnect wiring and I / O pads. An advantage of the fan-out WLP process is that the I / O pads associated with the die can be redistributed to a larger area compared to the die's own surface. Therefore, the number of I / O pads packaged with the die can be increased.

[0003] Fan-out WLP packages can be used to package a single die, multiple dies side-by-side, or multiple dies in a vertically configured package-on-package (POP) configuration. The POP configuration in a fan-out WLP is achieved through interconnect features (e.g., vias) that vertically connect multiple dies. For example, with integrated fan-out InFO technology, the InFO_POP architecture contains DRAM dies connected via InFO through-InFO vias (TIVs), while the InFO_M package contains one or more dies that can be placed side-by-side, such as logic dies and memory dies.

[0004] Overlap metrology is used to monitor and control vertical alignment in various semiconductor manufacturing processes. Overlap metrology typically specifies the accuracy (e.g., vertical alignment) of the alignment of a first patterned layer or features thereon relative to a second patterned layer located at a different vertical level than the first patterned layer. Overlap error refers to the misalignment between a first portion on the first patterned layer and a second portion on the second patterned layer. Overlap error metrology (e.g., measurement) can be based on the offset between the first and second portions, or between the actual position of the first portion and a target position of the first portion. The target position can be determined based on advanced process control (APC) in wafer processing.

[0005] In fan-out WLP (Wafer-Loop Layer), tested and good dies are positioned onto a carrier wafer. Interconnect feature layers are formed to connect the dies to associated I / O pads and the various interconnect layers themselves. Interconnects are formed using wafer-level processes, where photoresist and photolithography processes are used similarly to those in front-end wafer fabrication. Therefore, it is necessary to manage the vertical alignment between or within consecutive interconnect layers. Summary of the Invention

[0006] This invention provides a method for managing overlap alignment, comprising: determining a tool signal relating to a wafer processing tool forming a first feature on a first wafer; determining a first value of an overlap measurement of the first feature; generating a first dataset containing the tool signal and the first value of the overlap measurement; generating a classification based on the first dataset that correlates the tool signal and the overlap measurement, the classification including a base classification and a meta-classification; and using the classification to estimate a second value of the overlap measurement of a second feature on a second wafer.

[0007] This invention provides a system for managing overlap alignment, comprising: a wafer processing tool operable to form a plurality of features on a wafer, the plurality of features including a first subset of features located in a first region on the wafer and a second subset of features located in a second region on the wafer; a metrology tool operable to measure a plurality of overlap metrology values ​​of the plurality of features on the wafer; an overlap modeling tool operable to generate a plurality of estimated overlap metrology values ​​of the plurality of features based on tool signal data of the wafer processing tool; and a process control tool operable to adjust the operation of the wafer processing tool based on an adjustment value based on the first region and an adjustment value based on the second region, the adjustment value based on the first region being the estimated overlap metrology value based on the first feature subset, and the adjustment value based on the second region being the estimated overlap metrology value based on the second feature subset.

[0008] This invention provides a computing system, including: a processor; and a storage unit having executable instructions stored thereon, the executable instructions configuring the processor, when executed by the processor, to perform actions including: receiving data regarding tool signals of a wafer processing tool forming a first feature on a first wafer; receiving data of overlap measurement of the first feature; generating a data pool including the data of the tool signals and the data of the overlap measurement; learning a stacking classification including a first classification and a second classification, the first classification and the second classification being learned based on a first data subset and a second data subset from the data pool; and using the stacking classification to estimate an overlap measurement of a second feature on a second wafer. Attached Figure Description

[0009] The various aspects of this disclosure will be best understood in conjunction with the accompanying drawings, based on the following detailed description. In the drawings, unless the context otherwise indicates, the same reference numerals identify similar elements or actions. The size and relative position of elements in the drawings are not necessarily drawn to scale. In fact, the dimensions of individual features may be arbitrarily increased or decreased for clarity of explanation.

[0010] Figure 1 It is an instance system used to manage overlapping alignments.

[0011] Figure 2 It is an instance error source dataset.

[0012] Figure 3 It is an instance overlap error dataset.

[0013] Figure 4 It is an instance machine learning unit.

[0014] Figure 5 These are instance operations of the machine learning unit.

[0015] Figure 6 It is an example of obtaining tool signal data.

[0016] Figure 7 It is a table of instances of signal data for plotting tools.

[0017] Figure 8 It is an instance process for managing overlapping alignments. Detailed Implementation

[0018] The following disclosure provides numerous different embodiments or instances for implementing various features of the described subject matter. Specific examples of components and arrangements are described below to simplify this description. These components and arrangements are, of course, merely examples and are not intended to be limiting. For example, in the following description, the formation of a first feature over or on a second feature may include embodiments where the first and second features are formed in direct contact, and may also include embodiments where additional features may be formed between the first and second features so that the first and second features are not in direct contact. Additionally, reference numerals and / or letters may be repeated in various instances of this disclosure. This repetition is for simplicity and clarity and does not in itself indicate a relationship between the various embodiments and / or configurations discussed.

[0019] Furthermore, for ease of description, spatial relative terms such as “below,” “under,” “lower,” “above,” and “upper” may be used herein to describe the relationship between one element or feature and another element(s) shown in the diagrams. In addition to the orientations depicted in the diagrams, the spatial relative terms are intended to cover different orientations of the device during use or operation. The device may be oriented in other ways (rotated 90 degrees or in other orientations), and the spatial relative descriptive terms used herein shall be interpreted accordingly.

[0020] In the following description, certain specific details are set forth to provide a thorough understanding of various embodiments of this disclosure. However, those skilled in the art will understand that this disclosure may be practiced without these specific details. In other instances, well-known structures associated with electronic components and manufacturing techniques have not been described in detail to avoid unnecessarily obscuring the description of embodiments of this disclosure.

[0021] Unless the context otherwise requires, throughout the specification and the foregoing claims, the word “comprising” and its variations (e.g., “comprises / comprising”) shall be interpreted in an open-ended, inclusive sense, meaning “including but not limited to”.

[0022] For example, the use of ordinal numbers such as first, second, and third does not necessarily imply a hierarchical meaning of order, but can simply distinguish multiple examples of actions or structures.

[0023] Throughout this specification, references to "an embodiment" or "an embodiment" mean that a particular feature, structure, or characteristic described in connection with that embodiment is included in at least one embodiment. Therefore, the phrases "in one embodiment" or "in an embodiment" appearing in various places throughout this specification do not necessarily refer to the same embodiment. Furthermore, in one or more embodiments, a particular feature, structure, or characteristic may be combined in any suitable manner.

[0024] As used in this specification and the appended claims, unless otherwise expressly stated, the singular forms “a / an” and “the” include plural indicators. It should also be noted that, unless otherwise expressly stated, the term “or” is generally used to mean “and / or”.

[0025] This disclosure describes techniques for managing vertical alignment or overlap in semiconductor manufacturing using machine learning. For example, the disclosed techniques are used to evaluate and manage the alignment of interconnect features in a fan-out WLP process. For example, in back-end wafer-level packaging, degraded tooling can cause a reduced ability to form features at intended locations on the wafer, resulting in overlap alignment errors. For example, alignment tools (e.g., aligners) are used to determine the location of exposure when patterning a mask layer for forming InFO through-hole (TIV) vias. Degraded alignment tools may have an increased variation in the tool output signal, resulting in an increased positional shift of the TIV formed by the alignment tool. Therefore, the TIV may be partially or completely misaligned with the corresponding connection feature (e.g., a bump) of the package to which the TIV should be connected. The disclosed techniques use machine learning to analyze the tool signal (e.g., the tool output signal) of the alignment tool relative to overlap measurement of the same tool, such that the tool signal is correlated with the overlap measurement value. The learned correlation, referred to as "classification," is used to predict the overlap measurement of features on the workpiece processed by the alignment tool. For example, the tool signal when the alignment tool processes the workpiece is obtained and applied to the learned classification to predict the overlap measurement of features on the workpiece.

[0026] Tool signals encompass any characteristics, properties, or parameters of the tool as it processes a workpiece to form features on it. Tool signals may include tool output signals, tool operating status signals, and tool positioning signals. Tool output signals may include the electrical characteristics of the tool's output signal, the physical characteristics of the tool's output signal, or a combination of both. Tool signals can be measured as the distinction between the tool's signal and a reference signal of the tool. In some embodiments, tool signals are classified as source factors for overlap measurements. However, tool signals do not necessarily cause or produce overlap measurements. Tool signals may only be statistically correlated with overlap measurements. These statistical correlations may be sufficient for predicting overlap measurements and correcting overlap errors, but they do not necessarily indicate any causal relationship between tool signals and overlap measurements.

[0027] Tool-related overlap factors include, but are not limited to, alignment output signal characteristics, light source position, mask position, lens distortion, scan direction, step position at exposure points, wafer stage position, or other tool signals. Tool signals can be detected by sensors embedded in or outside the tool.

[0028] Overlap measurement data can also be obtained. Overlap measurement data may include measured overlap alignment offset information, such as overlap error or acceptable overlap shift. For example, categories of overlap error include, for example, the magnitude of overlap alignment offset on the x and y axes, rotational offset between corresponding features, or overlap error, for example, the position relative to the wafer or relative to the die.

[0029] Overlapping source factor data (e.g., tool signals) and overlapping metric data are collected as corresponding pairs and stored in a data pool. In the data pool, one overlapping source factor data point corresponds to one overlapping metric data point, and these data points together form a data entry. In some embodiments, data entries in the data pool are, for example, randomly assigned to a training dataset, a validation dataset, and a test dataset. The training dataset is used to learn or fit parameters representing the relationship between the source factors and the overlapping metrics. The validation dataset is used to tune or adjust the classification parameters learned based on the training dataset. The test dataset is used to evaluate the generalization or performance of the final classification. In some embodiments, a stacked classification is used to represent the relationship between overlapping source factors and overlapping metrics. In these embodiments, the training dataset is used to train the parameters of the base classification. The validation dataset is used to train the meta-classifier. Specifically, overlapping source data in the validation dataset is input into the base classification, and actual overlap error data in the validation dataset is used as output to train the meta-classifier. After the parameters of both the base classification and the meta-classifier have been obtained through training, the test dataset is used to evaluate the generalization or performance of the stacked classification containing the base classification and the meta-classifier.

[0030] In some embodiments, a training dataset is fed into a machine learning model to learn parameters representing the relationship between source factors (e.g., instrumental signals) and overlapping measures. Random forest models, neural network models, inverse distance weighting k-nearest neighbor regression (IDW-kNN) models, or other suitable machine learning models can be used to correlate the overlapping error source factors and overlapping measures. As a result of the machine learning, a regression model function or decision tree is obtained.

[0031] The classification parameters can be validated using a validation dataset. The classification parameters can be validated in various ways, all of which are within the scope of this disclosure. For example, estimated overlap measures based on overlapping source data in the validation dataset can be compared with corresponding actual measures in the validation dataset. The difference between the estimated and actual measures is used to adjust the classification parameters. Following the validation process, the classification parameters are finalized.

[0032] The test dataset is used to evaluate whether the determined regression model function fits the data in the test dataset. For example, in the case of a linear regression function, the R-squared value is used to determine whether the final classification fits the data in the test dataset. The R-squared value indicates the scattering of the data points in the test dataset around the fitted regression line, where the linear regression line assesses the change between each data point and the predicted value. In some embodiments, a threshold may be used to determine R. 2Whether the value meets a threshold. In another example, a threshold is used to evaluate each predicted overlap value and the actual overlap measurement value in a test dataset. The threshold can be determined based on the overlap tolerance requirements of features or layers formed in the wafer processing procedure. For instance, in a fan-out WLP, a threshold of approximately 0.1 micrometers can be selected to determine whether the estimated overlap measurement value meets the actual measurement data of the interconnect features on the fan-out WLP.

[0033] After the final classification has been determined to be a good fit, it is used to predict the overlap measurement values ​​for features on the workpiece (e.g., interconnect layers formed in a fan-out WLP packaging process). Corrective adjustments to tooling settings or other parameters, or settings used when forming features on the workpiece, can be determined based on the predicted overlap measurement values.

[0034] In the disclosure herein, the overlapping metrology process is illustrated relative to a stepper machine as an example processing tool for generating metrology errors, which does not imply limitation of the scope of this disclosure. For example, an example stepper machine is a mask alignment machine. The disclosed techniques can also be used to predict overlapping metrology relative to other bump-forming alignment tools, such as die-attach-to-die shifting metrology or ball-mount template alignment metrology. Furthermore, the disclosed virtual overlapping metrology techniques can be applied relative to front-end semiconductor process alignment tools or processes. Deviation tool signals (e.g., the difference between a detected characteristic of the stepper machine and a reference characteristic) are used as example tool signals to describe the technique. Other tool signals are also possible, and all such tool signals are included within the scope of this disclosure.

[0035] Figure 1 It is the instance overlap management system 100. For example... Figure 1 As illustrated, system 100 includes a wafer fabrication field system 110, an input dataset 120, a big data unit 130, a machine learning unit 140, and an output dataset 150. The wafer fabrication field system 110 includes a wafer processing toolset 112 (e.g., a lithography toolset 112), a process log 114, and a metrology toolset 116. The input dataset 120 includes overlap error data 122, error source data 124, and other data 126. In addition to the wafer processing data obtained and maintained by the process log 114, the other data 126 may include historical wafer processing data, historical error source data, and other wafer processing data. The machine learning unit 140 may include an alignment control unit 142 and a verification unit, or work in conjunction with the alignment control unit 142 and the verification unit. The output dataset 150 includes error prediction data 152, corrective adjustment data 154, and other output data.

[0036] In operation, wafer processing toolkit 112 is configured to process a wafer (e.g., a carrier wafer on which a die is positioned) in an example fan-out WLP process to form an interconnect layer over the die. Wafer processing toolkit 112 may be a photolithography tool and includes, for example, a stepper (e.g., a stepper and repeat camera), a wafer stage or chuck, and stage tools, as well as other suitable tools. The stepper passes light through a mask, thereby forming an image of the mask pattern. The image is focused and reduced by a lens and projected onto the surface of a wafer coated with photoresist. The stepper operates in a step-repeating manner, wherein the pattern on the mask is repeatedly exposed to the surface of the wafer in a grid pattern. The stepper moves the wafer via the wafer stage, as a step size from one shooting position to another. The stepping movement of the wafer can be reciprocated and lateral, with the grid below the stepper's lens. In fan-out WLP, the stage picks up the die and positions it at designated grid points or areas on the carrier wafer. In some cases, more than one bench tool is used to position the die onto the carrier wafer.

[0037] Process log 114 is configured to monitor, measure, determine, and record wafer processing parameters related to toolset 112 and the wafer processed by toolset 112. Some or all of the wafer processing parameter data are identified as contributing factors or related to overlap errors and are merged into error source data 124.

[0038] After the wafer is processed by wafer processing toolset 112 via layers or features on layers, metrology toolset 116 is configured to measure the wafer to determine overlap metrology data, such as whether overlap error exists on the wafer relative to layers or features on layers. The metrology data also includes details of the overlap error. In some embodiments, the overlap error metrology data is categorized as the location of the overlap error on the wafer or a die, the offset dimension of the overlap error on the x-axis and / or y-axis, or the rotation angle of the overlap error. Some or all of the measurement results from metrology toolset 116 are combined into overlap error data 122. Cases without overlap error are also useful and can be collected as part of overlap error data 122 for the purpose of modeling the correlation between error source factors and overlap error categories.

[0039] Big Data Unit 130 is configured to collect overlap error data 122, error source data 124, and other data 126, and is configured to combine or coordinate various datasets for further analysis. For example, Big Data Unit 130 links different data categories of overlap error data 122, error source data 124, and other data 126 in various ways relative to wafer layers, locations on wafer layers, structures on wafer layers, or multiple vertical structures designed to overlap vertically or otherwise connect or couple to each other. In the description herein, "feature" on the wafer is used to refer to any part of the wafer. For example, a feature on the wafer could be an interconnect structure. Overlap error data 122 is the overlap error of the interconnect structure, and the corresponding error source data 124 is a tool signal obtained or recorded with respect to the wafer processing toolset 112 that forms the interconnect structure.

[0040] Big Data Unit 130 also coordinates collected data regarding measurement dimensions, units of measurement, time scales, tolerance thresholds, etc., so that data entries can be used in the same analysis process. Big Data Unit 130 also incorporates collected data regarding missing data entries and data entry interpolation. Other data processing, coordination, or combination techniques are also possible and included within Big Data Unit 130.

[0041] Other data 126 may be historical data about overlap error data or error source data about previously processed wafers. Such historical data is processed in the same or similar manner as the overlap error data 122 and error source data 124 in the big data unit 130.

[0042] Machine learning unit 140 is an artificial intelligence machine learning unit configured to perform a learning process using data provided by big data unit 130 to infer classifications that can predict overlap errors of features on a semiconductor wafer based on measured processing parameters (e.g., error source data) of an instance stepper machine used to perform feature lithography. Overlap error data 122 and error source data 124 can be divided into three datasets: a training dataset, a validation dataset, and a test dataset. Machine learning unit 140 can use the different datasets to train a base classifier, a meta-classifier, and test the stacking adaptability of the base classifier and meta-classifier in predicting overlap errors as described herein.

[0043] Alignment control unit 142 is configured to control overlap alignment on wafer processing (e.g., wafer processing of interconnects on a fan-out WLP) using a regression model function or overlap estimation / prediction results. Specifically, alignment control unit 142 may apply processing log information of the wafer processing operation as input to the regression model function to generate alignment control output data 150. The alignment control output data can then be used to control or adjust the wafer processing. For example, among other things, alignment control output data 150 includes error prediction data 152 and corrective adjustment data 154, etc. Error prediction data 152 indicates the predicted / estimated overlap measurement in one or more wafer positions. Corrective adjustment data 154 indicates a change in one or more error source factors in error source data 124 to eliminate estimated / predicted overlap errors.

[0044] In one embodiment, the alignment control unit 142 is part of the machine learning unit 140. In other embodiments, the alignment control unit 142 is a unit independent of the machine learning unit 140 and includes a mechanism for automatically adjusting or causing adjustment of one or more of the wafer processing tool settings or wafer position to eliminate overlap errors.

[0045] The validation unit of machine learning unit 140 is configured to validate or adjust the parameters of a regression model function in the validation dataset. The regression model function can be validated in various ways. For example, estimated overlapping measures based on overlapping source factors in the validation dataset can be compared with corresponding actual measures in the validation dataset. The difference between the estimated and actual measures is used to adjust the classification parameters of the regression model function. After the validation process, the classification parameters of the regression model function are finalized.

[0046] The test unit is configured to evaluate whether the defined regression model function fits the data in the test dataset. For example, when the defined regression model function is a linear regression function, the R-squared value Rsquared is... 2 (R 2 = Explained Change / Total Change) is used to determine whether the final regression model fits the data in the test dataset. For example, a threshold can be used to determine R0. 2 The value is large enough to indicate that the final regression model function is sufficient to fit the test dataset. For example, a threshold is used to evaluate each predicted overlap value and actual overlap measurement value in the test dataset. The threshold can be determined based on overlap tolerance requirements for features or layers on the workpiece. For example, in a fan-out WLP, a threshold of approximately 0.1 micrometers can be chosen to determine whether the estimated overlap measurement value meets the actual measurement data for the interconnect features above the die encapsulated in the fan-out WLP.

[0047] Figure 2Example error source data 124 is illustrated. Error source data 124 includes a subset 210 of tool signal error source factors, a subset 250 of wafer / die alignment error source factors, and a subset 280 of context error source factors. Tool signal subset 210 contains data on tool characteristics or parameters or settings that affect overlap alignment, or are related to overlap alignment of features on the workpiece. In an embodiment, tool alignment subset 210 includes tool output signal data 212, wafer stage position data 214, step position data 216, mask button position data 218, exposure position data 220, or other suitable tool signal data. Wafer / die alignment subset 250 contains data on the position of the wafer on the wafer stage or the die on the wafer that affects overlap alignment (in the case of fan-out WLP). Context error source subset 280 contains contextual information about overlap metrology measurements. With contextual information included, different types of wafers and different wafer processing parameters can be identified and considered during machine learning. In an embodiment, regression model functions are determined or trained separately for different types of wafers and / or different wafer processing parameters. For example, maintain different training datasets for overlap error source factors and overlap errors for each type of wafer and / or each wafer processing parameter set. Update the different training datasets separately to maintain a fixed number of data entries in each training dataset.

[0048] In some embodiments, tool output signal data 212 is deviation data representing the difference between the detected tool signal value and the reference tool signal value. For example, tool output signal data 212 is the difference between the detected tool output signal value and the reference tool output signal value; wafer stage position data 214 is the difference between the detected wafer stage portion and the reference wafer stage position; step position data 216 is the difference between the detected step position and the reference step position; mask button position data 218 is the difference between the detected mask button position and the reference mask button position; and exposure position data 220 is the difference between the detected exposure position and the reference exposure position.

[0049] The reference tool signal value can be predetermined or determined dynamically. In some embodiments, the optimal historical signal value of the tool can be dynamically determined and used as the reference tool signal value. The optimal historical signal value can be determined among various categories of tool signals.

[0050] The wafer stage position data 214 can be measured by the position of the actuators anchoring the wafer stage. For example, the wafer stage may include three actuators for anchoring the wafer stage's x-axis, y-axis, and z-axis positions. The positions of the x-axis, y-axis, and z-axis actuators are recorded and entered to indicate the wafer stage position.

[0051] Step position data 216 represents the position of the stepper.

[0052] Mask button position data 218 indicates the position of the mask anchored by the mask button.

[0053] Exposure position data 220 indicates the location or area on the wafer exposed to illumination light.

[0054] The stage position 214, step position 216, mask button position 218, or exposure position 220 can be represented as positions on the x-axis, y-axis in the horizontal plane, and / or the z-axis in the vertical plane. In some embodiments, the stage position 214, step position 216, mask button position 218, or exposure position 220 can each be represented as a positional error of the corresponding target position. In measuring the positional error, x-axis offset, y-axis offset, z-axis offset, and rotation (angle) error can be used. In some embodiments, the target position can be used as a reference position.

[0055] In wafer / die alignment subset 250, wafer offset data 252 indicates the offset of the wafer on the wafer stage. The wafer offset includes offsets on the x-axis and y-axis. In embodiments, the wafer offset is determined by wafer alignment marks included on the wafer or other suitable mechanisms.

[0056] The wafer rotation data 254 indicates the rotation angle of the wafer on the wafer stage. In an embodiment, the wafer rotation is determined by a wafer alignment mark or other suitable mechanism.

[0057] Die offset data 256 indicates the amount of offset of the die on the carrier wafer. Die offset includes x-axis offset and y-axis offset. In an embodiment, die offset is determined by die alignment marks included on the die or other suitable mechanisms.

[0058] Die rotation data 258 indicates the rotation angle of the die on the carrier wafer. In an embodiment, die rotation is determined by die alignment marks or other suitable mechanisms.

[0059] Context data 280 may include focus depth 282, exposure duration 284, step speed 286, illumination setting 288, illumination source 290, EGA position 292, field position 294, metering position 296, or other context data. In some embodiments, the context data is used to ensure that the learned regression model function between the tool signal and the overlap error metering is specific to the type of wafer, process, characteristics, or other context factors.

[0060] It should be understood that Figure 2 The example error source data categories listed are examples and do not limit the scope of this disclosure. Other factors that contribute to the occurrence and / or magnitude of overlap errors may also be used and are included in the error source data 124, all of which are included in this disclosure.

[0061] Figure 3 The example overlap error data is shown as 122. Figure 3 As illustrated, the overlap error data 122 includes four overlap error categories: overlap measurement position data 310, x-axis overlap error data 320, y-axis overlap error data 330, and overlap rotation (angle) data 340. Overlap measurement position data 310 indicates the location of the overlap error on the wafer or die. x-axis overlap error data 320 and y-axis overlap error data 330 indicate the magnitude of the overlap error on the x-axis or y-axis, respectively. Overlap rotation (angle) data 340 indicates the overlap error where the actual measurement of the feature deviates from the target measurement by an angle.

[0062] Various methods can be used to define overlap error. In one embodiment, overlap error is determined as the difference between the actual measurement and the target measurement of a feature. In another embodiment, overlap error is determined based on the alignment between corresponding features (e.g., upper and lower features configured to be vertically aligned with each other). Other methods for determining overlap alignment accuracy or overlap error are also possible and are included within the scope of this disclosure.

[0063] Figure 4 Example of Machine Learning Unit 140. See also... Figure 4 The machine learning unit 140 includes: a processing unit 410, such as a computer processor or processing power allocated to the machine learning unit 140 in a virtual machine application; a storage unit 420 having a machine learning application 430 stored thereon; a communication unit 440 configured to communicate with other computers or machines linked to the machine learning unit 140 in a distributed computing environment; an interfacing unit 450 configured for input, output, and user interaction; and other components 460.

[0064] Application 430 includes executable instructions that, when executed by processing unit 410, configure processing unit 410 to implement training set generation module 432, learning module 434, and prediction module 438. Learning module 434 includes a basic learning module 435, a meta-learning module 436, and a testing module 437. In embodiments, executable instructions dedicated to implementing training set generation module 432, learning module 434, and prediction module 438 are stored in a separate dedicated space of storage unit 420 or stored in a separable / resolvable manner to facilitate easy identification of the executable instructions for these modules through parsing or indexing.

[0065] Figure 5The diagram illustrates the operational structure of an instance of the learning module 434 in application 430. In operation, the training set generation module 432 receives data from the big data unit 130 and generates three disjoint datasets in operation 510: a training dataset, a validation dataset, and a test dataset. Each of the training, validation, and test datasets contains similar categories of data entries containing error source data and overlapping error data. In some embodiments, the learning module 434 uses one or more of the training, validation, and test datasets to implement a multi-layered stacking of the learning process.

[0066] In instance operation 520, the basic learning module 435 uses a training dataset to train multiple basic classifications. Each of the multiple basic classifications represents the correlation (e.g., regression or decision tree) between error source data and overlapping error data. In some embodiments, the basic learning unit 435 uses a random forest model to train the parameters of each of the multiple basic classifications in the training dataset. For example, error source data (e.g., process parameters of the wafer processing toolkit 112 when processing semiconductor wafers) in the training dataset is used as input, and corresponding overlapping error data (e.g., measured overlapping data of features on the semiconductor wafer) in the training dataset is used as output to train the parameters of the basic classifications under the random forest model. The random forest model finds a classification representing the magnitude or mean prediction of the overlapping errors of the multiple classifications among multiple classifications (or decision trees) that link the error source data and the overlapping error data respectively. For example, in the random forest model, a large number of decision trees (e.g., ...) are trained. Figure 5 The diagram illustrates 500 decision trees (tree 1 to tree 500), each guiding a "path" from the input of an overlapping source to the output of an overlapping prediction. The parameters of each tree are determined or trained using a training dataset. In some embodiments, the magnitude, median, or mean of the outputs of a large number of decision trees responding to the same input can be considered as the output of a random forest model.

[0067] In instance operation 530, after the parameters of the base classification have been determined, the meta-learning module 436 trains the meta-classifier using a validation dataset. In some embodiments, error source data from the validation dataset (e.g., process parameters of the wafer processing toolset 112 when processing semiconductor wafers) are fed into the training base classification to predict overlap errors. The predicted or estimated overlap errors of the meta-classifier X or input data, classified as meta-classifier X, and the actual overlap measurement data from the validation dataset (e.g., measured overlap data of features on the semiconductor wafer), classified as meta-classifier Y or output data, are used together to train the meta-classifier. In some embodiments, the meta-learning module 436 uses inverse distance weighted k-nearest neighbor regression (IDW-kNN) as a model to represent the relationship between the predicted overlap error (or meta-X) generated by the base classification and the actual overlap error (or meta-Y) in the validation dataset. The parameters of the IDW-kNN meta-classifier are thus obtained.

[0068] In instance operation 540, test module 437 determines whether the stacked classification, which includes the base classification and the meta-classification, fits the test dataset. That is, test module 437 determines whether the predicted overlap error is consistent with or deviates from the actual measured overlap error. For example, error source data from the test dataset is fed into the base classification to obtain meta-X data. The meta-X data is then fed into the meta-classification to obtain predicted or estimated overlap error data, i.e., meta-Y. The predicted overlap error data is compared with the corresponding actual overlap measure data in the test dataset to assess whether the predicted overlap error is consistent with or deviates from the actual measured overlap error. In some embodiments, the R-squared value is calculated based on the predicted overlap error and the corresponding actual measured overlap error, and a threshold analysis is performed to determine whether the stacked classification fits the test dataset. Using the test dataset instead of the training dataset and / or validation dataset to evaluate the fitness of the stacked classification helps avoid overfitting problems.

[0069] If the stacked classification is determined to be well-fitted, it is ready to be used to predict overlap measurement error. If the stacked classification is determined to be poorly-fitted, one or more of the parameters of the base classification or meta-classifier can be adjusted to improve the fitness score of the stacked classification. For example, new training or validation data can be used to retrain the parameters of the base classification and / or meta-classifier.

[0070] In prediction operation 550, after the parameters of the base classification and meta-classification have been determined, the base classification and meta-classification are used together as a stack of classifications to predict overlap errors in the actual semiconductor chip manufacturing process. For example, prediction module 438 uses stack classification to predict or estimate overlap measurements of features on the wafer based on measured process parameters of the wafer processing operation that fabricates the wafer. Specifically, for example, error source data (e.g., tool signals) is obtained from processing log 114 and applied as input to train the stack classification to predict overlap measurement data as output. Prediction module 438 may also predict corrective adjustments to wafer processing parameters (e.g., tool alignment settings and / or wafer / die alignment factors, one or more) to eliminate or reduce the estimated / predicted overlap errors.

[0071] Overlap errors can be predicted across the entire wafer. Within the wafer, overlap errors may follow a consistent pattern or vary between / among different fields on the wafer. Using the stepping device used in wafer processing, wafer processing settings can be adjusted to specifically correct overlap errors in individual fields on the wafer. However, there are cases where overlap errors in two or more fields on the wafer cannot be corrected individually and are correlated under wafer processing operations. Decision-making operations can be performed to determine overall corrective adjustments to the wafer processing operations, thereby balancing overlap alignment between / among two or more fields.

[0072] In some embodiments, corrective adjustments are determined based on regions on the wafer. For example, features on the wafer are grouped into multiple subsets contained in corresponding regions on the wafer surface. These regions are annular and concentric. The predicted overlap error for each subset of features is averaged (e.g., mean or median) to obtain the average overlap error of the region. Corrective adjustments for the region are determined based on this average overlap error.

[0073] The result of the overlap prediction operation 550 is output to the alignment control unit 142 to control the wafer processing operation accordingly. For example, error prediction data 152 can be used to manually adjust wafer processing parameter settings to avoid or reduce overlap error problems. Corrective adjustment data 154 can be used by the operator or machine to automatically adjust wafer processing parameter settings, thereby improving overlap alignment accuracy. The adjusted processing parameters and related overlap measurement data are again fed into the machine learning unit 140 for further training or refinement of the classification, as they together represent new data entries. That is, the machine learning process in the machine learning unit 140 can be a dynamic process that continuously updates the classification using new data entries.

[0074] Figure 6This illustrates an example process for obtaining tool signal data as overlapping source data, which can be used in training datasets, validation datasets, and / or test datasets. As an illustrative example, Figure 6 The process of obtaining an instance output signal 1416 of a stepper machine (or alignment tool) as an instance of process toolset 112 is illustrated, and such instance does not limit the scope of this disclosure. Figure 6 As illustrated in (a), the example wafer 600 includes a plurality of alignment dies 602 (six are shown for illustration) designated for adjusting overlap alignment. Each alignment die 602 contains a plurality of alignment marks 604 arranged along the x-axis or y-axis. Each alignment mark 604 contains a plurality of alignment mark lines 606.

[0075] The output signal characteristics of a stepper motor affect exposure alignment in photolithography, and thus the position of features on the semiconductor wafer. Therefore, the output signal characteristics can be used as error source data. Figure 6 As illustrated in (b), the signal waveform of the alignment output signal on each alignment mark 604 is obtained as an example of the output signal characteristics. For illustration, Figure 6 (b) illustrates the signal waveforms of the two alignment marks (mark 3, mark 4) of the alignment die 602. Each waveform 610 contains multiple peaks 612 corresponding to the alignment mark line 606 (for illustration of the four peak points 612).

[0076] Ideally, the peaks of the waveform should align with the alignment marks. The alignment or misalignment between the waveform peaks and the corresponding alignment marks then serves as an indication of overlap error in the stepper machine's output signal. For example, as... Figure 6 As shown in (c), one or more of the peak points 612 are converted into position coordinates 614, 614(1), 614(2), 614(3), and 614(4) for illustration purposes. The position coordinates 614 obtained from the alignment signal output (referred to as "signal position coordinates") are compared with the corresponding position coordinates 616 of the alignment mark line 606 obtained from the Graphical Database System (GDS) data (referred to as "GDS position coordinates"). Specifically, the signal position coordinates 614(1), 614(2), 614(3), and 614(4) are compared with the GDS position coordinates 616(1), 616(2), 616(3), and 616(4) to determine the offset value between the output signal and the alignment mark line. The offset value is referred to as the signal characteristic.

[0077] Figure 7 This illustrates an example process for creating data entries containing the characteristics of the tool's output signal and the measured overlapping metrological data. Figure 7 In (a), for each alignment mark 604, multiple signal characteristic data 714, signal characteristic data 714(1), signal characteristic data 714(2), signal characteristic data 714(3), and signal characteristic data 714(4) are obtained for illustration purposes. Each of the signal characteristics 714(1), 714(2), 714(3), and 714(4) indicates the alignment or misalignment between the output signal and the alignment mark line (e.g., the difference between the signal position coordinates 614(1), 614(2), 614(3), and 614(4) and the corresponding GDS position coordinates 616(1), 616(2), 616(3), and 616(4).

[0078] Figure 7(b) illustrates data from multiple exposures performed by an alignment tool (e.g., shooting points 18, 16, 25, 26, 36, 35, 46, and 48 of the alignment tool). For each shooting point 18, 16, 25, 26, 36, 35, 46, and 48, overlap measurement data on the y-axis, OVL Box R / D, OVL Box L / U, and mean Y are obtained as the corresponding alignment output signal. OVL Box R / D and OVL Box L / U are the overlap errors on the y-axis obtained from two locations on the wafer. The mean Y is the average of OVL Box R / D and OVL Box L / U. For each of shooting points 18, 16, 25, 26, 36, 35, 46, and 48, tool signal data 724 (724(1), 724(2), 724(3), 724(4)) is also obtained. Specifically, the tool signal data 724 for each alignment mark 10, 12, 2, 4, 6, and 8 is obtained as the difference between the signal characteristics 714 of the specific shooting point and the signal characteristics 714 of the reference shooting point (here, shooting point 35). For example, for alignment mark 10, the signal characteristics 714(1), 714(2), 714(3), and 714(4) of shooting point 18 are compared with the signal characteristics 714(1), 714(2), 714(3), and 714(4) of reference shooting point 35, thereby generating tool signal values ​​"-15", "-18", "-18", and "-18" for four alignment mark lines. The "+" or "-" of tool signal 724 indicates a randomly specified direction of deviation (e.g., larger or smaller). It should be noted that for reference shooting point 35, since the corresponding signal characteristics themselves do not deviate, all tool signals are "0".

[0079] In some embodiments, the reference shooting point is selected as the shooting point of the aligner with the minimum overlap error. As an illustrative example, Figure 7The mean y-axis overlap error of shooting point 35 in (b) is "0.010066702", which is the smallest among shooting points 18, 16, 25, 26, 36, 35, 46, and 48. In some embodiments, the reference shooting point can be dynamically selected. For example, if new data is added to the data pool (containing shooting points better than the current reference shooting point 35), then the better shooting point can be selected as the new reference shooting point. Correspondingly, the signal characteristics 714 of the new reference shooting point will be used to generate a new tool signal 724 for each shooting point in the data pool.

[0080] Figure 8 This is a flowchart of instance operation process 800. In instance operation 810, process log 114 obtains tool signal data from the wafer processing tool (e.g., an aligner) regarding the formation of interconnect features on the first wafer.

[0081] In instance operation 820, the metrology toolset 116 obtains overlapping metrology data regarding interconnect features on the first wafer.

[0082] In instance operation 830, training set generation module 432 generates a dataset containing tool signal data and overlapping measurement data about the first wafer.

[0083] In instance operation 840, machine learning module 434 uses the dataset to generate training classifications that represent the correlation between instrumental signal data and overlapping econometric data. In some embodiments, the training classifications may comprise a stack of base classifications and one or more meta-classifications.

[0084] In instance operation 850, prediction module 438 predicts overlapping metrics on the second wafer based on the trained classification. For example, a tool signal about the second wafer can be applied to the trained classification to generate the predicted overlapping metrics.

[0085] In instance operation 860, alignment control unit 142 adjusts tool signals or other processing parameters when processing a second wafer, for example, by means of corrective adjustment data 154 based on predicted overlap measurements.

[0086] Finally, the training classification is updated based on the measured overlap measurement data and tool signal data of the second chip.

[0087] The technology is described using a fan-out WLP process as an example. It should be understood that the technology is similarly applicable to front-end wafer processing and other back-end wafer processing operations.

[0088] This disclosure can be further understood through the following description of embodiments.

[0089] In some embodiments, a method for managing overlap alignment includes: determining a tool signal relating to a wafer processing tool forming a first feature on a first wafer; determining a first value of an overlap measurement of the first feature; generating a first dataset comprising the tool signal and the first value of the overlap measurement; generating a classification based on the first dataset that correlates the tool signal and the overlap measurement, the classification comprising a base classification and a meta-classification; and using the classification to estimate a second value of an overlap measurement of a second feature on a second wafer.

[0090] In a relevant embodiment, the output of the basic classification is the input of the meta-classification.

[0091] In a related embodiment, the first dataset includes a training dataset and a validation dataset, the base classification is generated based on the training dataset, and the meta-classification is generated based on the validation dataset.

[0092] In a related embodiment, the method further includes using a test dataset to determine the fitness score of the classification, the test dataset being a part of the first dataset.

[0093] In a related embodiment, the tool signal is the difference between the signal of the wafer processing tool when it forms the first feature on the first wafer and the reference signal of the wafer processing tool.

[0094] In a related embodiment, the reference signal is a historical signal of the wafer processing tool forming a feature on the wafer with the smallest overlap error in history.

[0095] In relevant embodiments, the reference signal is determined dynamically.

[0096] In a relevant embodiment, the basic classification is a random forest model classification.

[0097] In a relevant embodiment, the meta-classification is inverse distance weighted k-nearest neighbor regression.

[0098] In related embodiments, the method further includes adjusting the operation of the wafer processing tool based on the estimated second value of the overlap measurement.

[0099] In a related embodiment, the operation of adjusting the wafer processing tool is based on estimated overlap measurements of multiple regions on the second wafer.

[0100] In a related embodiment, the estimated overlap metric of one of the plurality of regions is the average of one or more estimated second values ​​of the overlap metric on one or more second features within the region.

[0101] In some embodiments, a system for managing overlap alignment includes: a wafer processing tool operable to form a plurality of features on a wafer, the plurality of features including a first subset of features located in a first region on the wafer and a second subset of features located in a second region on the wafer; a metrology tool operable to measure a plurality of overlap metrology values ​​of the plurality of features on the wafer; an overlap modeling tool operable to generate a plurality of estimated overlap metrology values ​​of the plurality of features based on tool signal data of the wafer processing tool; and a process control tool operable to adjust the operation of the wafer processing tool based on an adjustment value based on the first region and an adjustment value based on the second region, the adjustment value based on the first region being the estimated overlap metrology value based on the first feature subset and the adjustment value based on the second region being the estimated overlap metrology value based on the second feature subset.

[0102] In a related embodiment, the overlap modeling tool is operable to generate a classification that correlates the tool signal data with overlap measurements of features on the wafer.

[0103] In related embodiments, the classification includes multiple basic classifications and meta-classifications, and the outputs of the multiple basic classifications serve as the inputs of the meta-classifications.

[0104] In relevant embodiments, the first region and the second region are concentric with each other.

[0105] In a related embodiment, the adjustment value based on the first region is the average of the estimated overlapping metrics of the first feature subset, and the adjustment value based on the second region is the average of the estimated overlapping metrics of the second feature subset.

[0106] In some embodiments, a computing system includes a processor and a storage unit having executable instructions stored thereon. When executed by the processor, the executable instructions configure the processor to perform actions including: receiving data regarding tool signals of a wafer processing tool forming a first feature on a first wafer; receiving data of overlap measurements of the first feature; generating a data pool containing the tool signal data and the overlap measurement data; learning a stack classification comprising a first classification and a second classification, the first and second classifications being learned based on a first subset of data and a second subset of data from the data pool; and using the stack classification to estimate an overlap measurement of a second feature on a second wafer.

[0107] In a related embodiment, the first classification correlates the tool signal with the overlap measurement.

[0108] In a related embodiment, learning the second classification includes: generating estimated overlapping measurement data by applying tool signal data from the second data subset to the first classification; and learning the second classification based on the estimated overlapping measurement data and the overlapping measurement data from the second data subset.

[0109] The various embodiments described above can be combined to provide other embodiments.

[0110] Based on the detailed description above, these and other changes can be made to the embodiments. Generally, the terminology used in the appended claims should not be construed as limiting the claims to the specific embodiments disclosed in this specification and claims, but should be understood to include all possible embodiments and the full scope of equivalents obtained under the ownership of the claims. Therefore, the claims are not limited by this disclosure.

Claims

1. A method for managing overlapping alignment, comprising: Determine the tool signal for the wafer processing tool that forms the first feature on the first wafer; Determine a first value for the overlap measurement of the first feature; Generate a first dataset containing the tool signal and the first value of the overlapping measurement; Based on the first dataset, a classification is generated that relates the tool signal and the overlap measurement. The classification includes a base classification and a meta-classification, wherein the base classification is trained to output an estimated overlap measurement value, and the meta-classification is trained to take the estimated overlap measurement value of the base classification as input and the actual overlap measurement as output. The base classification and the meta-classification are trained using different subsets of data from the first dataset. as well as The classification is used to estimate a second value of the overlap measurement of the second feature on the second wafer.

2. The method of claim 1, wherein in training the meta-classifier, the output of the base classification of the data subset used to train the meta-classifier is used as the input of the meta-classifier.

3. The method according to claim 1, wherein the first dataset is randomly assigned to a training dataset and a validation dataset, the base classification is generated based on the training dataset, and the meta-classification is generated based on the validation dataset.

4. The method of claim 3, further comprising using a test dataset to determine the fitness score of the classification, the test dataset being a portion of the first dataset.

5. The method of claim 1, wherein the tool signal is the difference between the signal of the wafer processing tool when forming the first feature on the first wafer and a reference signal of the wafer processing tool.

6. The method of claim 5, wherein the reference signal is a historical signal of the wafer processing tool forming on the wafer with a historically smallest overlap error.

7. The method of claim 6, wherein the reference signal is determined dynamically.

8. The method according to claim 1, wherein the basic classification is a random forest model classification.

9. The method according to claim 1, wherein the meta-classification is inverse distance weighted k-nearest neighbor regression.

10. The method of claim 1, further comprising adjusting the operation of the wafer processing tool based on the estimated second value of the overlap measurement.

11. The method of claim 10, wherein the operation of adjusting the wafer processing tool is based on estimated overlap measurements of a plurality of regions on the second wafer.

12. The method of claim 11, wherein the estimated overlap metric of one of the plurality of regions is the average of one or more estimated second values ​​of the overlap metric on one or more second features within the region.

13. A system for managing overlapping alignments, comprising: A wafer processing tool operable to form a plurality of features on a wafer, the plurality of features including a first subset of features located in a first region on the wafer and a second subset of features located in a second region on the wafer; A metrology tool operable to measure multiple overlapping metrological values ​​of the plurality of features on the wafer; An overlap modeling tool is operable to generate multiple estimated overlap measures of the plurality of features based on tool signal data of the wafer processing tool, wherein the overlap modeling tool is operable to generate a classification that correlates the tool signal data and the overlap measures of the features on the wafer, wherein the classification comprises multiple base classifications and meta-classifications, the outputs of the multiple base classifications being the inputs of the meta-classifications; as well as A process control tool operable to adjust the operation of the wafer processing tool based on an adjustment value based on a first region and an adjustment value based on a second region, wherein the adjustment value based on the first region is an estimated overlap measurement value based on a first feature subset, and the adjustment value based on the second region is an estimated overlap measurement value based on a second feature subset.

14. The system of claim 13, wherein the first region and the second region are concentric with each other.

15. The system of claim 13, wherein the adjustment value based on the first region is the average of the estimated overlapping metrics of the first feature subset, and the adjustment value based on the second region is the average of the estimated overlapping metrics of the second feature subset.

16. A computing system, comprising: processor; as well as A storage unit having executable instructions stored thereon, which, when executed by the processor, configure the processor to perform actions including the following operations: Receive data regarding tool signals from a wafer processing tool used to form a first feature on the first wafer; Receive the overlap measurement data of the first feature; Generate a data pool containing the data of the tool signal and the data of the overlapping measurement; The learning includes a stacked classification comprising a first classification and a second classification, the first classification and the second classification being learned based on a first subset of data from the data pool and a second subset of data different from the first subset of data, wherein the first classification is trained to output an estimated overlap measurement value, and the second classification is trained to take the estimated overlap measurement value of the first classification as input and the actual overlap measurement as output. as well as The stacking classification is used to estimate the overlap measurement of the second feature on the second wafer.

17. The computing system of claim 16, wherein the first classification correlates the tool signal and the overlapping measurement.

18. The computing system of claim 17, wherein learning the second classification comprises: The estimated overlapping measurement data is generated by applying the tool signal data from the second data subset to the first classification; and The second classification is learned based on the estimated overlapping metrics and the overlapping metrics of the second data subset.

Images