Semiconductor Manufacturing Metrology and Process Control

Abstract

A semiconductor metrology method is provided that optimizes travel and measurement time, tool stability, and tool repeatability. [Solution] The semiconductor metrology system includes a spectral acquisition tool configured to collect a baseline scatterometry spectrum on a first semiconductor wafer target using a first measurement protocol and to collect a variation set of scatterometry spectra embodying the spectral variability on a second semiconductor wafer target for various predetermined sources of spectral variability; a reference metrology tool that collects parameter values ​​of the first semiconductor wafer target using the second measurement protocol; and a training unit that predicts values ​​of the production semiconductor wafer target based on the spectra and trains a predictive model using machine learning to minimize an associated loss function incorporating a spectral variability term.

Description

[Technical field]

[0001] This invention relates to semiconductor manufacturing metrology and process control. [Background technology]

[0002] The shrinking dimensions and increasing complexity of today's semiconductor manufacturing processes are pushing metrology of such processes to their limits, making it extremely challenging to keep the metrology tools within specifications dictated by the stringent process limits. The accuracy, process robustness, precision, matching and other uncertainties involved in the metrology results are very difficult to achieve with current methods. Furthermore, metrics such as limited throughput (TPT) and within-wafer (WiW) sampling rates pose particular challenges given process control requirements. Finally, it is becoming increasingly difficult to obtain external reference data to train and / or test model-based metrology solutions.

[0003] Today, these difficulties are largely mitigated by hardware improvements that optimize "move and measure" (MAM) time, tool stability and tool repeatability, which also optimizes TPT and sampling rates. Process robustness is currently handled by searching for the most stable metrology recipe possible when creating the recipe (i.e., measurement protocol), but this can come at the expense of performance. Summary of the Invention

[0004] In its embodiments, the present invention presents a machine learning (ML) method based on supervised learning. In particular, given a data set of optical signals Si for each measurement sample i (e.g. for each die on a set of semiconductor wafers) and reference values ​​of parameters Pi, ML derives a yet to be measured signal S <new>P <predicted>A model is established to predict P. Model complexity can vary, and the choice of model will depend on the type and size of the data set used to train the model, its intrinsic sensitivity of S to P, and the type and amplitude of noise in the reference measurement relative to P.

[0005] In a first aspect of the present invention, there is provided a semiconductor metrology method comprising: collecting a baseline set of spectra on a first set of semiconductor wafer targets using a spectra acquisition tool and according to a first measurement protocol; collecting values ​​of predetermined parameters for the first set of semiconductor wafer targets using an optical metrology tool and according to a second measurement protocol for each of one or more predetermined sources of spectra variability; collecting a variability set of spectra on a second set of semiconductor wafer targets corresponding to the first set of semiconductor wafer targets using the spectra acquisition tool and according to the first measurement protocol, where the variability set of spectra embodies the spectra variability; and using the collected set of spectra and parameter values ​​to train a predictive model using machine learning and minimize a loss function associated with the predictive model, where the predictive model predicts production spectra for a third set of semiconductor wafer targets. the production spectra are collected using the spectral acquisition tool and according to the first measurement protocol, and the loss function is configured to be used to predict values ​​of any of the predetermined parameters using a loss function that incorporates, for each of the one or more predetermined sources of spectral variability,This is minimized by a term representing the spectral variability.

[0006] In another aspect of the invention, the predetermined source of spectral variability includes tool variability.

[0007] In another aspect of the invention, collecting the variation spectrum includes using a plurality of identical spectral acquisition tools to collect a variation spectrum from a selected one of the semiconductor wafer targets.

[0008] In another aspect of the invention, the predetermined source of spectral variability includes measurement repeatability.

[0009] In another aspect of the invention, collecting the variation spectrum includes collecting the variation spectrum from a selected one of the semiconductor wafer targets using the spectrum acquisition tool at a plurality of different times.

[0010] In another aspect of the present invention, the first and second measurement protocols differ in any of the number of channels, the illumination angle, the target, and the signals acquired from the same target.

[0011] In another aspect of the invention, the method further includes collecting production scatterometric spectra during production of a production semiconductor wafer and generating a predicted value of any of the predetermined parameters based on the production scatterometric spectra using a predictive model.

[0012] Another aspect of the invention further includes providing input to a semiconductor manufacturing tool to control operation of the semiconductor manufacturing tool during manufacturing of the manufactured semiconductor wafer.

[0013] In another aspect of the invention, there is provided a semiconductor metrology system comprising: a spectral acquisition tool configured to collect a baseline set of scatterometry spectra on a first set of semiconductor wafer targets according to a first measurement protocol, and for each of one or more predetermined sources of spectral variability, collect a variance set of scatterometry spectra embodying the spectral variability on a second set of semiconductor wafer targets corresponding to the first set of semiconductor wafer targets according to the first measurement protocol; an optical metrology tool configured to collect values ​​of predetermined parameters of the first set of semiconductor wafer targets according to a second measurement protocol; and a training tool configured to use the collected set of spectra and parameter values ​​to train a predictive model using machine learning and to minimize a loss function associated with the predictive model, wherein the predictive model is configured to be used to predict values ​​of any of the predetermined parameters using production spectra of a third set of semiconductor wafer targets, the production spectra being collected according to the first measurement protocol using the spectral acquisition tool, and the loss function being minimized by incorporating a term representing the spectral variability, for each of one or more predetermined sources of spectral variability.

[0014] In another embodiment of the present invention, the predetermined source of spectral variability includes tool variability.

[0015] In another embodiment of the invention, the spectral acquisition tool is configured to collect the variation spectrum from a selected one of the semiconductor wafer targets using multiple identical spectral acquisition tools.

[0016] In another embodiment of the invention, the predetermined source of spectral variability comprises measurement repeatability.

[0017] In another embodiment of the invention, the spectral acquisition tool is configured to collect the variation spectrum from a selected one of the semiconductor wafer targets using the spectral acquisition tool at a number of different times.

[0018] In another embodiment of the present invention, the first and second measurement protocols differ in any of the number of channels, the illumination angle, the target and the signal acquired from the same target.

[0019] In another embodiment of the invention, the spectral acquisition tool is configured to collect a production scatterometry spectrum during production of a production semiconductor wafer, and further comprises a prediction unit configured to use the prediction model to generate a predicted value of any of the predetermined parameters based on the production scatterometry spectrum.

[0020] In another embodiment of the invention, the system further includes a processing control unit configured to provide input to a semiconductor manufacturing tool based on the predicted value to control operation of the semiconductor manufacturing tool during production of the manufactured semiconductor wafers.

[0021] Aspects of the present invention will be more fully understood and appreciated from the following detailed description taken in conjunction with the accompanying drawings, in which: [Brief description of the drawings]

[0022] [Figure 1A] 1 is a simplified schematic diagram of a system for semiconductor metrology and process control, according to one embodiment of the present invention. [Figure 1B] 1 is a simplified schematic diagram of a system for semiconductor metrology and process control, according to one embodiment of the present invention. [Diagram 2] 3 is a simplified flow chart of an exemplary method of operation of the system of FIGS. 1 and 2 , operating in accordance with various embodiments of the present invention. [Figure 3A] 1 is a simplified graph illustrating experimental results of the present invention illustrating an investigation of control elements across multiple layers and multiple parameters per layer. [Figure 3B] 1 is a simplified graph illustrating experimental results of the present invention illustrating an investigation of control elements across multiple layers and multiple parameters per layer. [Figure 4A] 1 is a simplified graph representing experimental results of the present invention showing how the DOE on polishing time was learned by the predictive model of the present invention. [Figure 4B] 1 is a simplified graph representing experimental results of the present invention showing how the DOE at the intended remaining thickness was learned by the predictive model of the present invention. [Figure 5A] 1 is a simplified schematic diagram of a system architecture for practicing an embodiment of the present invention. [Figure 5B] 1 is a simplified schematic diagram of a system architecture for practicing an embodiment of the present invention. [Figure 6A] 1 is a simplified graph illustrating experimental results of the present invention showing an example of the improved accuracy performance of the predictive model of the present invention. [Figure 6B] 1 is a simplified graph illustrating experimental results of the present invention showing an example of the improved accuracy performance of the predictive model of the present invention. [Figure 7] 1 is a simplified graph illustrating experimental results of the present invention related to training a predictive model of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0023] 1A and 1B, each of which is a simplified schematic diagram of a semiconductor metrology system constructed and operative in accordance with one embodiment of the present invention. In FIG. 1A, a spectral acquisition tool 100 is used to collect scatterometric spectra 102 of multiple targets (e.g., structures, dies) 104 on one or more reference semiconductor wafers 106 in accordance with the prior art. The spectral acquisition tool 100 collects the scatterometric spectra 102, referred to herein as baseline spectra 102, in accordance with a first predetermined measurement protocol proximate a selected processing step during the manufacture of the reference semiconductor wafers 106, such as immediately after completion of a given etch step. Spectral acquisition tool 100 can be any type of known tool capable of collecting a scatterometry spectrum of a semiconductor wafer target, such as a Spectral Ellipsometer (SE), Spectral Reflectometer (SR), Polarized Spectral Reflectometer, or any other Optical Critical Dimension (OCD) metrology tool. The first predetermined measurement protocol used by spectral acquisition tool 100 preferably incorporates scatterometry measurements from two or more channels of information.

[0024] A reference metrology tool 108, such as a Critical Dimension Scanning Electron Microscope (CD-SEM), Atomic Force Microscope (AFM), cross-section Tunneling Electron Microscope (TEM), electric metrology tool, Critical Dimension Atomic Force Microscope (CD-AFM), X-RAY metrology tool, or optical metrology tool, is used to measure the baseline spectra 102 of reference semiconductor wafers 106 at a second predetermined measurement protocol proximate to the same selected processing step during fabrication of reference semiconductor wafers 106 at which spectrum acquisition tool 100 collected baseline spectra 102 according to the prior art. 102), measurements 110 of predetermined parameters of the targets 104 on a reference semiconductor wafer 106 are collected. The predetermined parameters can be any type of known parameters of the semiconductor wafer, such as those related to physical and chemical properties, material properties, electrical properties, and geometric properties of the semiconductor wafer structure.

[0025] The first measurement protocol used by the spectrum acquisition tool 100 and the second measurement protocol used by the reference metrology tool 108 preferably differ from each other in one or more of the following respects:

[0026] Differences in information channels, for example when one protocol uses multiple signals from multiple illumination channels with different illumination angles, while the other protocol uses only the signal acquired during normal illumination of a given sample.

[0027] Differences in positions within the same sample, for example when one protocol uses signals obtained from spot target positions within area A of a given sample, while the other protocol uses signals obtained from spot target positions within area B of the same sample.

[0028] Different signal combinations, for example when one protocol uses multiple overlapping signals acquired through the same illumination channel from multiple spot target locations, while the other uses a single signal acquired from a single spot target location.

[0029] In addition to the baseline spectrum 102, the spectral acquisition tool 100 collects one or more sets of scatterometry spectra 112, referred to herein as variability spectra 112, on one or more reference semiconductor wafers 106' (which may be the reference semiconductor wafer 106 or a different semiconductor wafer) in accordance with a first predetermined measurement protocol near the same selected processing step during the manufacture of the reference semiconductor wafer 106 at which the spectral acquisition tool 100 collected the baseline spectrum 102. The spectral acquisition tool 100 collects each set of variability spectra 112 associated with a predetermined source of spectra variability, such that the set of spectra variability 112 embodies the spectra variability. For example, if a given source of spectral variability is related to tool variability, the spectral acquisition tool 100 collects a set of variability spectra 112 by measuring given samples (e.g., collecting spectra of structures within a particular die on a particular wafer) using multiple identical tools (i.e., same model from the same manufacturer) using the same measurement protocol. Similarly, if a given source of spectral variability is related to measurement repeatability, the spectral acquisition tool 100 collects a set of variability spectra 112 by measuring given samples at different times (e.g., collecting spectra of structures within a particular die on a particular wafer). The spectral acquisition tool 100 collects any number of sets of variability spectra 112 associated with any number of given sources of spectral variability.

[0030] The training unit 114 is configured to train a predictive model 116 by performing machine learning (ML) to identify correspondences between the scatterometry spectra, i.e., baseline spectrum 102 and the various sets of variation spectra 112 collected by the spectrum acquisition tool 100 and measurements 110 of the predetermined parameters of the targets 104 on the reference semiconductor wafer 106 while minimizing a loss function associated with the model. The predictive model 116 is configured to be used to predict values ​​of any of the predetermined parameters, such as using scatterometry spectra of the semiconductor wafer targets during high-volume manufacturing (HVM) processing, where the spectra are collected using the spectrum acquisition tool according to a first measurement protocol. An exemplary loss function is expressed as follows:

[0031]

number

[0032] ·Loss accuracy is a loss term that takes into account the baseline spectrum 102 and the measurement values ​​110.

[0033] ·X is the spectrum collected by the above spectrum acquisition tool.

[0034] · y is the value of a given parameter resulting from measurement using a reference metrology tool.

[0035] ·y^(X) (y^ is a symbol with a ^ above the alphabet y, and so on) is a predictive model that provides a prediction of y using the spectrum X.

[0036] · Total sum refers to the sum of unique physical samples (e.g., various dies on a wafer).

[0037] Variance (y^) across something refers to the variance of the predicted y^(X) calculated for a sample (e.g. a specific die on a wafer) across a set of measurements performed on the sample, reflecting a specific source of variability. For example, Variance (y^) across repetitions represents the variance of the predicted y^(X) across 10 repeated measurements performed on the same die, when the source of variability is measurement repeatability.

[0038] In the example loss function above, the square of the loss accuracy can be replaced by any positive power and / or |yy^(X)| 2 Instead of , any monotonically increasing function can be used. Regarding the type of loss variability, any metric that represents the specific variability can be used, not necessarily the variance, but any higher statistical moment which describes the variability, etc. That is, for example, the variance of y (which is <(y^-<y^> ) 2 Instead of using <(y^-<y^> ) p > (where p>0) or y^ and<y^> Any other moment of a monotonically increasing function of the difference between

[0039] The parameters Λ represent the hyperparameters of the predictive model 116 and determine the relative importance of suppressing the variability across the variability terms versus the importance of the accuracy term. These are preferably modified during ML model training by methods of validation and cross-validation to optimize a user-defined recipe ranking that reflects a given specification, such as the specification of a user of the system of Figures 1A and 1B, in terms of accuracy, reproducibility, tool matching, and other features important to the user.

[0040] Predictive model 116 is preferably provided for use with a process controller configured to control the fabrication of semiconductor devices on semiconductor wafers, and is described below with reference to FIG. 1B.

[0041] In FIG. 1B, spectrum acquisition tool 100 or another similar or identical spectrum acquisition tool, spectrum acquisition tool 100', is used during a manufacturing process, such as during a high volume manufacturing (HVM) process for fabricating semiconductor devices on a semiconductor wafer, to collect scatterometry spectra of a production semiconductor wafer 106'. Spectrum acquisition tool 100' collects scatterometry spectra proximate to the same selected processing step during fabrication of reference semiconductor wafers 106 at which spectrum acquisition tool 100 collected baseline spectra according to a first predetermined measurement protocol. A prediction unit 118 applies prediction model 116 to the spectra collected from the production semiconductor wafer 106' to generate a prediction associated with any predetermined parameter described above with reference to FIG. 1 of any target of the production semiconductor wafer 106' at that processing step. The process control unit 120 may be any known process control hardware and / or software for controlling a process for manufacturing semiconductor devices on a semiconductor wafer and is configured to control the production of the manufactured semiconductor wafer 106' or subsequent manufactured semiconductor wafers in accordance with conventional techniques by providing inputs to any known semiconductor manufacturing (fabrication) tools 122 (e.g., lithography tools, etch tools, deposition tools, etc.) that control the operation of the tools during the production of the manufactured semiconductor wafer 106', wherein the inputs are determined according to a predetermined protocol using the predicted values.

[0042] Thus, the systems of Figures 1A and 1B are used to improve metrics such as tool stability and tool repeatability by adding additional data during ML training of the predictive model 116 that represents measurements of the same physical die and wafer, but from multiple tools (to optimize matching) and / or at multiple time points (to optimize repeatability). These data can be obtained relatively inexpensively as they do not require reference values ​​for parameters. Once these data are obtained, the predictive model 116 can be used to predict the reference values ​​while being optimized to have matching and repeatability within specifications. This is known as α·Lrepeatability + β T2T This is achieved by supplementing the standard loss function of the predictive model 116, which typically aims to minimize the agreement with a reference (Loss = Loss reference) that balances accuracy and productivity using prior coefficients α and β of the hyperparameters of these terms, with loss terms reflecting additional metrics such as: Table 1 below shows experimental results of the present invention showing an improvement in repeatability across different "layers" (i.e. different measurement scenarios of different physical parameters in different types of semiconductor manufacturing steps), while accuracy is reasonably uncompromised.

[0043] [Table 1]

[0044] It should be appreciated that self-consistency merits such as repeatability, T2T, and wafer smoothness all improve strongly the more the model is regularized. This effect can be used to find the optimum point of the regularization scheme that balances accuracy and these other performance metrics. The hyperparameters that determine the improvement of the additional metrics can be automatically selected using any suitable technique while maintaining the same accuracy.

[0045] Referring to FIG. 2, FIG. 2 is a simplified flow chart of an exemplary method of operation of the system of FIG. 1A and FIG. 1B operating in accordance with one embodiment of the present invention. In the method of FIG. 2, baseline spectra of a plurality of reference semiconductor wafer targets are collected according to a first predetermined measurement protocol (step 200). Measurements of predetermined parameters of the reference semiconductor wafer targets are collected according to a second predetermined measurement protocol (step 202). A set of one or more variability spectra is collected on the plurality of reference semiconductor wafer targets according to a first predetermined measurement protocol, the variability set embodying the spectral variability (step 204). A predictive model is trained by performing machine learning (ML) to identify correspondence between the collected spectra and the measurements of the predetermined parameters while minimizing a loss function associated with the model (step 206). Production scatterometry spectra of production semiconductor wafer targets are collected according to the first predetermined measurement protocol (step 208). The predictive model is applied to the production spectrum to generate a predicted value associated with any predetermined parameter of the produced semiconductor wafer target (step 210). The production of the produced semiconductor wafer or subsequent produced semiconductor wafers is controlled by providing inputs to a semiconductor manufacturing tool, the inputs being determined according to a predetermined protocol using the predicted values ​​(step 212).

[0046] Sampling Enhancement In one embodiment of the present invention, enhanced sampling is achieved by setting the signal S to a set of signals measured on a relatively small number of semiconductor wafer dies, such as 10-15 dies per wafer, and training the prediction model 116 to predict the parameter P of a larger number of dies, such as a complete wafer map. The parameters used by the process control unit 122 can be wafer model parameters, such as the wafer mean or the polynomial parameters in a radial description of the wafer map. Table 2 below presents experimental results of the present invention and shows the accuracy of the wafer mean of a CMP thickness parameter in a BEOL application. The ground truth for the above measurements was the RCWA interpretation of a NOVA600® Advanced Optical CD Metrology tool available from Nova Measuring Instruments Limited, Rehovot, Israel, measuring a complete wafer map of 65 dies. As shown in the table, using the spectra of 13 dies together with ML to predict the average of a full wafer map of 65 dies resulted in a modest error of ~2.0A-2.6A, which is very close to the benchmark error chosen as the error obtained in the wafer average calculated on the same dies used to collect the spectra.

[0047] [Table 2]

[0048] Better sampling and improved sampling accuracy can also be achieved by biasing the sampling strategy based on measurements performed on the current layer and previous layers, whereby sparse sampling at different layers is optimized to improve full wafer map prediction for the current layer when measurements from previous layers are used. Such methods can be used to predict parameters as described above, or to predict raw spectra itself for unmeasured dies using variants of autoencoders and other deep learning upsampling methods. According to this embodiment of the invention, full wafer information is provided by measuring a small set of dies on the wafer.

[0049] Optimizing throughput by reducing the number of channels being measured In one embodiment of the invention, metrology throughput (TPT) can be optimized by starting with a metrology recipe with many measurement channels, obtaining reference values ​​using regular physical modeling, and searching for a small number of channels (e.g. only two regular channels) with good TPT, one of which can be "transferred" to a solution with a good balance between accuracy and throughput. This effect is illustrated in Figures 3A and 3B, which represent experimental results of the invention showing the investigation of 22 control elements across seven layers and multiple parameters per layer. In Figures 3A and 3B, the reference values ​​are provided by either physical modeling or external reference tools, such as CDSEM or TEM. As shown in Figures 3A and 3B, many channels can be transferred to a single channel without significant penalties in accuracy (Figure 3A) and repeatability (Figure 3B). This can increase the TPT by a factor of 3 to 4.

[0050] According to this embodiment of the invention, the training unit 114 automatically selects an optimal channel that is constrained to have a user-defined or otherwise predetermined desired throughput.

[0051] Methods of handling a limited amount of reference data Direct Processing Control In one embodiment of the invention, tool knob control solutions, such as those used in Advanced Process Control (APC) systems used in semiconductor manufacturing, are provided for two levels of variation sources: within a processed wafer (batch) and wafer-to-wafer variation (batch-to-batch). ML can be used to directly control a process parameter by first acquiring a set of signals S measured over a range of recipe parameters of the process tool (e.g., a DOE on a knob controlling CMP time or etch temperature). A predictive model 116 is then trained to learn and calibrate the OCD signal to the tool knob. The predictive model 116 can then be used to predict effective knob settings that can be used as feedback to the tool knob.

[0052] Given two groups of signals with two different DOE levels, the training unit 114 uses ML to extract the spectral response to knob changes, i.e., the spectral sensitivity to the knob, using the DOE levels as a reference. The training unit 114 tracks the spectral differences between the groups, and the nominals of the groups are preferably as close as possible so that the knob effect dominates. Future processes can then be adjusted according to the learned relationship to achieve the desired target results of the parameters or to adjust for process and tool drift to prevent potential failures. This effect is illustrated in Figures 4A and 4B, which represent experimental results of the present invention showing how the DOE was learned by the predictive model 116 using IM spectra at polishing time (Figure 4A) and intended remaining thickness (Figure 4B). Figures 4A and 4B show that the "effective knob" captures the WIW variation (i.e., the model predicts the range of effective knob settings for each wafer) despite the wafer-level uniformity of the DOE. To verify that this effective knob technique is useful, it is presented against an external ground truth of the target remaining thickness measured by a physical algorithm using a NOVA T600® MMSR Advanced Optical CD Metrology tool, available from Nova Measuring Instruments Limited, Rehovot, Israel. Such a model can be used for feedback to a processing tool equipped with a WIW knob. As shown in Figure 4A, the predicted effective polishing time ("effective knob") is highly correlated with the ground truth polishing layer thickness, while Figure 2B shows the predicted polishing thickness effect against the reference (reference).

[0053] Data augmentation Simple ML algorithms do not contain prior knowledge of the underlying physics they are trying to explain. This means that their success is based on the availability of data, more specifically the availability of reference data. In one embodiment of the invention, the size of the training data set used to train the predictive model 116 is increased by means of generative models, which are a class of ML methods that have the ability of describing a highly complex probability distribution of data. A simple physical analogy is the Boltzmann weight in statistical physics or the path integral formalism for Euclidean quantum theories. In both cases, probability functions are defined that describe the behavior of the degrees of freedom of the system (e.g. spin in statistical physics or electrons in quantum electrodynamics). These probability functions have an explicit form and can be used to generate realizations of the physics involved. In fact, in many areas of physics, such formalisms are used in large scale Markov chain Monte-Carlo simulations to calculate physical properties of materials. Generative models in data science can be explicit or implicit, where the probability functions are explicitly stated (similar to Boltzmann weights), or where the algorithm "learns" the statistics of an empirical data set and can then use those same statistics to generate new data examples.Examples of implicit methods include Variational Auto Encoders and Generative Adversarial Networks. In one embodiment of the invention, these methods are used to augment and diversify the reference dataset, thereby optimizing the performance of the training unit 114 and allowing the introduction of prior information and constraints between various features of the machine learning model that reflect the underlying physics in the predictive model 116.

[0054] Solving process robustness To address processing robustness, in one embodiment of the present invention, a combination of ML algorithms and big data systems is employed to create, monitor and control the recipe as follows:

[0055] Recipe Training: It allows users to query training data over months, or even years, in a big data database that can span hundreds of spectral types ("measurement recipes"), possibly automatically using ML.

[0056] Recipe verification: Cross-validation and blind testing methodologies are used to generate customer feedback on recipe quality.

[0057] Recipe Monitoring: The training data is used to generate monitor defense indices along with the specifications for these indices. These indices are then monitored inline, flagging any degradation in the accuracy of the recipe, for example by employing SEMI standard SPC rules.

[0058] Recipe Update: When the monitor flags a metrology recipe for repair, an automatic retraining sequence is initiated and a new recipe is automatically suggested along with its validation results.

[0059] In this embodiment shown in FIG. 5A, the outer circle represents the reference tool, which can be a high-end, channel-rich optical metrology tool like the NOVA T600®-MMSR, or a non-optical reference like XPS, SEM or TEM. Typically, these tools have a high COO and a low TPT, which also reflects important productivity challenges like destructiveness, tool matching and repeatability. Precision-wise, these tools are characterized by a high information content (e.g., high-resolution imaging tools and multi-channel optical tools). The inner circle represents a fast and compact OCD tool like the NOVA i550® normal incidence channel integrated metrology (IM) tool available from NOVA Measuring Instruments Limited, Rehovot, Israel, with a high throughput, the ability to measure every wafer inline at the required sampling rate (number of dies), and an ML algorithm running inline that takes the Sinner-circle as input and outputs a prediction about the Preference.

[0060] System architecture FIG. 5B illustrates an architecture for implementing one or more embodiments of the invention described herein, where a computer cluster includes a web server layer (e.g., NodeJS™) and a big data layer (e.g., a Hadoop™ cluster available from Apache Software Foundation, Forest Hill, Maryland). The big data layer is used to store and process data in a scalable and distributed manner. The web server layer allows users to specify processing jobs in the control system and view reports and execution status of the control system. Both layers are designed using a micro-services approach to achieve scalability, high availability, and load balancing. Message queues are used to communicate between the services to reduce tight coupling. Various data pre-processing pipelines are employed to improve the performance of the big data layer.

[0061] Data size and long-term process robustness The use of big data systems allows an increase in the amount of training data available for training the predictive model 116, allowing the predictive model 116 to be improved and updated in real time, which eliminates accuracy limitations due to data management overhead and ensures robustness for long-term processing. Figures 6A and 6B show two examples of improved accuracy performance of the predictive model 116 applied to the FEOL layer. Here, the reference is the physical model developed on the NOVA T600®-MMSR, and the predictive model 116 uses two normal channels for inference. The plots show how the accuracy of the predictive model 116, defined as 1σ of the blind test results over ~2000 dies, depends on the number of wafers (each wafer has ~15 dies) used to train the model. As shown in Figure 6A, increasing the training size from 5 to 200 wafers can improve the accuracy by 45% (13A → 6A). Saturation on such large samples is not common, and we often see good results on tens on sites provided a DOE is present. Figure 6B shows a different situation, where the accuracy of the predictive model 116 is strongly dependent on the data used to train it, and not just on the amount of data used. Specifically, Figure 6b shows that when trained on the ~20 wafers closest in time to the blind test wafers (shown as curve 600), the model converged to the same level of accuracy as when trained on the 100 wafers furthest from the blind test (shown as curve 602). Between these curves is a benchmark curve (shown as curve 604), which describes a random shuffle of the data so as to 'lose' memory of the important of time.This indicates that the type of wafer is important and this application would benefit from the dynamic updating of ML recipes that the big data system of this embodiment is capable of.

[0062] The recipe update step described above is performed automatically, as is the selection of the data used to perform the retraining. This selection may use all available data or may be based on logic that considers the value of the monitor flag for each wafer and / or the temporal or spatial proximity of a given wafer to other wafers being processed. For example, a different weight p can be set for each wafer in the data available for retraining. The weight p may depend on the flag value for that wafer, the spectrum itself, and / or other metadata characteristics such as wafer or lot ID. This is shown in Figure 7.

[0063] Any aspect of the invention described in this specification may be implemented in accordance with conventional techniques in computer hardware and / or computer software embodied in a non-transitory computer-readable medium, computer hardware including one or more computer processors, computer memory, I / O devices, and network interfaces that interoperate in accordance with conventional techniques.

[0064] It should be understood that the term "processor" or "device" as used in this specification is intended to include any processing device, such as, for example, one that includes a CPU (Central Processing Unit) and / or other processing circuitry. It should also be understood that the term "processor" or "device" may refer to more than one processing device, and that various elements associated with a processing device may be shared by other processing devices.

[0065] The term "memory" as used in this specification is intended to include memory that cooperates with a processing unit or CPU, such as RAM, ROM, fixed memory devices (e.g., hard drives), removable memory devices (e.g., disks), flash memory, etc. Such memory may be considered a computer-readable storage medium.

[0066] Furthermore, the term "input / output device" or "I / O device" as used in this specification is intended to include, for example, one or more input devices (e.g., a keyboard, mouse, scanner, etc.) for inputting data into a processing unit, and / or one or more output devices (e.g., speakers, displays, printers, etc.) that cooperate with a processing unit to present results.

[0067] Embodiments of the invention include systems, methods, and / or computer program products including a computer-readable storage medium (or media) having computer-readable program instructions for causing a processor to perform aspects of the invention.

[0068] A computer readable storage medium may be a tangible device capable of holding and storing instructions for use by an instruction execution device. A computer readable storage medium may be, for example, but not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof. A non-exhaustive list of more specific examples of computer readable storage media includes portable computer diskettes, hard disks, random access memory (RAM), read only memory (ROM), erasable programmable read only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disk read only memory (CD-ROM), digital versatile disks (DVD), memory sticks, floppy disks, mechanically encoded devices such as punch cards or ridge-in-groove structures having instructions recorded thereon, and any suitable combination thereof. A computer readable storage medium as used in this specification should not be construed as a transitory signal itself, such as an electric signal transmitted over a wire, such as an electric wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., a light pulse passing through a fiber optic cable), or an electric signal transmitted over a wire.

[0069] The computer readable program instructions described herein can be downloaded from the computer readable storage medium to each computing / processing device or to an external computer or storage device over a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, fiber optics, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface of each computing / processing device receives the computer readable program instructions from the network and forwards the computer readable program instructions for storage in a computer readable storage medium within the respective computing / processing device.

[0070] The computer readable program instructions for carrying out the operations of the present invention may be either assembler instructions, instruction set architecture (ISA) instructions, machine instructions, machine dependent instructions, microcode, firmware instructions, state setting data, or object code written in any combination of one or more programming languages, including object oriented programming languages ​​such as Java, Smalltalk, C++, traditional procedural programming languages ​​such as the "C" programming language or similar programming languages. The computer readable program instructions may be executed entirely on the user's computer, partly on the user's computer as a standalone software package, partly on the user's computer and partly on a remote computer, or entirely on a remote computer or server. In the latter scenario, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet Service Provider). In some embodiments, electronic circuitry including, for example, a programmable logic circuit, a field programmable gate array (FPGA) or a programmable logic array (PLA) may execute computer readable program instructions by utilizing state information of the computer readable program instructions to personalize the electronic circuitry to perform aspects of the invention.

[0071] Aspects of the invention are described herein with reference to flowchart and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart and / or block diagrams, and combinations of blocks in the flowchart and / or block diagrams, can be implemented by computer readable program instructions.

[0072] These computer readable program instructions can be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing device to manufacture a machine that executes the instructions via the processor of the computer or other programmable data processing device to create means for performing the functions / acts specified in the flowcharts and / or one or more blocks. These computer readable program instructions can also be stored on a computer readable storage medium to instruct a computer, programmable data processing device, and / or other device to function in a particular manner. A computer readable storage medium includes an article of manufacture that includes instructions stored thereon that implement aspects of the functions / acts specified in the flowcharts and / or one or more blocks.

[0073] Computer-readable program instructions may also be loaded into a computer, other programmable data processing apparatus, or other device to cause the computer, other programmable apparatus, or other device to perform a series of operational steps to generate a computer-implemented process. The instructions executing on the computer, other programmable apparatus, or other device perform the functions / acts identified in the flowcharts and / or one or more block diagrams.

[0074] The flowcharts and block diagrams in the drawings illustrate the architecture, functions, and operations that can be implemented in the systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in the flowcharts or block diagrams can represent a module, segment, or portion of computer instructions, which includes one or more executable computer instructions for implementing a particular logical function. In some other implementations, the functions shown in the blocks may operate out of the order shown in the drawings. For example, two blocks shown in succession may in fact be executed substantially simultaneously, or the blocks may be executed in the reverse order depending on the functionality involved. It should also be noted that each block of the flowcharts and block diagrams, as well as combinations of blocks, can be implemented by dedicated hardware-based and / or software-based systems that perform specific functions or operations.

[0075] The description of various embodiments of the present invention is presented for illustrative purposes and is not intended to be exhaustive or limited to the disclosed embodiments. For example, the systems and methods described herein are applicable to any structure on a semiconductor wafer. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.< / predicted> < / new>

Claims

1. a spectral acquisition tool configured to collect a baseline set of scattering spectra of the semiconductor wafer target according to a first measurement protocol; a reference metrology tool configured to provide values ​​of the predetermined parameter of the semiconductor wafer target according to a second measurement protocol different from the first measurement protocol; a web server layer connectable to said spectrum acquisition tool and said reference measurement tool; And a training unit configured to create training set data for training a predictive model that predicts values ​​of the predetermined parameters using machine learning from the collected set of spectra and values ​​of the predetermined parameters, the training unit being configured by a big data layer that stores and processes data in a scalable and distributed manner; The spectral acquisition tool is configured as a plurality of integrated measurement (IM) tools; the plurality of integrated metrology (IM) tools are in-line monitors, and a baseline set of scattering spectra of all of the semiconductor wafer targets in production are collected in-line; High volume semiconductor process control and measurement systems.

2. 2. The mass production semiconductor process control metrology system of claim 1, wherein a training unit configured to generate training set data from the collected set of spectra and values ​​of the predetermined parameters increases a size of the training set data using one or more generative models.

3. 3. The mass production semiconductor process control metrology system of claim 2, wherein any of the one or more generative models uses a predetermined probability function that provides a probability distribution of data in the training set to generate new data examples using the predetermined probability function, thereby increasing the size of the training set data.

4. 4. The mass production semiconductor process control metrology system of claim 3, wherein said predetermined probability function is explicitly stated.

5. 3. The mass production semiconductor process control metrology system of claim 2, wherein one or more of the generative models uses a predefined algorithm to determine statistics of the data in the training set and increase the size of the training set data by generating new data examples having the same statistics.

6. 6. The mass production semiconductor process control metrology system of claim 5, wherein any of the one or more generative models is a variational autoencoder.

7. 6. The mass production semiconductor process control and metrology system of claim 5, wherein the one or more generative models use generative adversarial networks.

8. 3. The mass production semiconductor process control metrology system of claim 2, further comprising introducing into the predictive model any information and constraints between various features of any of the one or more generative models that reflect underlying physics of the semiconductor wafer targets.

9. 2. The mass production semiconductor process control metrology system of claim 1, wherein the predetermined parameters relate to any of physical and chemical characteristics, material properties, electrical properties, and geometric properties of the structure of the semiconductor wafer target.

10. 2. The mass production semiconductor process control metrology system of claim 1, wherein the spectral acquisition tool is one of a spectral ellipsometer (SE), a spectral reflectometer (SR), a polarized spectral reflectometer, and an optical critical dimension (OCD) metrology tool.

11. 2. The mass production semiconductor process control metrology system of claim 1, wherein the reference metrology tool is one of a critical dimension scanning electron microscope (CD-SEM), an atomic force microscope (AFM), a cross-section tunneling electron microscope (TEM), an electrical metrology tool, a critical dimension atomic force microscope (CD-AFM), and an X-ray metrology tool.

12. 2. The high volume semiconductor process control metrology system of claim 1, wherein said integrated metrology (IM) tool is configured as a normal incidence channel integrated tool.

13. 2. The high volume semiconductor process control metrology system of claim 1, wherein the reference metrology tool comprises a multi-channel optical metrology tool.

14. 2. The mass production semiconductor process control metrology system of claim 1, wherein the reference metrology tool obtains values ​​for predetermined parameters of the semiconductor wafer targets using RCWA interpretation.

Images