Machine Learning and Deep Learning Methods for Spectral-Based Measurement and Process Control

Abstract

A system and method for advanced process control (APC) in semiconductor manufacturing includes, for each of a plurality of wafer sites, receiving a pre-process set of scatterometry training data measured before performing a process step, receiving a corresponding post-process set of scatterometry training data measured after performing the process step, and receiving a set of process control knob training data indicative of process control knob settings applied during performance of the process step; generating a machine learning model that correlates variability in the pre-process set of scatterometry training data and the corresponding process control knob training data to the corresponding post-process set of scatterometry training data; and training the machine learning model to recommend changes in the process control knob settings to correct for the variability in the pre-process scatterometry data.

Description

[Technical Field]

[0001] This invention relates generally to the field of optical inspection of integrated circuit wafer patterns, and more particularly to algorithms for silicon wafer manufacturing. [Background technology]

[0002] Integrated circuits (ICs) are fabricated on semiconductor wafers through multiple steps of deposition, modification, and removal of thin layers. Modern semiconductor manufacturing processes can require over 1000 such process steps. Advanced process control (APC) aims to optimize process tool settings to reduce the overall manufacturing variability. Process tool settings, hereafter also referred to as process "knobs," can include all aspects of process control, including process settings such as spin-on film, thermal oxide growth, chemical vapor deposition (CVD), physical vapor deposition (PVD), electroplating, wafer temperature, chamber pressure, and polishing pressure.

[0003] Under certain conditions, traditional process control methods can no longer meet the ever-increasing levels of precision required in semiconductor manufacturing. Recent advances in hardware development have introduced a wider range of process knobs, causing traditional APC solutions to have shortcomings when applied to high-dimensional knob spaces. Another challenge with APC is that multiple production lines may use multiple manufacturing routes, which can lead to large variations in production results due to time-scale drift. Attempts to improve APC by applying machine learning techniques have been described. For example, International Publication WO 2021 / 030833 by Drori et al., entitled "Model Based Control of Wafer Non-Conformity," describes the creation of several types of neural networks that correlate process parameters with measurement data.

[0004] Multiple process steps in semiconductor manufacturing create layered structures ("stacks") with optical properties similar to diffraction gratings. Optical critical dimension (OCD) metrology uses these optical properties to measure the critical dimensions (CD) and material properties of a pattern at a site on a wafer ("wafer site"). (Hereinafter, CD and material properties are also referred to as "pattern parameters"). CD can include the height, width, and pitch of the stack. As described in Dixit et al., "Sensitivity analysis and line edge roughness determination of 28-nm pitch silicon fins using Mueller matrix spectroscopic ellipsometry -based optical critical dimension metrology," J. Micro / Nanolith. MEMS MOEMS. 14(3), 031208 (2015), which is incorporated herein by reference, CD can also include side wall angle (SWA), spacer width, spacer pull-down, epitaxial proximity, footing / undercut, over-fill / under-fill in 2D (HKMG), 3D profile (FinFETs), and line edge roughness (LER).

[0005] Scatterometric data (also referred to herein as "spectral data") is typically acquired as reflected light radiation that indicates the optical characteristics of patterns at a wafer site. U.S. Patent No. 6,476,920 to Scheiner and Machavariani, "Method and Apparatus for Measurements of Patterned Structures," incorporated herein by reference, describes the development of an "optical model," also known as a "physical model," that estimates scatterometric data measured during spectroscopic testing from predetermined pattern parameters. Optical models can also be designed to perform the opposite (or "inverse") function of estimating pattern parameters based on measured scatterometric data. Typically, optical models are applied to optical coherence tomography (OCD) measurements to determine whether patterns at a wafer site are manufactured to the correct specifications. Hereinafter, the more general term "OCD model" refers to both physical models developed from optical principles and machine learning models known in the art.

[0006] Exemplary scatterometry tools for measuring (obtaining) scatterometry data (e.g., spectrograms) include spectral ellipsometers (SE), spectral reflectometers (SR), polarized spectral reflectometers, and other optical critical dimension (OCD) measurement tools. Such tools are incorporated into currently available OCD measurement systems. One such OCD measurement system is the NOVA T600® Advanced OCD Metrology tool, commercially available from Nova Measuring Instruments Ltd., Rehovot, Israel, which measures pattern parameters at designated wafer sites, or "in-die." Other methods for measuring critical dimensions (CD) include interferometry, X-ray Raman spectroscopy (XRS), X-ray diffraction (XRD), and pump-probe tools. Some examples of such tools are disclosed in U.S. Patent Nos. 10,161,885, 10,054,423, 9,184,102, and 10,119,925, as well as in pending international application WO 2018 / 211505, all of which are assigned to the applicant and are incorporated herein by reference in their entireties.

[0007] High-precision methods for measuring pattern parameters that do not rely on the optical models described above include wafer measurements using instruments such as CD-Scanning Electron Microscopy (CD-SEM), atomic force microscopes (AFM), cross-sectional tunneling microscopes (TEM), or X-ray metrology tools. These methods are typically expensive and time-consuming compared to optical and machine learning modeling methods. Hereafter, pattern parameters measured with such tools are referred to as "reference parameters."

[0008] The embodiments of the present invention disclosed below help overcome the shortcomings of current APC methods. It should be understood that the background and context discussion contained in this specification is provided solely for the purpose of generally providing a context for the present disclosure. Much of this disclosure presents the work of the present inventors, and the fact that such work is set forth in the Background Art section or presented as context elsewhere in this specification does not constitute an admission that it is prior art. Summary of the Invention [Means for solving the problem]

[0009] Embodiments of the present invention provide systems and methods for machine learning-based advanced process control (APC) in semiconductor manufacturing, the systems and methods including receiving, for each of a plurality of wafer sites, a pre-process set of scatterometric training data measured before performing a process step, a corresponding post-process set of scatterometric training data measured after performing the process step, and a set of process control knob training data indicative of process control knob settings to be applied during performing the process step. A machine learning model can then correlate variability in the pre-process set of scatterometric training data and the corresponding process control knob training data with the corresponding post-process set of scatterometric training data, thereby training the machine learning model to recommend changes in process control knob settings to compensate for variability in the pre-process scatterometric data.

[0010] Embodiments of the present invention may further include applying the machine learning model to generate recommended values ​​for process control knobs during semiconductor manufacturing.

[0011] In a further embodiment, the post-processed set of scatterometry training data may be correlated to one or more target, post-process pattern parameters by an optical model. Additionally or alternatively, the post-processed set of scatterometry training data may be correlated to one or more target, post-process pattern parameters by a second machine learning model.

[0012] The process control knob settings can include settings for one or more of a process step duration, a pedestal edge ring height, a temperature distribution over multiple control zones of a pedestal, and a process chamber pressure.

[0013] The pre-process and post-process sets of scatterometry data may be indicative of one or more pattern parameters at each wafer site, including one or more of a critical dimension, a feature depth, a feature height, and a feature pitch.

[0014] The process step may be one or more of a deposition, etching, or polishing operation.

[0015] Generating the machine learning model may include training a neural network (NN) including multiple encoder layers followed by a bottleneck latent layer and then followed by at least one decoder layer, wherein a pre-processed set of scatterometry training data is applied as a model input, a corresponding post-processed set of scatterometry training data is applied as a model output, a plurality of process control knob training data is applied as auxiliary inputs to the NN at any one of the multiple encoder layers, and a plurality of process control knob training data is applied as auxiliary outputs to any one of the at least one decoder layer. A loss function for backpropagation of the NN may be configured to maximize a similarity between the output of the NN and the post-processed set of scatterometry training data. This loss function may be a squared error loss function. The machine learning model may also include a calibration step following the NN that calibrates the post-processed set of scatterometry training data to predicted, post-process pattern parameters. This calibration may be performed by an OCD model.

[0016] In a further embodiment, optimizing the machine learning model may include minimizing the difference between the target, post-process pattern parameters and the predicted, post-process pattern parameters.

[0017] The loss function for backpropagation of the above NN auxiliary output can express the quality of similarity between the auxiliary output and the process knob training data. This loss function is a squared error loss function.

[0018] In a further embodiment, generating the machine learning model may include determining a maximum covariance between the post-process set of scatterometry training data and a corresponding set of process control knob training data to generate latent variables; subtracting the process control knob training data from the latent variables to generate corresponding residuals representing the contribution of the process control knob training data to the variability of the post-process scatterometry training data; calibrating the set of pre-process scatterometry data to the corresponding residuals to determine knob value estimators of variation in the pre-process scatterometry data; and optimizing the machine learning model to determine the process control knob recommendations from the knob value estimators.

[0019] In some embodiments, the multiple wafer sites are located on multiple wafers. The multiple sets of pre-process and post-process scatterometry training data may also be measured by two or more measurement channels.

[0020] For a better understanding of various embodiments of the present invention and to show how they may be carried into effect, reference is made, by way of example, to the accompanying drawings, in which structural details of the invention are shown to provide a basic understanding of the invention, and the description, when read in conjunction with the drawings, will make apparent to those skilled in the art how several forms of the invention may be embodied in practice. [Brief explanation of the drawings]

[0021] [Figure 1] 1 is a schematic diagram of a system for machine learning-based APC with self-supervised representation learning, according to an embodiment of the present invention. [Figure 2] 1 is a flowchart illustrating a process for machine learning-based APC by self-supervised representation learning, according to an embodiment of the invention. [Figure 3] 1 is a schematic diagram of a process for training and applying a machine learning model for APC according to an embodiment of the invention. [Figure 4A] 1 is a schematic diagram of a process for training and applying a machine learning model for APC according to an embodiment of the invention. [Figure 4B] 1 is a schematic diagram of a process for training and applying a machine learning model for APC according to an embodiment of the invention. [Figure 5] 1 is a schematic diagram of a process for training and applying a machine learning model for APC according to an embodiment of the invention. [Figure 6] 1 is a schematic diagram of a process for training and applying a machine learning model for APC according to an embodiment of the invention. DETAILED DESCRIPTION OF THE INVENTION

[0022] Embodiments of the present invention provide systems and methods for generating machine learning (ML) models for advanced process control (APC) in semiconductor manufacturing. Machine learning (ML), including deep learning (DL) algorithms, can be a powerful tool in the design of APC and measurement systems. These algorithms can adapt to multidimensional spaces and automatically update according to process demands. However, the success of data-driven control systems depends on the availability of accurate data for training. In semiconductor manufacturing, such "reference parameters" are a costly resource. Below, we describe a method for APC that facilitates ML and DL techniques without relying on such reference parameters.

[0023] 1 is a schematic diagram of a system 10 for semiconductor manufacturing including advanced process control (APC) in accordance with an embodiment of the present invention. The purpose of APC is to reduce the variability of parameters manufactured at sites on wafers ("wafer sites").

[0024] System 10 can be a production line for producing and monitoring wafers 12. Wafers 12 are fabricated with wafer sites 14 and have measurable pattern parameters, including one or more of critical dimensions, feature depth, feature height, feature pitch, and other parameters described in the Background. Typically, a wafer has multiple sites, or "dies," designed to have the same pattern (i.e., the same pattern design is used to fabricate all of the patterns). For each wafer site 14, a set of multiple pattern parameters can typically be measured. Hereinafter, this set of multiple parameters is referred to as a vector p It can also be represented by (a symbol with a right-pointing arrow above p) (hereinafter, similarly, vector symbols will be represented by underlines in the main text of the specification), and each element of the vector is one of multiple parameters CD.

[0025] The system 10 can include a wide range of process control "tools," shown as process control knob settings 16, that control process conditions. The process control knob settings (also referred to herein as "knob settings" or "knob values") can control, for example, the temperature distribution on the pedestal 18 on which the wafer is mounted during processing. Additional knob settings are typically provided to control additional process parameters, including the duration of a process step, the height of the pedestal edge ring, the temperature distribution across multiple control zones on the pedestal, and the process chamber pressure. Manufacturing variations introduce small variations in pattern parameters, which vary from wafer to wafer and from site to site across a single wafer after each process step. As described in more detail below, embodiments of the present invention provide methods and systems for determining changes to be made to knob settings to reduce parameter variations across wafer sites and wafer-to-wafer sites. The changes made to the knob settings correct for parameter variations caused by previous process steps. Knob settings established to enrich the dataset used for model training can be referred to as "Design of Experiment" (DOE) knob settings. If recommended variations for knob settings are determined, such recommended knob settings are referred to as k Write "recommended."

[0026] The system 10 includes a light source 20, which generates a light beam 22 in a predetermined wavelength band. The light beam 22 is reflected from a wafer pattern at the wafer site 14 toward a spectrophotometric detector 26 (shown as reflected or "scattered" light 24). In some configurations, the light source and spectrophotometric detector are included in an OCD measurement system 30 (e.g., an ellipsometer or spectrophotometer). The structure and operation of the measurement system 30 can be of any known type, such as those disclosed in U.S. Pat. Nos. 5,517,312, 6,657,736, and 7,169,015, and International Publication WO 2018 / 211505, all of which are assigned to the present applicant and incorporated by reference in their entireties. Typically, the measurement system 30 includes additional components (not shown), such as light directing optics including an objective lens, a beam splitter, and a beam deflector with mirrors. Additional components of such systems may include an imaging lens, a polarizing lens, a variable aperture stop, and a motor. The operation of such elements is typically automated by a computerized controller, which may include I / O devices and may be configured to perform data processing tasks such as generating scatterometry data 32 (also referred to herein as "measurement signals").

[0027] The scatterometry data 32 generated by the measurement system 30 typically includes various types of plotted data 34 and can be represented in vector form (e.g., a spectrogram in which the data points are measurements of reflected light intensity "I" at various light wavelengths, or a mapping of reflected irradiance versus angle of incidence). As discussed above, variations between scatterometry data sets indicate variations in pattern parameters at each wafer site. In a typical OCD measurement, the range of light measured covers the visible light spectrum and may also include wavelengths in the ultraviolet and infrared regions. A typical spectrogram output for an OCD measurement can have 245 data points and cover a wavelength range from 200 to 970 nm.

[0028] The measurement signal (i.e., the scatterometry data) contains noise from a variety of sources. This noise can be the result of temperature and air pressure fluctuations that occur during the measurement process, or it can be variations in the conditions of the measurement system 30, such as variations in the optical system alignment, variations in the determination of the location of a given wafer site on the wafer sample, and differences in the physical and optical conditions of different measurement systems used by manufacturers. The presence of such noise sources in the scatterometry data can affect the determination and, consequently, lead to mismatches in the desired target parameters.

[0029] In an embodiment of the present invention, a computer system including machine learning (ML) tools known in the art, referred to herein as an ML modeling system 40, can be configured to train an ML model of OCD metrology. The training feature set (also referred to as feature input) used by the ML modeling system can include a set of scatterometry data 34 before and after a given process step is performed, and data indicative of process control knob settings 16 to be applied during the process step. After training, the ML model is used to recommend process control knob settings to achieve target pattern parameters.

[0030] A process step can include any type of automated process that affects the wafer pattern, such as etching, deposition, or polishing. In further embodiments, the term "process step" can include multiple sub-steps with independent knob settings. A knob vector can include settings for these multiple sub-steps. The ML modeling system 40 can operate independently from the measurement system 30 or can be integrated with the measurement system.

[0031] APC control systems aim to minimize process variation in the post-process parameters (i.e., after a process has been performed) that are the result of variations in pre-process inputs and non-uniformity in the process tool. These systems calibrate the controlled tool knobs to compensate for wafer-site variations that affect process uniformity and the achievement of target parameters. Such variations can occur at the wafer level (die-to-die), lot level (wafer-to-wafer), or lot-to-lot (lot-to-lot). For example, the chemical-mechanical chemical polishing (CMP) process is a key process technology that is repeated dozens of times along long semiconductor manufacturing production lines. CMP tools remove material from thick layers to form desired thicknesses according to a desired design. Because of the large number of CMP process steps and multiple manufacturing processes, wafer variations occur and must be corrected by process control. Similarly, an etcher tool is a device that selectively removes insulating or metallic material added during deposition. Correcting wafer-level variations requires a process tool's ability to apply a within-wafer spatial setting knob, i.e., an entire wafer map of knobs, rather than just a single value per wafer. High-end etcher tools provide such controllability, for example, through temperature settings that can be used as knobs to correct this within-wafer variation. Embodiments of the present invention can determine control parameters (e.g., knob settings) for additional semiconductor manufacturing processes related to material deposition, removal, and patterning, such as chemical vapor deposition (CVD), physical vapor deposition (PVD), electroplating, wafer temperature, chamber pressure, polishing pressure, and photolithography.

[0032] The pattern parameters at wafer sites can be measured by optical critical dimension (OCD) signals collected by a multi-channel metrology tool. To accurately learn the sensitivity and response of knob settings to wafer condition variations, as well as the desired post-target parameters, the training set for the above ML model requires multiple variations in process conditions that can capture the range of knob setting effects on the results. That is, for the purpose of training the ML model, a process step can be repeated on different wafers by applying various knob settings that differ slightly from the DOE knob values. Such variations provide a means to capture the effects of such variations. In the following, we define the set of knob values ​​applied to a given process step to generate training data as k It is expressed as training.

[0033] In what follows, we refer to the set of scatterometry data generated by the spectrophotometer as the scatterometry vector S where each element of the above vector represents a data point of the scatterometry data.

[0034] 2 is a flowchart illustrating a computer-implemented process 200 for generating a machine learning model for semiconductor manufacturing APC, according to an embodiment of the present invention. Process 200 may be implemented by ML modeling system 40 described above. A first step 214 involves receiving multiple sets of scatterometry data for training the machine learning model. The set of scatterometry data measured from a given wafer pattern before a given processing step is received. S On the other hand, each set of scatterometry data measured from the same wafer pattern after the given process step is written as S Typically, in order to effectively perform subsequent machine learning training, a large number of S pre and correspondingS Further, in step 216, sets of process control knob data are obtained, each set being a knob vector k DOE:t Each set of process control knob data indicates one or more control parameters employed during a given process step performed on a measured wafer pattern, and the data set S pre and S post Generate a pair of

[0035] Next, in step 220, the machine learning model k DOE:t , S pre and S post The data set is trained to predict the appropriate knob settings (i.e., when pre-process scatterometry data indicates wafer pattern parameter variations). k recommended A model is generated to recommend the best possible settings for the desired wafer pattern. By changing the knob settings, wafer patterns with less variability can be produced. As described below, several types of machine learning models can be effective in achieving this goal.

[0036] In step 220, in production, the measured S pre By inputting the parameter, the machine learning model is applied to get the corresponding recommended knob settings to reduce the variation of the post-process pattern parameters. k recommended can be generated.

[0037] FIG. 3 is a schematic diagram of an exemplary machine learning model 300. In this model, the effect (or "signature") of the knob setting is first separated or isolated from other sources of variability in the post-process scatterometry data. To achieve such separation, ML techniques that optimize a measure of correlation or covariance between spaces can be employed. For example, a partial least squares (PLS) algorithm finds pairs of components in two spatial data sets and determines the largest percentage of covariance between them. Here, two spaces, specifically, post-process scatterometry data as X and DOE knob setting as Y, are used. k DOE As shown in step 310, X and Y are decomposed into latent variables based on the maximum covariance principle.

[0038] Next, step 320 extracts the rank 1 variables of this latent space to represent the "total effective knob."

[0039] Assuming that the knob settings (based on DOE values) are the dominant factor determining the target output, then in step 330, the knob settings can be subtracted from the "total effective knob" (typically expressed as a vector). The residual of this subtraction represents a residual contribution of the knob settings to the target parameter. This results in a "residual" effective knob, which can be used in the next step 340, where the pre-process scatterometry data is calibrated to this residual. As a result, the incoming variability (i.e., S preThis gives us a trained estimator that can predict knob values ​​that represent the variability of the

[0040] 4A and 4B are schematic diagrams of a process for training and applying an ML model for APC. As shown in FIG. 4A, a training process 400 is applied to train an exemplary ML model 410. The ML model 410 is then applied in an inference process 450 during wafer production, as shown in FIG. 4B. The illustrated ML model 410 illustrates an unsupervised learning, APC machine learning model that may be based on a deep neural network, specifically an encoder-decoder model 410, in accordance with an embodiment of the present invention. As shown in FIG. 4A, the structure of such a network 410 may consist of three parts, shown schematically as an encoder section 420, a "bottleneck" (BN) middle section 425, and a decoder section 430. The encoder 420 compresses the dimensionality of the pre-process spectrum (i.e., pre-process scatterometry data) into a latent structure, and the decoder 430 decompresses the latent structure into a post-process spectrum.

[0041] The compressed midsection 425 of the network (the "bottleneck" denoted as "BN") typically contains at least two layers: a layer representing the pre-process reduced dimensionality and a layer representing the post-process spectral dimensionality. The layer between those two layers represents the transfer of the neural network process between the two latent spaces.

[0042] In addition to the main network, an auxiliary input is provided for process control knob settings. k trainingThis input intersects the main network at any layer of the encoder (for example, in the figure, it intersects at the end of the encoder). The second addition to the main network is an auxiliary output linked to the post-latent layer in the midsection.

[0043] Dual loss functions are used in training the ML model 410. The first loss function 440 is the loss function for the reconstructed or predicted spectrum. S post:predicted The main network output, denoted as , and the measured post-process scatterometry data S post:measured The spectral loss expresses the quality of similarity between the two. This "similarity" can be measured, for example, by a mean squared error loss function.

[0044] The second loss function 445 is a function of the auxiliary output, shown as the target parameter knob setting, and the implemented process control knob setting. k train That is, the second loss function 445 can represent the quality of the similarity between the auxiliary output and the process knob training data.

[0045] During network training, the combined loss function effectively minimizes both spectra and knob loss terms. To successfully establish a link between the process knobs and the network's auxiliary output "neurons," the training set must contain a designed enhancement of incoming variation using well-controlled knobs, as well as a designed bias in the process knob values. This variation in knob settings is k training It is shown as follows.

[0046] After the network is successfully trained, the trained ML model 410 is applied to find recommended knob values ​​for subsequently processed samples of wafer patterns to reduce the variability of the target parameters for these samples, as shown in process 450 of FIG. 4B. To summarize the process, the ML training determines nominal target auxiliary neuron values, and for each wafer pattern, the auxiliary neuron values ​​are determined as a function of the auxiliary input knob values. Then, for each wafer pattern, knob values ​​are recommended that satisfy the desired target parameters.

[0047] FIG. 5 illustrates the process of training a machine learning model 500, which directly transforms representations of wafer patterns before and after process steps and then converts the post-process signals (i.e., scatterometry data) S post ) as the target (controlled) parameter P post The model 500 includes an encoder-decoder neural network 510, which has an encoder layer 520 that reduces the dimensionality of the pre-processed signal (i.e., pre-processed scatterometry data), a bottleneck layer 540, and a decoder layer 550. The input layer converts the pre-processed signal into a signal that is calibrated by knob settings. k trainingThese two inputs can be combined in many ways, but are typically tied together (as described above with respect to Figure 4, using the knob setting k training (can also be introduced at an internal layer of the encoder). The encoder transforms the input into an optimally reduced bottleneck, for example by fully connected layers, or convolutional and / or pooling layers. The decoder augments the bottleneck layer representation through an arbitrary set of deep network layers (usually symmetric to the encoder, but not necessarily). The output is a set of processed scatterometry data. S post The loss function 540 can be set as the difference between the scatterometry data predicted by the network and the measured set of processed scatterometry data.

[0048] The second part of the ML 500 is a metrology interpretation function, which refers to the calibration of the network output (post-process scatterometry data) to the wafer parameters to be controlled. Typically, this calibration by the OCD model 570 uses a second loss term and uses labeled data. The second loss term 545 can be set as the difference between the predicted parameter (output of the OCD model 570) and the measured target parameter.

[0049] The two loss terms compete in the gradient direction during training convergence, which means that an additional hyperparameter in the form of these loss weights must be tuned.

[0050] Figure 6 shows the application of ML model 500 in production, i.e., for inference, the knob recommendation value k recommendedAn optimization step can be applied to bring the model output as close as possible to the target parameters. k recommended can be achieved by minimizing the distance metric D. The metric D is defined as: 1) the pre-processing spectrum S pre and knob settings k 1) the predicted values ​​of the pattern parameters produced by the model based on the above (hereinafter referred to as P^ (the symbol P with a ^ above it)), and 2) the target values ​​of the pattern parameters. P target This is the difference between the above. k recommended The value of can be estimated by the following formula:

[0051]

number

[0052] As shown in Figure 6, the operation of the ML model during production involves inputting a new set of pre-process scatterometry data and by keeping the knob input node(s) free for optimization. A runtime optimization step searches for knob values ​​that minimize the difference D between the model's predicted output parameters and fixed desired target parameters. The inverse of the OCD model 570 (model -1 Note that the parameter D (denoted as 670) is applied to convert the desired target parameters into the form of a scatterometry vector. The knob value that achieves the minimization of D is the recommended knob value.

[0053] Feedforward machine learning models, such as ML Model 410 or 510, can also be trained with additional data inputs, enhancing the model's ability to account for variations such as tool-to-tool variability and process time drift. The above method compensates for variability resulting from the heterogeneity of multiple possible manufacturing routes and process steps. Additional factors can introduce further variability, affecting process control performance and stability. Such variability can arise from the tool's momentary physical state, for example, as the physical parts of active process tools tend to erode over time. Physical processes such as erosion, moving parts, sidewall deposition, and residues all contribute to dynamic tool health, performance, and noise characteristics. Routine inspections, performed periodically to clean or replace parts and recalibrate settings, can reduce such degradation. This means that the inspection cycle point for each manufacturing instance is a critical factor affecting performance. Additionally, each tool / chamber in the process is in a unique time position relative to the inspection point, which introduces another source of variation into the process known as "tool-to-tool" variation (or "chamber-to-chamber" depending on the tool module setting). A record of this time information can be kept in the tool log. By accounting for such time information in the model input, these sources of time drift and tool-to-tool variation can be reduced.

[0054] These variations can be continuously tracked and addressed using small modifications to process tool settings according to feedback from previous performance, such as polishing pressure in a CMP process. These "APC settings" can be implemented frequently depending on the stability of the process. ML training that takes these tool records (i.e., the specific instances of process tools employed, and their inspection schedules) as input extends the ability of the ML model to handle these variations, both within wafer and between wafers.

[0055] It should be understood that the process elements shown or described in this specification are preferably implemented by one or more computers in computer hardware and / or computer software embodied in non-transitory computer-readable media through conventional techniques such as the employment of computer processors, memory, I / O devices, and network interfaces coupled via a computer bus or other connection configuration.

[0056] Unless otherwise stated, the terms "processor" and "device" are intended to include any processing device, such as, for example, one that includes a CPU (Central Processing Unit) and / or other processing circuitry (e.g., a GPU), and may refer to multiple processing devices. Various elements associated with a processing device may be shared by other processing devices.

[0057] The term "memory" as used in this specification is intended to include memory associated with a processor or CPU, such as, for example, RAM, ROM, fixed memory devices (e.g., hard drives), removable memory devices (e.g., diskettes, tapes), flash memory, etc. Such memory may be considered a computer-readable storage medium.

[0058] Additionally, the phrase "input / output device" or "I / O device" may include one or more input devices (e.g., keyboard, mouse, scanner, HUD, etc.) for inputting data into a processing unit, and / or one or more output devices (e.g., speakers, displays, printers, HUD, AR, VR, etc.) for presenting results by the processing unit.

[0059] Embodiments of the invention may include systems, methods, and / or computer program products. The computer program product may include a computer-readable storage medium (or media) having computer-readable program instructions thereon for causing a processor to perform aspects of the invention.

[0060] 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 disc read-only memory (CD-ROM), digital versatile disk (DVD), Blue-Ray, magnetic tape, holographic memory, memory sticks, floppy disks, mechanically encoded devices having instructions recorded on punch cards, ridge structures in grooves, etc., and any suitable combination thereof. As used in this specification, computer-readable storage media is not to be construed as being a transitory signal per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide or other transmission medium (e.g., light pulses passing through a fiber optic cable), or electrical signals transmitted over wires.

[0061] The computer-readable program instructions described in this specification can be downloaded from a computer-readable storage medium to each computing / processing device or to an external computer or external 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, optical transmission fiber, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium of the respective computing / processing device.

[0062] The computer-readable program instructions for carrying out the operations of the present invention can be either source code or object code written in any combination of one or more programming languages, including 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++, and traditional procedural programming languages ​​such as the "C" programming language or similar programming languages. The computer-readable program instructions can be executed completely on the user computer, partially on the user computer, as a standalone software package, partially on the user computer, partially on a remote computer, or completely on a remote computer or server. In the latter scenario, the remote computer can be connected to the user computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or can 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) can utilize state information of computer-readable program instructions to personalize the electronic circuitry and execute the computer-readable program instructions to carry out aspects of the invention.

[0063] Where aspects of the invention are described in this specification with reference to flowchart illustrations 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 illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0064] These computer-readable program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus to produce a machine whereby the instructions, when executed by the processor of the computer or other programmable data processing apparatus, create means for performing the functions / acts specified in the flowchart and / or block diagram blocks. These computer-readable program instructions can be stored on a computer-readable storage medium that can direct a computer, programmable data processing apparatus, and / or other device to function in a particular manner, whereby a computer-readable storage medium having instructions stored therein can also comprise an article of manufacture containing instructions that implement aspects of the functions / acts specified in the flowchart and / or block diagram blocks.

[0065] The computer-readable program instructions may 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, whereby the instructions executing on the computer, other programmable apparatus, or other device perform the functions / actions specified in the flowchart and / or block diagram blocks.

[0066] Any flowcharts and block diagrams included herein illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of instructions, which comprises one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions noted in the blocks may occur out of the order depicted in this specification. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending on the functionality involved. It should also be noted that each block of the block diagrams and / or flowchart diagrams, and combinations of blocks in the block diagrams and / or flowchart diagrams, may be implemented by a special-purpose hardware-based system that performs the specified functions or actions or executes a combination of special-purpose hardware and computer instructions.

[0067] The descriptions of various embodiments of the present invention are presented for illustrative purposes and are not intended to be exhaustive or limiting of the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terms used in this specification have been selected to best explain the principles of the embodiments, practical applications or technical improvements over commercially available technology, or to enable those skilled in the art to understand the embodiments disclosed in this specification.

Claims

1. 1. A system for advanced process control (APC) in semiconductor manufacturing, comprising one or more processors having one or more associated non-transitory memories, the non-transitory memories containing instructions for execution by the one or more processors, the instructions comprising: receiving, for each of a plurality of wafer sites, a pre-process set of scatterometry training data measured before performing a process step, a corresponding post-process set of scatterometry training data measured after performing said process step, and a set of process control knob training data indicative of process control knob settings applied during performance of said process step; generating a machine learning model that correlates variability in the pre-process set of scatterometry training data and the corresponding process control knob training data to the corresponding post-process set of scatterometry training data, and training the machine learning model to recommend changes to process control knob settings to compensate for variability in the pre-process set of scatterometry training data.

2. 10. The system of claim 1, further comprising applying the machine learning model to generate process control knob setting recommendations during semiconductor manufacturing.

3. 10. The system of claim 1, wherein the machine learning model is a first machine learning model, and the post-process set of scatterometry training data is correlated to one or more target post-process pattern parameters by an optical model or by a second machine learning model.

4. 10. The system of claim 1, wherein the process control knob settings include settings for one or more of a process step duration, a pedestal edge ring height, a temperature distribution across multiple control zones of the pedestal, and a process chamber pressure.

5. 10. The system of claim 1, wherein each of the pre-process and post-process sets of scatterometry data is indicative of one or more pattern parameters at a respective wafer site, including one or more of a critical dimension, a feature depth, a feature height, and a feature pitch.

6. 10. The system of claim 1, wherein the process step is one or more of a deposition operation, an etching operation, and a polishing operation.

7. 2. The system of claim 1, wherein generating the machine learning model comprises training a neural network (NN) including a plurality of encoder layers connected to a bottleneck latent layer and then connected to at least one decoder layer, wherein the pre-processed set of scatterometry training data is applied as a model input, the corresponding processed set of scatterometry training data is applied as a model output, a plurality of sets of the process control knob training data are applied as auxiliary inputs intersecting the NN at any one of the plurality of encoder layers, and the plurality of process control knob training data are applied as auxiliary outputs linking to any one of the at least one decoder layer.

8. 8. The system of claim 7, wherein a loss function for backpropagation of the NN maximizes the similarity between the model output from the NN and the processed set of scatterometry training data.

9. The system of claim 8 , wherein the loss function is a squared error loss function.

10. 8. The system of claim 7, wherein the machine learning model includes a calibration step following the neural network, which calibrates the post-process set of scatterometry training data to predicted post-process pattern parameters, and wherein the calibration is performed by an OCD model.

11. 11. The system of claim 10, wherein optimizing the machine learning model comprises minimizing a difference between the post-process set of scatterometry training data and the predicted post-process pattern parameters.

12. 8. The system of claim 7, wherein a loss function for backpropagation of the auxiliary output of the NN represents a quality of similarity between the auxiliary output and the set of process control knob training data.

13. The system of claim 12 , wherein the loss function is a squared error loss function.

14. 10. The system of claim 1, wherein the plurality of wafer sites are located on a plurality of wafers.

15. 10. The system of claim 1, wherein the multiple sets of pre-process and post-process scatterometry training data are measured by two or more measurement channels.

16. 1. A method for advanced process control (APC) in semiconductor manufacturing, comprising: receiving, for each of a plurality of wafer sites, a pre-process set of scatterometry training data measured before performing a process step, a corresponding post-process set of scatterometry training data measured after performing said process step, and a set of process control knob training data indicative of process control knob settings applied during performance of said process step; generating a machine learning model that correlates variability in the pre-process set of scatterometry training data and the corresponding process control knob training data to the corresponding post-process set of scatterometry training data, and training the machine learning model to recommend changes to process control knob settings to compensate for variability in the pre-process set of scatterometry training data.

17. A non-transitory machine-accessible storage medium having stored thereon instructions that, when executed by a machine, cause the machine to: receiving, for each of a plurality of wafer sites, a pre-process set of scatterometry training data measured before performing a process step, a corresponding post-process set of scatterometry training data measured after performing said process step, and a set of process control knob training data indicative of process control knob settings applied during performance of said process step; generating a machine learning model that correlates variability in the pre-process set of scatterometry training data and the corresponding process control knob training data with the corresponding post-process set of scatterometry training data, and training the machine learning model to recommend changes to process control knob settings to compensate for variability in the pre-process set of scatterometry training data.

Images