Increased resolution of semiconductor manufacturing data collection instruments using machine learning

Abstract

A data set is stored in memory circuitry, the data set indicating a state of a semiconductor manufacturing process or a semiconductor structure manufactured thereby. Features in the data set are resolvable within limits constrained by a data resolution. A machine learning model is also stored in the memory circuitry, the machine learning model including parameters having respective assigned values constrained by a model training process. Processor circuitry communicatively coupled to the memory circuitry generates an output data set from the data set in accordance with the machine learning model such that features in the output data set are resolvable within limits constrained by an output data resolution, the output data resolution being finer than a data resolution of the data set.

Description

[0001] Cross-reference to related applications

[0002] This application claims the benefit of priority to U.S. Provisional Patent Application No. 62 / 819,115, filed March 15, 2019, entitled “Deep-learning Enhanced Spectrometer Resolution Techniques”, pursuant to 35 USC §119(e), the entire disclosure of which is incorporated herein by reference.

[0003] background Technical Field

[0004] This disclosure relates to techniques for improving the imaging and spectral resolution of images and spectral data used to monitor semiconductor processing or features formed on semiconductor substrates. Specifically, this disclosure provides new techniques for applying low-resolution images or spectral data to trained artificial neural networks to generate high-resolution images or spectral data. Background Technology

[0005] Broadband compact spectrometers are widely used in semiconductor plasma etching processes to collect light emission spectra for various process control purposes, such as endpoint detection. The signal-to-noise ratio (SNR) of a spectrometer system defines the quality of the information upon which control signals are based. SNR is typically highly proportional to the spectrometer's throughput. Typically, spectrometer performance is quantified by several key parameters, including: spectral bandwidth, the detectable wavelength range; spectral resolution, the finest spectral features that can be resolved; and optical throughput, the percentage of light collected by the detector.

[0006] Higher resolution spectrometers typically sacrifice throughput and spectral (wavelength) range to achieve higher resolution results. Similarly, spectrometers designed for higher throughput and wider spectral range usually have lower resolution. Currently, most high-resolution spectrometers lack the ability to resolve finer details within spectral data while maintaining the throughput and spectral range of low-resolution spectrometers, especially when instrument cost is a limiting design factor. Many existing spectrometer designs involve trade-offs between key design parameters. A simple and effective way to improve spectral resolution is to use a narrower slit as the spectrometer inlet, which effectively narrows the "image" of each wavelength of the diffraction spectrum on the charge-coupled device (CCD) detector; however, this comes at the cost of reduced light passing through the slit and the corresponding information. Overcoming the trade-off between resolution and throughput is difficult through physical spectrometer design alone. Summary of the Invention

[0007] One objective of this disclosure is to mitigate the aforementioned trade-offs between the physical design parameters associated with conventional spectrometers.

[0008] Another objective of this disclosure is to enable the use of low-resolution spectrometers to collect spectral data and to resolve finer details within spectral data or images without sacrificing throughput or spectral range.

[0009] In one aspect of the present invention, an apparatus is provided comprising a memory circuit system storing a dataset indicating the state of a semiconductor manufacturing process or a semiconductor structure manufactured therefrom. Features in the dataset are discriminable within limits limited by data resolution. The memory circuit system also stores a machine learning model including parameters having assigned values ​​constrained by a model training process. A processor circuit is communicatively coupled to the memory circuit system and generates an output dataset from the dataset according to the machine learning model, such that features in the output dataset are discriminable within limits limited by an output data resolution finer than the data resolution of the dataset. Attached Figure Description

[0010] Figure 1 This is a schematic block diagram of an example system in which the concepts of the present invention can be implemented.

[0011] Figure 2 This is a schematic diagram of an artificial neural network (ANN) through which the present invention can be implemented.

[0012] Figure 3 This is a schematic diagram of an example spectrometer that can be used to implement the present invention.

[0013] Figure 4 This is a schematic diagram of the spectrometer inlet slit and its image 420 used to simulate data in embodiments of the present invention.

[0014] Figure 5 This is a graph of low-resolution spectral data from underfilled and overfilled embodiments of the low-resolution spectrometer used in the embodiments of the present invention.

[0015] Figure 6 This is a graph of high-resolution spectral data from underfilled and overfilled embodiments of the high-resolution spectrometer used in the embodiments of the present invention.

[0016] Figure 7 This is a flowchart illustrating an exemplary training process through which the present invention can be implemented.

[0017] Figure 8 This is a flowchart illustrating an exemplary production process through which the present invention can be implemented. Detailed Implementation

[0018] The concept of the invention is best described through certain embodiments thereof, which are described in detail herein with reference to the accompanying drawings, in which similar reference numerals refer to similar features. It should be understood that, when used herein, the term "invention" is intended to refer to the inventive concept implied by the embodiments described below, and not merely to the embodiments themselves. It should further be understood that the general inventive concept is not limited to the exemplary embodiments described below, and the following description should be read with this understanding.

[0019] The term “exemplary” is used herein to mean “as an example, instance, or illustration.” Any embodiment of a construction, process, design, technique, etc., specified herein as exemplary is not necessarily to be construed as being more preferred or advantageous than other such embodiments. The particular quality or suitability of the examples indicated herein as exemplary is neither intentional nor should be inferred.

[0020] Furthermore, this document contains mathematical expressions, and the principles conveyed therefrom will be considered to be thoroughly described using these expressions. It will be understood that, in the use of mathematical operations, this is a convenient description for explaining the fundamental principles, and unless otherwise stated, no other purpose is implied or should be inferred. Overall, it will become clear from this disclosure how the mathematical operations herein relate to the invention and the intent of embodiments of the principles implied by the mathematical expressions; those skilled in the art will recognize that many techniques are used to implement the physical manifestations of principles expressed mathematically.

[0021] This disclosure relates to “constructing” or generating high-resolution (HR) datasets from data collected by a low-resolution (LR) spectrometer or image capture device. In other words, the techniques disclosed herein enable HR spectrometers to operate with an extended wavelength range and higher optical throughput than LR spectrometers. Furthermore, variations of similar techniques can be used to generate higher-resolution image data from lower-resolution images.

[0022] The invention described herein combines deep learning techniques with relatively low-resolution spectral data collected, for example, by conventional optical spectrometer designs. It uses predictive models built upon the analysis of paired low-resolution and high-resolution datasets (e.g., experimentally or simulated) associated with features on a semiconductor process or substrate to obtain higher-resolution data. In this way, high-resolution data on semiconductor process or substrate features can be acquired using lower-cost hardware and faster cycle times compared to using existing relatively high-resolution spectrometers or imaging systems and techniques. For example, in spectral embodiments, key criteria for defining the quality of a spectral dataset include bandwidth (i.e., the detectable wavelength range), resolution (i.e., the finest spectral feature that can be resolved), and optical throughput (i.e., the percentage of light collected by the detector). In most cases, these factors are balanced through trade-offs in the physical design of the spectrometer to achieve predictable and meaningful results (e.g., patterns, features, or process conditions) in the shortest possible time. For example, higher-resolution spectral datasets are obtained at the expense of higher throughput and wavelength range. Therefore, it is advantageous to obtain higher-resolution data without making these sacrifices. In addition to spectral embodiments, deep learning techniques can also be applied to other low-resolution data (e.g., images) to decipher higher-resolution information or characteristics of the analyzed semiconductor process or substrate features.

[0023] One way to achieve this advantage is to construct a machine learning (ML) model capable of generating high-resolution datasets from lower-resolution datasets. Ideally, the high-resolution data generated by this model should provide information that would be difficult to discern from low-resolution data without processing it through the ML model. This ML model can be prepared by analyzing low-resolution and high-resolution training data collected experimentally or through simulation to determine statistical transformations or relationships that can be used to train an artificial neural network (ANN) (such as a convolutional neural network (CNN)). For example, a CNN with several layers can be trained to statistically correlate low-resolution data with high-resolution data regarding a process or feature. Low-resolution and high-resolution data can be collected on similar or slightly different processes or features to improve the robustness of the statistical relationship between the datasets. In this way, the ML model is improved to accurately predict high-resolution results, and the model can be optimized by comparing the modeled results with actual high-resolution data. After validating the accuracy and precision of the model, the trained network can be applied to low-resolution data to generate corresponding high-resolution data or results related to semiconductor processes or features. Two spectrometers can be used to generate corresponding low-resolution and high-resolution datasets representing process or substrate features. However, in other embodiments, low-resolution and high-resolution datasets can be generated by simulating measured spectra to approximate actual process conditions. In this way, simulated spectral data can be optimized to simulate specific spectral characteristics based on known or observed phenomena, allowing training datasets to include fully simulated low and high datasets, combinations of simulated and experimental training pairs, or training pairs comprising both simulated and experimental data.

[0024] Figure 1 This is a schematic block diagram of an example system 100 through which the concepts of the present invention may be implemented. For purposes of description and not limitation, system 100 may be constructed or otherwise configured for semiconductor manufacturing, wherein a significant portion of the manufacturing process occurs in one or more sealed chambers, represented by chamber 5. Chamber 5 may include one or more ports (represented by port 7) through which the processes occurring within chamber 5 can be monitored, such monitoring may be performed by means of spectroscopic techniques, optical imaging, or other techniques known to those skilled in the art. Such internal processes may include material deposition and etching, both of which may utilize plasmas having a composition depending on the process. The composition within chamber 5 may be evaluated by, for example, spectroscopic techniques. Other processes may be monitored or inspected by optical or electronic imaging. As those skilled in the art will recognize upon reviewing this disclosure, the techniques implemented in embodiments of the present invention are applicable to optically acquired images and other acquired images in spectroscopic form.

[0025] System 100 may include processor 150, which includes a memory circuitry 120 for storing data and processor instruction code, a processor circuitry 130 for executing data and control processes, and a user interface circuitry 140 through which a human user interfaces with system 100 for, among other things, data analysis and control. Figure 1 As shown, memory circuitry 120 can be constructed or otherwise configured to store datasets, for example, in data storage location 122, machine learning models, for example, in model storage location 124, and processor instruction code, for example, in code storage location 126. The processor instruction code may include, for example, ML engine code in ML engine code storage location 127 (when the ML engine is not hardwired in a dedicated circuitry) and process control code, for example, in process controller code storage location 129. Processor circuitry 130 can be constructed or otherwise configured to execute the processor instruction code stored in memory circuitry 120, such as as shown with ML engine 132 and process controller 134.

[0026] System 100 may also include data acquisition mechanisms, typically represented by data acquisition mechanisms 105l and 105h. As described above, the invention can be implemented to improve the resolution of various data types, thus, instrument 105l, representing a low-resolution data acquisition instrument that generates low-resolution data 110l, and instrument 105h, representing a high-resolution data acquisition instrument that generates high-resolution data 110h, represent any number of devices that can be used to collect specialized data (e.g., image data, spectral data, etc.). It will be understood that although the following description primarily relates to spectral datasets, other data may also be used in embodiments of the invention.

[0027] Experimental or simulated data can be collected or generated under various conditions or variables to establish or define a predictable relationship between low-resolution and high-resolution datasets 110l and 110h. This relationship can be based on common elements between process conditions, feature types, or detection techniques. However, the relationship can also be assigned based on differences or dissimilar elements between datasets 110l and 110h (e.g., device setup / configuration (slit size, field of view, aperture size)). In practice, datasets 110l and 110h can be collected or generated according to various device setup / configuration options to obtain a resolution distribution for the observed process conditions or features. This training relationship can group one or more low-resolution datasets 110l to one or more high-resolution datasets 110h associated with process condition characteristics, including but not limited to process-related chemical composition, temperature, density, and / or energy or signal. Furthermore, the relationship can be extended to process conditions implemented in chamber 5 (e.g., setpoints), which may include, but are not limited to, temperature, pressure, gas flow rate, applied power, applied voltage, rotational speed, translational speed, liquid distribution rate, and liquid distribution volume. Furthermore, process conditions can include a range of results resulting from distributions or profiles arranged to correspond to specific characteristics. In addition to process conditions, spectral data can represent one or more features on the substrate. In this case, feature characteristics can include, but are not limited to, feature thickness, width, height, orientation, roughness, flatness, angle, coverage, alignment, and / or geometric profile. Additionally, feature characteristics can refer to the area between features, distance between features, or the orientation of features relative to each other (e.g., feature overlap, exposed area, alignment). Any of the above applications in generating experimental or simulated LR and HR datasets can be used to train artificial neural networks using the techniques described herein.

[0028] System 100 can operate in two (2) operating modes: a training mode, in which a machine learning model is configured to improve data resolution; and a production mode, in which data resolution is improved based on the configured (trained) model. In production mode, an LR dataset 110l (image, spectrum) is acquired via an LR data acquisition instrument 105l (imager, spectrometer) and provided to a processor 150. The processor 150 can pass the LR dataset 110l to an ML engine 132, which can process the data (as described below) based on a previously trained ML model to improve the data resolution. This improved dataset can be presented on a user interface 140, on which production personnel can analyze the dataset. In one embodiment, the improved dataset is provided to a process controller 134, which analyzes the data to determine whether a specific target process state has been reached. For example, analysis of the improved dataset can reveal specific gas components in chamber 5 that indicate a process endpoint. At such an endpoint, the process controller 134 can instruct or otherwise command a process change or abort all processes occurring in chamber 5.

[0029] Figure 2 This is a schematic diagram of an artificial neural network (ANN) 200 through which the present invention can be implemented. The present invention is not limited to a specific ANN implementation, as long as the goal of improving the data resolution of the output data is achieved. In the example shown, the ANN 200 is constructed or otherwise configured as a pyramid residual neural network, which includes an input convolutional layer 210 and multiple residual blocks B1 to B2. K (Represented by residual blocks 220 and 230), and output convolutional layer 240. In ANN 200, 1) the size of the internal feature map or the number of feature map channels is gradually increased from the input to the output of the ANN, and 2) identity mapping is implemented in a shortcut that solves the gradient vanishing / exploding problem.

[0030] exist Figure 2In an exemplary embodiment, LR datasets 205a to 205m (represented herein by way of LR dataset 205) can be provided to an input convolutional layer 210, which, among other possible operations, maps each dataset 205 to a corresponding feature map 215a to 215m (represented herein by way of feature maps 215). Feature maps 215 can be provided to residual blocks 220 (which are the same as all K residual blocks in ANN 200), which can include a first convolutional layer 260a, a first activation layer 265a, a second convolutional layer 260b, and a second activation layer 265b. Additionally, each feature map 215 can be provided to summing nodes 225a to 225n (represented herein by way of summing nodes 225) via corresponding shortcuts 217a to 217n (represented herein by way of shortcuts 217). Therefore, the output of each summing node can be given by the following equation:

[0031]

[0032] Among them, X k X is the input of the k-th residual block. k+1 It is its output, ψ is the activation function (e.g., Corrected Linear Unit (ReLU)), and * is the convolution operator. and This is the trainable convolutional kernel for the k-th block. Note that the bias term has been omitted from the above expression. The feature map generated by each convolutional layer can be given by the following equation:

[0033] γ k，j =∑ i φ k，i *w k，i，j +β k，j Ω,

[0034] Where, γ k，j It is the output feature map of the j-th convolutional layer of the k-th residual block, φ k，i w is the input feature map of the i-th convolutional layer of the k-th residual block. k，i，j yes The i-th kernel, β k，j Ω is the trainable bias term, and Ω is a matrix of the feature map size with all terms set to 1.

[0035] Feature map 215 is processed by residual block 220 and passed to the next residual block, namely residual block 230, where these feature maps receive similar processing. However, each residual block can generate feature maps (and / or associated feature map channels) with more dimensions than the previous residual block. For example, the input to residual block 220 may include N channels 227a to 227n, and the output of residual block 220 may include P channels 232a to 232p, where P > N. This dimensionality mismatch can be overcome, for example, by zero-padding the shortcut connections (e.g., shortcut connections 227a to 227n that bypass residual block 230).

[0036] As the dataset progresses through ANN 200 and its dimensions gradually increase, increasingly finer features can be discerned block by block, achieving an improvement in resolution relative to the original dataset 205. The processed feature maps proceed to the output convolutional layer 240, where, among other optional processing operations, these feature maps are transformed into output datasets 245a through 245m (represented herein as output dataset 245). If ANN 200 has already been trained, output dataset 245 is provided to organizations outside the ANN, such as for data analysis or process control. However, if ANN 200 is being trained, output dataset 245 is compared with corresponding high-resolution training datasets 250a through 250m (represented herein as HR training dataset 250) selected to correspond to the data contained in the original LR dataset 205. That is, each original LR dataset 205 is matched with the HR training dataset 250 such that the same features are presented in both, but at different resolutions. During training, the original LR dataset 205 is transformed into the HR dataset according to a randomized initialization function. The transformed HR dataset is appropriately compared with the corresponding HR training dataset (e.g., through differencing) to determine an error metric. This error metric is then used to modify the ANN parameters, as described above. and β k，j This is done to minimize (or maximize) an objective function such as the training loss function. The modified parameters can be backpropagated through ANN 300 (represented by backpropagation channel 257), and the process can be repeated with new parameters to produce additional HR output datasets. This process iterates until the training objective has been achieved, such as when the trainable network parameters converge or when further resolution improvements are no longer possible.

[0037] LR dataset 205 may include LR spectrometer data, and HR dataset 250 may include HR spectrometer data. In addition to the spectral data itself, information related to the observed processes or features associated with the spectral data can be used to label the HR and LR datasets. Depending on the type of model optimization required to generate predictable HR results based on inputs from random but similar LR datasets, such labeling can be used to isolate specific aspects of the HR / LR datasets or to allow some aspects of the HR / LR datasets to overlap. HR / LR datasets can be grouped into training pairs (which may include combinations of one or more HR / LR datasets) and used to train ANN 200 based on identifying relationships between the HR / LR training pairs. At a high level, training is an iterative process where HR data generated according to the ML model is compared with the HR dataset used to optimize the statistical relationships between the datasets, using ML techniques.

[0038] In another embodiment, the LR dataset 205 includes image data that can be used to train an ML model. In this case, the LR dataset 205 is represented by LR images, while the HR dataset 250 is represented by HR images. Both types of images can represent input data collected by the low-resolution data acquisition instrument 105l and the high-resolution data acquisition instrument 105h as described above. In addition to the image data, information related to the observed process or features associated with the image data can be used to label the HR and LR datasets, allowing the labeling information to highlight differences or similarities between the datasets. Depending on the type of model optimization required to generate predictable HR results based on inputs from random but similar LR datasets, the labeling information can be used to isolate specific aspects of the HR / LR datasets or to make certain aspects of the HR / LR datasets overlap. The HR / LR datasets are organized into training pairs (which may include one or more combinations of HR / LR datasets) and used to train an ANN 200 using ML techniques based on identifying relationships between the HR / LR training pairs, thereby enabling the ML model to predict or create the HR dataset 245 using the LR dataset 205.

[0039] It should be understood that the training examples provided above are merely one of many possibilities for training the ANN of this invention. In one embodiment, the dataset generated by a data acquisition instrument or data generator can be partitioned, for example, into: a training set for training learning techniques, a validation set for evaluating the model architecture, and a test set for determining whether the model has been appropriately generalized to datasets beyond those used for training. Those skilled in the art of machine learning will recognize how to use such a partitioned dataset without needing the specific details set forth herein, and will also recognize other machine learning techniques for implementing this invention without departing from its spirit and intended scope.

[0040] Return again Figure 1 System 100 may include a high-resolution data generator 115, through which simulated datasets are generated for model training purposes. In addition to image and spectral data, simulated spectral model data can also be used to train the ANN 200. However, training pairs can be derived from both low-resolution and high-resolution simulations, both of which are derived from data similar to the following... Figure 3 The estimated LR and HR spectrometer data of the spectrometer design shown.

[0041] Figure 3 This is a schematic diagram of an example spectrometer 300 that can be used to implement the present invention. The spectrometer 300 may follow a Czerny-Turner design and includes an entrance slit 310, a collimating mirror 320, a diffraction grating 330, a folded mirror 340, and an array detector 350. Light (including...) Figure 1 The light generated during the electrochemical process in chamber 5 enters the spectrometer 300 through the entrance slit 310 and is reflected from the collimating mirror 320 onto the diffraction grating 330, where the incident light is resolved to its spectral components. These components are reflected from the folded mirror 340 onto the array detector 350, which generates an electrical signal indicating the spectral content of the analyzed light. Essentially, a spectral resolution image of the entrance slit 310 is generated on the array detector 350, which can be used with... Figure 1 The HR data generator 115 implements various optical system design / analysis techniques to simulate this imaging.

[0042] Figure 4 This is a schematic diagram of the entrance slit 310 and the image 420 generated by the optical system of the spectrometer 300. In the illustrated embodiment, twelve (12) imaging points 412a to 412l (represented herein as imaging points 412) on the entrance slit 310 were selected for imaging. Using optical design / analysis techniques, the imaging points 412 were transferred to the image 420 generated on the array detector 350 for each of the multiple selected wavelengths, as represented by image region 425. Note that for a fixed-size slit, the sharpness of the imaging points 412 is affected by wavelength-dependent diffraction.

[0043] Model data for LR and HR spectrometers can be generated, simulating light from a calibration lamp in the 300nm to 900nm spectrum with similar spectral intensities. For demonstration purposes, two different spectrometer designs were modeled for LR and HR implementations to form two training pairs. The LR dataset in the training pairs was modeled using a 23µm slit and 1024 CCDs, and two additional LR datasets were generated by varying the placement of the grating relative to the slit. The underfilled embodiment represents a spectrometer design where the grating is positioned closer to the slit than the overfilled embodiment. In this way, the LR training dataset covers a wider range of LR spectrometer designs and / or capabilities and indicates performance (e.g., SNR) differences between the two different LR designs. For example, Figure 5 The study demonstrates the intensity differences between LR spectrometer designs, where the underfilled embodiment has a higher and sharper profile than the overfilled embodiment.

[0044] Overfilled and underfilled LR datasets can be paired with corresponding overfilled and underfilled HR datasets. The example HR dataset is generated using a 10µm slit and a 2048×256 pixel CCD by varying the raster distance between the underfilled and overfilled datasets. Figure 6 The differences between HR spectrometer designs are demonstrated by highlighting the intensity differences. In contrast to the LR dataset, the overfilled dataset results in a higher intensity signal compared to the underfilled dataset.

[0045] Once the ANN 200 has been trained and validated, the model can be applied to applicable low-resolution datasets to output estimated high-resolution data corresponding to the low-resolution input. In contrast, the trained model receives low-resolution data (without requiring a corresponding high-resolution data) and applies that data to generate the corresponding high-resolution output. Low-resolution data can include, but is not limited to, image or spectral data related to features on semiconductor processing techniques or semiconductor substrates.

[0046] Within the semiconductor space, deep learning techniques can be applied to images ranging from macroscopic to microscopic applications, depending on the size of the observed features. In macroscopic applications, the observed features can range from the entire substrate surface area to a relatively large surface area, allowing for the observation and monitoring of process trends at a macroscopic level. For example, a large surface area could include substrate chamfers or edge exclusion regions (<5 mm) extending near the substrate periphery. Similarly, regions adjacent to edge exclusion regions and extending up to 20 mm toward the substrate center can be observed to monitor process execution near the substrate edge, where process uniformity issues relative to the substrate center are known. Therefore, comparing images of the substrate center and edge can be beneficial for observing or detecting process deviations, or for adjusting the process to address in-situ or impending process non-uniformities. In some cases, depending on the application, the center-to-edge relationship can be used to train low-to-high-level models. In macroscopic applications, the image capture device can be a digital system or an analog-to-digital system that produces pixelated images for analysis. In one embodiment, the image capture device can be a CCD camera. Conversely, microscopic application embodiments are used to observe relatively small features on a die or smaller areas within a die, compared to macroscopic applications. For example, microscopic application images can have a scale for controlling and monitoring processes with atomic-level precision. Atomic-level control over transistor gate structures (e.g., FinFETs), transistor components (e.g., contacts, isolators), or back-end process (BEOL) features (e.g., metal line spacing, vias) is extremely important. In one embodiment, microscopic application images can be generated by scanning electron microscopy (SEM), which uses a focused electron beam to scan the substrate surface / features, thereby providing information on the substrate's morphology and composition. Alternatively, optical emission spectroscopy techniques can be used to generate images for monitoring semiconductor processes, including but not limited to plasma processing.

[0047] Figure 7 This is a flowchart illustrating an exemplary training process 700 through which the present invention can be implemented. It is assumed that process 700 has access to a training corpus comprising matching pairs of LR and HR datasets (e.g., spectral data acquired from LR and HR spectrometers described herein). The training corpus may be appropriately partitioned or otherwise divided into training bodies 705a, validation bodies 705b, and test bodies 705c. Dividing the corpus into training, validation, and test bodies can follow such techniques known in the field of machine learning. In fact, process 700, in flowchart form, represents one of many model building / training paradigms that can be used in embodiments of the present invention. In the following description, this model paradigm refers to the above references. Figure 2 The paradigm of description. What will be understood is... Figure 7The order of operations and data flow in this example are for ease of explanation; other orders of operations and data flows can also be used to implement model building / training.

[0048] like Figure 7 As shown, training dataset 705a may include the raw LR dataset 707a, which can be obtained through a low-resolution data acquisition instrument. In operation 710, the LR dataset 707a is processed according to the ANN to produce the HR output dataset 715. In operation 720, the differences between the HR output dataset 715 and the training dataset 709a, which is obtained through a high-resolution data acquisition instrument or simulated through a high-resolution data generator, are determined. In operation 725, the ANN parameters (weights, biases) are determined to maximize or minimize the objective function. In operation 730, the determined ANN parameters are backpropagated through the ANN.

[0049] Once the training set is exhausted, process 700 can transition to operation 735, thereby determining whether the model is overfitting the data or, conversely, generalizing to a given specification. For this purpose, a validation set 705b, comprising the original LR dataset 707b and the HR validation set 709b, can be applied to the trained model in a manner similar to that described above in conjunction with training set 705a. However, this time the focus is on the model's architecture, as represented by so-called hyperparameters, rather than its predictive efficiency. Therefore, if it is found in operation 735 that the model is not generalizing well, one or more of its hyperparameters can be modified via operation 740. For example, the reference above can be removed. Figure 2 The residual blocks of the model, or the order of operation blocks within each residual block, can be revised, etc. Those skilled in the art will recognize that many techniques can be applied to remedy overfitting in embodiments of the invention.

[0050] If it is determined in operation 735 that the model generalizes well, process 700 can transition to operation 750, thereby determining whether the trained model meets performance criteria. For this purpose, test body 705c, comprising the original LR dataset 707c and the HR test set 709c, can be applied to the trained model in a manner similar to that described above, combining training body 705a and validation body 705b. However, this time the focus is on the model's performance, which is measured by specific performance criteria in terms of accuracy (i.e., how close the predicted HR dataset from the model is to the real HR dataset generated from actual measurements of the observations) and sensitivity (i.e., the proportion of features correctly identified in this way in HR predictions (by the model)). If such criteria are met, process 700 can terminate. Otherwise, process 700 can transition to operation 740, thereby modifying the ANN hyperparameters again. Process 700 can then transition to operation 710 and continue from there.

[0051] Figure 8 This is a flowchart illustrating an exemplary production process 800 through which the present invention can be implemented. Production process 800 can be executed when the system is in production mode. At operation 810, the LR dataset 805 is used to predict the corresponding HR output dataset 815, which is the output of the ANN described above. Process 800 can then transition to operation 820, where the process state is determined based on the HR output dataset 815. If it is determined in operation 825 that the target process state has been reached, production process 800 terminates. If the target process state has not been reached, process 800 can transition to operation 810 using newly acquired LR data and continue from there.

[0052] One embodiment includes a method for training a neural network to provide high-resolution data from measured low-resolution data. The training method may include storing a dataset associated with a semiconductor process or feature on a semiconductor substrate within a processing chamber in a tangible computer-readable storage medium. The dataset may include a coarse (low-resolution) dataset representing the process or feature, and a detailed (high-resolution) dataset representing the process or feature, which differs from the coarse dataset. In one example, the difference between the coarse and detailed datasets includes a wider or narrower range of information about at least one characteristic of the process or feature. Computer processing devices are used to assign training relationships between instances of the coarse and detailed datasets, and neural network techniques are used to create a model to associate the coarse dataset with the detailed dataset, at least in part, based on the training relationships and differences between the coarse and detailed datasets. In one example, for the at least one characteristic, the detailed dataset includes data at a higher resolution compared to the coarse dataset. A characteristic of the process may be the chemical composition of the process, the temperature of the process, the density of the process, and / or the electrical energy associated with the process. A characteristic of the process may include the distribution or profile of chemical composition, temperature, density, and / or electrical energy. The characteristics of a feature may include measured physical features such as the feature's thickness, width, height, orientation, roughness, flatness, angle, coverage, alignment, and / or geometric profile.

[0053] The dataset may include images, spectral representations, or modular simulations of a process or feature. In one embodiment, the dataset is at least partially based on images, spectral representations, and / or simulations of a feature or process. Neural network techniques may be at least partially based on convolutional neural network techniques, and the dataset may be generated by a spectrometer. In some embodiments, a coarse dataset is generated by a coarse spectrometer having a coarse slit for allowing light or energy to enter the coarse spectrometer, and a detailed dataset is generated by a detailed spectrometer including a detailed slit for allowing light or energy to enter the detailed spectrometer, wherein the detailed slit includes an opening or gap smaller than the corresponding opening or gap of the coarse slit. The dataset may also be generated by an image capturing device, such as a charge-coupled device camera. Alternatively, the image capturing device may include a microscope and a camera.

[0054] The training relation assignment is based, at least in part, on common elements present during the collection of coarse and detailed datasets. These common elements are process conditions of the semiconductor process or feature types of features on the semiconductor substrate. Process conditions may include temperature, pressure, chemical composition, gas flow rate, applied power, applied voltage, rotational speed, translational speed, liquid dispensing rate, and liquid dispensing volume. Feature types may include patterns in or on the substrate. In some examples, feature types include feature thickness, width, height, length, orientation, roughness, flatness, angle, coverage, alignment, and / or geometric profile. Feature types may also include critical dimensions of features, surface area of ​​features, exposed surface area of ​​features, overlap between features, distance between features, orientation between features, and / or alignment between features. The training relation assignment of this method may include pairing a single coarse dataset and a single detailed dataset collected during the process or on a feature.

[0055] One embodiment includes a method during which a dataset relating to a feature on a semiconductor process or semiconductor substrate is stored in a tangible computer-readable storage medium. The dataset may include a coarse (low-resolution) dataset representing the process or feature. The coarse dataset is applied to a detailed resolution model based at least in part on the differences between the low-resolution dataset and the high-resolution dataset relating to the process or feature. The detailed (high-resolution) dataset representing the process or feature is generated using neural network techniques, and differs from the coarse dataset, wherein these differences are based at least in part on a larger amount of information about at least one characteristic of the process or feature. The neural network techniques may be based at least in part on convolutional neural network techniques. For the at least one characteristic, the detailed dataset may include higher resolution data compared to the coarse dataset, and the at least one characteristic may include: the chemical composition of the process, the temperature of the process, the density of the process, and / or the electrical energy associated with the process; the distribution, uniformity, or profile of the chemical composition, the temperature, the density, and / or the electrical energy; and the thickness, width, height, orientation, roughness, flatness, angle, coverage, alignment, and / or geometric profile of the feature.

[0056] The dataset may include an image or spectral representation of a process or feature, and may be at least partially based on such an image or spectral representation. The dataset may be generated by a spectrometer; in some embodiments, a coarse dataset is generated by a coarse spectrometer including a coarse slit for allowing light or energy to enter the coarse spectrometer, and a detailed dataset is generated by a detailed spectrometer including a detailed slit for allowing light or energy to enter the detailed spectrometer. The detailed slit may have a larger opening or gap than the corresponding opening or gap of the coarse slit.

[0057] Data sets can also be generated by image capture devices, such as those that include charge-coupled device cameras with or without microscopes.

[0058] The training relation assignment can be based at least in part on common features that exist during the collection of coarse and detailed datasets, and also includes sending adjustments to the matching processing chamber, wherein the common features can be process conditions of the semiconductor process, feature type of features on the semiconductor substrate, temperature, pressure, chemical composition, gas flow rate, applied power, applied voltage, rotational speed, translational speed, liquid dispensing rate, and liquid dispensing volume.

[0059] Feature types can be patterns in a substrate, patterns on a substrate, feature thickness, width, height, length, orientation, roughness, flatness, angle, coverage, alignment, geometric profile, feature critical size, feature surface area, feature exposed surface area, feature overlap, feature distance, feature orientation and / or feature alignment.

[0060] The memory circuitry system 120 can be implemented by any number of any type of conventional or other memory or storage devices, and can be volatile (e.g., RAM, cache, flash memory, etc.) or non-volatile (e.g., ROM, hard disk, optical storage device, etc.), and includes any suitable storage capacity. Data in the system (e.g., images, spectra, model parameters, etc.) is stored in the storage area of ​​the memory circuitry system 120.

[0061] The processor circuitry 130 may be, for example, one or more data processing devices that execute instructions for process logic stored in memory, such as a microprocessor, microcontroller, system-on-a-chip (SoC), or other fixed or programmable logic. The processor itself may be multiprocessor and may have multiple CPUs, multiple cores, multiple dies including multiple processors, etc. Additionally, the processor circuitry 130 may include a dedicated machine learning circuitry. Display rendering hardware may be part of the processor circuitry or may be, for example, a separate graphics processing unit (GPU).

[0062] User interface circuitry system 140 may include any suitable display, screen, or monitor capable of displaying information to a user, such as a tablet computer screen or a monitor attached to a computer workstation. User interface circuitry system 140 may include input devices, which may include any suitable input devices, such as a keyboard, mouse, touchpad, touch input tablet computer, touchscreen, camera, microphone, remote control, speech synthesizer, etc. Multiple output devices may also be included, such as speakers, headphones, audio output ports, etc.

[0063] As those skilled in the art will understand, aspects of the present invention can be implemented as systems, methods, or computer program products. Therefore, aspects of the present invention can take the form of entirely hardware embodiments, entirely software embodiments (including firmware, resident software, pseudocode, etc.), or combined software and hardware embodiments generally referred to herein as “circuit,” “module,” or “system.” Furthermore, aspects of the present invention can also take the form of computer program products embodied in one or more computer-readable media having computer-readable program code implemented thereon.

[0064] Any combination of computer-readable media may be used. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable medium can be, for example, but not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or apparatuses, or any suitable combination of the foregoing. More specific examples (not an exhaustive list) of computer-readable storage media will include: electrical connections having one or more wires, portable computer disks, hard disks, solid-state drives, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, phase-change memory devices, or any suitable combination of the foregoing. In the context of this document, a computer-readable storage medium can be any tangible medium that can contain or store a program for use by or in conjunction with an instruction execution system, device, or apparatus.

[0065] Computer-readable signal media may include propagated data signals having computer-readable program code implemented therein (e.g., in baseband or as part of a carrier wave). Such propagated signals may take any of a variety of forms, including, but not limited to, electromagnetic, optical, or any suitable combination thereof. Computer-readable signal media may be any computer-readable medium that is not a computer-readable storage medium and can communicate, propagate, or deliver a program (for use by or connection to an instruction execution system, device, or apparatus). Program code embodied on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, wired, fiber optic cable, radio frequency (RF), etc., or any suitable combination thereof.

[0066] Computer program code used to perform operations relating to various aspects of the present invention can be written in any combination of one or more programming languages, such as object-oriented programming languages ​​(e.g., Java, Python, Smalltalk, C++, etc.) or conventional programming languages ​​(e.g., the "C" programming language or similar programming languages). The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter case, the remote computer can be connected to the user's 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).

[0067] The various functions of System 100 can be distributed in any number of software modules or units, processing or computer systems and / or circuit systems in any way, wherein the computers or processing systems can be located locally to each other or remotely to each other and communicate via any suitable communication medium (e.g., LAN, WAN, intranet, Internet, hardwired, modem connection, wireless, etc.).

[0068] Various aspects of the present invention are described 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 in the flowchart illustrations and / or block diagrams, and any combination of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions and executed by processor circuitry system 130 to produce a machine, such that the instructions create means for implementing the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams.

[0069] The flowcharts and block diagrams in the accompanying drawings 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 code, including (multiple) executable instructions for implementing a specific logical function. It should also be noted that in some alternative implementations, the functions shown in the blocks may not occur in the order shown in the figures. For example, depending on the functions involved, two blocks shown successively may be executed substantially simultaneously, or sometimes in reverse order. It will also be noted that each block of the block diagram and / or flowchart, and combinations of blocks in the block diagram and / or flowchart, may be implemented by a system based on dedicated hardware or a combination of dedicated hardware and computer instructions that performs a specific function or action.

[0070] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used herein, the singular forms “a,” “an,” and “the” are intended to include the plural forms as well, unless the context explicitly indicates that the plural forms are not included. It will be further understood that when the terms “comprises” and / or “comprising” are used in this specification, they specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more features, integers, steps, operations, elements, components, and / or groups thereof.

[0071] Any and all means or steps plus functional elements in the following claims, corresponding structures, materials, operations, and equivalent substitutions, are intended to include any structure, material, or operation used to perform the function in combination with other elements specifically pointed out in the claims.

[0072] The above description is intended to illustrate possible embodiments of the present invention and is not intended to be limiting. Many variations, modifications, and alternatives will become apparent to those skilled in the art upon review of this disclosure. For example, components equivalent to those shown and described may be replaced, individually described elements and methods may be combined, and elements described as discrete may be distributed across many components. Therefore, the scope of the invention should not be determined by reference to the above description but rather by reference to the appended claims and their equivalents.

Claims

1. An apparatus comprising: a data acquisition instrument that observes a state of a semiconductor manufacturing process and generates a data set accordingly, wherein the data acquisition instrument is a spectrometer that generates electromagnetic spectrum data as the data set; a memory circuitry that stores: the data set that is indicative of the state of the semiconductor manufacturing process or a semiconductor structure manufactured thereby, features in the data set being distinguishable within limits constrained by a data resolution; a machine learning model that comprises parameters having assigned respective values constrained by a model training process; and a training data set that contains features that are distinguishable within limits constrained by a training resolution that is greater than the data resolution; a processor circuitry that is communicatively coupled to the memory circuitry and that: generates an output data set from the data set according to the machine learning model such that features in the output data set are distinguishable within limits constrained by an output data resolution that is finer than the data resolution of the data set, wherein the spectrometer comprises an entrance slit and a detector array, the training data set comprises spectrum data acquired while varying a distance between the entrance slit and the detector array, and wherein the machine learning model comprises functional blocks to which the processor circuitry applies the data set sequentially, each of the functional blocks increasing a dimensionality of the data set compared to a preceding functional block, wherein the machine learning model is a pyramid residual convolutional neural network, wherein each functional block comprises a plurality of convolutional layers that convolve corresponding ones of the parameters with feature maps based on the data set, wherein each functional block comprises a plurality of activation layers that limit a passage of convolved feature maps through to a succeeding layer in the functional block according to an activation function, wherein the activation function is a rectified linear unit activation function.

2. The apparatus of claim 1, wherein, the processor circuitry implements a process controller by which the semiconductor manufacturing process is controlled based on the output data set.

3. The apparatus of claim 1, wherein: the processor circuitry distinguishes the output data set from the training data set in response to initiation of a training process and modifies the parameters based on the distinction to minimize or maximize an objective function.

4. The apparatus of claim 3, wherein, the processor circuitry implements a data generator that generates the training data set according to optical properties of the data acquisition instrument.

5. The apparatus of claim 3, wherein, the training data set comprises spectrum data acquired from the spectrometer.

6. A system for processing materials for semiconductor manufacturing, the system comprising: a chamber that contains the materials, the chamber comprising a port through which the semiconductor manufacturing process is monitored; a data acquisition instrument that is communicatively coupled to the port of the chamber and that generates a data set indicative of a state of the semiconductor manufacturing process, features in the data set being distinguishable within limits constrained by a data resolution of the data acquisition instrument, wherein the data acquisition instrument is a spectrometer that generates electromagnetic spectrum data as the data set; and a memory circuitry that stores: the data set that is indicative of the state of the semiconductor manufacturing process or a semiconductor structure manufactured thereby, features in the data set being distinguishable within limits constrained by a data resolution; a machine learning model that comprises parameters having assigned respective values constrained by a model training process; and a training data set that contains features that are distinguishable within limits constrained by a training resolution that is greater than the data resolution; a processor circuitry that is communicatively coupled to the memory circuitry and that: generates an output data set from the data set according to the machine learning model such that features in the output data set are distinguishable within limits constrained by an output data resolution that is finer than the data resolution of the data set, wherein the spectrometer comprises an entrance slit and a detector array, the training data set comprises spectrum data acquired while varying a distance between the entrance slit and the detector array, and wherein the machine learning model comprises functional blocks to which the processor circuitry applies the data set sequentially, each of the functional blocks increasing a dimensionality of the data set compared to a preceding functional block, wherein the machine learning model is a pyramid residual convolutional neural network, wherein each functional block comprises a plurality of convolutional layers that convolve corresponding ones of the parameters with feature maps based on the data set, wherein each functional block comprises a plurality of activation layers that limit a passage of convolved feature maps through to a succeeding layer in the functional block according to an activation function, wherein the activation function is a rectified linear unit activation function. the processor circuitry implements a process controller by which the semiconductor manufacturing process is controlled based on the output data set. processor circuitry communicatively coupled to the data acquisition instrument and generating an output dataset from the dataset in accordance with a machine learning model such that features in the output dataset can be resolved within limits constrained by an output data resolution that is finer than a data resolution of the dataset, the machine learning model including parameters having assigned respective values constrained by a model training process, wherein the spectrometer includes an entrance slit and a detector array, the training dataset includes spectral data acquired while varying a distance between the entrance slit and the detector array, and wherein the machine learning model includes functional blocks to which the processor circuitry applies the dataset sequentially, each of the functional blocks increasing a dimensionality of the dataset compared to a preceding functional block, wherein the machine learning model is a pyramid residual convolutional neural network, wherein each functional block includes a plurality of convolutional layers that convolve corresponding ones of the parameters with feature maps based on the dataset, wherein each functional block includes a plurality of activation layers that limit passage of convolved feature maps through to a succeeding layer in the functional block in accordance with an activation function, wherein the activation function is a rectified linear unit activation function.

7. The system of claim 6, wherein, the spectrometer is optically coupled to a port of the chamber and generates spectra as the dataset, another dataset being high resolution spectra having a finer resolution than the spectra generated by the spectrometer, a feature of a physical design of the spectrometer being that bandwidth and optical throughput are achieved without diminution by virtue of the manner in which the spectra are processed by the machine learning model.

8. The system of claim 6, wherein, the processor circuitry implements a process controller by which the semiconductor manufacturing process is controlled based on the output dataset.

9. The system of claim 6, wherein, the process controller analyzes the high resolution spectra to determine a state of the semiconductor manufacturing process and generates control signals that modify the semiconductor manufacturing process to force occurrence of a different state thereof. the process controller analyzes the high resolution spectra to determine a state of the semiconductor manufacturing process and generates control signals that modify the semiconductor manufacturing process to force occurrence of a different state thereof.

Images