Enhanced Resolution in Semiconductor Manufacturing Data Acquisition Instruments Using Machine Learning
Machine learning techniques address the trade-off in spectrometer design by generating high-resolution spectral data from low-resolution inputs, improving resolution without sacrificing throughput or spectral range, benefiting semiconductor manufacturing processes.
Patent Information
- Application Number
- JP2021555451
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2019-03-15
- Filing Date
- 2020-03-13
- Publication Date
- 2024-12-05
- Estimated Expiration
- 2040-03-13
AI Technical Summary
Conventional spectrometers face a trade-off between spectral resolution and throughput, with higher resolution typically sacrificing throughput and spectral range, which is challenging to overcome with physical design alone.
Utilizing machine learning techniques, specifically artificial neural networks, to process low-resolution spectral data and generate high-resolution data sets, balancing spectral resolution, throughput, and spectral range without physical design compromises.
Enables the generation of high-resolution spectral data with lower-cost hardware and faster cycle times, maintaining throughput and spectral coverage, enhancing the ability to resolve finer details in semiconductor manufacturing processes.
Smart Images

Figure 0007594717000001 
Figure 0007594717000002 
Figure 0007594717000003
Abstract
Description
[Technical field]
[0001] CROSS-REFERENCE TO RELATED APPLICATIONS This application claims the benefit of priority under 35 U.S.C. §119(e) from U.S. Provisional Patent Application No. 62 / 819,115, entitled “Deep-learning Enhanced Spectrometer Resolution Techniques,” filed March 15, 2019, the entire disclosure of which is incorporated herein by reference.
[0002] The present disclosure relates to techniques for imaging and spectral resolution improvement for images and spectral data used to monitor semiconductor processing or features formed on semiconductor substrates. In particular, the present disclosure provides novel techniques for applying low resolution image or spectral data to a trained artificial neural network to generate high resolution image or spectral data. [Background technology]
[0003] Broadband compact spectrometers are widely used in semiconductor plasma etch processes to collect optical 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 from which control signals are generated. The SNR is usually highly proportional to the throughput of the spectrometer. Typically, spectrometer performance is quantified by a few critical parameters including the spectral bandwidth, i.e., the range of wavelengths that can be detected; the spectral resolution, i.e., the finest spectral features that can be resolved; and the optical throughput, i.e., the fraction of light collected by the detector.
[0004] Higher resolution spectrometers will typically sacrifice throughput and spectral (wavelength) range to obtain high resolution results. Similarly, spectrometers designed for higher throughput and wider spectral range will typically have lower resolution. The ability to resolve finer details in the spectral data while maintaining the throughput and spectral range of lower resolution spectrometers is not currently available in most high resolution spectrometers, especially when the cost of the instrument is a limiting design factor. Many spectrometer designs available today involve trade-offs between design-critical parameters. A simple and effective way to improve the spectral resolution is to use a narrower slit as the spectrometer entrance, which effectively narrows the "image" of each wavelength of the diffraction spectrum on the charge-coupled device (CCD) detector, but at the expense of reduced light passing through the slit and the corresponding information. The trade-off between resolution and throughput is very difficult to overcome by physical spectrometer design alone. Summary of the Invention [Problem to be solved by the invention]
[0005] One objective of the present disclosure is to mitigate the above-mentioned trade-offs with physical design parameters associated with conventional spectrometers.
[0006] Another object of the present disclosure is to enable the use of a lower resolution spectrometer to collect spectral data and to be able to resolve finer details in the spectral data or images without sacrificing throughput or spectral range. [Means for solving the problem]
[0007] In one aspect of the inventive concept, an apparatus is provided that includes a memory circuit that stores a dataset indicative of a state of a semiconductor manufacturing process or a semiconductor structure produced thereby. Features in the dataset are identifiable to an extent limited by a data resolution. The memory circuit also stores a machine learning model that includes parameters having respective values assigned to be constrained by a model training process. A processor circuit is communicatively coupled to the memory circuit and generates an output dataset from the dataset according to the machine learning model such that features in the output dataset are identifiable to an extent limited by an output data resolution that is finer than the data resolution of the dataset. [Brief description of the drawings]
[0008] [Figure 1] 1 is a schematic block diagram of an example system in which the concepts of the present invention may be embodied; [Diagram 2] FIG. 1 is a schematic diagram of an artificial neural network (ANN) in which the present invention may be embodied; [Diagram 3] 1 is a schematic diagram of an example spectrometer in which the present invention may be embodied; [Figure 4] FIG. 4 is a schematic diagram of a spectrometer entrance slit and its image 420 for simulating data in an embodiment of the present invention. [Diagram 5] 1 is a graph of low resolution spectral data for underfill and overfill implementations of a low resolution spectrometer used in embodiments of the present invention. [Figure 6] 1 is a graph of high resolution spectral data for underfill and overfill implementations of a high resolution spectrometer used in embodiments of the present invention. [Figure 7] FIG. 1 is a flow diagram illustrating an exemplary training process in which the present invention may be implemented. [Figure 8] FIG. 1 is a flow diagram illustrating an exemplary production process in which the present invention may be embodied. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0009] The inventive concept is best explained through certain embodiments of the invention, which are described in detail herein with reference to the accompanying drawings, in which like reference numerals refer to like features throughout. It should be understood that the term invention, as used herein, is intended to encompass the inventive concept underlying the embodiments described below, and not merely the embodiments themselves. It should be further understood that the general inventive concept is not limited to the exemplary embodiments described below, and that the following description should be read in this light.
[0010] The word exemplary is used herein to mean "serving as an example, instance, or illustration." Any embodiment of a structure, process, design, technique, or the like, referred to herein as exemplary is not necessarily to be construed as preferred or advantageous over other such embodiments. No particular qualities or suitability of embodiments shown herein as exemplary is intended or should be inferred.
[0011] Additionally, this specification includes mathematical expressions, and the principles communicated should be understood to be fully explained using mathematical expressions. Where mathematics is used, it should be understood that it is for the purpose of concise explanation of the underlying principles being explained, and no other purpose should be suggested or inferred unless otherwise expressed. It will be apparent from the entirety of this disclosure how the mathematics herein relates to the invention, and that those skilled in the art will recognize numerous techniques for implementing the physical manifestation of the principles expressed mathematically, where embodiments of the principles underlying the mathematical expressions are intended.
[0012] This disclosure relates to "constructing" or generating high resolution (HR) data sets 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 at the extended wavelength range and high light throughput of LR spectrometers. Furthermore, variations of similar techniques can be used to generate higher resolution image data from lower resolution images.
[0013] The invention described herein combines deep learning techniques with relatively low resolution spectral data collected, for example, by conventional optical spectrometer designs, to obtain higher resolution data using predictive models built based on analyzing pairs of low and high resolution data sets (e.g., empirical or simulated) related to semiconductor processes or features on semiconductor substrates. In this manner, high resolution data of semiconductor process or substrate features may be obtained with lower cost hardware and faster cycle times than with existing relatively high resolution spectrometer or imaging systems and techniques. For example, in a spectral embodiment, the critical criteria that define the quality of the spectral data set include bandwidth (i.e., 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, the aforementioned factors are balanced against each other by tradeoffs in the physical design of the spectrometer to achieve predictable and meaningful results (e.g., patterns, features, or process conditions) for semiconductor manufacturing in the shortest possible time. For example, obtaining a higher resolution spectrometer data set is achieved at the expense of higher throughput and wavelength range. It would therefore be advantageous to obtain higher resolution data without sacrificing them. In addition to spectral embodiments, deep learning techniques may be applied to other low resolution data (e.g., images) to decipher higher resolution information or characteristics of the semiconductor process or substrate features being analyzed.
[0014] One approach to achieving this advantageous capability is to build machine learning (ML) models capable of generating high-resolution data sets from lower resolution data sets. Ideally, the model-generated high-resolution data may provide information that is not readily discernible from the low-resolution data unless such low-resolution data is processed by the ML model. The ML model may be prepared by analyzing low-resolution and high-resolution training data collected experimentally or by simulation to determine statistical transformations or relationships that may be used to train an artificial neural network (ANN), such as a convolutional neural network (CNN). For example, multiple layers of CNNs may be trained to statistically relate low-resolution data to high-resolution data for a process or feature. Low-resolution and high-resolution data may be collected for similar or slightly different processes or features to improve the robustness of the statistical relationships between the data sets. In this way, ML models may be developed to accurately predict high-resolution results and optimized by comparing the modeled results to the actual high-resolution data. After validating the accuracy and precision of the model, the trained network may be applied to the low-resolution data to generate corresponding high-resolution data or results for a semiconductor process or feature. Two spectrometers may be used to generate corresponding low and high resolution data sets representative of process or substrate characteristics. However, in other embodiments, the low and high resolution data sets may be generated by simulation of measured spectra to approximate actual process conditions. In this manner, the simulated spectral data may be optimized to simulate certain characteristics of the spectra based on known or observed phenomena, such that the training data sets may include fully simulated low and high data sets, combinations of simulated and experimental training pairs, or training pairs including simulated and experimental data.
[0015] FIG. 1 is a schematic block diagram of an example system 100 in which the inventive concepts may be embodied. For purposes of explanation and not limitation, system 100 may be constructed or configured for semiconductor manufacturing, with much of the manufacturing processing occurring in one or more sealed chambers, representatively shown as chamber 5. Chamber 5 may include one or more ports, representatively shown as port 7, by which processes occurring inside chamber 5 may be monitored, including by spectroscopy, 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 that are of process-dependent configuration. Components within chamber 5 may be evaluated, for example, by spectroscopy. Other processes may be monitored or examined using optical or electronic imaging. As will be recognized by those skilled in the art upon review of this disclosure, embodiments of the inventive concepts implement techniques applicable to spectroscopic optically acquired images and other acquired images.
[0016] The system 100 may include a processor 150 including a memory circuit 120 in which data and processor instruction code are stored, a processor circuit 130 in which data and control processes are executed, and a user interface circuit 140 in which a human user interfaces with the system 100, particularly for purposes of data analysis and control. As shown in FIG. 1, the memory circuit 120 may be constructed or configured to store data sets in a data store location 122 or the like, machine learning models in a model store location 124 or the like, and processor instruction code in a code store location 126 or the like. The processor instruction code may include ML engine code in an ML engine code store location 127 or the like (when the ML engine is not hardware implemented in a dedicated circuit), and process control code in a process controller code store location 129 or the like. The processor circuit 130 may be constructed or configured to execute the processor instruction code stored in the memory circuit 120, as shown in the ML engine 132 and the process controller 134.
[0017] System 100 may further include data acquisition mechanisms, shown in data acquisition devices 105l and 105h, respectively. As indicated above, the present invention may be embodied to increase the resolution in a variety of different data types, with devices such as device 105l representing a low-resolution data acquisition device generating low-resolution data 110l and device 105h representing a high-resolution data acquisition device generating high-resolution data 110h representing any number of devices that may be used to collect application-specific data, such as, for example, image data, spectral data, etc. Although the following description is primarily directed to spectral data sets, it should be understood that other data may be used in embodiments of the present invention.
[0018] Experimental or simulated data may be collected or generated under various conditions or variables to establish or define a predictable relationship between the low-resolution data set 110l and the high-resolution data set 110h. The relationship may be based on commonalities between process conditions, feature types, or detection techniques. However, the relationship may be assigned based on differences or dissimilarities between the data sets 110l and 110h, such as device settings / configurations (slit size, field of view, aperture size). In fact, the data sets 110l and 110h may be collected or generated along various device setting / configuration options to obtain a resolution distribution for the observed process conditions or features. The training relationship may group one or more low-resolution data sets 110l into one or more high-resolution data sets 110h associated with a process condition characteristic. The process condition characteristic may include, but is not limited to, chemical composition, temperature, density, and / or energy or signal associated with the process. In addition, the relationship may extend to the process conditions (e.g., set points) implemented in chamber 5, which may include, but are not limited to, temperature, pressure, gas flow rate, applied power, applied voltage, rotation speed, conversion speed, liquid dispense rate, liquid dispense amount. Furthermore, the process conditions may include a range of results configured as a distribution or profile for each characteristic. In addition to the process conditions, the spectral data may represent one or more features on the substrate. In this case, the feature characteristics include, but are not limited to, thickness, width, height, orientation, roughness, flatness, angle, coverage, alignment, and / or geometric profile of the features. In addition, the feature characteristics may refer to area, distance between features, or orientation of features relative to each other (e.g., overlap of features, exposed area, alignment). Any of the above applications in which experimental or simulated LR and HR data sets are generated may be used to train an artificial neural network using the techniques described herein.
[0019] The system 100 may be operated in two operating modes: a training mode in which a machine learning model is configured to increase the data resolution, and a production mode in which the data resolution is increased based on a pre-configured (trained) model. In the production mode, an LR data set 110l (image, spectrum) is acquired through an LR data acquisition device 105l (imager, spectrometer) and provided to a processor 150. The processor 150 may communicate the LR data set 110l to an ML engine 132. The ML engine 132 may process the data according to a pre-trained ML model as described below to increase the resolution of the data. The increased resolution data set may be presented on a user interface 140, where it may be analyzed by a generator. In one embodiment, the increased resolution data set is provided to a process controller 134, which analyzes the data to determine whether a particular target process state has been reached. For example, analysis of the increased resolution data set may reveal certain gas components in chamber 5 that indicate a process endpoint. At such an end point, the process controller 134 may command or direct a process transition or interruption of any process occurring within the chamber 5 .
[0020] 2 is a schematic diagram of an artificial neural network (ANN) 200 in which the present invention may be embodied. The present invention is not limited to a particular ANN implementation, so long as it achieves the goal of increasing data resolution in the output data. In the illustrated embodiment, the ANN 200 includes an input convolutional layer 210 and a number of residual blocks B1-B2, shown at 220 and 230, respectively. K and an output convolutional layer 240. In the ANN 200, 1) the size of the internal feature map or the number of feature map channels gradually increases from the input to the output of the ANN, and 2) an identity mapping that addresses the vanishing / exploding gradient problem is achieved in a shortest path.
[0021] In the example embodiment of FIG. 2, LR data sets 205a-205m, each referred to herein as an LR data set 205, may be provided to an input convolutional layer 210. The input convolutional layer 210 maps each data set 205 to a corresponding feature map 215a-215m, each referred to herein as a feature map 215, among other possible operations. The feature maps 215 may be provided to a residual block 220. The residual block 220 may include a first convolutional layer 260a, a first activation layer 265a, a second convolutional layer 260b, and a second activation layer 265b, similarly for all K residual blocks in the ANN 200. Additionally, each feature map 215 may be provided to a summation node 225a-225n, each referred to herein as a summation node 225, in a respective shortcut path 217a-217n, each referred to herein as a shortcut path 217. As a result, the output of each summation node is X k+1 =X k +ψ{ψ(X k *W k (1) )*W k (2)} where X k is the input to the kth residual block, and X k+1 is its output. ψ is the activation function (e.g., rectified linear unit (ReLU)). * is the convolution operator, and W k (1) and W k (2) is the trainable convolution kernel of the kth block. It should be noted that the bias term has been omitted from the above equation. The feature map generated by each convolution layer is Gamma k,j =Σ i φ k,i *w k,i,j +β k,j Ω where γ k,j is the output feature map of the jth convolutional layer of the kth residual block. k,iis the input feature map for the kth residual block to the ith convolutional layer. k,i,j is W k (j) is the i-th kernel of β k,j is a trainable bias term, and Ω is a matrix of feature map sizes with all entries set to unity.
[0022] The feature map 215 is processed by the residual block 220 and transmitted to the next residual block, i.e., residual block 230, where it is similarly processed. However, each residual block may generate a feature map (and / or associated feature map channel) of greater dimension than that generated by the previous residual block. For example, the input to the residual block 220 may include N channels 227a-227n, and the output of the residual block 220 may include P channels 232a-232p, where P>N. This dimensionality mismatch may be overcome, for example, by zero-padding shortcut connections that bypass the residual block 230, such as shortcut connections 227a-227n.
[0023] As the dataset progresses through the ANN 200 and its dimensionality gradually increases, finer features may be identified for each residual block, affecting an increase in resolution beyond the original dataset 205. The processed feature maps proceed to the output convolutional layer 240 where, among other optical processing operations, they are converted into output datasets 245a-245m, each referred to herein as an output dataset 245. When the ANN 200 is being trained, the output dataset 245 is provided to mechanisms external to the ANN for data analysis or process control, etc. On the other hand, when the ANN 200 is being trained, the output dataset 245 is compared to high resolution training datasets 250a-250m, each referred to herein as an HR training dataset 250, which have been selected to correspond to data contained in the original LR dataset 205. That is, each original LR dataset 205 is matched to the HR training dataset 250 such that the same features are present in both, even though at different resolutions. During training, the original LR data set 205 is transformed into an HR data set according to a randomly initialized function. The transformed HR data set is then suitably compared, such as by difference, to the corresponding HR training data set to determine an error measure. Using such an error measure, the W k (j) and β k,j The ANN parameters, such as, are modified to minimize (or maximize) an objective function, such as a training loss function. The modified parameters may be backpropagated through the ANN 300, each shown in a backpropagation channel 257. The process may be repeated with the new parameters to generate another HR output data set. The process iterates until the training goal is reached, such as when the trainable network parameters converge or when no improvement in resolution is achieved.
[0024] The LR dataset 205 may include LR spectrometer data, and the HR dataset 250 may include HR spectrometer data. In addition to the spectral data itself, the HR and LR datasets may be labeled with information related to observed processes or features related to the spectral data. Such labeling may be used to isolate specific aspects of the HR / LR datasets or to overlap certain aspects of the HR / LR datasets, depending on the type of model optimization that requires generating predictable HR results based on random but similar LR dataset inputs. The HR / LR datasets may be collated into training pairs, which may include one or more combinations of HR / LR datasets, and are used to train the ANN 200 based on identifying relationships between the HR / LR training pairs. At a high level, training is an iterative process in which HR data generated according to an ML model is compared to the HR dataset with the aim of optimizing the statistical relationships between the datasets using ML techniques.
[0025] 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 an LR image and the HR dataset 250 is represented by an HR image. The two images may represent input data collected by the low resolution data acquisition device 105l and the high resolution data acquisition device 105h as described above. In addition to the image data, the HR and LR datasets may be labeled with information related to the observed process or features related to the image data, so that the label information may be used to highlight differences or similarities between the datasets. The label information may be used to isolate certain aspects of the HR / LR datasets or to overlap certain aspects of the HR / LR datasets, depending on the type of model optimization that requires generating predictable HR results based on inputs of random but similar LR datasets. The HR / LR datasets are collated into training pairs, which may include one or more combinations of the HR / LR datasets, and are used to train the ANN 200 using ML techniques based on identifying relationships between the HR / LR training pairs, thereby enabling the ML model to predict or generate the HR dataset 245 using the LR dataset 205.
[0026] It should be understood that the training example provided above is just one of many possibilities for training the ANN of the present invention. In one embodiment, the data set generated by the data acquisition device or by the data generator can be partitioned into, for example, a training set used to train the learning technique, a validation set used to evaluate the model architecture, and a test set used to determine whether the model generalizes properly beyond the data set used for training. Those skilled in the art of machine learning will recognize how to use such partitioned data sets without the specific details described herein, as well as other machine learning techniques to practice the present invention without departing from its spirit and intended scope.
[0027] Returning again to Figure 1, the system 100 may include a high-resolution data generator 115 that generates simulated data sets for model training purposes. In addition to image and spectral data, simulated spectral model data may be used to train the ANN 200. However, training pairs may be derived from low-resolution and high-resolution simulations that are estimates of LR and HR spectrometer data from a spectrometer similar to the design shown in Figure 3 below.
[0028] FIG. 3 is a schematic diagram of an example spectrometer 300 in which the present invention may be embodied. Spectrometer 300 may follow a Czerny-Turner design and includes an entrance slit 310, a collimating mirror 320, a diffraction grating 330, a folding mirror 340, and an array detector 350. Light, including that produced in the electrochemical process occurring in chamber 5 of FIG. 1, enters spectrometer 300 through entrance slit 310 and reflects off collimating mirror 320 onto diffraction grating 330, where the impinging light is resolved into its spectral components. These components reflect off folding mirror 340 onto array detector 350, which generates an electrical signal indicative of the spectral content of the light under analysis. In essence, a spectrally resolved image of entrance slit 310 is generated on array detector 350, and such imaging may be simulated using various optical system design / analysis techniques implemented by HR data generator 115 of FIG. 1.
[0029] 4 is a schematic diagram of entrance slit 310 and its image 420 as produced by the optical system of spectrometer 300. In the illustrated embodiment, twelve image points 412a-412l on entrance slit 310, each referred to herein as image points 412, are selected for imaging. Through optical design / analysis techniques, image points 412 are brought onto an image 420 produced on array detector 350, each shown in image area 425 for each of several selected wavelengths. Note that for a fixed size slit, the clarity of image points 412 is affected by wavelength dependent diffraction.
[0030] Model data can be generated for LR and HR spectrometers simulating light from a calibration lamp spanning the spectrum from 300 nm to 900 nm with similar intensity in the spectrum. For demonstration purposes, two different spectrometer designs are modeled for the LR and HR implementations to form two sets of training pairs. The LR data set of the training pair was modeled with a 23 um slit and a 1024 CCD, and two additional LR data sets were generated by varying the grating placement relative to the slit. The underfilled embodiment represents a spectrometer design in which the grating is located closer to the slit than the overfilled embodiment. In this way, the LR training data set covers a wider range of LR spectrometer designs and / or capabilities and shows the difference in performance (e.g., SNR) between the two different LR designs. For example, FIG. 5 shows the intensity difference between the LR spectrometer designs, with the underfilled embodiment having a higher and sharper profile than the overfilled embodiment.
[0031] Overfilled and underfilled LR data sets can be paired with corresponding overfilled and underfilled HR data sets. The example HR data sets were generated using a 10um slit and a 2048x256 pixel CCD with varying grating distances between the underfilled and overfilled data sets. Figure 6 illustrates the differences between the HR spectrometer designs by highlighting the intensity differences. In contrast to the LR data set, the overfilled data set yielded a higher intensity signal compared to the underfilled data set.
[0032] Once the ANN 200 has been trained and validated, the model may be applied to an applicable low-resolution data set to output estimated high-resolution data corresponding to the low-resolution input. In contrast to the training method, the trained model receives low-resolution data without a high-resolution counterpart and applies the data to the model to generate a corresponding high-resolution output. The low-resolution data may include, but is not limited to, image or spectral data relating to a semiconductor processing process or features on a semiconductor substrate.
[0033] Within the semiconductor space, deep learning techniques can be applied to images ranging between macro and micro applications depending on the size of the features being observed. In macro applications, the observed features can range between the entire substrate surface area to relatively large surface area regions to observe and monitor process trends at a macro level. For example, the large surface area regions can include the substrate bevel or edge exclusion regions (<5 mm) that extend around the periphery of the substrate. Similarly, an area adjacent to the edge exclusion region that extends up to 20 mm toward the center of the substrate can be observed to monitor process performance close to the substrate edge. This area is known for process uniformity issues relative to the substrate center. Thus, comparing images of the substrate center and edge is beneficial to observe or detect process deviations or to adjust the process to account for in-situ or upcoming process non-uniformities. In some cases, the center-edge relationship can be used to train low to high models depending on the application. In macro applications, the image capture device can be a digital or analog-digital system that generates pixelated images for analysis. In one embodiment, the image capture device can be a CCD camera. In contrast, micro application embodiments are for observing relatively small features across a die or small areas within a die, as compared to macro applications. For example, micro application images may be at a scale to control and monitor processes with atomic layer precision. Atomic layer control of transistor gate structures (e.g., FinFET), transistor components (e.g., contacts, isolation), or interconnect line (BEOL) features (e.g., metal line pitch, vias) is essential. In one embodiment, micro application images may be generated by a scanning electron microscope (SEM) that scans the substrate surface / features with a focused beam of electrons, providing topological and compositional information of the substrate. Alternatively, images may be generated using optical emission spectroscopy to monitor semiconductor processes, including but not limited to plasma processing.
[0034] FIG. 7 is a flow diagram illustrating an exemplary training process 700 in which the present invention may be embodied. It is assumed that the process 700 accesses a training corpus of data including matched pairs of LR and HR data sets, for example, spectral data acquired from the LR and HR spectrometers described herein. The training corpus may be appropriately segmented or partitioned into training volumes 705a, validation volumes 705b, and test volumes 705c. Partitioning the corpus into training volumes, validation volumes, and test volumes may follow such techniques known in the machine learning field. In fact, the process 700 represents in flow schematic form one of many model building / training paradigms that may be used in embodiments of the present invention. In the following description, model paradigm refers to the one described above with reference to FIG. 2. It should be understood that the order of operations and data flows in FIG. 7 is for the sake of simplicity of explanation, and other operations and data flows may also be used to perform model building / training.
[0035] As shown in Fig. 7, the training volume 705a may include an original LR data set 707a, which may be acquired by a low-resolution data acquisition device. In operation 710, the LR data set 707a is processed according to an ANN to generate an HR output data set 715. In operation 720, the difference between the HR output data set 715 and the training data set 709a, acquired by a high-resolution data acquisition device or simulated by a high-resolution data generator, is 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 back-propagated through the ANN.
[0036] Once the training volume is exhausted, the process 700 may transition to operation 735, whereby it is determined whether the model is overfitting the data or generalizing to the given specifications. To this end, a validation volume 705b, including the original LR dataset 707b and the HR validation set 709b, may be applied to the trained model in a manner similar to that described above with the training volume 705a. This time, however, we focus on the architecture of the model, expressed through so-called hyperparameters, rather than its predictive efficiency. Thus, if it is found in operation 735 that the model does not generalize well, one or more of its hyperparameters may be modified by operation 740. For example, residual blocks of the model described above with reference to FIG. 2 may be removed, or the order of the operation blocks within each residual block may be revised, etc. Those skilled in the art will recognize many techniques that may be applied to correct overfitting in embodiments of the present invention.
[0037] If, in operation 735, it is determined that the model generalizes sufficiently, the process 700 may transition to operation 750, whereby it is determined whether the performance criteria are met by the trained model. To this end, a test volume 705c including the original LR dataset 707c and the HR test set 709c may be applied to the trained model in a manner similar to that described above with the training volume 705a and the validation volume 705b. This time, however, we focus on the performance of the model, measured by specific performance criteria of accuracy, i.e. how close the HR dataset predicted from the model is to the true HR dataset generated by the actual measurement of the observations, and sensitivity, i.e. the proportion of features in the HR prediction (through the model) that are so correctly identified. If such criteria are met, the process 700 may end. If not, the process 700 may transition to operation 740, whereby the ANN hyperparameters are again modified. The process 700 may then transition to operation 710, and continue from that point on.
[0038] FIG. 8 is a flow diagram illustrating an exemplary generation process 800 in which the present invention may be embodied. The generation process 800 may be executed when the system is in a generation mode. In operation 810, the LR data set 805 is used to predict a corresponding HR output data set 815, which is the output of the ANN described above. The process 800 may then transition to operation 820, whereby a process state is determined from the HR output data set 815. If the target process state has been achieved, as determined in operation 825, the generation process 800 ends. If the target process state has not been achieved, the process 800 may transition to operation 810 with newly acquired LR data and continue from that point.
[0039] 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 relating to a semiconductor process or a feature on a semiconductor substrate in a processing chamber in a tangible computer readable memory. The dataset may include a coarse (low resolution) dataset representative of the process or feature, and a detailed (high resolution) dataset representative of the process or feature, the detailed dataset being different from the coarse dataset. In one example, the difference between the coarse dataset and the detailed dataset includes a broader or narrower range of information for at least one characteristic of the process or feature. A computer processing device is used to assign training relationships between instances of the coarse dataset and the detailed dataset, and neural network techniques are used to generate a model that correlates the coarse dataset to the detailed dataset based at least in part on the training relationships and the difference between the coarse dataset and the detailed dataset. In one example, the detailed dataset includes higher resolution data for at least one characteristic compared to the coarse dataset. The characteristic of the process may be a chemical composition of the process, a temperature of the process, a density of the process, and / or an electrical energy associated with the process. The process characteristics may include chemical composition, temperature, density, and / or electrical energy distribution or profile. The feature characteristics may include measured physical characteristics such as thickness, width, height, orientation, roughness, flatness, angle, coverage, alignment, and / or geometric profile of the feature.
[0040] The data set may include an image, a spectral representation, or a modular simulation of a process or feature. In an embodiment, the data set is based at least in part on an image, a spectral representation, and / or a simulation of a feature or process. The neural network technique may be based at least in part on a convolutional neural network technique, and the data set may be generated by a spectrometer. In some embodiments, the coarse data set is generated by a coarse spectrometer having a coarse slit that allows light or energy to enter the coarse spectrometer, and the detailed data set is generated by a detailed spectrometer including a detailed slit that allows light or energy to enter the detailed spectrometer, where the detailed slit includes an opening or gap that is smaller than the corresponding opening or gap of the coarse slit. The data set may also be generated by an image capture device, such as a charge-coupled device camera. Alternatively, the image capture device may include a microscope and a camera.
[0041] The assignment of the training relationship is based at least in part on common elements that exist between the collection of the coarse data set and the detailed data set. The common elements are process conditions of the semiconductor process or feature types of the features on the semiconductor substrate. The process conditions may include temperature, pressure, chemical composition, gas flow rate, applied power, applied voltage, rotation speed, translation speed, liquid dispense rate, liquid dispense amount. The feature types may include patterns in the substrate or patterns on the substrate. In some examples, the feature types include thickness, width, height, length, orientation, roughness, flatness, angle, coverage, alignment, and / or geometric profile of the features. The feature types may also include critical dimensions of the features, surface area of the features, exposed surface area of the features, overlap between features, distance between features, orientation between features, and / or alignment between features. The assignment of the training relationship of the method may include pairing a single coarse data set and a single detailed data set collected during the process or on the features.
[0042] One embodiment includes a method in which a dataset relating to a semiconductor process or feature on a semiconductor substrate is meanwhile stored in a tangible computer readable memory. The dataset may include a coarse (low resolution) dataset representative of the process or feature. The coarse dataset is applied to a fine resolution model based at least in part on a difference between the low resolution dataset and the high resolution dataset for the process or feature. A fine (high resolution) dataset representative of the process or feature is generated using neural network techniques, the fine dataset differing from the coarse dataset, where the difference is based at least in part on a greater amount of information for 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. The fine dataset may include higher resolution data for at least one characteristic compared to the coarse dataset. The characteristics 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, temperature, density, and / or electrical energy, the thickness, width, height, orientation, roughness, flatness, angle, coverage, alignment, and / or the geometric profile of the feature.
[0043] The dataset may include an image or spectral representation of a process or feature, which may be based at least in part on an image or spectral representation of the feature or process. The datasets may be generated by spectrometers, and in an embodiment, the coarse dataset is generated by a coarse spectrometer that includes a coarse slit that allows light or energy to enter the coarse spectrometer, and the detailed dataset is generated by a detailed spectrometer that includes a detailed slit that allows 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.
[0044] The data set may also be generated by an image capture device, such as one including a charge-coupled device camera with or without a microscope.
[0045] The training relationship assignment may be based at least in part on common elements present in the collection of the coarse data set and the fine data set, where the common elements may be process conditions of the semiconductor process, feature types of the features on the semiconductor substrate, temperatures, pressures, chemical compositions, gas flow rates, applied powers, applied voltages, rotational speeds, conversion speeds, liquid dispense measures, and liquid dispense volumes, and further includes sending adjustments to a matching process chamber.
[0046] The feature type may be a pattern in a substrate, a pattern on a substrate, a thickness, a width, a height, a length, an orientation, a roughness, a flatness, an angle, a coverage, an alignment, a geometric profile of a feature, a critical dimension of a feature, a surface area of a feature, an exposed surface area of a feature, an overlap between features, a distance between features, an orientation between features, and / or an alignment between features.
[0047] The memory circuit 120 may be implemented with any amount of any type of conventional or other memory or storage device, may be volatile (e.g., RAM, cache, flash, etc.) or non-volatile (e.g., ROM, hard disk, optical storage, etc.), and may include any suitable storage capacity. Data within the system (e.g., images, spectra, model parameters, etc.) is stored in the memory circuit 120 storage area.
[0048] The processor circuitry 130 may be one or more data processing devices that execute instructions for process logic stored in memory, such as, for example, a microprocessor, microcontroller, system on chip (SOC), or other fixed or programmable logic. The processor may itself be multi-processor and may have multiple CPUs, multiple cores, multiple dies containing multiple processors, etc. In addition, the processor circuitry 130 may include dedicated machine learning circuitry. Display rendering hardware may be part of the processor circuitry or may be, for example, a separate graphics processor unit (GPU).
[0049] User interface circuitry 140 may include any suitable display, screen, or monitor capable of displaying information to a user, such as the screen of a tablet or a monitor attached to a computer workstation. User interface circuitry 140 may include input devices, which may include any suitable input device, such as a keyboard, a mouse, a track pad, a touch input tablet, a touch screen, a camera, a microphone, a remote control, a voice synthesizer, etc. Output devices may include, for example, speakers, headphones, an audio output port, etc.
[0050] As will be appreciated by one of ordinary skill in the art, aspects of the present invention may be embodied as a system, method, or computer program product. Accordingly, aspects of the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment (including firmware, resident software, microcode, etc.), or an embodiment combining software and hardware aspects, all of which may be generally referred to herein as a "circuit," "module," or "system." Additionally, aspects of the present invention may take the form of a computer program product embodied in one or more computer readable medium(s) having computer readable program code embodied thereon.
[0051] Any combination of one or more computer readable media may be utilized. The computer readable medium may be a computer readable signal medium or a computer readable storage medium. The computer readable medium may be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples (non-exhaustive list) of computer readable storage media include an electrical connection having one or more wires, a portable computer diskette, a hard disk, a solid state disk, a random access memory (RAM), a read only memory (ROM), an erasable programmable read only memory (EPROM or flash memory), an optical fiber, a portable compact disk read only memory (CD-ROM), an optical storage device, a magnetic storage device, a phase change memory storage device, or any suitable combination of the foregoing. In the context of this document, a computer readable storage medium may be any tangible medium that can contain or store a program for use by or in connection with an instruction execution system, apparatus, or device.
[0052] A computer-readable signal medium may include a propagated data signal having computer-readable program code embodied therein, for example in baseband or as part of a carrier wave. Such a propagated signal may take any of a variety of forms, including but not limited to electromagnetic, optical, or any suitable combination thereof. A computer-readable signal medium may be any computer-readable medium that can communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device, other than a computer-readable storage medium. The program code embodied on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, wired wiring, fiber optic cable, RF, etc., or any suitable combination of the foregoing.
[0053] Computer program code for carrying out operations for aspects of the present invention may be written in any combination of one or more programming languages, e.g., object-oriented programming languages such as Java, Python, Smalltalk, C++, or conventional procedural programming languages such as the "C" programming language or similar programming languages. The program code may run completely 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 completely on a remote computer or server. In the latter scenario, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN), or a wide area network (WAN), or the connection may be made to an external computer (e.g., through the Internet using an Internet Service Provider).
[0054] The various functions of system 100 may be distributed in any manner among any amount of software modules or units, processes or computer systems and / or circuits, where the computers or processing systems may be located locally or remotely from one another and may communicate via any suitable communications medium (e.g., LAN, WAN, intranet, Internet, hardware or modem connections, wireless, etc.).
[0055] 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 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 program instructions and executed by processor circuitry 130 to produce machines, such that the instructions produce means for performing the functions / acts specified in the blocks of the flowchart illustrations and / or block diagrams.
[0056] The flowcharts and block diagrams in the figures 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 the flowcharts or block diagrams may represent a module, segment, or portion of code, including one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions described in the blocks may occur out of the order described in the figures. For example, two blocks shown in succession may in fact be executed substantially simultaneously, or the blocks may be executed in the reverse order depending on the functionality involved. It should also be noted that each block of the block diagrams and / or flowchart diagrams, and combinations of blocks in the block diagrams and / or flowchart diagrams, may be implemented by a dedicated hardware-based system that performs the specified functions or operations, or a combination of dedicated hardware and computer instructions.
[0057] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. As used herein, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It is further to be understood that the terms "comprises" and / or "comprising," as used herein, indicate 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.
[0058] The corresponding structure, material, acts, and equivalents of any and all means-plus-function or step-plus-function elements in the following claims are intended to include any structure, material, or act for performing a function in combination with other claimed elements as specifically claimed.
[0059] The above description is intended to illustrate possible implementations of the inventive concept, but is not intended to be limiting. Many variations, modifications, and alternatives will be apparent to those skilled in the art upon review of this disclosure. For example, components equivalent to those shown and described may be substituted accordingly, elements and methods described separately may be combined, and elements described as separate may be distributed across multiple components. Thus, the scope of the invention should be determined not with reference to the above description, but with reference to the appended claims, along with their full scope of equivalents.
Claims
1. An apparatus comprising: a data acquisition device that observes conditions of a semiconductor manufacturing process and generates a data set accordingly, the data acquisition device being a spectrometer that generates electromagnetic spectrum data as the data set; 1. A memory circuit, comprising: a data set indicative of a state of the semiconductor manufacturing process or a semiconductor structure being manufactured, wherein features within the data set are identifiable to an extent limited by data resolution; a machine learning model that is a pyramidal residual convolutional neural network, the machine learning model including parameters having respective values that are assigned to be constrained by a model training process; and A training data set that includes features that are identifiable to an extent limited by a training resolution that is higher than the data resolution. A memory circuit for storing the a processor circuit communicatively coupled to the memory circuit, the processor circuit generating an output dataset from the dataset in accordance with the machine learning model, features in the output dataset being identifiable to an extent limited by an output data resolution that is finer than the data resolution of the dataset; and having the spectrometer includes an entrance slit and a grating; the training data set comprises low resolution training data and high resolution training data acquired from the spectrometer, and the grating is positioned closer to the entrance slit when the low resolution training data is acquired than when the high resolution training data is acquired.
2. The apparatus of claim 1 , wherein each functional block includes a plurality of convolutional layers that convolve a corresponding one of the parameters with a feature map based on the dataset.
3. 3. The apparatus of claim 2, wherein each function block includes a plurality of activation layers that restrict passage of a convolved feature map to a subsequent layer within the function block according to an activation function.
4. The apparatus of claim 3 , wherein the activation function is a rectified linear unit activation function.
5. The apparatus of claim 1 , wherein the processor circuit implements a process controller in which the semiconductor manufacturing process is controlled based on the output data set.
6. The apparatus of claim 1 , wherein the training data set comprises spectral data obtained from the spectrometer.
7. An apparatus comprising: a data acquisition device for observing a state of a semiconductor manufacturing process and generating a data set accordingly, the data acquisition device having a coarse spectrometer and a fine spectrometer for generating electromagnetic spectrum data as the data set; 1. A memory circuit, comprising: a data set indicative of a state of the semiconductor manufacturing process or a semiconductor structure being manufactured, wherein features within the data set are identifiable to an extent limited by data resolution; a machine learning model that is a pyramidal residual convolutional neural network, the machine learning model including parameters having respective values that are assigned to be constrained by a model training process; and a training data set including features that are identifiable to an extent limited by a training resolution that is higher than the data resolution, the training data set including the electromagnetic spectrum data acquired from the spectrometer. A memory circuit for storing the a processor circuit communicatively coupled to the memory circuit, the processor circuit generating an output dataset from the dataset in accordance with the machine learning model, features in the output dataset being identifiable to an extent limited by an output data resolution that is finer than the data resolution of the dataset; and having the coarse spectrometer includes a coarse slit, the fine spectrometer has a fine slit with an aperture smaller than a corresponding aperture of the coarse slit, and the training data set includes low resolution training data generated by the coarse spectrometer and high resolution training data generated by the fine spectrometer.
8. 1. A system for processing materials for semiconductor manufacturing, comprising: a chamber containing a material, the chamber including a port through which the process is monitored; a data acquisition device communicatively coupled to the port of the chamber and configured to generate a data set indicative of a state of the process, features in the data set being identifiable to an extent limited by a data resolution of the data acquisition device, the data acquisition device being a spectrometer that generates electromagnetic spectrum data as the data set; a processor circuit communicatively coupled to the data acquisition device, the processor circuit generating an output dataset from the dataset according to a machine learning model, features in the output dataset being identifiable to an extent limited by an output data resolution that is finer than the data resolution of the dataset, the machine learning model being a pyramidal residual convolutional neural network including parameters having respective values assigned to be constrained by a model training process; having the processor circuit has a training data set including features that are identifiable to an extent limited by a training resolution that is higher than the data resolution, the training data set including the electromagnetic spectrum data acquired from the spectrometer; the spectrometer includes an entrance slit and a grating; The system, wherein the training data set comprises low resolution training data and high resolution training data acquired from the spectrometer, and the grating is positioned closer to the entrance slit when the low resolution training data is acquired than when the high resolution training data is acquired.
9. 10. The system of claim 8, wherein the spectrometer is optically coupled to the port of the chamber, the other data set is a high resolution spectrum having a finer resolution than that produced by the spectrometer, and the spectrometer has a physical design characterized by a bandwidth and optical throughput that is not reduced through processing the spectrum by the machine learning model.
10. The system of claim 9 , wherein the spectrometer includes an entrance slit and a detector array that define the bandwidth and the optical throughput.
11. The system of claim 10 , wherein the training data set comprises the spectral data acquired with varying distances between the entrance slit and the detector array.
12. The system of claim 8 , wherein the processor circuit implements a process controller in which the process is controlled based on the output data set.
13. 13. The system of claim 12, wherein the process controller analyzes the output data set to determine the state of the process and generates control signals to modify the process to force a different state.
Citation Information
Patent Citations
Method and apparatus for thickness analysis of complex layered structures
JP2006504092A
System and method for calibration of optical signal in semiconductor processing system
JP2018091836A
Neural network based hermite interpolator for scatterometry parameter estimation
US20100017351A1
Non-transitory computer-readable medium and method for monitoring a semiconductor fabrication process
US20200090304A1
Method and apparatus for enhancing data resolution
US7149262B1
