Deep learning model in high-mix semiconductor manufacturing

Abstract

This disclosure relates to deep learning models in highly hybrid semiconductor manufacturing. Techniques for applying neural network deep learning models to manufacturing strategies for highly hybrid semiconductor processes such as deposition, chemical mechanical polishing (CMP), etching, photolithography, and electroplating are disclosed. Training modes and normal operating modes of the manufacturing strategies are described.

Description

[0001] Cross-referencing of related patent applications

[0002] This application claims priority to U.S. Provisional Patent Application No. 63 / 247,904, filed September 24, 2021, the entire contents of which are incorporated herein by reference. Technical Field

[0003] This disclosure relates in general to applying deep learning networks, such as recurrent neural network (RNN) deep learning models, to manufacturing strategies for highly hybrid semiconductor fabrication, such as deposition, chemical mechanical polishing (CMP), etching, photolithography, and electroplating. Background Technology

[0004] Semiconductor manufacturing typically utilizes linear process models to handle parameters such as deposition time. For example, a linear process model of layer thickness versus process time can be used to calculate process time. However, linear process models may have limitations in their application and use.

[0005] Over the past decade, primarily driven by the dramatic expansion of data and computing power availability, artificial intelligence (AI) and machine learning (ML) technologies have entered many different fields and disruptively changed our lives and problem-solving methods. For example, these technologies improve online search and shopping results, customize advertising, tailor news delivery, and even drive cars. More recently, AI / ML's ability to autonomously learn from data and quickly find patterns and correlations has made it applicable to metrology and inspection in the semiconductor manufacturing industry. However, adopting AI / ML technologies in semiconductor manufacturing remains challenging because building and deploying deep learning models in semiconductor manufacturing environments is not a simple, straightforward process, but involves complex steps, especially in highly hybrid manufacturing environments. Summary of the Invention

[0006] This disclosure describes techniques for applying recurrent neural network (RNN) deep learning models to manufacturing strategies for highly hybrid semiconductor fabrication, such as deposition, chemical mechanical polishing (CMP), etching, photolithography, and electroplating.

[0007] This disclosure describes a method for setting at least one process parameter for manufacturing a semiconductor device. The method includes: receiving contextual information about the at least one process parameter; inputting the contextual information into a machine learning network; receiving a predicted value of the at least one process parameter from the machine learning network; setting the at least one process parameter based on the predicted value to manufacture the semiconductor device; receiving a measurement result of the semiconductor device associated with the at least one process parameter; and feeding the measurement result back into the machine learning network.

[0008] This disclosure also describes a control system including at least one hardware processor. The control system further includes at least one memory storing instructions that, when executed by the at least one hardware processor, cause the at least one hardware processor to perform operations including: receiving context information about at least one process parameter; inputting the context information into a machine learning network; receiving a predicted value of at least one process parameter from the machine learning network; setting the at least one process parameter based on the predicted value to manufacture a semiconductor device; receiving a measurement result of the semiconductor device associated with the at least one process parameter; and feeding the measurement result back to the machine learning network.

[0009] This disclosure also describes a machine storage medium embodying instructions that, when executed by a machine, cause the machine to perform operations including: receiving context information about at least one process parameter; inputting the context information into a machine learning network; receiving a predicted value of at least one process parameter from the machine learning network; setting at least one process parameter based on the predicted value to manufacture a semiconductor device; receiving a measurement result of the semiconductor device associated with at least one process parameter; and feeding the measurement result back to the machine learning network. Attached Figure Description

[0010] The accompanying drawings illustrate only exemplary embodiments of this disclosure and should not be construed as limiting its scope.

[0011] Figure 1 Example portions of a manufacturing system according to some examples of this disclosure are shown.

[0012] Figure 2 The framework of a run-to-run strategy for RNN-based deep learning models is shown, based on some examples of this disclosure.

[0013] Figure 3 This is a flowchart of a method for training an RNN model based on some examples of this disclosure.

[0014] Figure 4 Examples of selected features of a deposition rate model according to some examples of this disclosure are shown.

[0015] Figure 5 Examples of one-hot encoding of string data according to some examples of this disclosure are shown.

[0016] Figure 6 A flowchart illustrating a method for manufacturing a semiconductor device in normal operating mode, according to some examples of this disclosure, is shown.

[0017] Figure 7 Simulation results of deposition controllers according to some examples of this disclosure are shown.

[0018] Figure 8 A block diagram is shown illustrating an example of a machine that can perform any one or more of the techniques (e.g., methods) described herein. Detailed Implementation

[0019] Recurrent Neural Network (RNN) deep learning models can be applied to manufacturing strategies in highly hybrid semiconductor fabrication processes such as deposition, chemical mechanical polishing (CMP), etching, photolithography, and electroplating. RNN models can be trained and then used during runtime within the manufacturing strategy. The training and normal operation modes of the manufacturing strategy are described.

[0020] First, the RNN model is configured and trained to compute process parameters during training. During training, relevant test data can be collected. The collected data may include diverse features (or model inputs). In some examples, the features may be independent of each other. Based on these features, the collected data can include sufficient variability for model extraction. The collected data can be preprocessed. For example, string data can be converted into numerical data. Additionally, the collected data can be scaled so that these features are evaluated by their scale rather than their absolute values. Furthermore, the model's hyperparameters can be tuned.

[0021] Secondly, in normal operation (runtime) mode, the trained RNN model can be deployed during the manufacturing process to calculate and set process parameters. Relevant data can be collected for each run. The data can be processed and filtered, and then fed back to the RNN model for updates. The updated RNN model can then predict process parameters in semiconductor manufacturing processes with improved accuracy. The RNN modeling techniques described herein can be used to calculate machine settings for different processes, such as deposition time, exposure dose, and coverage settings.

[0022] Figure 1 An example portion of a manufacturing system 100 is shown that can be used to perform one or more of the techniques shown and described herein. The manufacturing system 100 can be used to manufacture and inspect a substrate 102, such as a semiconductor device.

[0023] Manufacturing system 100 may include one or more tools 110 (also referred to as machines) to perform manufacturing processes on substrate 102. In some examples, tool 110 may include a chemical vapor deposition (CVD) machine for depositing one or more layers on substrate 102. In some examples, tool 110 may include polishing tools for performing CMP, an irradiation source for emitting a radiation beam (e.g., electromagnetic waves) to project onto substrate 102 for photolithography, electroplating tools for electroplating substrates, and / or other suitable tools for semiconductor manufacturing.

[0024] Manufacturing system 100 may include metrology instrument 112 (also referred to as inspection instrument) to measure various features or characteristics on substrate 102. These features and characteristics may include, for example, film thickness measurements, critical dimension (CD) measurements (x, y, and / or z dimensions) of features formed on the substrate, pitch of line space features formed on the substrate, overlay offset from one layer to another on the substrate, or many other measurements or characteristics known to those skilled in the art. The inspection instrument can be used to inspect the consistency of features formed on or otherwise present on the substrate. For example, the object of inspection may be the substrate on which an integrated circuit die is to be manufactured, the location of the die, the location of unwanted particulate matter or other unwanted or unplanned features, etc.

[0025] Tool 110 and measuring instrument 112 may be coupled to control system 120. Control system 120 may include a computer processing unit (CPU) 122, a graphics processing unit (GPU) 124, a field-programmable gate array (FPGA) 131 (or other suitable accelerators such as a data processing unit (DPU), RNN, artificial neural network (ANN), etc.), memory 126, a display 128, input devices 130, and a communication interface 132 (e.g., a high-performance network (HPC)). Control system 120 may also include front-end circuitry, such as transmit signal chains, receive signal chains, switching circuitry, digital and analog circuitry, etc. The transmit signal chain provides control signals to tool 110. The receive signal chain receives process parameter measurements from measuring instrument 112.

[0026] The front-end circuitry can be coupled to and controlled by one or more processor circuits (such as CPU 122, GPU 124, and FPGA 131). CPU 122 can be provided as one or more multi-core processors. GPU 124 and FPGA 131 can be used to accelerate data processing and the performance of the machine learning networks described herein. The techniques shown and described herein can be executed by CPU 122 working in conjunction with GPU 124 to achieve faster processing.

[0027] CPU 122 and GPU 124 may be coupled to memory 126, such as to execute instructions that cause control system 120 to perform one or more operations, such as manufacturing control, processing or storing data related to inspection, or otherwise perform the techniques shown and described herein. Control system 120 may be communicatively coupled to other parts of system 100, such as using wired or wireless communication interface 132.

[0028] For example, the performance of one or more technologies shown and described herein can be achieved onboard to the control system 120 or using other processing or storage facilities such as computing facilities 140 (e.g., general-purpose computing devices, such as servers, cloud processing systems, data warehouses, laptops, tablets, smartphones, desktop computers, etc.). For example, processing tasks that would be undesirably slow or beyond the capabilities of the control system 120 to perform onboard to the control system 120 can be performed remotely (e.g., on a separate system), such as in response to a request from the control system 120. Similarly, storage of inspection data or intermediate data can be achieved using remote facilities communicatively coupled to the control system 120. The control system 120 may also include a display 128 (e.g., for presenting configuration information or results) and input devices 130 (e.g., including one or more of a keyboard, trackball, function keys or soft keys, mouse interface, touchscreen, stylus, etc., for receiving operator commands, configuration information, or responses to queries).

[0029] The control system 120 can utilize an RNN model to predict and set the process parameters used by tool 110. The RNN model can be provided via computing facility 140 (e.g., a cloud processing system). The processing results can be measured by metrology instrument 112 to update the RNN model. The RNN model can be built using run-to-run (R2R) algorithms and software scripts. For example, the RNN model can be built offline in a computer programming script such as Python to predict process parameters (e.g., deposition rate) that will be used in the manufacturing process using historical production data. During runtime, a deposition control strategy (or other processing strategy) can invoke the RNN script to trigger model predictions, which are then used to calculate the process time settings for each batch that will produce the desired target layer thickness value. The actual layer thickness of this batch can then be measured using metrology instrument 112 and fed back to control system 120 to update the RNN model.

[0030] Figure 2 A framework 200 for an R2R strategy based on an RNN deep learning model is illustrated. Framework 200 may include a tool 110, a metrology instrument 112, a control system 120, and an RNN model 202. During runtime, tool 110 may provide contextual information to control system 120 regarding the process (such as deposition) to be performed by tool 110. Control system 120 may utilize RNN model 202 to generate process parameter predictions (e.g., deposition rate predictions), which are subsequently used to operate tool 110 to perform the process (e.g., deposition). Metering instrument 112 may measure the outcome of the process (e.g., film thickness), which may then be used to update RNN model 202.

[0031] As described in this paper, building and deploying deep learning models is not a simple, straightforward process, but rather involves complex steps, especially in highly hybrid manufacturing environments. Therefore, Subject Expertise (SME) can be employed to build reliable and effective deep learning models that can be used for R2R control as described in this paper. This is because, in order to build a working model, there should be relevance and meaning in the dataset that can be extracted, and there should be sufficient features (or inputs to the model), preferably independent of each other, to be incorporated into the model. Feature selection can include manufacturing environment, equipment hardware parameters, consumable usage, and upstream parameter data that can affect the output values ​​of the model (or labels). The manufacturing environment can include the tool ID used to process the batch, the layer the batch is currently in, and the final product ID (e.g., memory, CPU, etc.), as the same processing can vary depending on different combinations of these manufacturing environments. Equipment hardware parameters can include maintenance information (e.g., number of runs since the last maintenance), machine / tool ​​configuration such as current and voltage settings. Upstream parameter data can include data from the same batch after upstream processing, such as thickness before deposition, etching, or polishing steps.

[0032] Initially, an RNN model is built and trained, and then deployed for use during runtime. Figure 3 A flowchart of method 300 for training an RNN model is shown. At operation 302, historical data can be retrieved. In some examples, historical data may be provided by a device automation program. In some examples, historical data may be stored in a database and retrieved by an R2R system. The database may be updated based on runtime operations, thus data can be retrieved before runtime operations for updates.

[0033] However, simply feeding data into deep learning algorithms without proper feature selection generally will not yield valuable results. Large datasets alone do not guarantee successful ML models. Instead, as discussed in this paper, selecting data containing relevant features with sufficient variability leads to better results for model extraction.

[0034] At operation 304, a subset of historical data can be selected. The selected data can be based on features related to process parameters associated with the RNN model. Subject matter expertise (SME) can be used to reduce large datasets to a select set of data to build reliable and effective deep learning models for R2R control. To build a reliable and accurate model, relevant and meaningful features can be extracted from the dataset. A minimum number of features (e.g., at least two) are extracted and incorporated into the model as input. These features can be independent of each other. For example, routes and products can be considered non-independent. A route is a list of steps performed on a batch of products, and routes are typically dedicated to a single product line. Therefore, selecting routes and products as features provides recurring contextual information to the model.

[0035] Feature selection can include manufacturing environment, equipment hardware parameters, consumable usage, and upstream parameter data that can affect the model's output values. Manufacturing environment can include the tool ID used to process the batch, the layer the batch is currently in, and the final product ID (e.g., memory, CPU, etc.), as the same processing can vary depending on different combinations of these manufacturing environments. Equipment hardware parameters can include maintenance information (e.g., number of runs since the last maintenance), machine / tool ​​configuration such as current and voltage settings. Upstream parameter data can include data from the same batch after upstream processing, such as thickness before deposition, etching, or polishing steps.

[0036] Feature selection can be performed using data analysis. For example, linear or nonlinear regression techniques can be used to identify the most relevant features for a given process parameter that will be predicted by an RNN model. Analysis of variance (ANOVA) can be performed to estimate the variance of the process parameter based on various features. Features that are directly related to the variance can be identified and selected.

[0037] Figure 4 An example of selected features for a deposition rate model is shown. Here, feature 402 may include machine / tool ​​402.1, layer 402.2, and product 402.3. Additionally, output (i.e., label 404) may also be included. Here, the observed deposition rate is included as part of the selected data used to train the model. The observed deposition rate may have been measured using metrology tools in previous runs.

[0038] return Figure 3 At operation 306, a filtering operation can be performed on the dataset. The filtering operation removes outliers from the dataset so that they do not adversely affect the model. The filtering operation can include the Tukey filter (also known as the Tukey window), the Grubb filter, the basic limit filter, and / or other suitable filters.

[0039] Preprocessing can be performed on the dataset. Selected data can be preprocessed before being used by the RNN model. For example, in a highly mixed manufacturing environment (e.g., a foundry), the same processing can be performed on multiple tools to build different products with multiple layers; the same processing can be performed on each different combination of these contextual attributes (e.g., machine, product, layer, etc.) to produce different results from each other.

[0040] In addition to numerical data, categorical data such as strings can also be fed into machine learning models. However, most machine learning algorithms typically only process numerical data, such as data matrices. At operation 308, string data can be converted to numerical data. In some examples, one-hot encoding can be used to convert categorical data (e.g., string data) into numerical data. Figure 5 An example of one-hot encoding for converting string data is shown. Here, table 502, which includes selected features of categorical data, is converted into a matrix 504 of numerical values. For example, machine types (CVD-01, CVD-02) are encoded into columns 0 and 1 of matrix 504.

[0041] return Figure 3 At operation 310, data can be scaled. For example, numerical values ​​in a matrix can be scaled or normalized so that these values ​​are evaluated based on proportion rather than absolute value. Scaled data can be used to achieve better performance of machine learning models by evaluating features by their proportion rather than their absolute value (e.g., size). For example, upstream parameter data may include multiple values ​​for different steps (e.g., thickness before a polishing step); these values ​​can be scaled or normalized. Scaling (or normalization) ensures that the numerical values ​​of a feature are properly weighted and expressed as a percentage. In some examples, the minimum and maximum values ​​of a feature can be determined. Based on the minimum and maximum values, the numerical values ​​of the feature can be normalized so that they fall within the range of 0 (minimum value) and 1 (maximum value).

[0042] At operation 312, the set of variables known as hyperparameters of the RNN model can be adjusted. Hyperparameters can be configured in training mode, and their values ​​can be used to control the learning process and significantly affect model performance. Hyperparameters can include the number of hidden layers in the neural network, the number of neurons per layer, batch size, time intervals, dropout, etc. Batch size refers to the size of records fed into the model at one time (e.g., 100 records at a time). Time intervals refer to the number of times the same data is fed into the model. Dropout refers to the percentage of data removed in each time interval. For example, if 100 records are fed into the model in 10 batches of 10 records each in the first time interval with a dropout rate of 20%, then in the second time interval, 80 of the 100 records are fed into the model in 8 batches of 10 records each. A random selection technique can be used to select the 20 records removed between the first and second time intervals.

[0043] In some examples, at least a portion of hyperparameter tuning can be performed manually by data science practitioners, for example, using trial and error. In some examples, the optimal values ​​for hyperparameters for a given problem may be unknown; therefore, the technique may use rules of thumb, replicate values ​​used for other problems, and / or search for optimal values ​​through trial and error (e.g., iterative processes). For example, when designing an RNN model for deposition rates in a CVD process, hyperparameters from a previous model designed for similar processes such as physical vapor deposition (PVD) can be used. The techniques described above can be used to further tune PVD hyperparameters. Hyperparameter tuning can also utilize automated techniques such as grid search to determine optimal values.

[0044] After the machine learning network has completed the training process, the RNN model can be used in normal operation (runtime) mode to set process parameters during manufacturing. Figure 6 A flowchart of a method 600 for manufacturing a semiconductor device in normal operating mode is shown. In some examples, method 600 may be performed by a manufacturing system 100 having a trained RNN model as described herein (e.g., using framework 200 and method 300).

[0045] At operation 602, the manufacturing tool / machine can provide contextual information to the control system about the operation to be performed by the manufacturing tool / machine. In the deposition example, the manufacturing tool can provide information about tool identification, product identification, layer number, etc. At operation 604, the control system can input the received contextual information into a trained RNN model. At operation 606, the RNN model can generate predicted process parameters (e.g., deposition rate prediction).

[0046] At operation 608, the control system can set the process parameters of the manufacturing tool (e.g., deposition time) based on predicted process parameters. At operation 610, the manufacturing tool can perform the operation (e.g., deposition) using the set process parameters. At operation 612, a metering instrument can measure the result of the operation (e.g., film thickness). At operation 614, the measurement result can then be fed back to the RNN model to update the model. In some examples, the measurement result values ​​can be filtered and preprocessed, as described herein (e.g., method 500). For example, a filtering operation can be performed on the measurement results to remove outliers that may be caused by malfunctioning operations. In some examples, a detection system can be used to detect malfunctioning operations, and if a malfunctioning operation is detected, the results from the malfunctioning operation can be discarded without being fed back to the RNN model. For example, a fault detection and classification (FDC) system can be used to detect indicators of malfunctioning operations. For example, an FDC system can monitor tool temperature, gas flow rate, etc., which may be indicators of malfunctioning operations.

[0047] In some examples, measurement results can be fed into the RNN in batches. Measurement results from a set of product batches can be saved and then fed into the RNN in batches to update the model. The batch size can be the same as the batch size adjusted in the hyperparameter tuning operation used to train the RNN model. In some examples, measurement results can be monitored, and if the quality of the predicted process parameters begins to deteriorate, the RNN model updates can be stopped. The RNN model can then be reset.

[0048] Figure 7 Simulation results are shown for a deposition controller employing an RNN deep learning model versus a conventional linear model using an exponentially weighted moving average (EWMA) tuning algorithm. Simulations using production data demonstrate that the deposition controller supported by the RNN deep learning model (31% reduction in root mean square error (RMSE)) performs comparably to the conventional R2R strategy using a linear model and EWMA tuning algorithm (32% reduction in RMSE). Therefore, in addition to use in highly mixed production environments, machine learning solutions can provide complementary capabilities for complex, time-sensitive scenarios where accurate physical or statistical models are not yet readily available.

[0049] Some examples of using the aforementioned machine learning networks are described for predicting deposition rates in CVD processes. For clarity, the techniques using machine learning networks described herein can be applied to other semiconductor manufacturing processes, such as, but not limited to, predicting etch rates in etching processes, polishing rates in CMP, offset adjustments in photolithography, and plating rates in electroplating.

[0050] The techniques shown and described herein can be used as follows: Figure 1 The manufacturing system 100 shown is used as part of or in whole or otherwise in accordance with the following text. Figure 8 The aforementioned machine 800 is used to execute this. Figure 8 A block diagram is shown illustrating an example of a machine 800 including any one or more of the technologies (e.g., methods) described herein. In various examples, machine 800 may operate as a stand-alone device or may be connected (e.g., networked) to other machines.

[0051] In a networked deployment, machine 800 can operate as a server machine, a client machine, or both in a server-client network environment. In one example, machine 800 can act as a peer-to-peer (P2P) (or other distributed) network environment. Machine 800 can be a personal computer (PC), tablet device, set-top box (STB), personal digital assistant (PDA), mobile phone, web device, network router, switch, or bridge, or any machine capable of executing instructions specifying the actions to be taken by that machine (sequentially or otherwise). Furthermore, although only a single machine is shown, the term "machine" should also be understood to include any collection of multiple machines, such as those in cloud computing, Software as a Service (SaaS), or other computer cluster configurations, that individually or jointly execute one or more sets of instructions to perform any of the methods described herein.

[0052] The examples described herein may include logic or multiple components or mechanisms, or may be operated by logic or multiple components or mechanisms. A circuit system is a collection of circuits implemented in a tangible entity including hardware (e.g., simple circuits, gates, logic, etc.). The membership of a circuit system may change over time and with potential hardware variations. A circuit system includes members that can perform a specified operation individually or in combination during operation. In one example, the hardware of the circuit system may be immutably designed to perform a specific operation (e.g., hardwired). In one example, the hardware including the circuit system may include variable-connected physical components (e.g., execution units, transistors, simple circuits, etc.) including computer-readable media that are physically modified (e.g., magnetically, electronically, such as via a change in physical state or a transformation of another physical property) to encode instructions for a specific operation. When connecting physical components, the underlying electrical characteristics of the hardware composition may, for example, change from insulating to conductive or vice versa. Instructions enable embedded hardware (e.g., execution units or loading mechanisms) to form members of a circuit system in the hardware via variable connections to perform a specific operation during operation. Therefore, when the device is operational, the computer-readable medium is communicatively coupled to other components of the circuit system. In one example, any physical component may be used in more than one member of more than one circuit system. For instance, in operation, an execution unit may be used at one point in a first circuit of a first circuit system and reused at different times by a second circuit of the first circuit system or by a third circuit of the second circuit system.

[0053] Machine 800 (e.g., a computer system) may include a hardware-based processor 801 (e.g., a central processing unit (CPU), a graphics processing unit (GPU), a hardware processor core, or any combination thereof), main memory 803, and static memory 805, some or all of which may communicate with each other via interconnect 830 (e.g., a bus). Machine 800 may also include a display device 809, an input device 811 (e.g., an alphanumeric keypad), and a user interface (UI) navigation device 813 (e.g., a mouse). In one example, display device 809, input device 811, and UI navigation device 813 may include at least some portions of a touchscreen display. Machine 800 may additionally include a storage device 820 (e.g., a drive unit), a signal generation device 817 (e.g., a speaker), a network interface device 850, and one or more sensors 815, such as a global positioning system (GPS) sensor, a compass, an accelerometer, or other sensors. Machine 800 may include output controller 819, such as a serial controller or interface (e.g., Universal Serial Bus (USB)), a parallel controller or interface, or other wired or wireless (e.g., infrared (IR) controller or interface, near field communication (NFC), etc.), which are coupled to transmit information to or control one or more peripheral devices (e.g., printer, card reader, etc.).

[0054] Storage device 820 may include a machine-readable medium on which one or more sets of data structures or instructions 824 (e.g., software or firmware) embodying or utilized by any or more of the techniques or functions described herein are stored. During execution of instructions 824 by machine 800, the instructions may also reside wholly or at least partially in main memory 803, in static memory 805, in mass storage device 807, or in a hardware-based processor 801. In one example, one or any combination of the hardware-based processor 801, main memory 803, static memory 805, or storage device 820 may constitute a machine-readable medium.

[0055] Although machine-readable media is considered as a single medium, the term “machine-readable media” can include a single medium or multiple media (e.g., a centralized or distributed database, and / or associated caches and servers) configured to store one or more instructions 824.

[0056] The term "machine-readable medium" can include any medium capable of storing, encoding, or carrying instructions executable by machine 800 and causing machine 800 to perform any one or more of the technologies disclosed herein, or any medium capable of storing, encoding, or carrying instructions used by or associated with such instructions. Examples of non-limiting machine-readable media can include solid-state memory, as well as optical and magnetic media. Therefore, machine-readable media are not transiently propagating signals. Specific examples of large-scale machine-readable media can include: non-volatile memory, such as semiconductor memory devices (e.g., electrically programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM)) and flash memory devices; magnetic or other phase-change or state-changing memory circuits; magnetic disks, such as internal hard disks and removable disks; magneto-optical disks; and CD-ROM and DVD-ROM disks.

[0057] Instruction 824 can also transmit or receive over communication network 821 using a transmission medium via network interface device 850 using any of a variety of transmission protocols (e.g., Frame Relay, Internet Protocol (IP), Transmission Control Protocol (TCP), User Datagram Protocol (UDP), Hypertext Transfer Protocol (HTTP), etc.). Example communication networks may include local area networks (LANs), wide area networks (WANs), packet data networks (e.g., the Internet), mobile phone networks (e.g., cellular networks), conventional telephone (POTS) networks, and wireless data networks (e.g., Wi-Fi). The Institute of Electrical and Electronics Engineers (IEEE) 802.22 standard series, known as The IEEE 802.26 series of standards, the IEEE 802.27.4 series of standards, peer-to-peer (P2P) networks, etc. In one example, network interface device 850 may include one or more physical jacks (e.g., Ethernet, coaxial, or telephone jacks) or one or more antennas to connect to communication network 821. In one example, network interface device 850 may include multiple antennas to perform wireless communication using at least one of single-input multiple-output (SIMO), multiple-input multiple-output (MIMO), or multiple-input single-output (MISO) technologies. The term "transmission medium" should be understood to include any intangible medium capable of storing, encoding, or carrying instructions executable by machine 800, and includes digital or analog communication signals or other intangible media to facilitate communication of such software.

[0058] Various precautions

[0059] Each of the above non-limiting aspects may be independent or may be arranged or combined in various ways with one or more of the other aspects or other topics described in this document.

[0060] The above detailed description includes reference to the accompanying drawings, which form a part of this detailed description. The drawings illustrate, by way of example, specific embodiments in which the invention may be practiced. These specific embodiments are also generally referred to as "examples." Such examples may include elements other than those shown or described. However, the inventors also contemplate examples that provide only those elements shown or described. Furthermore, the inventors contemplate examples using any combination or arrangement of those elements (or one or more aspects thereof) relative to a particular example (or one or more aspects thereof) or relative to other examples (or one or more aspects thereof) shown or described in the text.

[0061] In the event of any inconsistency between the usage in this document and any other document incorporated herein by reference, the usage in this document shall prevail.

[0062] In this document, the terms “a” or “an” are used as commonly found in patent documents to include one or more, independent of any other instances or uses of “at least one” or “one or more.” In this document, unless otherwise specified, the term “or” is used to indicate non-exclusivity or to make “A or B” include “A but not B,” “B but not A,” and “A and B.” In this document, the terms “comprising” and “in which” are used as their commonly understood English equivalents to the corresponding terms “including” and “wherein.” Furthermore, the terms “comprising” and “including” are open-ended, meaning that a system, apparatus, article, composition, formulation, or process that includes elements other than those listed after such terms in one aspect is still considered to fall within the scope of that aspect. Furthermore, the terms “first,” “second,” and “third,” etc., are used merely as labels and are not intended to impose numerical requirements on their objects.

[0063] The methods described herein may be implemented, at least in part, by a machine or computer. Some examples may include computer-readable or machine-readable media encoded with instructions for configuring an electronic device to perform the methods described in the examples above. Specific implementations of such methods may include code, such as microcode, assembly language code, high-level language code, etc. Such code may include computer-readable instructions for performing various methods. The code may form part of a computer program product. Furthermore, in one example, such as during execution or at other times, the code may be tangibly stored on one or more volatile, non-transitory, or non-volatile tangible computer-readable media. Examples of such tangible computer-readable media may include, but are not limited to, hard disks, removable disks, removable optical disks (e.g., high-density disks and digital video disks), magnetic tape cassettes, memory cards or sticks, random access memory (RAM), read-only memory (ROM), etc.

[0064] The above description is intended to be illustrative and not limiting. For example, the examples above (or one or more aspects thereof) may be used in combination with each other. Other specific embodiments may be used, as will be apparent to those skilled in the art after reading the above description. A summary of the specification is provided to enable the reader to quickly determine the essence of the technical disclosure. It should be understood at the time of submission that it is not intended to interpret or limit the scope or meaning of any aspect. Furthermore, in the above specific embodiments, various features may be combined together to simplify the disclosure. This should not be construed as meaning that any unclaimed disclosed features are necessary for any claim. Rather, the subject matter of the invention may be present in all features of fewer specific embodiments than those in a particular disclosure. Therefore, the following aspects are thus incorporated into the specific embodiments as examples or specific embodiments, wherein each aspect is independently as a separate specific embodiment, and it is contemplated that such specific embodiments may be combined or arranged with each other in various combinations or arrangements.

Claims

1. A method of setting at least one process parameter for fabricating a semiconductor device, the method comprising: receiving contextual information about the at least one process parameter for a fabrication machine to perform a process to fabricate the semiconductor device; inputting the contextual information into a machine learning network; receiving a predicted value of the at least one process parameter from the machine learning network; setting the at least one process parameter for the fabrication machine based on the predicted value for performing the process to fabricate the semiconductor device; receiving a measurement from a metrology instrument of the process on the semiconductor device associated with the at least one process parameter; and feeding back the measurement to the machine learning network, wherein the machine learning network is trained based on a plurality of features identified by performing a regression technique to estimate a variance of the at least one process parameter to identify the plurality of features that have a direct relationship with the variance, wherein the plurality of features includes at least one fabrication environment feature, at least one equipment hardware parameter feature, and at least one upstream parameter data feature.

2. The method of claim 1, wherein the machine learning network comprises a recurrent neural network (RNN) deep learning model.

3. The method of claim 1, wherein the machine learning network is trained to generate the predicted value using a training process, the training process comprising: retrieving data associated with historical production; selecting a subset of the data based on a plurality of features related to the at least one process parameter, wherein the plurality of features are independent; and inputting the subset into the machine learning network for training.

4. The method of claim 3, wherein the training process further comprises: screening the subset of the data to remove outliers.

5. The method of claim 3, wherein the training process further comprises: converting string data in the subset of the data to numerical values.

6. The method of claim 3, wherein the training process further comprises: scaling the subset of the data to generate normalized data.

7. The method of claim 3, wherein the training process further comprises: adjusting at least one hyperparameter associated with the machine learning network.

8. A control system, the control system comprising: at least one hardware processor; and at least one memory storing instructions that, when executed by the at least one hardware processor, cause the at least one hardware processor to perform operations comprising: receiving contextual information about at least one process parameter for a fabrication machine to perform a process to fabricate a semiconductor device; inputting the contextual information into a machine learning network; receiving a predicted value of the at least one process parameter from the machine learning network; setting the at least one process parameter for the fabrication machine based on the predicted value for performing the process to fabricate the semiconductor device; ​ ​ receiving, from a metrology instrument, a measurement associated with the at least one process parameter for the process on the semiconductor device; and feeding back the measurement to the machine learning network, wherein the machine learning network is trained based on a plurality of features identified by performing a regression technique to estimate a variance of the at least one process parameter to identify the plurality of features that are directly related to the variance, wherein the plurality of features includes at least one manufacturing environment feature, at least one equipment hardware parameter feature, and at least one upstream parameter data feature.

9. The control system of claim 8, wherein the machine learning network comprises a recurrent neural network (RNN) deep learning model.

10. The control system of claim 8, wherein the machine learning network is trained to generate the predicted value using a training process, the training process comprising: retrieving data associated with historical production; selecting a subset of the data based on a plurality of features related to the at least one process parameter, wherein the plurality of features are independent; and inputting the subset to the machine learning network for training.

11. The control system of claim 10, wherein the training process further comprises: screening the subset of the data to remove outliers.

12. The control system of claim 10, wherein the training process further comprises: converting string data in the subset of the data to numerical values.

13. The control system of claim 10, wherein the training process further comprises: scaling the subset of the data to generate normalized data.

14. The control system of claim 10, wherein the training process further comprises: adjusting at least one hyperparameter associated with the machine learning network.

15. A machine storage medium embodying instructions that, when executed by a machine, cause the machine to perform operations comprising: receiving contextual information about at least one process parameter for a manufacturing machine to perform a process to manufacture a semiconductor device; inputting the contextual information to a machine learning network; receiving, from the machine learning network, a predicted value of the at least one process parameter; setting the at least one process parameter for the manufacturing machine based on the predicted value for performing the process to manufacture the semiconductor device; receiving, from a metrology instrument, a measurement associated with the at least one process parameter for the process on the semiconductor device; and feeding back the measurement to the machine learning network, wherein the machine learning network is trained based on a plurality of features identified by performing a regression technique to estimate a variance of the at least one process parameter to identify the plurality of features that are directly related to the variance, wherein the plurality of features includes at least one manufacturing environment feature, at least one equipment hardware parameter feature, and at least one upstream parameter data feature.

16. The machine storage medium of claim 15, wherein the machine learning network comprises a recurrent neural network (RNN) deep learning model.

17. The machine storage medium of claim 15, wherein the machine learning network is trained to generate the predicted values using a training process, the training process comprising: retrieving data associated with historical production; selecting a subset of the data based on a plurality of features related to the at least one process parameter, wherein the plurality of features are independent; and inputting the subset into the machine learning network for training.

18. The machine storage medium of claim 17, wherein the training process further comprises: screening the subset of the data to remove outliers.

19. The machine storage medium of claim 17, wherein the training process further comprises: converting string data in the subset of the data to numerical values.

20. The machine storage medium of claim 17, wherein the training process further comprises: scaling the subset of the data to generate normalized data.

Images