Method and non-transitory machine-readable storage medium for virtual measurement of conditions proximate to a substrate with physics-informed compressed sensing

Abstract

The method described herein includes the following steps: measuring a subset of attribute values ​​within a manufacturing chamber during processing of a substrate within the manufacturing chamber. The method further includes the step of determining the attribute values ​​within the manufacturing chamber at a position removed from the measurement location. The method further includes the step of performing a correction action based on the determined attributes.

Description

Technical Field

[0001] This specification relates to the use of a machine learning model based on physical information compression sensing to virtually measure the properties of adjacent substrates in a manufacturing chamber. Prior Technology

[0002] Chambers are used in many types of processing systems. Examples of chambers include etching chambers, deposition chambers, and annealing chambers, among others. Typically, a substrate, such as a semiconductor wafer, is placed in the chamber, and the conditions within the chamber are set and maintained to process the substrate. To fully characterize and predict the effects of the processing conditions, measurements of property values ​​at the substrate location are ideal. However, performing measurements at the substrate location is often inconvenient or impossible. Many systems have more convenient (though less immediately useful) locations to measure the values ​​of relevant operating properties, such as temperature, gas flow rate, and pressure, etc. Summary of the Invention

[0003] The following is a simplified overview of this application to provide a basic understanding of some aspects of it. This [Summary] is not a broad summary of the contents of this application. This [Summary] is not intended to identify key or essential elements of this application, nor is it intended to describe any category of specific embodiments of this application or any category of the scope of the claims. The sole purpose of this [Summary] is to present some concepts of this application in a simplified form as a prelude to the [Simplified Embodiments] presented later.

[0004] In one embodiment of this application, the method includes the following steps: measuring values ​​of one or more properties at multiple locations within a manufacturing chamber. Measurements are performed using sensors placed within the chamber during processing on the substrate. The method further includes the step of providing the measured attribute values ​​as input to a trained machine learning model. The trained machine learning model is able to infer attribute values ​​at other locations within the chamber, including the substrate location, using the measurements at the sensor locations. The method further includes the step of performing an action based on the attributes inferred by the trained machine learning model.

[0005] In another embodiment of this application, the method includes the following steps: training a machine learning model. The method further includes the following steps: providing training input data to the machine learning model. The training input data includes measurements of attribute values ​​acquired at locations within a manufacturing chamber while the substrate is being processed. At least most of the measurements are performed at locations not immediately adjacent to the substrate. The method further includes the following steps: providing target output data to the machine learning model. The target output includes a mapping of attribute values ​​in a chamber region, including the substrate location. The method further includes the following steps: providing input and output datasets to the machine learning model to train the model to predict the mapping of attribute values ​​in a chamber region given measurements from multiple locations within the chamber.

[0006] In another embodiment of this application, a system for implementing methods (such as those described above) is disclosed. The system includes a computer-readable storage medium. The system further includes a processing device. The system is configured such that the computer-readable storage medium contains instructions to cause the processing device to perform the method when executed. The method executed by the processing device includes the step of receiving measured values ​​of attribute values ​​from sensors placed within the manufacturing chamber during manufacturing in a manufacturing chamber. The method executed by the processing device further includes the step of using a trained machine learning model to determine attribute values ​​in a region of the manufacturing chamber, including a substrate location. The method executed by the processing device further includes the step of performing a correction action or sending an alarm to a user device based on the determined attribute values. Simple Explanation of the Diagram

[0007] The application is illustrated in the accompanying diagrams by way of example rather than limitation.

[0008] Figure 1 is a block diagram illustrating an exemplary system (exemplary system architecture) according to certain embodiments.

[0009] Figure 2 is a block diagram of an exemplary dataset generator for creating datasets for a machine learning model, according to certain embodiments.

[0010] Figure 3 is a block diagram illustrating a system for generating output data (e.g., compressed sensing data) according to certain embodiments.

[0011] Figures 4A to 4C are flowcharts of methods for associating compressed sensing prediction data with certain embodiments to induce correction actions.

[0012] Figure 5 depicts a cross-sectional view of a manufacturing chamber (e.g., a semiconductor wafer manufacturing chamber) according to some embodiments.

[0013] Figure 6 is a block diagram illustrating a computer system according to certain embodiments. Implementation

[0014] This article describes a technique for virtually measuring attribute values ​​at substrate locations within a manufacturing chamber using physically based compression sensing technology. Manufacturing equipment is used to produce substrates such as semiconductor wafers. The properties of these substrates are controlled by the state in which they have been processed. Accurately understanding the properties within the manufacturing chamber (especially in the immediate vicinity of the substrates) during operation allows for the prediction of finished product properties, consistent production of substrates with the same properties, and the customization of processing parameters to account for variations in chamber design and construction.

[0015] In some cases, sensors can be placed at the location of the substrate, or sensors capable of measurement from a distance (e.g., IR temperature sensors) can be used. In some cases, this arrangement is impractical or impossible. When it is not desirable to directly measure the property values ​​of adjacent substrates (e.g., on the substrate), empirical methods can be used to produce consistent substrates with selected characteristics. This approach involves experimentation, which can be costly in terms of time, material usage, energy, and gas consumption. Furthermore, because manufacturing chambers are reused, their properties tend to deviate due to material buildup, component aging, changes during cleaning, and the like. Any acquired knowledge mapping input processing parameters to output substrate properties degrades with deviations in chamber properties. Another problem with this conventional approach relates to the differences between chambers. Any new chamber design requires a new set of experiments to determine a new mapping between input parameters and finished product properties; these mappings also degrade in quality as chambers age. Moreover, manufacturing chambers are subject to tolerances when they are built, meaning that each chamber may potentially function differently, even from nominally identical chambers. Not only is experimentation required for each chamber type to accurately understand which input parameters lead to which products, but each individual chamber may also be subject to this treatment, and the understanding of these properties can benefit from continuous updates. Any attempt to produce a different type of substrate (whose properties exceed the scope of previously collected knowledge about chamber operation) further exacerbates this problem.

[0016] The method and apparatus of this application address at least some of the shortcomings of conventional methods. This application allows for virtual measurements at a substrate using compressed sensing based on physical information. In some embodiments, the method of this application includes the following steps: using a computer-aided physical model. The physical model is configured to include all features of the manufacturing chamber of interest, including geometry, material properties, and airflow, etc. The physical model is then run multiple times, which may reflect a range of processing parameters. For example, the power supplied to the heater, the gas mixture used, the gas flow rate and pressure, hardware parameters such as substrate placement, and the like can all be varied. The output of these calculations using the physical model is a set of mappings indicating steady-state property values ​​within the manufacturing chamber. This set of mappings can span a significant portion of the parameter space related to substrate production.

[0017] In some embodiments, the method of this application further includes the step of: constructing a baseline set using the output of a physics-based model to describe possible combinations of attribute values ​​in a manufacturing chamber. The simplest conceptual representation of an attribute in a chamber is typically assigning the value of each attribute to each point in the chamber. This can be considered as a weighted combination of members of the baseline set, where each member of the baseline set includes an attribute value (e.g., 1 Kelvin temperature) at a specific point in the chamber and an attribute assigned a null value (e.g., zero Kelvin temperature) at every other point in the chamber. A complete mapping of attribute values ​​for the chamber is then established by adding a weighted combination of each baseline state to assign non-zero temperatures to each point in the chamber.

[0018] This conceptually simple representation is generally not the most convenient. Instead, a reference set with more complex members can be used, where the properties in the chamber are represented as a weighted sum of reference state members. With careful selection of the reference set, fewer members (rather than those used in the simple case) can be used to adequately describe the property values ​​in the chamber, where each member of the reference set makes a meaningful contribution. In some embodiments, the method of this application includes the step of constructing a reference set from the output of a physics-based model such that a weighted sum of fewer reference states is used to adequately describe the property values ​​in the chamber within a parameter space region relevant to substrate production. This reference set with reduced dimensionality is called a reduced-order model (ROM). In some embodiments, the ROM can be formed by performing an appropriate orthogonal decomposition on the output of the physics-based model.

[0019] In some embodiments, the method of this application further includes the steps of: training a machine learning model as input to a finite subset of attribute values ​​corresponding to attribute values ​​at multiple locations within the manufacturing chamber; training the machine learning model to generate a mapping of attribute values ​​as output, this mapping extending to locations beyond the input locations, including attribute values ​​proximate to the substrate. In some embodiments, a sparse regression model is used to train the machine learning model. Alternatively, other types of machine learning models, such as neural networks and support vector machines, may be used.

[0020] In some embodiments, the method of this application further includes the step of providing measurements of a subset of attribute values ​​within the manufacturing chamber as input to a trained machine learning model during processing. The method further includes the step of receiving predictions of attributes of adjacent substrates as output from the trained machine learning model. This application also enables the use of a system including processing means and memory to perform methods for physically based compressed sensing and virtual measurement.

[0021] The approach described in this application results in a technical advantage over conventional methods. Without compromising processing quality, by arranging the measuring device at a location adjacent to the substrate, accurate predictions of the property values ​​of the adjacent substrate during processing (e.g., temperature values ​​at points on the substrate) can be made. Furthermore, by using a reference set informed by a physically based model of the chamber to inform the temperature mapping, the disclosed method can be used to interpret variations between chambers or variations within a chamber over time.

[0022] Figure 1 is a block diagram illustrating an exemplary system 100 (exemplary system architecture) according to certain embodiments. System 100 includes a client device 120, manufacturing equipment 124, a sensor 126, a metering device 128, a compression sensing server 112, and data storage 140. The compression sensing server 112 may be part of the compression sensing system 110. The compression sensing system 110 may further include server machine 170 and server machine 180.

[0023] Sensor 126 can provide sensor data 142 associated with manufacturing equipment 124 (e.g., associated with the production of a corresponding product such as a wafer by manufacturing equipment 124). For example, sensor data 142 can be used for equipment health and / or product health (e.g., product quality). Manufacturing equipment 124 may produce products according to a formula or perform operations over a period of time. In some embodiments, sensor data 142 may include one or more of temperature (e.g., heater temperature), spacing (SP), pressure, high-frequency radio frequency (HFRF), electrostatic chuck voltage (ESC), current, (e.g., flow rate of one or more gases), power, voltage, etc. Sensor data 142 may include historical sensor data 144 and current sensor data 146. Manufacturing equipment 124 may be configured according to manufacturing data 150. Manufacturing data 150 may be associated with or indicate manufacturing parameters, such as hardware parameters of the manufacturing equipment (e.g., settings or components of manufacturing equipment 124 (e.g., size and type, etc.)) and / or processing parameters of the manufacturing equipment. Manufacturing data 150 may include historical manufacturing data 152 and / or current manufacturing data 154. Manufacturing data 150 may indicate input settings of the manufacturing apparatus (e.g., heater power and airflow, etc.). Sensor data 142 and / or manufacturing data 150 (e.g., equipment readings during product processing) may be provided while the manufacturing apparatus 124 is performing a manufacturing process. Sensor data 142 may be different for each product (e.g., each wafer).

[0024] In some embodiments, sensor data 142, measurement data 169, and / or manufacturing data 150 may be processed (e.g., via client device 120 and / or via compression sensing server 112). Processing of sensor data 142 may include generating features. In some embodiments, features are patterns (e.g., slope, width, height, and peak values, etc.) in sensor data 142, measurement data 169, and / or manufacturing data 150, or combinations of values ​​from sensor data 142, measurement data 196, and / or manufacturing data 150 (e.g., power derived from voltage and current, etc.). Sensor data 142 may include features, and these features may be used by compression sensing component 114 to perform signal processing and / or to obtain prediction data 168 and / or compression sensing data 166, and may be used to perform correction actions. Compression sensing data 166 may be any data associated with compression sensing system 110, such as predicted sensor data at a substrate location during manufacturing processes.

[0025] Each instance (e.g., group) of sensor data 142 may correspond to a product (e.g., wafer), a set of manufacturing equipment, a type of substrate produced by the manufacturing equipment, and combinations thereof. Similarly, each instance of measurement data 169 and manufacturing data 150 may correspond to a product, a set of manufacturing equipment, a type of substrate produced by the manufacturing equipment, and combinations thereof. The data storage may further store information associated with different data types in a group, such as information indicating that a group of sensor data, a group of measurement data, and / or a group of manufacturing data are all associated with the same product, manufacturing equipment, and substrate type, etc.

[0026] In some embodiments, the compressed sensing system 110 may use supervised machine learning to generate compressed sensing data 166 (e.g., target output including virtual sensor data provided in the compressed sensing system 110, etc.).

[0027] Client device 120, manufacturing equipment 124, sensor 126, metering device 128, compression sensing server 112, data storage 140, server machine 170 and server machine 180 can be coupled to each other via network 130 to generate compression sensing data 166 and / or prediction data 168, and optionally to perform correction actions.

[0028] In some embodiments, network 130 is a public network providing client device 120 with access to the compression sensing server 112, data storage 140, and / or other publicly available computing devices. In some embodiments, network 130 is a private network providing client device 120 with access to manufacturing equipment 124, sensor 126, metering device 128, data storage 140, and / or other privately available computing devices. Network 130 may include one or more wide area networks (WANs), local area networks (LANs), wired networks (e.g., Ethernet), wireless networks (e.g., 802.11 networks or Wi-Fi networks), cellular networks (e.g., LTE networks), routers, hubs, switches, server computers, cloud computing networks, and / or combinations thereof.

[0029] Client device 120 may include computing devices such as personal computers (PCs), laptops, mobile phones, smartphones, tablets, laptops, network-connected televisions ("smart TVs"), network-connected media players (e.g., Blu-ray players), set-top boxes, over-the-top (OTT) streaming devices, and junction boxes, etc. Client device 120 may include a calibration action component 122. Calibration action component 122 may (e.g., via a graphical user interface (GUI) displayed on client device 120) receive user input of instructions associated with manufacturing equipment 124. In some embodiments, calibration action component 122 transmits instructions to compression sensing system 110, receives output from compression sensing system 110 (e.g., compression sensing data 166), determines a calibration action based on the output, and performs the calibration action.

[0030] In some embodiments, the compression sensing system 110 may further include a prediction component 116. The prediction component 116 may acquire data retrieved from the compression sensing component 114 to generate prediction data 168. In some embodiments, the prediction component 116 provides the prediction data 168 to a client device 120, and the client device 120 initiates a correction action via a correction action component 122 based on the prediction data 168. In some embodiments, the correction action component 122 obtains sensor data 142 associated with the manufacturing equipment 124 (e.g., current sensor data 146) from data storage 140, and provides the sensor data 142 associated with the manufacturing equipment 124 (e.g., current sensor data 146) to the compression sensing system 110.

[0031] In some embodiments, the calibration action component 122 stores sensor data 142 in data storage 140, and the compressed sensing server 112 retrieves the sensor data 142 from data storage 140. In some embodiments, the compressed sensing server 112 may store the outputs of trained machine learning models(s) 190 (e.g., compressed sensing data 166) in data storage 140, and the client device 120 may retrieve the outputs from data storage 140. In some embodiments, the calibration action component 122 receives instructions for calibration actions from the compressed sensing system 110 and performs the calibration actions. Each client device 120 may include an operating system that allows a user to perform one or more of the following: generating, viewing, or editing data (e.g., instructions associated with manufacturing equipment 124 and calibration actions associated with manufacturing equipment 124, etc.).

[0032] In some embodiments, metering data 192 corresponds to historical attribute data of a product (e.g., a product manufactured using manufacturing parameters associated with historical sensor data 144 and historical manufacturing data 152), and prediction data 168 is associated with predicted attribute data (e.g., a product to be manufactured or a product already manufactured in the state recorded in current sensor data 146 and / or current manufacturing data 154). In some embodiments, prediction data 168 is predicted metering data (e.g., virtual metering data) of a product to be manufactured or already manufactured based on the state recorded as current sensor data 146 and / or current manufacturing data 154. In some embodiments, prediction data 168 is or includes indications of anomalies (e.g., abnormal products, abnormal components, abnormal manufacturing equipment, and abnormal energy usage, etc.) and / or one or more causes of anomalies. In some embodiments, prediction data 168 includes indications of changes or drifts over time in some components such as manufacturing equipment 124, sensors 126, and metering devices 128. In some embodiments, prediction data 168 includes indications of the end of the lifespan of components such as manufacturing equipment 124, sensors 126, and metering devices 128.

[0033] Performing manufacturing processes that result in defective products can be costly in terms of time, energy, products, components, manufacturing equipment 124, and the cost of identifying and discarding defective products. By inputting sensor data 142 (e.g., manufacturing parameters for manufacturing products), receiving output from compressed sensor data 166, and performing correction actions based on the compressed sensor data 166, system 100 can have the technical advantage of avoiding the costs of producing, identifying, and discarding defective products.

[0034] Performing manufacturing processes that cause component failure in manufacturing equipment 124 can be costly in terms of downtime, product damage, equipment damage, and the need for rapid replacement of components. By inputting sensor data 142 (e.g., manufacturing parameters currently in use or to be used in manufacturing products), receiving the output of compressed sensing data 166, comparing compressed sensing data sets from the same chamber over time to diagnose drifting or faulty components (recorded as predictive data 168), and performing corrective actions based on predictive data 168 (e.g., predicted operational maintenance such as component replacement, handling, and cleaning), system 100 offers the technological advantage of avoiding costly issues such as one or more unexpected component failures, unplanned downtime, productivity losses, unexpected equipment failures, and product scrap. Monitoring the performance of components (such as manufacturing equipment 124, sensors 126, metering devices 128, etc.) over time can provide indications of component degradation.

[0035] For products that may have high production costs, manufacturing parameters may be suboptimal. These high costs stem from increased resource consumption (e.g., energy, coolant, and gas), increased production time, increased component failures, and an increased number of defective products. By inputting sensor data 142 into a trained machine learning model 190, receiving the output of compressed sensor data 166, and (e.g., based on compressed sensor data 166) performing corrective actions to update manufacturing parameters (e.g., setting optimal manufacturing parameters), system 100 can have the technological advantage of using optimal manufacturing parameters (e.g., hardware parameters, processing parameters, and optimal design) to avoid the high costs arising from suboptimal manufacturing parameters.

[0036] Corrective actions may be associated with one or more of the following: computational process control (CPC), statistical process control (SPC) (e.g., SPC for determining electronic components in a controlled process, SPC for predicting component lifespan, and SPC for comparing 3-sigma plots), advanced process control (APC), model-based process control, preventative operation and maintenance, design optimization, manufacturing parameter updates, manufacturing recipe updates, feedback control, machine learning modifications, or the like.

[0037] In some embodiments, the corrective action includes providing an alarm (e.g., an alarm to stop or not perform manufacturing processing if compression sensing data 166 or prediction data 168 indicates a predicted anomaly (such as an anomaly in the product, component, or manufacturing equipment 124). In some embodiments, the corrective action includes providing feedback control (e.g., modifying manufacturing parameters in response to compression sensing data 166 or prediction data 168 indicating an anomaly). In some embodiments, the corrective action includes providing machine learning (e.g., modifying one or more manufacturing parameters based on prediction data 168). In some embodiments, the execution of the corrective action includes causing an update to one or more manufacturing parameters.

[0038] Manufacturing parameters may include hardware parameters (e.g., component replacement, use of certain components, replacement of processing wafers, and firmware updates, etc.) and / or processing parameters (e.g., temperature, pressure, flow rate, speed, current, voltage, gas flow rate, and lift speed, etc.). In some embodiments, corrective actions include inducing preventative operational maintenance (e.g., replacing, processing, and cleaning components of manufacturing equipment 124, etc.). In some embodiments, corrective actions include inducing design optimization (e.g., updating manufacturing parameters, manufacturing processes, and manufacturing equipment 124, etc., to optimize the product). In some embodiments, corrective actions include updating the formulation (e.g., placing manufacturing equipment 124 in idle mode, sleep mode, and preheating mode, etc.).

[0039] Compression sensing server 112, server machine 170 and server machine 180 may each include one or more computing devices, such as rack server, router computer, server computer, personal computer, mainframe computer, laptop computer, tablet computer, desktop computer, graphics processing unit (GPU), accelerator application-specific integrated circuit (ASIC) (e.g., tensor processing unit (TPU)), etc.

[0040] The compressed sensing server 112 may include a compressed sensing component 114. In some embodiments, the compressed sensing component 114 may receive current sensor data 146 and / or current manufacturing data 154 (e.g., received from client device 120, retrieved from data storage 140) and generate an output (e.g., compressed sensing data 166) to perform a correction action associated with manufacturing equipment 124 based on the current data. In some embodiments, the compressed sensing component 114 may use one or more trained machine learning models 190 to determine the output to perform the correction action based on the current data.

[0041] The compression sensing server 112 may also include a prediction component 116. The prediction component 116 can be used to generate prediction data 168. In some embodiments, the prediction data 168 includes indications of component aging and degradation, etc., of the manufacturing equipment 124. In some embodiments, the prediction component 116 may combine the compression sensing data 166 with virtual measurements of the internal state of the manufacturing chamber to determine predictions of wafer properties. In some embodiments, the prediction component 116 includes a machine learning model.

[0042] In some embodiments, sensor data 142 is provided to a trained machine learning model 190. The machine learning model 190 is trained to output attribute values ​​(e.g., attribute values ​​at the substrate location within the manufacturing chamber) based on the input sensor data 142. In some embodiments, model 190 may utilize a sparse regression algorithm. In some embodiments, machine learning model 190 takes temperature data from multiple sensors arranged at locations within the manufacturing chamber as input. In some embodiments, at least most of the sensors are not located at the substrate location. For example, temperature sensors may be positioned at various locations near gas inlets, exhaust ducts, substrate supports, and chamber walls, etc. Model 190 then generates temperature data at all points within the chamber as output. In some embodiments, model 190 may generate attribute values ​​at regions of interest within the chamber (which do not include attribute values ​​at other regions of interest) as output; the regions of interest within the chamber include locations at least adjacent to the substrate. Generating data relating only to a subset of locations within the chamber provides the technical advantage of reliably virtually measuring attribute values ​​on the substrate using fewer sensors 126.

[0043] In some embodiments, model 190 generates a set of coefficients as output. These coefficients serve as weights to generate a weighted summation of the members of a reference set. The weighted summation of the reference set indicates the mapping of attribute values ​​in the manufacturing chamber. In an embodiment, the reference set used by model 190 is generated by a reduced-order model generator 174 of server machine 170. The reference set can be saved as reduced-order reference data 164 in data storage 140. The reference set can be selected using a small number of members of the reference set (e.g., a small number of non-zero coefficients output by model 190) in a manner that represents the mapping of attribute values ​​in the manufacturing chamber for different manufacturing input parameters. The parameter value that determines the number of non-zero coefficients output by model 190 can be determined during the training of model 190.

[0044] In some embodiments, the reduced-order model generator 174 generates a reference set based on the output of a physics-based model 176. In some embodiments, the physics-based model 176 is used to solve equations describing the steady-state conditions of a manufacturing chamber under various input parameters (power supplied to the heater, gas mixture, gas flow rate, etc.). The selected input parameters may span a parameter space region considered relevant to substrate production. For each variation of the input data, the physics-based model 176 generates a mapping of attribute values ​​corresponding to predicted attribute values ​​in a steady-state state within the manufacturing chamber under the given input data. The data corresponding to the physics-based model may be stored in data storage 140 as physics-based model data 160, which includes storing the resulting attribute value mappings as model attribute mapping data 162. The reduced-order model generator 174 takes as input a set of attribute value mappings generated by the physics-based model 176 (which can be retrieved from the model attribute mapping data 162 in data storage 140). The reduced-order model generator 174 generates a reference set describing the set of attribute value mappings. The generated reference set is stored as reduced-order reference data 164 in data storage 140. The reduced-order model generator 174 can generate the reference set using an appropriate orthogonal decomposition algorithm.

[0045] Historical sensor data 144 can be combined with current sensor data 146 and manufacturing data 150 to detect drift, changes, and aging of components in manufacturing equipment 124. Predictive component 116 can use the combination and comparison of these data types to generate predictive data 168. In some embodiments, predictive data 168 includes data predicting the lifespan of components such as manufacturing equipment 124, sensors 126, etc.

[0046] Combining sparse regression and physics-based models offers a technological advantage over other techniques. Accurate physics-based models provide a reliable mapping of property values ​​within a manufacturing chamber. However, manufacturing chambers cannot always be perfectly modeled due to factors such as manufacturing tolerances and component aging. For example, heaters may provide slightly less or more energy than expected, gas flow regulators may not allow for precise flow rate selection, and contact between surfaces within the chamber may be less than ideal. Users may be unaware of such variations and may not be able to capture them in a purely physics-based model. These limitations can be overcome by using a machine learning sparse regression model that performs virtual measurements via compressed sensing. The machine learning model is trained on a variety of input parameters spanning a region of the parameter space. Interpolation can be used to explain variations in the actual state within the parameter space used to train the machine learning model (due to inaccuracies in the model as described above), and extrapolation can be used to explain settings used to train the machine learning model that are somehow outside the parameter space region. In this way, unexpected variations in the operating state of the chamber can be incorporated into the results of the compressed sensing system 110.

[0047] In some embodiments, the compression sensing component 114 receives current sensor data 146 and can perform preprocessing such as extracting patterns from the data or combining the data into new composite data. The compression sensing component 114 can then provide the data as input to a trained machine learning model 190. The compression sensing component 114 can receive a mapping of attribute values ​​from the trained machine learning model 190, including values ​​at substrate locations. The compression sensing component 114 can then cause a calibration action. The calibration action may include sending an alarm to a client device 120. The calibration action may also include updating manufacturing parameters of the manufacturing equipment 124. The calibration action may also include generating predictive data 168 indicating chamber or instrument drift, aging, or failure.

[0048] Data storage 140 may be memory (e.g., random access memory), a drive (e.g., hard disk drive and flash drive), a database system, or another type of component or device capable of storing data. Data storage 140 may include multiple storage components (e.g., multiple drives or multiple databases) that can span multiple computing devices (e.g., multiple server computers). Data storage 140 may store sensor data 142, manufacturing data 150, physics-based model data 160, compressed sensing data 166, predictive data 168, and measurement data 169. Sensor data 142 may include historical sensor data 144 and current sensor data 146. Sensor data may include time trajectories of sensor data during the manufacturing process, correlations between data and physical sensors, preprocessed data such as averages and composite data, and data indicating changes in sensor performance over time (i.e., many manufacturing processes). Manufacturing data 150 and measurement data 169 may contain similar characteristics. Historical sensor data 144 and historical manufacturing data 152 may be historical data (e.g., at least a portion of data used to train machine learning model 190). Current sensor data 146 may be current data (e.g., at least a portion of data to be input into training machine learning model 190 after historical data) for generating compressed sensor data 166 (e.g., for performing correction actions). Measurement data 169 may be measurement data for manufacturing substrates, as well as sensor data, manufacturing data, and model data corresponding to those products. Measurement data 169 may be used to design processes for manufacturing additional substrates. Predictive data 168 may include predictions of what measurement data can be generated from a set of input parameters. Predictive data 168 may also include data indicating the aging and failure of components of system 100 over time.

[0049] In some embodiments, the compressed sensing system 110 further includes server machine 170 and server machine 180. Server machine 170 includes a dataset generator 172, which is capable of generating datasets (e.g., a set of data inputs and a set of target outputs) to train, validate, and / or test machine learning models 190. Some operations of the dataset generator 172 are described in detail below with reference to Figures 2 and 4A. In some embodiments, the dataset generator 172 may divide historical data (e.g., historical sensor data 144, historical manufacturing data 152) and physical model data (e.g., model attribute mapping data 162, reduced-order base data 164) into a training set (e.g., 60 percent of the data), a validation set (e.g., 20 percent of the data), and a test set (e.g., 20 percent of the data). In some embodiments, the compressed sensing system 110 (e.g., via compressed sensing component 114) generates multiple sets of features. For example, the first set of features may correspond to sensor data of a first type (e.g., from the first set of sensors, a first combination of values ​​from the first set of sensors, and a first pattern of values ​​from the first set of sensors), the sensor data of the first type corresponding to each dataset (e.g., training set, validation set, and test set), and the second set of features may correspond to sensor data of a second type (e.g., from a second set of sensors different from the first set of sensors, values ​​of a second combination different from the first combination, and a second pattern different from the first pattern), the sensor data of the second type corresponding to each dataset.

[0050] Server machine 180 includes a training engine 182, a verification engine 184, a selection engine 185, and / or a testing engine 186. Engines (e.g., training engine 182, verification engine 184, selection engine 185, and testing engine 186) can refer to hardware (e.g., circuitry, dedicated logic, programmable logic, microcode, and processing devices, etc.), software (e.g., instructions running on a processing device, general-purpose computer system, or dedicated machine), firmware, microcode, or a combination thereof. Training engine 182 may be able to train machine learning model 190 using one or more sets of features associated with a training set from dataset generator 172. Training engine 182 may generate multiple trained machine learning models 190, each trained machine learning model 190 corresponding to a different set of features from the training set (e.g., sensor data from a different set of sensors). For example, a first trained machine learning model may have been trained using all features (e.g., X1-X5), a second trained machine learning model may have been trained using a first subset of features (e.g., X1, X2, X4), and a third trained machine learning model may have been trained using a second subset of features that partially overlap with the first subset of features (e.g., X1, X3, X4, and X5). Data set generator 172 may receive the output of a trained machine learning model (e.g., 190), collect that data into training, validation, and test datasets, and use these datasets to train the second machine learning model.

[0051] The validation engine 184 may be able to validate the trained machine learning model 190 using a corresponding feature set from the validation set generated by the data set generator 172. For example, a first set of features from the validation set may be used to validate a first trained machine learning model 190 trained using a first set of features from the training set. The validation engine 184 may determine the accuracy of each trained machine learning model 190 based on the corresponding feature set from the validation set. The validation engine 184 may discard trained machine learning models 190 whose accuracy does not meet a threshold accuracy. In some embodiments, the selection engine 185 may be able to select one or more trained machine learning models 190 whose accuracy meets the threshold accuracy. In some embodiments, the selection engine 185 may be able to select the trained machine learning model 190 with the highest accuracy among the trained machine learning models 190.

[0052] The testing engine 186 may be able to test the trained machine learning model 190 using corresponding sets of features from the test set generated by the data set generator 172. For example, the first set of features from the test set may be used to test the first trained machine learning model 190 trained using the first set of features from the training set. The testing engine 186 may determine the trained machine learning model 190 with the highest accuracy among all trained machine learning models based on the test set.

[0053] Machine learning model 190 may refer to a model artifact created by training engine 182 using a training set that includes data inputs and corresponding target outputs (correct answers to the corresponding training inputs). Patterns can be found in the dataset that map the data inputs to the target outputs (correct answers), and machine learning model 190 provides mappings that capture these patterns. In some embodiments, machine learning model 190 uses a sparse regression algorithm to reconstruct attribute values ​​on a substrate in a manufacturing chamber based on multiple values ​​provided by sensors arranged in the manufacturing chamber, where at least most of the sensors are not located at the substrate location. In some embodiments, machine learning model 190 may use LASSO regression to generate attribute values ​​in the manufacturing chamber.

[0054] The compressed sensing component 114 can provide current sensor data 146 to a trained machine learning model 190 and can run the trained machine learning model 190 on the input to obtain one or more outputs. The compressed sensing component 114 can determine (e.g., extract) compressed sensing data 166 from the output of the trained machine learning model 190, and can determine (e.g., extract) confidence data from the output; the confidence data indicates that the compressed sensing data 166 is a confidence level of an accurate predictor of a process associated with input data of a product produced or to be produced using the manufacturing equipment 124 under the current sensor data 146. The prediction component 116 can determine prediction data 168, which includes predictions of finished substrate properties and effective lifespan predictions of components, sensors 126, or metering equipment 128 of the manufacturing equipment 124 based on the output of model 190. The prediction component 116 can determine confidence data; the confidence data indicates that the prediction data 168 is a confidence level of an accurate predictor of a specified event. The compression sensing component 114, prediction component 116, or correction action component 122 may use confidence data to determine whether to initiate a correction action associated with the manufacturing equipment 124 based on the compression sensing data 166 and / or prediction data 168.

[0055] Confidence data may include or indicate a confidence level. As an example, prediction data 168 may indicate the properties of a finished wafer given a set of manufacturing inputs. Confidence data may indicate that prediction data 168 is an accurate prediction of the product associated with at least a portion of the input data. In one example, the confidence level is a real number between 0 and 1, where 0 indicates disbelief that prediction data 168 is an accurate prediction of the product processed based on the input data, and 1 indicates absolute confidence that prediction data 168 accurately predicts the properties of the product processed based on the input data. In response to confidence data indicating a confidence level below a threshold level for a predetermined number of instances (e.g., percentage of instances, frequency of instances, total number of instances, etc.), prediction component 116 may retrain a trained machine learning model 190 (e.g., based on current sensor data 146, current manufacturing data 154, etc.).

[0056] For illustrative and non-limiting purposes, this application describes the use of one or more trained machine learning models 190 to train historical data (e.g., historical sensor data 144, historical manufacturing data 152) and to input current data (e.g., current sensor data 146, current manufacturing data 154) into one or more trained machine learning models 190 to determine compressed sensing data 166. In other embodiments, heuristic or rule-based models are used to determine compressed sensing data (e.g., in the absence of a trained machine learning model, parameter values ​​controlling the sparsity of the fit may be specified (rather than trained)). Compressed sensing component 114 and prediction component 116 may monitor historical sensor data 144, historical manufacturing data 152, and measurement data 169. Any information described with respect to data input 210 of FIG2 may be monitored or otherwise used in heuristic or rule-based models.

[0057] In some embodiments, the functionality of client device 120, compression sensing server 112, server machine 170, and server machine 180 may be provided by a smaller number of machines. For example, in some embodiments, server machine 170 and server machine 180 may be integrated into a single machine, while in some other embodiments, server machine 170, server machine 180, and compression sensing server 112 may be integrated into a single machine. In some embodiments, client device 120 and compression sensing server 112 may be integrated into a single machine.

[0058] Generally, functions described in one embodiment as being performed by client device 120, compressed sensing server 112, server machine 170, and server machine 180 may also be performed on compressed sensing server 112 in other embodiments, if appropriate. Furthermore, functions belonging to a particular component may be performed by different or multiple components operating together. For example, in some embodiments, compressed sensing server 112 may determine correction actions based on prediction data 168. In another example, client device 120 may determine prediction data 168 based on the output from a trained machine learning model.

[0059] Furthermore, different or multiple components can operate together to perform the function of a specific component. One or more of the compression sensing server 112, server machine 170, or server machine 180 can be accessed as a service provided to other systems or devices via an appropriate application programming interface (API).

[0060] In this embodiment, a "user" may be referred to as a single individual. However, other embodiments of this application cover a "user" as an entity controlled by a plurality of users and / or automation sources. For example, a group of individual users collectively acting as a group of administrators may be considered a "user".

[0061] The embodiments of this application can be applied to data quality assessment, feature enhancement, model evaluation, virtual measurement (VM), predictive maintenance (PdM), and limit optimization, and so on.

[0062] Although embodiments of this application are discussed in relation to generating compressed sensing data 166 and predicted data 168 to perform calibration operations in a manufacturing facility (e.g., a semiconductor manufacturing facility), the embodiments can also be generally applied to improved data processing by utilizing physical information-based virtual measurement and compressed sensing. The embodiments can typically be applied to characterization and monitoring based on different types of data.

[0063] Figure 2 is a block diagram of an exemplary dataset generator 272 (e.g., dataset generator 172 of Figure 1) for creating datasets for machine learning models (e.g., model 190 of Figure 1) according to certain embodiments. The dataset generator 272 may be part of the server machine 170 of Figure 1. In some embodiments, the system 100 of Figure 1 includes multiple machine learning models. In this case, each model may have a separate dataset generator; or, the models may share a dataset generator.

[0064] As shown in Figure 2, a system 200 including a dataset generator 272 (e.g., dataset generator 172 of Figure 1) creates a dataset for a machine learning model (e.g., model 190 of Figure 1). The dataset generator 272 can create the dataset using retrieved data as output from a computer-aided physics-based model. In some embodiments, the dataset generator 272 creates training inputs by selecting a subset of points in the physics-based model (e.g., a subset of points in the model attribute mapping data 162 of Figure 1, and the output from the physics-based model 176 of Figure 1). For example, the output of the physics-based model could be a mapping of temperatures at all locations within a chamber. The dataset generator 272 can form a training set by extracting temperature values ​​from multiple of these locations. The locations used as training inputs correspond to the locations of sensors (e.g., sensor 126 of Figure 1) within the manufacturing chamber. The placement of these sensors may have been selected by the user or may have been output by a machine learning model trained to optimize sensor placement within the manufacturing chamber for compression sensing. In some embodiments, training input 210 is analog data representing data that can be acquired by physical sensors, including analog sensor datasets 262A-262Z. Dataset generator 272 also generates a target output 220 for training a machine learning model. The target output includes a mapping of attribute values ​​230 within the manufacturing chamber, including attribute values ​​remote from sensor locations within the chamber. In some embodiments, the target output 220 includes a mapping of attributes within the chamber adjacent to substrate locations. The mapping of attribute values ​​230 may include values ​​throughout the chamber, values ​​in regions of high interest, and values ​​only adjacent to the substrate, etc. The input training data 210 and the target output data 220 are provided to a machine learning model (e.g., model 190 of FIG. 1).

[0065] The target output 220, represented in various ways, is within the scope of this application. The mapping of attribute value 230 may be a virtual measurement of attribute values ​​located away from the virtual sensor location; or the mapping of attribute value 230 may be coefficients indicating the weights to be applied to the reference set members in the reduced-order model to generate the virtual measurement.

[0066] Referring to FIG2, in some embodiments, dataset generator 272 generates datasets (e.g., training sets, validation sets, and test sets); the datasets include one or more data inputs 210 (e.g., training inputs, validation inputs, and test inputs) and may include one or more target outputs 220 corresponding to the data inputs 210. The datasets may also include mapping data that maps the data inputs 210 to the target outputs 220. The data inputs 210 may also be referred to as "features," "attributes," or "information." In some embodiments, dataset generator 272 may provide the datasets to training engine 182, validation engine 184, or test engine 186 of FIG1, wherein the datasets are used to train, validate, or test the machine learning model 190 of FIG1. ​​Some embodiments for generating training sets may be further described with reference to FIG4A.

[0067] In some embodiments, the data set generator 272 may generate a first data input corresponding to a first set of analog sensor data 262A to train, validate, or test a first machine learning model, and the data set generator 272 may generate a second data input corresponding to a second set of analog sensor data 262B to train, validate, or test a second machine learning model.

[0068] In some embodiments, the data set generator 272 may perform operations on one or more of the data input 210 and the target output 220. The data set generator 272 may extract patterns from data (slope and curvature, etc.), combine data (average and feature output, etc.), or divide analog sensors into multiple groups to train individual models.

[0069] The data input 210 and target output 220 used for training, validating, or testing machine learning models may include information for a specific manufacturing chamber (e.g., a specific semiconductor wafer manufacturing chamber). The data input 210 and target output 220 may also include information for a specific manufacturing chamber design (e.g., all chambers for this design).

[0070] In some embodiments, the information used to train the machine learning model may come from a specific type of manufacturing equipment (e.g., manufacturing equipment 124 of FIG. 1) in a manufacturing facility with specific characteristics, and the information used to train the machine learning model allows the trained machine learning model to determine the outcome of a specific set of manufacturing equipment 124 based on inputs of current sensor data (e.g., current sensor data 146) associated with one or more components sharing a specific set of characteristics. In some embodiments, the information used to train the machine learning model may be available from components of two or more manufacturing facilities, and may allow the trained machine learning model to determine the outcome of a component based on inputs from one manufacturing facility.

[0071] In some embodiments, after generating a dataset and using the dataset to train, validate, or test a machine learning model, the machine learning model can be further trained, validated, tested, or tuned.

[0072] Figure 3 is a block diagram illustrating a system 300 for generating output data (e.g., compressed sensing data 166 of Figure 1) according to certain embodiments. System 300 can be used to virtually measure the property values ​​of adjacent substrates during manufacturing processes using inputs from multiple sensors disposed remotely on the substrate and compression sensing techniques.

[0073] Referring to Figure 3, at box 310, system 300 (e.g., a component of the compressed sensing system 110 of Figure 1) (e.g., via dataset generator 172 of server machine 170 of Figure 1) performs computer-aided data segmentation of physical model data 360 (e.g., extracting analog sensor data 362 from the complete model dataset) to generate training set 302, validation set 304, and test set 306. For example, the training set may be 60% of the analog data, the validation set may be 20% of the analog data, and the test set may be 20% of the analog data.

[0074] At box 312, system 300A (e.g., via training engine 182 of FIG. 1) performs model training using training set 302. System 300 can use multiple sets of features from training set 302 (e.g., a first set of features including a set of simulated sensors in training set 302 and a second set of features including different sets of simulated sensors in training set 302, etc.) to train multiple models. For example, system 300 can train machine learning models to generate a first trained machine learning model using the first set of features from the training set, and generate a second trained machine learning model (e.g., data different from the data used to train the first machine learning model) using the second set of features from the training set. In some embodiments, the first trained machine learning model and the second trained machine learning model can be combined to generate a third trained machine learning model (e.g., the third trained machine learning model may be a better predictor than the first or second trained machine learning model itself). In some embodiments, the feature sets used for comparing models can overlap (e.g., one model can be trained using simulated sensors 1-15, and a second model can be trained using simulated sensors 10-20). In some embodiments, hundreds of models can be generated, including models and combinations of models with various feature arrangements.

[0075] At block 314, system 300 performs model validation using validation set 304 (e.g., via validation engine 184 of FIG. 1). System 300 can use the corresponding feature set of the validation set to validate each trained model 304. For example, validation set 304 can use the same subset of analog sensors used in training set 302 but for different input states. In some embodiments, system 300A can validate hundreds of models generated at block 312 (e.g., models and model combinations with various feature permutations, etc.). At block 314, system 300 can determine the accuracy of each of one or more trained models (e.g., via model validation), and can determine whether one or more trained models have an accuracy that meets a threshold accuracy. In response to the determination that no trained model has an accuracy that meets the threshold accuracy, the flow returns to block 312, where system 300 performs model training using different feature sets of the training set. In response to the determination that one or more trained models have an accuracy that meets the threshold accuracy, the flow continues to block 316. System 300 may (e.g., based on a validation set) discard trained machine learning models whose accuracy is below a threshold accuracy.

[0076] At box 316, system 300 (e.g., via selection engine 185 of FIG1) performs model selection to determine which of one or more trained models that meet a threshold accuracy has the highest accuracy (e.g., selected model 308, based on validation in box 314). In response to the determination that two or more trained models meeting the threshold accuracy have the same accuracy, the process can return to box 312, where system 300 performs model training using a further refined training set corresponding to a further refined feature set to determine the trained model with the highest accuracy.

[0077] At block 318, system 300 performs model testing using test set 306 (e.g., via test engine 186 of FIG. 1) to test the selected model 308. System 300 can use a first set of features in the test set (e.g., simulated sensors 1-15) to test the first trained machine learning model to determine if the first trained machine learning model meets a threshold accuracy (e.g., based on the first set of features in test set 306). In response to the selected model 308's accuracy not meeting the threshold accuracy (e.g., the selected model 308 overfits training set 302 and / or validation set 304, and the selected model 308 is not suitable for other datasets such as test set 306), the process continues to block 312, where system 30A performs model training (e.g., retraining) using different training sets that may correspond to different feature sets or a reorganization of a substrate split into training, validation, and test sets. In response to the determination based on test set 306 that the selected model 308's accuracy meets the threshold accuracy, the process continues to block 320. At least in box 312, the model can learn to simulate patterns in sensor data for prediction, and in box 318, the system 300 can apply the model to the remaining data (e.g., test set 306) to test the predictions.

[0078] In block 320, system 300 uses a trained model (e.g., selected model 308) to receive current sensor data 346 (e.g., current sensor data 146 of FIG1), and system 300 determines (e.g., extracts) actions from the trained model and compressed sensor data 366 (e.g., compressed sensor data 166 of FIG1) to perform actions (e.g., perform correction actions associated with manufacturing equipment 124 of FIG1, provide warnings to client device 120 of FIG1, etc.).

[0079] In some embodiments, retraining of the machine learning model occurs by providing additional data to further train the model. Current sensor data 352 may be provided at block 312. Additional physics-based model data 361 may also be provided. This data may differ from the data initially used to train the model by combining input parameters not part of the original training with input parameters outside the parameter space traversed by the original training, or by updating the data to reflect chamber-specific knowledge (e.g., differences from an ideal chamber due to manufacturing tolerances, aging components, etc.). The selected model 308 may be retrained based on this data.

[0080] In some embodiments, one or more of actions 310 to 320 may occur in various sequences and / or together with other actions not presented and described herein. In some embodiments, one or more of actions 310 to 320 may not be performed. For example, in some embodiments, one or more of the data partitioning in block 310, model validation in block 314, model selection in block 316, or model testing in block 318 may not be performed.

[0081] Figures 4A to 4C are flowcharts of methods 400A to 400C associated with generating compressed sensing data to induce correction actions according to certain embodiments. Methods 400A to 400C may be executed by processing logic; the processing logic may include hardware (e.g., circuitry, dedicated logic, programmable logic, microcode, and processing devices, etc.), software (e.g., instructions running on a processing device, general-purpose computer system, or dedicated machine), firmware, microcode, or a combination thereof. In some embodiments, methods 400A to 400C may be partially executed by the compressed sensing system 110. Method 400A may be partially executed by the compressed sensing system 110 (e.g., server machine 170 and dataset generator 172 of Figure 1 and dataset generator 272 of Figure 2). According to embodiments of this application, the compressed sensing system 110 may use method 400A to generate a dataset to perform at least one of the following steps: training a machine learning model, validating a machine learning model, or testing a machine learning model. Method 400B may be executed by server machine 180 (e.g., training engine 182, etc.). Method 400C may be executed by compression sensing server 112 (e.g., compression sensing component 114). In some embodiments, non-transitory storage media storage instructions, when executed by a processing device (e.g., compression sensing system 110, server machine 180, compression sensing server 112, etc.), cause the processing device to perform one or more of methods 400A to 400C.

[0082] For ease of explanation, methods 400A to 400C are depicted and described as a series of operations. However, the operations according to this application may occur in various sequences and / or simultaneously with other operations not presented and described herein. Furthermore, it is not necessary to perform all the illustrated operations to implement methods 400A to 400C according to the disclosed object. In addition, those skilled in the art will understand and appreciate that methods 400A to 400C may alternatively be represented by state diagrams or events as a series of interrelated states.

[0083] Figure 4A is a flowchart of a method 400A for generating a dataset for a machine learning model according to certain embodiments, the machine learning model being used to generate compressed sensing data (e.g., compressed sensing data 166 of Figure 1).

[0084] Referring to Figure 4A, in some embodiments, at block 401, the processing logic of method 400A initializes the training set T to an empty set.

[0085] At block 402, the processing logic generates a first data input (e.g., a first training input, a first verification input); the first data input may include sensor data (e.g., analog sensor data 262 of FIG2), etc. In some embodiments, the first data input may include a first set of features for a data type, and the second data input may include a second set of features for a data type (e.g., those features as described with respect to FIG3).

[0086] At block 403, the processing logic generates a first target output for one or more data inputs (e.g., a first data input). In some embodiments, the first target output is an attribute value received as the output of a physics-based model. In some embodiments, the target output may be coefficients indicating a weighted additive combination of reference set members in a reduced-order model.

[0087] At block 404, the processing logic optionally generates mapping data indicating the input / output mapping. The input / output mapping (or mapping data) may refer to data inputs (e.g., one or more of the data inputs described herein), the target output of the data inputs, and the relationship between the data inputs(s) and the target output.

[0088] At box 405, in some embodiments, the processing logic adds the mapping data generated at box 404 to the dataset T.

[0089] At block 406, the processing logic branches based on whether the dataset T is sufficient to perform at least one of the following steps: training the machine learning model 190 of Figure 1, validating the machine learning model 190 of Figure 1, and / or testing the machine learning model 190 of Figure 1. If so, execution proceeds to block 407; otherwise, execution returns to block 402 to continue. It should be noted that in some embodiments, the sufficiency of the dataset T may be determined simply based on the number of inputs mapped to the outputs in some embodiments, while in other embodiments, the sufficiency of the dataset T may be determined based on one or more other criteria besides (or, instead of) the number of inputs (e.g., a measure of the diversity of data examples, and accuracy, etc.).

[0090] At block 407, the processing logic provides a dataset T (e.g., to server machine 180 of FIG1) to train, validate, and / or test the machine learning model 190. In some embodiments, dataset T is a training set, and dataset T is provided to training engine 182 of server machine 180 to perform training. In some embodiments, dataset T is a validation set, and dataset T is provided to validation engine 184 of server machine 180 to perform validation. In some embodiments, dataset T is a test set, and dataset T is provided to test engine 186 of server machine 180 to perform testing. In the case of a sparse regression machine learning model, input values ​​of a given input / output mapping (e.g., values ​​associated with data input 210 of FIG2) are input to the sparse regression model, and output values ​​of the input / output mapping (e.g., values ​​associated with target output 220 of FIG2) are stored as the target output. Furthermore, a baseline set representing the reduced-order model is provided to the sparse regression model. The machine learning model learns to sparsely fit the coefficients of the reduced-order model to the input sensor data, meaning many coefficients are zero. This regression is performed by penalizing both the inaccuracy of the fit and the number of non-zero coefficients.

[0091] For illustrative purposes and not for limitation, the mechanism of sparse regression is briefly discussed below. A sparse regression model is used to recover a list of attribute values ​​at various locations within a chamber; this list can be represented as a vector x. The reference set can be represented as a matrix P, and the problem reduces to finding a vector a that satisfies x = Pa. Consider a vector representing the known attribute values ​​in the chamber due to the sensors positioned at those locations; this vector is called y. We can say y = Cx, where C is a sparse matrix of 1s and 0s, indicating which data points of the chamber attribute value x are known. The problem can then be expressed as CPa = y, which is an overdetermined system (sensors y have more non-zero reference set member coefficients a). Sparse regression is a method for finding solutions for such systems.

[0092] On multiple training sets, the machine learning model selects appropriate parameter values ​​to control the number of non-zero coefficients in the reduced-order model. Following block 407, at least one of the following steps may be performed on the machine learning model (e.g., machine learning model 190): training using the training engine 182 of server machine 180, validating using the validation engine 184 of server machine 180, or testing using the test engine 186 of server machine 180. The trained machine learning model may be implemented by the compression sensing component 114 (of the compression sensing server 112) to generate compression sensing data 166, thereby performing signal processing or performing correction actions associated with the manufacturing equipment 124.

[0093] Figure 4B shows a method 400B for training a machine learning model (e.g., model 190 of Figure 1) to determine the compression of sensing data to induce the execution of a correction action.

[0094] Referring to Figure 4B, at block 410 of method 400B, the processing logic receives data from the physics-based model. The data indicates attribute values ​​of a specific arrangement of input settings (processing parameters and hardware settings, etc.) used by the manufacturing chamber in a steady state. The physics-based model data can be represented as coefficients of a baseline set in a reduced-order model. The model data can be retrieved by the processing logic from memory (e.g., data storage 140 in Figure 1) (rather than directly from the physics-based model). In some cases, multiple machine learning models may be part of a single composite machine learning model. In this case, training one component of this composite model may involve receiving the output from another component of the model as training input to the component of the model to be trained.

[0095] At block 412, the processing logic generates a dataset for the analog sensor. This dataset may be a subset of data points from the physical model data received at block 410. Specific data points for the analog sensor can be determined by deciding which data points correspond to locations within the manufacturing chamber where the sensor can be placed. This data is mapped to a physical model dataset, which is a subset of the physical model dataset.

[0096] At box 414, the processing logic provides the simulated sensor dataset as training input to the machine learning model. The processing logic then provides the machine learning model with physically based model data as the target output. The machine learning model is then trained to provide virtual sensing of attribute values ​​in the manufacturing chamber by performing sparse regression on the sensor data, thereby constructing a mapping of attribute values ​​in the chamber (especially attribute values ​​of adjacent substrates).

[0097] At box 416, the machine learning model can be retrained using additional data from the physics-based model. In one embodiment, the training of the machine learning model is continuously updated to interpret or predict drift in manufacturing equipment, sensors, metering devices, etc., to predict equipment failures, thereby reflecting changes to procedures or recipes, etc. The processing logic provides additional data from the physics-based model, different from the first data, as additional training input data. The processing logic provides simulated sensor data taken from the physics-based model data as the target output data.

[0098] Figure 4C illustrates a method 400C for determining compressed sensing data using a trained machine learning model (e.g., model 190 of Figure 1).

[0099] Referring to Figure 4C, at block 420 of method 400C, the processing logic provides current sensor data associated with the substrate manufacturing process to a trained machine learning model. The sensor data type may correspond to the type provided in block 412 of Figure 400B for training the machine learning model.

[0100] At block 422, the processing logic obtains one or more outputs indicating compression sensing data from a trained machine learning model. In some embodiments, the compression sensing data may be temperature data, which includes data representing the location of a neighboring substrate within the manufacturing chamber.

[0101] At block 424, the processing logic causes the execution of a correction action. In some embodiments, a correction action may be taken based on the output of a machine learning model after the output data has been further processed, for example, after the coefficients of the reduced-order model have been applied as weights to an additive combination of reference set members to recover the attribute values ​​in the chamber. The types of correction actions consistent with this application can vary widely. In some embodiments, the execution of a correction action may include one or more of the following correction actions, including: providing an alert to the user; interrupting the function of the manufacturing equipment; updating manufacturing parameters, including processing parameters and hardware parameters; planning to replace components of the manufacturing equipment; putting one or more components into a sleep or idle mode at specific times during product manufacturing to reduce energy consumption; replacing one or more components to reduce energy consumption; causing preventative maintenance; causing modifications to components (e.g., tightening mounting screws and replacing bindings, etc.); correcting sensor drift of sensors associated with the manufacturing equipment; correcting chamber drift and updating processing recipes, etc. Predictive data and / or corrective actions can indicate combinations that lead to an anomaly (e.g., a single item in a combination may not cause an anomaly on its own), such as combinations of components and combinations of manufacturing parameters.

[0102] Figure 5 depicts a cross-sectional view of a manufacturing chamber 500 (e.g., a semiconductor wafer manufacturing chamber) according to some embodiments of the present application. The manufacturing chamber 500 may be one or more of an etching chamber, a deposition chamber (including atomic layer deposition, chemical vapor deposition, physical vapor deposition, or plasma-enhanced forms thereof), an annealing chamber, etc. For example, the manufacturing chamber 500 may be a chamber for a plasma etcher and a plasma cleaner, etc. Examples of chamber components may include a substrate support assembly 504, an electrostatic chuck, a ring (e.g., a processing kit ring), chamber walls, a base, a nozzle 506, a gas distribution plate, a liner, a liner kit, a shroud, a plasma screen, a flow equalizer, a cooling base, a chamber viewing port, a chamber cover, and a nozzle, etc.

[0103] In one embodiment, manufacturing chamber 500 includes a chamber body 508 surrounding an internal volume 510 and a nozzle 506. In some chambers, the nozzle 506 may be replaced by a cap and a nozzle. The chamber body 508 may be made of aluminum, stainless steel, or other suitable materials. The chamber body 508 typically includes sidewalls 512 and a bottom 514.

[0104] An exhaust port 516 may be defined within the chamber body 508, and the exhaust port 516 may connect the internal volume 510 to a pump system 518. The pump system 518 may include one or more pumps and valves for evacuating and regulating the pressure of the internal volume 510 of the manufacturing chamber 500.

[0105] Nozzle 506 may be supported on the sidewall 512 of chamber body 508 or on top of chamber body. Nozzle 506 (or a cover in some embodiments) may be opened to allow access to the internal volume 510 of manufacturing chamber 500 and may provide a seal for manufacturing chamber 500 when nozzle 506 (or a cover in some embodiments) is closed. Gas panel 520 may be coupled to manufacturing chamber 500 to provide processing or cleaning gas to internal volume 510 via nozzle 506 (or cover and nozzle). Nozzle 506 may include a plurality of through gas delivery orifices. Examples of processing gases that may be used to process substrates in manufacturing chamber 500 include halogen-containing gases such as C2F6, SF6, SiCl4, HBr, NF3, CF4, CHF3, F2, Cl2, CCl4, BCl3, and SiF4, and other gases such as O2 or N2O. Examples of carrier gases include N2, He, Ar, and other gases that are inert to the processing gas (e.g., non-reactive gases).

[0106] The substrate support assembly 504 is disposed within the internal volume 510 of the manufacturing chamber 500 below the nozzle 506. In some embodiments, the substrate support assembly 504 includes a base 522 and a shaft 524. The substrate support assembly 504 supports the substrate during processing. In some embodiments, one or more heaters 526 and reflectors 528 are also arranged within the manufacturing chamber 500.

[0107] Understanding the property values ​​within the manufacturing chamber adjacent to the substrate is useful for controlling the processing results. In many cases, monitoring the state of the substrate location is impractical. In other cases, the condition of certain portions of the substrate can be monitored, but not the entire area. The method and system of this application provide a way to infer property values ​​at a substrate location without needing to place sensors at the substrate location.

[0108] In some embodiments, virtual measurements of the property values ​​of adjacent substrates are obtained by performing compression sensing (e.g., using the compression sensing system 110 of FIG1). When it is inconvenient to place sensors adjacent to substrate 502 in the manufacturing chamber 500 to monitor the state, there may be many other areas in the manufacturing chamber where sensors can be more easily placed. Such locations may include base 522, shaft 524, reflector 528, vent 516, and chamber body 508, etc.

[0109] With the aid of computer-aided physics-based models (e.g., physics-based model 176 in Figure 1), the measurement of attribute values ​​in these accessible locations allows for virtual measurement of the state at the substrate location. To utilize the physics-based model, a set of input parameters is defined, such as heater power, gas composition, gas flow rate, chamber pressure, and hardware parameters, etc. These different input states define a parameter space region, which can be understood as the entire landscape of possible input parameters in the manufacturing process. A set of input parameters can be selected to be tightly grouped in the parameter space around states useful for a specific process and a specific wafer; a set of input parameters can be selected to substantially span the entire region of the parameter space considered potentially useful for the substrate process; or another index can be used to select a set of input parameters.

[0110] The representation of the manufacturing chamber, along with processing parameters representing a set of possible states, is received by a physics-based model. The model then uses computational techniques to solve equations describing fluid flow, thermodynamics, heat transfer, etc., to recover a mapping of property values ​​within the manufacturing chamber. This mapping can be a representation of steady-state property values. This process is repeated for selected combinations of processing parameters. The resulting set of property value mappings spans a region of property value space within the manufacturing chamber, associated with a parameter space region defined by the selected input parameters.

[0111] In some embodiments, the information output by the physics-based model can be represented in a way that is more convenient for virtual measurements. A reference set consisting of members is generated, which, when combined in a weighted summation, can substantially recreate any portion of the attribute value space spanned by the attribute value mapping of the physics-based model output. This reference set is generated as a reduced-order model (e.g., by the reduced-order model generator 174 of Figure 1). The reduced-order model is constructed such that a relatively small number of reference set members can reconstruct most of the physics in the chamber, that is, reconstruct the attribute values ​​at points in the chamber that are substantially equal to those output by the physics-based model. The reference set can be constructed using an appropriate orthogonal decomposition algorithm. The specific range of "substantially equal" may vary depending on the application, but may be related to experimental errors, measurement errors, acceptable manufacturing tolerances, etc.

[0112] In some embodiments, the selection of points output by the physics-based model is of interest. For example, limiting virtual measurements to the boundaries between components in a manufacturing chamber 500 (e.g., ignoring attribute values ​​within the components) allows for the use of fewer sensors than those used to virtually measure attribute values ​​at all points within the chamber to make useful attribute value predictions. The region of interest could be a location where the attribute value variation in the output of a physics-based model with different processing parameter inputs exceeds a certain threshold.

[0113] Once data is retrieved from the physics-based model, the sensor is placed inside the chamber. Possible locations for sensor placement can be constrained by geometry and engineering issues, among other things. Information can be extracted from locations where attribute values ​​change with input values ​​to the physics-based model or input parameters of the manufacturing process. Information can also be extracted if the attribute values ​​at the sensor's location are irrelevant. The sensor location is selected taking these and any other constraints into account. In some embodiments, machine learning tools can be used to optimize the sensor location.

[0114] Unbound by theory, the number of sensors used to reconstruct the attribute value mapping in the manufacturing chamber can be estimated as follows: , where ns is the number of sensors, k is the number of reference set members used to reconstruct the attribute value map, and N is related to the number of locations in the chamber where the predicted attribute values ​​are to be obtained. In some implementations, the chamber region to be mapped and the number of reference set members used are adjusted to allow for compressed sensing using a feasible or convenient number of sensors.

[0115] During the manufacturing process, the state within the manufacturing chamber 500 is monitored by sensors. Data indicating measurements taken by the sensors is fed as input to a trained machine learning model (e.g., model 190 in Figure 1). The trained machine learning model is trained to reconstruct attribute values ​​in chamber regions not measured by the sensors (e.g., the location of the substrate). The trained machine learning model utilizes a sparse regression algorithm to determine the weights of multiple members of a reference set generated by a reduced-order model generator 174. In this way, the state of adjacent substrates can be virtually measured without physically placing sensors at the substrate locations. Physically based virtual measurement and compressed sensing offer advantages over other techniques. Differences between chambers due to component variations, drift, manufacturing tolerances, etc., are explained using sensor data projected into the attribute value space. Furthermore, these methods also allow for understanding of states beyond the space spanned by the output of the physical model.

[0116] In some embodiments, compression sensing data is then used to perform a correction action. The correction action may include sending an alarm to a client device. The correction action may include adjusting the processing recipe by adjusting at least one processing parameter. The correction action may be facilitated by providing the compression sensing data to another model. The second model may be used to monitor component health or drift and chamber aging, etc. The second model may take corrective actions based on the compression sensing data (e.g., recommending maintenance and recommending component replacement, etc.).

[0117] Figure 6 is a block diagram illustrating a computer system 600 according to certain embodiments. In some embodiments, the computer system 600 may be connected to other computer systems (e.g., via a network such as a local area network (LAN), intranet, extranet, or internet). The computer system 600 may operate as a server or client computer in a client-server environment, or as a peer computer in a peer-to-peer or distributed network environment. The computer system 600 may be provided by a personal computer (PC), tablet PC, set-top box (STB), personal digital assistant (PDA), cellular phone, network device, server, network router, switch, or bridge, or any means capable of executing (sequentially or otherwise) a set of instructions (which specify the actions to be taken by the device). Furthermore, the term "computer" should include any collection of computers that individually or collectively execute a set (or more sets) of instructions to perform any one or more of the methods described herein.

[0118] In another embodiment, the computer system 600 may include a processing device 602 that can communicate with each other via a bus 608, volatile memory 604 (e.g., random access memory (RAM)), non-volatile memory 606 (e.g., read-only memory (ROM) or electronically erasable programmable ROM (EEPROM)) and data storage device 618.

[0119] Processing device 602 may be provided by one or more processors, such as general-purpose processors (e.g., complex instruction set computing (CISC) microprocessors, reduced instruction set computing (RISC) microprocessors, very long instruction field (VLIW) microprocessors, microprocessors that implement other types of instruction sets or microprocessors that implement combinations of instruction set types) or special-purpose processors (e.g., application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), or network processors).

[0120] Computer system 600 may further include (e.g., coupled to network 674) a network interface device 622. Computer system 600 may also include a video display unit 610 (e.g., LCD), an alphanumeric input device 612 (e.g., keyboard), a cursor control device 614 (e.g., mouse), and a signal generation device 620.

[0121] In some embodiments, data storage device 618 may include a non-transitory computer-readable storage medium 624 (e.g., a non-transitory machine-readable storage medium) on which instructions 626 encoded for any or more of the methods or functions described herein may be stored; the non-transitory computer-readable storage medium 624 includes the instruction encoding components of FIG1 (e.g., compression sensing components 114 and model 190, etc.) and is used to implement the methods described herein.

[0122] Instruction 626 may also reside wholly or partially within volatile memory 604 and / or processing device 602 during execution by computer system 600; therefore, volatile memory 604 and processing device 602 may also constitute machine-readable storage media.

[0123] Although computer-readable storage media 624 is shown as a single medium in the illustrative example, the term "computer-readable storage media" should include single or multiple media (e.g., centralized or distributed databases, and / or associated caches and servers) that store one or more sets of executable instructions. The term "computer-readable storage media" should also include any tangible medium capable of storing or encoding a set of instructions for a computer to execute, which causes the computer to perform any one or more of the methods described herein. The term "computer-readable storage media" should include, but is not limited to, solid-state memory, optical media, and magnetic media.

[0124] The methods, components, and features described herein can be implemented by separate hardware components, or they can be integrated into the functionality of other hardware components such as ASICs, FPGAs, DSPs, or similar devices. Alternatively, the methods, components, and features can be implemented by firmware modules or functional circuitry within the hardware device. Furthermore, the methods, components, and features can be implemented in any combination of hardware devices and computer program components, or within a computer program.

[0125] Unless otherwise expressly stated, terms such as “receive,” “execute,” “provide,” “obtain,” “cause,” “access,” “determine,” “add,” “use,” and “train” refer to operations and processes performed or implemented by a computer system that manipulate data represented as physical (electronic) quantities in the computer system’s registers and memory, and convert this data into other data similarly represented as physical quantities in the computer system’s memory or registers (or other such information storage, transmission, or display devices). Furthermore, the terms “first,” “second,” “third,” “fourth,” etc., as used herein are intended as labels to distinguish different elements, and such terms may not have an ordering meaning depending on their numerical designation.

[0126] The examples described herein also relate to devices for performing the methods described herein. Such devices may be specifically configured to perform the methods described herein, or the devices may comprise a general-purpose computer system selectively programmed by a computer program stored in a computer system. Such computer program may be stored in a computer-readable physical storage medium.

[0127] The methods and illustrative examples described herein are not inherently related to any particular computer or other device. Various general-purpose systems can be used based on the teachings described herein, or it may prove convenient to construct more specialized devices to perform each of the methods described herein and / or each of the respective functions, norms, sub-norms, or operations in such methods. Various structural examples of these systems are illustrated above.

[0128] The above description is intended to be illustrative and not restrictive. Although this application has been described with reference to specific illustrative examples and embodiments, it will be appreciated that this application is not limited to the described examples and embodiments. The scope of this application should be determined by referring to the full scope of equivalents enjoyed by the appended claims and claims.

[0129] 100: System 110: Compression Sensing System 112: Compression Sensing Server 114: Compression Sensing Component 116: Prediction Component 120: Client device 122: Correction Motion Component 124: Manufacturing equipment 126: Sensor 128: Measuring equipment 130: Internet 140: Data Storage 142: Sensing Data 144: Historical Sensing Data 146: Current Sensing Data 150: Manufacturing materials 152: Historical Manufacturing Data 154: Current Manufacturing Data 160: Physics-based model data 162: Model Attribute Mapping Data 164: Reduced-order baseline data 166: Compressed Sensing Data 168: Forecast Data 169: Measurement data 170: Server Machine 172: Data Set Generator 174: Reduced-order model generator 176: Physics-Based Models 180: Server Machine 182: Training Engine 184: Verification Engine 185: Select Engine 186: Test Engine 190: Model 200: System 210: Data Input 220: Target Output 230: Attribute Value 262A~262Z: Analog Sensor Data Set 272: Data Set Generator 300: System 302: Training set 304: Validation set 306: Test Set 308: Selected Model 310~320: Box 352: Current sensor data 360: Physics-based model data 361: Additional Physics-Based Model Data 362: Analog Sensor Data 366: Compressed Sensing Data 400A:Method 401~407: Boxes 400B: Method 410~416: Boxes 400C: Method 420~424: Box 500: Manufacturing Chamber 502:Substrate 504: Support components 506: Sprayer Head 508: Main chamber 510: Internal volume 512: Sidewall 514: Bottom 516: Exhaust port 518: Pump System 520: Gas Panel 522: Base 524: Axis 526: Heater 528: Reflector 600: Computer System 602: Processing unit / processor 604: Volatile memory / Main memory 606: Non-volatile memory / static memory 608: Busbar 610: Image Display 612: Alphanumeric Input Device 614: Vernier control device 618: Data storage device 620: Signal generating device 622: Network Interface Device 624: Computer-readable storage media 626: Instruction 674: Network

[0130] Domestic storage information (please note in order of storage institution, date, and number) none Overseas storage information (please note in the order of storage country, institution, date, and number) none

Claims

1. A method for measuring properties of an adjacent substrate, the method comprising the steps of: during a process performed on a substrate in a manufacturing chamber, measuring a subset of property values ​​in the manufacturing chamber using a plurality of sensors arranged in the manufacturing chamber; and determining a mapping of property values ​​in the manufacturing chamber adjacent to the substrate based on inputting the measured values ​​of the subset of property values ​​in the manufacturing chamber into a trained machine learning model; and performing an action according to the mapping of property values, wherein performing the action comprises the steps of: taking a correction action or sending an alarm to a client device, wherein the correction action comprises the steps of: adjusting at least one processing parameter.

2. The method of claim 1, wherein the trained machine learning model is configured to represent attribute values ​​in the manufacturing chamber according to a reduced-order model, wherein the mapping of the attribute values ​​is determined as a weighted summation of a plurality of members of the reduced-order model.

3. The method as described in claim 2, wherein the trained machine learning model includes a sparse regression model, wherein a sparse regression algorithm using the regression model is used to determine the mapping of attribute values.

4. The method of claim 2, wherein the members of the mapping selected by the trained machine learning model for determining attribute values ​​are a subset of a reference set, the reference set being constructed such that a weighted summation of the reference set describes the attribute values ​​of interest in the manufacturing chamber.

5. The method described in claim 4, wherein the reference set has been determined based on the results of a computer-aided physics-based model that predicts attribute values ​​for a plurality of different processing parameter configurations throughout the manufacturing chamber.

6. The method of claim 1, wherein the mapping of the attribute values ​​includes temperatures at a plurality of locations on the substrate.

7. The method of claim 6, wherein the step of measuring the subset of attribute values ​​in the manufacturing chamber includes the step of: performing temperature measurements at a plurality of points in the manufacturing chamber, at least a majority of which are not on the substrate.

8. A method for measuring properties adjacent to a substrate, the method comprising the steps of: generating training data for a machine learning model, wherein the step of generating the training data comprises the steps of: identifying a first training input, the first training input including data indicating first attribute values ​​associated with a finite subset of attribute values ​​in a manufacturing chamber; and identifying a first target output for the first training input, wherein the first target output includes a mapping of attribute values ​​in a plurality of locations within the manufacturing chamber, the plurality of locations including more locations than the subset of attribute values ​​used as the first training input; and providing the training data to train the machine learning model on (i) a set of training inputs including the first training input and (ii) a set of target outputs including the first target output, wherein the trained machine learning model will receive a new input including data indicating second attribute values ​​associated with a finite subset of attribute values ​​in the manufacturing chamber, the second attribute values ​​being different from the first attribute values, and generating a new output based on the new input, the new output including a mapping of attribute values ​​adjacent to a substrate in the manufacturing chamber.

9. The method as described in claim 8, wherein each training input in the set of training inputs is mapped to a target output in the set of target outputs.

10. The method of claim 8, wherein the training inputs include data indicating a plurality of measurements of attribute values ​​taken in the chamber at a location not adjacent to the substrate, and the target output includes a prediction of the attribute value at that location on the substrate.

11. The method of claim 8, wherein the target output includes data indicating the results of a computer-aided, physics-based model describing attribute values ​​for a plurality of processing parameters in the manufacturing chamber.

12. The method of claim 11, wherein a subset of the results from the computer-aided physical model is used, the subset corresponding to a location in the chamber where the change in a value of an attribute of interest is above a certain threshold when compared with different manufacturing parameters under steady-state conditions.

13. The method of claim 8, wherein the output of the machine learning model includes weights to be applied to a benchmark set, thereby forming an additive combination in a reduced-order model.

14. The method as described in claim 8, wherein the machine learning model comprises a sparse regression model.

15. The method of claim 8, wherein the subset of attribute values ​​includes measurements of the state at a finite number of locations within the chamber.

16. The method of claim 15, wherein the finite number of locations are selected by inputting the results of a computer-aided physics-based model into a trained machine learning model, the output of which includes optimized locations for measurement.

17. A non-transitory machine-readable storage medium storing instructions, which, when executed by a processing device, cause the processing device to perform operations including: receiving data indicating measurements of a subset of attribute values ​​in a manufacturing chamber containing a substrate; determining a mapping of the attribute values ​​in the manufacturing chamber using a machine learning model, the mapping including the attribute values ​​adjacent to the substrate; performing at least one of the following: (i) adjusting manufacturing parameters associated with the manufacturing chamber based on the mapping of attribute values, or (ii) sending an alarm to a user device; and generating an output describing the mapping of attribute values ​​in the manufacturing chamber as coefficients for a weighted summation combination of multiple members of a reference set, the reference set being constructed based on the results of a computer-aided physical model of the chamber evaluated under a plurality of processing parameters.

18. The non-transitory machine-readable storage medium as claimed in claim 17, wherein the measurements of a subset of attribute values ​​in the manufacturing chamber include temperature measurements taken at a plurality of points in the manufacturing chamber, at least a majority of which are not on the substrate.

19. The non-transitory machine-readable storage medium as described in claim 17, wherein the mapping of attribute values ​​includes temperatures at a plurality of points in the manufacturing chamber, including at least points on the substrate.

Images