Estimation of process control knob

Abstract

The subject matter herein may be embodied in, among other things, a method, a system, and a computer-readable storage medium. The method may include receiving (i) sensor data indicative of a first state of an environment of a processing chamber in which a substrate is being processed after a chamber recovery procedure, and (ii) substrate process data indicative of a set of process parameter values ​​associated with performing a substrate processing procedure with a processing chamber having an environment in a second state prior to the chamber recovery procedure. The method further includes processing the sensor data and the substrate process data using one or more machine learning models to determine one or more outputs. The one or more outputs include an update to at least one process parameter value of the set of process parameter values. The update is associated with performing a substrate processing procedure with a processing chamber having an environment in the first state.

Description

TECHNICAL FIELD

[0001] The present specification generally relates to estimation of chamber control knobs after preventive maintenance (PM). More specifically, the present specification relates to estimating chamber control knob settings associated with chamber conditions after performing a chamber recovery procedure such as chamber seasoning. BACKGROUND ART

[0002] Substrate processing may include a series of processes for forming electrical circuits on a semiconductor, such as a silicon wafer, in accordance with circuit designs. These processes may be performed in a series of chambers. Successful operation of modern semiconductor manufacturing facilities may aim to facilitate a stable flow of wafers that are moved from one chamber to another while forming electrical circuits on the wafers. In the course of performing many substrate processes, the conditions of the processing chamber may be altered, and as a result, the processed substrates may not meet desired conditions and outcomes.

[0003] One such substrate process may include plasma etching. Plasma etching is a process that transfers the pattern of a mask material layer to another layer, such as a conductive material layer or a dielectric material layer, under the mask, which is achieved by removing these laminated materials from the surface of the wafer. Such processes inevitably generate different types of etching by-products, such as silicon oxide and organic polymers, depending on the laminated materials and etching chemistry. Some of the by-products deposit on the inner surfaces of the chamber in which the plasma etching process is performed. By-product deposition may affect etching performance, such as by depositing particles (e.g., flakes) on the substrate, or reacting with the plasma to affect process results.

[0004] To mitigate the effects of etching by-products, preventative maintenance such as chamber cleaning is sometimes used to periodically remove deposits from the chamber walls. In an example of chamber cleaning, the chamber is removed from production, and a cleaning plasma, such as CF4+O2 plasma, is introduced to clean the silicon oxide deposited during silicon etching. This plasma reacts with the deposits, and the products of this reaction are pumped out of the chamber. However, it has been observed that after such chamber cleaning, the clean chamber walls make the chamber unsuitable for immediate production wafer etching. Chamber seasoning is a procedure in which a series of substrates (e.g., blank silicon wafers) are etched to restore chamber conditions suitable for the production substrate process. After chamber seasoning, a thin layer of silicon oxide covers the chamber walls. The chamber is then returned to production wafer etching until the next chamber cleaning and seasoning are required. [Overview of the Initiative]

[0005] Methods and systems for estimating a control knob after preventive maintenance are described. In some embodiments, the method includes a processing device receiving sensor data indicating a first state of the environment of a processing chamber processing a substrate after a chamber recovery procedure performed on the processing chamber. In some embodiments, the method further includes receiving substrate process data indicating a set of process parameter values ​​related to performing a substrate processing procedure by a processing chamber having a second state environment prior to the chamber recovery procedure. The method further includes processing the sensor data and the substrate process data using one or more machine learning models (MLMs) to determine one or more outputs. These one or more outputs include a first update to at least one process parameter value of a set of process parameter values. This first update may be related to performing a substrate processing procedure by a processing chamber having a first state environment. The method further includes preparing a notification indicating the first update for presentation on a graphical user interface (GUI). The method further includes causing the processing chamber to perform a selection of a substrate processing procedure based on the first update.

[0006] In some embodiments, a method for training a machine learning model to predict updates to a set of process parameter values ​​includes a processing unit receiving training data. This training data includes first sensor data indicating a first state of the environment of a first processing chamber processing a substrate after a chamber recovery procedure performed on the first processing chamber. This training data may further include first substrate process data indicating a first set of process parameter values ​​associated with performing a first substrate processing procedure by the first processing chamber having a second state environment prior to the chamber recovery procedure. This training data may further include second substrate process data indicating a first update to at least one of the first set of process parameter values. This first update may be associated with performing a first substrate processing procedure by the first processing chamber having a first state environment. The method may further include a processing unit training one or more machine learning models (MLMs) using input data and target outputs. The input data includes first sensor data and first substrate process data. The target outputs may include second substrate process data. One or more trained MLMs may receive a new input having new sensor data. This new sensor data may indicate a third state of the environment of a new processing chamber processing a new substrate after a chamber recovery procedure. This new sensor data may further include a third substrate process data indicating a second set of process parameter values ​​related to performing a first substrate processing procedure by a new processing chamber having a fourth state environment prior to the chamber recovery procedure, in order to generate a new output based on the new input. This new output indicates a second update to at least one of the second set of process parameter values. The second update is related to performing the first substrate processing procedure by a new processing chamber having a third state environment.

[0007] In some embodiments, the method described above may be stored on a non-transient machine-readable storage medium containing instructions, which, when executed by a processing unit, cause the processing unit to perform an operation, which may include one or more steps described in the method described above.

[0008] The aspects and embodiments of this disclosure will be better understood from the detailed description and accompanying drawings set forth below. This detailed description and accompanying drawings are intended to illustrate aspects and embodiments as examples and are not intended to limit them. [Brief explanation of the drawing]

[0009] [Figure 1] This block diagram shows an exemplary system architecture in which embodiments of the present disclosure may function. [Figure 2A] This block diagram shows a process for performing control knob tuning, in which embodiments of the present disclosure may function. [Figure 2B] This block diagram shows a process for performing control knob tuning, in which embodiments of the present disclosure may function. [Figure 3] This block diagram shows a chamber monitoring and maintenance system in which embodiments of the present disclosure may function. [Figure 4A] This block diagram shows an exemplary modeling architecture for chamber control knob tuning according to several embodiments. [Figure 4B] This block diagram shows an exemplary data preparation architecture in several embodiments. [Figure 5] This is a block diagram of a chamber knob prediction system using stacked modeling according to an aspect of the present disclosure. [Figure 6] This is a flowchart of a method for updating a chamber control knob after a chamber preventive maintenance event, according to some embodiments of the present disclosure. [Figure 7]This figure illustrates the training stages of a machine learning system according to some embodiments of the present disclosure. [Figure 8] This figure shows a model training workflow and a model application workflow for chamber knob prediction and simulation according to an aspect of this disclosure. [Figure 9] This is a block diagram of an exemplary computing device operating according to one or more aspects of the present disclosure. [Modes for carrying out the invention]

[0010] Substrate processing may involve a series of processes for forming electrical circuits on a semiconductor, such as a silicon wafer, according to a circuit design. These processes may be carried out in a series of chambers. Successful operation of modern semiconductor manufacturing equipment may aim to facilitate a stable flow of wafers as they are moved from one chamber to another while electrical circuits are being formed on the wafers. During the execution of many substrate procedures, the conditions of the processing chambers may change, and as a result, the processed substrate may not meet the desired conditions or process results (e.g., limit dimensions, process uniformity, thickness dimensions, etc.).

[0011] Preventive maintenance procedures (e.g., chamber cleaning) are often used as part of the chamber recovery process to return the processing chamber to a state suitable for entering a substrate processing production mode (e.g., mass processing of substrates). Recovery procedures are often used after preventive maintenance procedures to prepare the chamber for production mode (e.g., "warm up" the chamber). A common recovery procedure conventionally used for chamber recovery includes processing chamber seasoning. Chamber seasoning is a procedure that involves processing a series of substrates (e.g., blank silicon wafers) to restore chamber conditions suitable for the production substrate process (e.g., substrates processed in the chamber have process results that meet desired threshold criteria) (e.g., coating the chamber walls). After chamber seasoning, the chamber may operate in production mode for a period of time until the next preventive maintenance and additional chamber seasoning are required, or until it is recommended to restore the processing chamber to its original state for other reasons.

[0012] As described above, the chamber seasoning procedure includes processing a series of substrates to prepare the chamber for performing a substrate processing procedure. To give a substrate having a target process result (e.g., a process result that satisfies threshold criteria associated with one or more process procedures or recipes), the substrate processing procedure may include a set of chamber conditions (e.g., temperature, pressure, vacuum conditions, plasma conditions, etc.) that facilitate the substrate processing (e.g., after one or more process procedures or recipes).

[0013] Conventionally, after a preventive maintenance procedure, a series of substrates (e.g., seasoning wafers) are processed in a processing chamber (e.g., as part of a chamber recovery procedure). After processing this series of substrates, the processing chamber control knob is tuned. As used herein, the processing chamber control knob may include hardware and / or software elements for controlling one or more processing chamber parameters related to processing one or more substrates in the processing chamber according to a process procedure (e.g., a process recipe). These processing chamber parameters may include, among other things, process parameter values ​​such as process temperature parameter values, process pressure parameter values, process vacuum parameter values, process plasma parameter values, and process time (e.g., process duration) parameter values. Conventional control knob tuning often involves changing one or more processing chamber parameters and processing a test sample. The process results of the test sample (e.g., limit dimensions, process uniformity, thickness dimensions, etc.) are evaluated (e.g., using measuring tools) and compared to target process results.

[0014] For example, a test sample may include a substrate that is processed in a processing chamber and evaluated (e.g., the process result is measured) to determine whether the test sample contains a process result that satisfies a threshold condition. If the condition is not met, additional tuning of the processing chamber control knob may be performed, and another test sample may be processed and evaluated to determine whether that other test sample contains a process result that satisfies the threshold condition. This process may be repeated until the chamber control knob is tuned (e.g., until processing a substrate (e.g., a test sample) using the tuned control knob results in a process result that satisfies the target condition). Conventional methods for handling chamber recovery in preparation for production, such as those described above, may require many iterations of control knob tuning, including processing and evaluating multiple test samples. Furthermore, conventional methods may introduce delays when returning the chamber to production mode by waiting for test sample measurements and evaluations, which are often necessary to determine whether the control is effectively tuned. The time required for recovery procedures and chamber control knob tuning can result in productivity losses because the tool or chamber is down (e.g., not operating in production mode). Furthermore, conventional methods do not provide in-chamber condition monitoring and rely on the processing and evaluation of test samples.

[0015] The aspects and embodiments of this disclosure address these and other shortcomings of existing technologies by providing methods and systems that can estimate chamber control knob settings without relying on the results of testing a test sample in various embodiments. In some embodiments, machine learning models may be used to determine the chamber control knob settings during and / or after the seasoning procedure. One or more of these aspects and / or embodiments may result in reduced downtime of the processing tool (e.g., processing chamber) and enable greater productivity (e.g., shorter downtime) of the manufacturing equipment.

[0016] In exemplary embodiments, a method and system for tuning control knobs after preventive maintenance are performed. In some embodiments, the method performed by the processing apparatus may include receiving sensor data indicating a first state of the environment of a processing chamber. In some embodiments, the processing chamber is processing a substrate after a chamber recovery procedure has been performed on the processing chamber. The first state of the environment corresponds to the state of the processing chamber after a chamber seasoning operation. The method may further include receiving substrate process data. In some embodiments, this substrate process data may indicate a set of process parameter values ​​associated with the processing chamber performing the substrate processing procedure. In some embodiments, this set of process parameter values ​​may include one or more chamber control knob settings associated with the post-preventive maintenance operation of the processing chamber. A second state of the environment may relate to the state of the processing chamber that occurs before the execution of the chamber recovery procedure. In some embodiments, the second state of the environment is the state immediately after a preventive maintenance operation has been performed on the processing chamber. In some embodiments, the second state of the environment is the state that occurs before any seasoning of the processing chamber.

[0017] In some embodiments, the method further includes processing sensor data and substrate process data using one or more machine learning models (MLMs). The method may use these one or more MLMs to determine one or more outputs. In some embodiments, these one or more outputs may include a first update to at least one process parameter value from a set of process parameter values. In some embodiments, this first update is related to a processing chamber performing a substrate processing procedure. The first update may include one or more adjustments to the chamber control knob settings (e.g., new chamber control values, correction amounts, etc.) related to post-preventive maintenance operations of the processing chamber (e.g., when the processing chamber is in a state that occurs after a chamber recovery procedure). In some embodiments, adjusting the chamber control knob settings according to the first update places the processing chamber in a condition suitable for production.

[0018] In some embodiments, one or more trained machine learning models receive a new input. This new input may include new sensor data indicating a third state of the environment of a new process chamber processing a new substrate after a chamber recovery procedure. This new input may include third substrate process data indicating a second set of process parameter values ​​(e.g., chamber control knob settings) associated with performing a first substrate processing procedure by a new processing chamber having a fourth state environment. The fourth state may be a state that occurs before the execution of a chamber recovery procedure (e.g., chamber seasoning). The one or more trained machine learning models may generate a new output based on the new input. This new output may indicate a second update (e.g., a second chamber control knob update) to at least one of the second set of process parameter values. The second update may be associated with a new processing chamber (e.g., in a third state) performing the first substrate processing procedure after a chamber recovery procedure.

[0019] In some embodiments, the method further comprises preparing a notification for presentation on a graphical user interface (GUI), wherein the notification is configured to instruct a first update. In some embodiments, the method further comprises causing a process chamber to perform a selection of a substrate processing procedure based on the first update. This selection of the substrate processing procedure may include updates to one or more process parameter values (e.g., process control knobs).

[0020] In some embodiments, the method further comprises determining a first endpoint detection condition. The first endpoint detection condition may correspond to the first update. For example, the first endpoint detection condition may relate to tuning of a process chamber control knob after a chamber recovery procedure. In certain embodiments, the first endpoint detection condition is an endpoint detection setpoint. In some embodiments, the first endpoint detection condition corresponds to an optical sensor measurement value of an environment of the process chamber. In some embodiments, the optical sensor measurement value is provided by an optical emission spectroscopy (OES) sensor and / or a reflectometry spectroscopy sensor. In some embodiments, the OES sensor measures an optical emission spectrum of a plasma introduced into the process chamber. The reflectometry spectroscopy sensor measures one or more spectra associated with light reflected at a surface of a substrate being processed. The optical emission spectrum of the plasma may indicate a condition of the environment within the process chamber.

[0021] In some embodiments, tuning of one or more process knobs may be completed based on process results of one or more process substrates that have process results meeting a target condition (e.g., within a tolerance of advanced process control (APC)). When the setting of the control knob places the process chamber in a condition suitable for production, additional tuning of the control knob may not be required. In some embodiments, the method further comprises causing the process chamber to perform the selection of the substrate processing procedure using the first endpoint detection condition.

[0022] In some embodiments, determining a first endpoint detection condition involves applying a deterministic model to a first update to generate the first endpoint detection condition. In some embodiments, the deterministic model includes a representation in which any input changes according to a specific formula and does not exhibit random variation. In some embodiments, the deterministic model may be generated using one or more historical process parameters and one or more historical endpoint detection conditions. In some embodiments, the historical endpoint detection conditions are related to the historical process parameters. In some embodiments, the deterministic model includes a mechanism model generated by performing at least linear exponential regression. In some embodiments, this linear exponential regression is performed using one or more historical process parameters and one or more historical endpoint detection conditions. This mechanism model may be based on a relationship between the historical process parameters and the historical endpoint detection conditions.

[0023] In some embodiments, the method described above may be stored on a non-transient machine-readable storage medium containing instructions, which, when executed by a processing unit, cause the processing unit to perform an operation, which may include one or more steps described in the method described above.

[0024] Figure 1 is a block diagram showing an exemplary system architecture 100 in which embodiments of the present disclosure may function. As shown in Figure 1, the system architecture 100 includes a manufacturing system 102, a measurement system 110, a client device 150, a data store 140, a server 120, and a machine learning system 170. The machine learning system 170 may be part of the server 120. In some embodiments, one or more components of the machine learning system 170 may be fully or partially integrated into the client device 150. The manufacturing system 102, the measurement system 110, the client device 150, the data store 140, the server 120, and the machine learning system 170 may each host one or more computing devices, including a server computer, a desktop computer, a laptop computer, a tablet computer, a notebook computer, a personal digital assistant (PDA), a mobile communication device, a mobile phone, a handheld computer, or a similar computing device.

[0025] The manufacturing system 102, measurement system 110, client device 150, data store 140, server 120, and machine learning system 170 may be connected to each other via a network 160 (for example, to perform the methodologies described herein). In some embodiments, the network 160 is a private network that provides each element of the system architecture 100 with access to each other and to other privately available computing devices. In some embodiments, the network 160 is a cloud-based network capable of performing cloud-based functions (for example, providing cloud service functions to one or more devices in the system). The network 160 may include one or more wide area networks (WANs), local area networks (LANs), wired networks (e.g., Ethernet networks), wireless networks (e.g., 802.11 networks or Wi-Fi networks), cellular networks (e.g., Long-Term Evolution (LTE) networks), routers, hubs, switches, server computers, cloud computing networks, and / or any combination thereof. Alternatively, or in addition to that, any of these elements of the system architecture 100 can be integrated together or otherwise combined without using network 160.

[0026] The client device 150 may be any personal computer (PC), laptop, mobile phone, tablet computer, netbook computer, network-connected television ("smart TV"), network-connected media player (e.g., Blu-ray player), set-top box, over-the-top (OTT) streaming device, operator box, cloud server, cloud-based system (e.g., cloud service device, cloud network device), or may include these. The client device 150 may also include a browser 152, application 154, and / or other tools described, and may be run by other systems of the system architecture 100. In some embodiments, the client device 150 may have access to the manufacturing system 102, the measurement system 110, the data store 140, the server 120 and / or the machine learning system 170, and may transmit (e.g., transmit and / or receive) sensor data, processed data, data classification, chamber recovery data (e.g., recovery profiles), process result data instructions, and / or inputs and outputs of various process tools (e.g., measurement tool 114, data preparation tool 116, endpoint detection tool 124, chamber knob tool 126 and / or chamber recovery components) at various stages of processing of the system architecture 100, as described herein.

[0027] As shown in Figure 1, the manufacturing system 102 includes a process tool 104, process procedures 106, and a process controller 108. The process controller 108 may coordinate the operation of the process tool 104 to perform one or more process procedures 106. For example, various process tools 104 may include dedicated chambers such as etching chambers, deposition chambers (including chambers for atomic layer deposition, chemical vapor deposition, plasma atomic layer deposition, or plasma chemical vapor deposition), annealing chambers, and / or other similar chambers. In another example, the machine may include a sample transport system for transporting samples between machines and between process steps (e.g., a selective compliance assembly robot arm (SCARA) robot, a transport chamber, a front opening pod (FOUP), a side storage pod (SSP), and / or other similar devices).

[0028] Process procedure 106, sometimes called a process recipe or process step, may include various specifications for performing an operation by the process tool 104. For example, process procedure 106 may include process specifications such as the start-up duration of the process operation, the process tool used for the operation, the temperature of the machine (e.g., chamber), flow rate, pressure, deposition sequence, and other similar items. In some embodiments, these process specifications are referred to as control knobs. In another example, the process procedure may include a transport command for transporting a sample to an additional process step, or a transport command for transporting a sample for measurement by the measurement system 110. One or more process procedures 106 include preventative maintenance procedures and seasoning procedures.

[0029] The process controller 108 may include devices designed to manage and / or coordinate the operation of the process tool 104. In some embodiments, the process controller 108 is associated with a process recipe or a set of process procedure instructions that, when applied as designed, result in a desired process outcome for the substrate process. For example, the process recipe may be related to processing the substrate to produce a target process outcome (e.g., limit dimensions, thickness, uniformity criteria, etc.). The process controller 108 can coordinate the execution of preventive maintenance procedures (e.g., chamber cleaning) and chamber recovery procedures (e.g., chamber seasoning) to return the chamber to a production state.

[0030] In some embodiments, one or more stations of the manufacturing system 102 may include sensors configured to generate and / or collect sensor data related to the process tool 104, process procedure 106, and / or process controller 108. The sensor data may include one or more values ​​such as temperature (e.g., heater temperature), spacing (SP), pressure, high-frequency radio frequency (HFRF), electrostatic chuck (ESC) voltage, current, flow rate, power, and voltage. The sensor data may relate to or indicate hardware parameters such as the settings or components of the manufacturing system 102 (e.g., size, type) or manufacturing parameters such as process parameters of the manufacturing system 102. The sensor data may be provided while the manufacturing system 102 is performing a substrate process. The sensor data may vary from substrate to substrate.

[0031] As shown in Figure 1, the measurement system 110 includes a measurement tool 114 and a data preparation tool 116. The measurement tool 114 may include various sensors for measuring process results (e.g., limit dimensions, thickness, uniformity, etc.) within the manufacturing system 102 and for detecting parameters. For example, limit dimensions can be measured using wafers processed in one or more processing chambers. The measurement tool 114 may further include devices for measuring process results of substrates processed using the manufacturing system. For example, process results such as limit dimensions and thickness measurements (e.g., film layers due to etching, deposition, etc.) of substrates processed according to a process recipe and / or operation performed by the process controller 108 can be evaluated. These measurements can also be used to measure chamber conditions throughout a recovery process, such as after a seasoning procedure involving processing a series of wafers (e.g., blank wafers) or after updating the processing chamber control knob.

[0032] The data preparation tool 116 may include process methodologies for extracting features and / or generating engineered data related to data measured by the measurement tool 114. In some embodiments, the data preparation tool 116 can identify correlations, patterns, and / or anomalies in the measured and / or process execution data. For example, the data preparation tool 116 may perform feature extraction, using a combination of measured data to determine whether a criterion is met. For example, the data preparation tool 116 may analyze multiple data points of relevant parameters (e.g., thickness, limit dimensions, defects, plasma conditions, etc.) to determine whether a rapid change occurred during the iteration of a seasoning procedure in the processing chamber. In some embodiments, the data preparation tool 116 performs normalization across various sensor data related to different process chamber conditions. Normalization may include processing incoming sensor data so that it looks similar across different chambers and sensors used to acquire the data.

[0033] In some embodiments, the data preparation tool 116 can perform one or more analyses on measurement data (e.g., acquired by the measurement tool 114), including process control analysis, univariate limit violation analysis, or multivariate limit violation analysis. For example, the data preparation tool 116 can perform statistical process control (SPC) by monitoring and controlling the process controller 108 using a statistically based methodology. For example, SPC can improve the efficiency and accuracy of preventive maintenance and / or processing chamber control knob tuning (e.g., by identifying data points located within and / or outside the control limits).

[0034] In some embodiments, the processing chamber can be measured throughout the entire seasoning procedure. In some embodiments, an increase in measurement data is acquired during a given seasoning procedure. For example, additional sensors can be activated while or immediately after processing a wafer, and / or the currently activated sensors may acquire additional data. In some embodiments, the process controller 108 may trigger measurements by the measurement tool 114 based on the actions performed by the process tool 104. For example, in response to the update of one or more processing chamber control knobs after a chamber recovery procedure, the process controller 108 may trigger the activation of one or more process results (e.g., of the measurement tool 114).

[0035] In some embodiments, the extracted features, generated synthetic / engineered data, and statistical analysis can be used in connection with the machine learning system 170 (for example, to train, verify, and / or test the machine learning model 190). In addition, and / or instead, the data preparation tool 116 can output data to the server 120 for use by either the endpoint detection tool 124 and / or the chamber knob tool 126.

[0036] The data store 140 may be memory (e.g., random access memory), a drive (e.g., a hard drive, a flash drive), a database system, or another type of component or device capable of storing data. The data store 140 may be in-situ (e.g., a local computer) or remote (e.g., a cloud-based system). The data store 140 may store one or more historical data 142, including old sensor data 144 and / or old control knob data 146 from previous chamber conditions, as well as process results of substrates processed under the relevant chamber conditions. The old sensor data 144 may be sensor data related to a previous execution of the process procedure (e.g., a previous run), and / or sensor data collected during a previous execution of the process procedure (e.g., a previous run). The old control knob data 146 may be control knob data related to a previous execution of the process procedure, and / or control knob data related to a previous seasoning operation of the processing chamber. In some embodiments, the historical data 142 may be used to train, verify, and / or test a machine learning model 190 of the machine learning system 170 (see, for example, the exemplary methodology in Figure 7). In some embodiments, the data store 140 stores endpoint detection data 148. The endpoint detection data 148 may indicate one or more endpoint detection setpoints for chamber control knob tuning.

[0037] Server 120 may include one or more computing devices such as rack-mount servers, router computers, server computers, personal computers, mainframe computers, laptop computers, tablet computers, and desktop computers. Server 120 may include an endpoint detection tool 124 and a chamber knob tool 126. Server 120 may include a cloud server or server capable of performing one or more cloud-based functions. For example, one or more operations of the endpoint detection tool 124 or the chamber knob tool 126 may be provided to a remote device (e.g., Server 120) using a cloud environment.

[0038] The chamber knob tool 126 may receive data (e.g., sensor data) from the process tool 104 and / or data (e.g., measurement data) from the data preparation tool 116, and generate chamber control knob updates related to preventive maintenance and / or chamber recovery procedures (e.g., seasoning procedures). The chamber knob tool 126 may generate chamber control knob update data (e.g., updates for one or more control knobs (e.g., process parameters)). In some embodiments, the chamber knob tool 126 receives chamber data (e.g., raw sensor data, processed sensor data) from the process tool 104 and / or processed data output from the data preparation tool 116. The chamber knob tool 126 may generate chamber control knob update data based on the chamber data received from the process tool 104 and / or processed data output from the data preparation tool 116. In some embodiments, the data preparation tool 116 receives raw sensor data from the measurement tool 114. This raw sensor data may be combined with engineered composite data from the data preparation tool 116. The chamber knob tool 126 may determine to update the chamber control knob (e.g., one or more chamber process parameters) to update the chamber conditions so that the substrate has a process result that satisfies threshold conditions. For example, the chamber knob tool may identify changes to the chamber control knob such that the substrate processed under the chamber conditions results in a substrate having a process result that satisfies threshold conditions associated with one or more process steps 106. The chamber control knob may be associated with process steps 106. These updates to the process steps may change the way in which the process tool 104 performs process steps 106. For example, the updates may include extending or shortening the process time of one or more process operations in process step 106.In some embodiments, the chamber knob tool 126 may use a machine learning model that receives input data such as chamber data from the process tool 104 and / or processed data output from the data preparation tool 116, and outputs a chamber control knob update to transition the processing chamber to production operation mode. As will be discussed later, this machine learning model may include, among other things, a bootstrap aggregation model, a random forest tree decision tree model, a partial least squares regression (PLS) model, a least absolute shrinkage and selection operator (LASSO) regression model, and / or a Ridge regression model. This machine learning model may include multiple models (e.g., stacked modeling) and may include ensemble modeling that utilizes a more reliable model for the final classification of the received data.

[0039] The endpoint detection tool 124 receives output data (e.g., chamber control knob update data) from the chamber knob tool 126 and determines the characteristics of the processing chamber (e.g., measurable / observable features) that serve as termination conditions for a substrate process or individual process step (e.g., executing multiple steps of a substrate process recipe). Based on the received chamber control knob update data, the endpoint detection tool 124 can determine termination conditions (e.g., measurable / observable features or characteristics of the chamber that indicate the termination of a process or process step). For example, the data output by the chamber knob tool 126 may include process parameters that indicate the termination time, such as process duration (e.g., sometimes called trim time, event time, process time, etc.) related to the chamber control knob tuning process. In some embodiments, the endpoint detection tool 124 and the chamber knob tool 126 may process the received data to determine the termination point (sometimes called an end point detection set point (EPD SP)) of a process procedure operation (e.g., a process procedure step). This endpoint may be an endpoint time, endpoint temperature, endpoint pressure, or any other endpoint parameter related to the process procedure. For example, the endpoint condition may be related to optical data such as a reflectance spectrum resulting from the reflection of light from the surface of the substrate being processed or a previously processed substrate. In some embodiments, the endpoint detection tool 124 includes a machine learning model that determines the endpoint of a process procedure operation using control knob data (e.g., chamber control knob update data (e.g., by the chamber knob tool 126)), synthesized and / or engineered data (e.g., from the data preparation tool 116), and / or general process parameter values ​​corresponding to the process procedure 106. For example, the endpoint detection tool may determine the endpoint of a chamber control knob tuning procedure.

[0040] In some embodiments, the chamber knob tool 126 and / or endpoint detection tool 124 may perform the functions described herein to prepare process parameters (e.g., control knobs) for production operating modes. For example, the associated endpoint conditions of the control knob update and chamber control knob tuning procedures may correspond to process parameters that result in a substrate having process results that satisfy target conditions (e.g., process uniformity, limit dimensions, etc.) when used to process the substrate in a processing chamber under the corresponding conditions.

[0041] As described above, some embodiments of the endpoint detection tool 124 and / or the chamber knob tool 126 may implement the methodology described for those tools using machine learning models. These associated machine learning models may be generated (e.g., trained, verified, and / or tested) using the machine learning system 170. The following exemplary description of the machine learning system 170 is given in the context of generating a machine learning model 190 associated with the chamber knob tool 126 using the machine learning system 170. However, it should be noted that this description is purely illustrative. Similar processing hierarchy structures and methodologies can be used when generating and executing machine learning models associated with the endpoint detection tool 124 and / or the chamber knob tool 126 individually and / or in combination with each other, which will be discussed further in relation to other embodiments.

[0042] The machine learning system 170 may include one or more computing devices such as a rack-mount server, router computer, server computer, personal computer, mainframe computer, laptop computer, tablet computer, desktop computer, cloud computer, cloud-based server, or a system stored on one or more clouds. The machine learning system 170 may also include a chamber knob component 194. In some embodiments, the chamber knob component 194 may use historical data 142 to determine a chamber control knob update (e.g., an update to one or more process parameters of a process procedure (e.g., process procedure 106)) in order to prepare the chamber for production using a chamber control knob update (e.g., an update to one or more process parameters). The chamber may be ready for production when a process procedure performed according to the updated process parameters results in one or more substrates having one or more process results that satisfy (e.g., nearly satisfy, and / or substantially satisfy) one or more target conditions (e.g., thickness, limit dimensions, sidewall angle, etc.). In some embodiments, the chamber knob component 194 may use a trained machine learning model 190 to determine a chamber control knob update. The trained machine learning model 190 may use historical data to determine a chamber control knob update.

[0043] In some embodiments, the trained machine learning model 190 is stored on server machine 192 and / or server 120. Server machine 180 and / or server machine 192 may store one or more recently trained machine learning models (e.g., to perform model execution management). For example, one or more models trained within a single period (e.g., within the last six months) may be stored on server machine 192, and one or more final models for predicting chamber conditions may be stored on server 120 (e.g., capable of working with the chamber knob tool 12).

[0044] In some embodiments, the machine learning system 170 further includes server machines 172 and 180. Server machines 172 and 180 may be one or more computing devices (e.g., rack-mount servers, router computers, server computers, personal computers, mainframe computers, laptop computers, tablet computers, desktop computers, cloud-based computing systems, etc.), data stores (e.g., hard disks, memory, databases), networks, software components, or hardware components.

[0045] The server machine 172 may include a dataset generator 174 that can generate datasets (e.g., a set of data inputs and a set of target outputs) for training, verifying, or testing a machine learning model. The dataset generator 174 may divide the historical data 142 into a training set (e.g., 60 percent of the historical data or any other portion of the historical data), a verification set (e.g., 20 percent of the historical data or some other portion of the historical data), and a test set (e.g., 20 percent of the historical data). In some embodiments, a chamber knob component 194 generates multiple sets of training data. For example, one or more sets of training data may each contain one of these datasets (e.g., a training set, a verification set, and a test set).

[0046] The server machine 180 includes a training engine 182, a verification engine 184, and a test engine 186. The training engine 182 may be capable of training a machine learning model 190 using one or more older sensor data 144 and older control knob data 146 from the historical data 142 (of the data store 140). In some embodiments, the machine learning model 190 may be trained using one or more outputs of the data preparation tool 116, the endpoint detection tool 124, and / or the chamber knob tool 126. For example, the machine learning model 190 may be a hybrid machine learning model that uses sensor data and / or mechanical features, such as feature extraction, mechanical modeling, and / or statistical modeling. The training engine 182 may generate multiple trained machine learning models 190, each trained machine learning model 190 corresponding to a different set of features in its respective training set.

[0047] The verification engine 184 may determine the accuracy of each trained machine learning model 190 based on a corresponding set of features in each training set. The verification engine 184 may discard trained machine learning models 190 that have an accuracy that does not meet a threshold accuracy. The test 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 (and optionally the verification set).

[0048] In some embodiments, training data is provided to train a machine learning model 190 so that the trained machine learning model can receive new inputs having new sensor data indicating a new state of a new processing chamber. The new output may indicate a chamber knob update in such a manner that a new substrate processed by a new processing chamber in a new state according to a first substrate processing procedure using chamber knob updates results in a new substrate having a new process result that satisfies threshold conditions.

[0049] The machine learning model 190 may refer to a model generated by the training engine 182 using a training set that includes data inputs and corresponding target outputs (historical results of the processing chamber under parameters related to the target inputs). Patterns in the dataset can be found that map the data inputs to target outputs (e.g., identifying connections between parts of sensor data and the resulting chamber control knob updates), and mappings that capture these patterns are provided to the machine learning model 190. The machine learning model 190 may use one or more of the following: logistic regression, parsing, decision trees, or support vector machines (SVMs). This machine learning may consist of a single level of linear or nonlinear operations (e.g., SVMs) and / or a neural network.

[0050] The chamber knob component 194 may provide current data (e.g., current sensor data related to the state of the processing chamber during a seasoning procedure) as input to a trained machine learning model 190, which may run on this input to obtain one or more outputs containing a set of values ​​that indicate an update to one or more chamber control knobs to satisfy a threshold condition. In some embodiments, this threshold condition indicates that a process procedure (e.g., process procedure 106) performed using the update to one or more chamber control knobs will result in one or more substrates having one or more process results that satisfy (e.g., nearly satisfy, and / or substantially satisfy) one or more target conditions (e.g., thickness, limit dimensions, sidewall angles, etc.). After the chamber control knob update, the chamber knob component 194 may be able to identify confidence data from this output indicating the confidence that the process parameters updated according to the chamber control knob update will satisfy the threshold condition for moving to production mode. In one non-restrictive example, this confidence level is a real number between 0 and 1, where 0 indicates no confidence in one or more chamber statuses, and 1 represents absolute confidence in a chamber status.

[0051] For illustrative purposes and not for limitation, aspects of this disclosure describe the training of a machine learning model using information about the historical data 142 and the use of the trained model. In other embodiments, a heuristic model or a rule-based model is used to determine the control knob update.

[0052] In some embodiments, the functions of the client device 150, server 120, data store 140, and machine learning system 170 may be provided by fewer machines than those shown in Figure 1. For example, in some embodiments, server machines 172 and 180 may be integrated into a single machine, and in some other embodiments, server machines 172, 180, and 192 may be integrated into a single machine. In some embodiments, the machine learning system 170 may be provided entirely or partially by server 120.

[0053] In general, functions described in one embodiment as being performed by the client device 150, data store 140, measurement system 110, manufacturing system 102, and machine learning system 170 may, in other embodiments, be performed on the server 120 where appropriate. Furthermore, functions belonging to a particular component may also be performed by different or multiple components working together.

[0054] In some embodiments, “User” may be represented as a single individual. However, other embodiments of the present disclosure include the case where “User” is an entity controlled by multiple users and / or an automated source. For example, a collection of individual users integrated as a group of administrators may be considered “User.”

[0055] Figures 2A-B are block diagrams illustrating processes 200A-B for performing control knob tuning, in which embodiments of the present disclosure may function.

[0056] Referring to Figure 2A, the chamber recovery process 200A may include performing preventive maintenance 202 on one or more processing tools. Preventive maintenance 202 may include any number of processes performed on one or more processing tools (e.g., processing chambers) to maintain the processing chamber conditions over time. In some embodiments, preventive maintenance 202 may use proactive preventive maintenance procedures having one or more of the following objectives, among others: minimizing processing equipment downtime; keeping maintenance costs lower (e.g., maintaining smaller repairs rather than preventing larger, less frequent repairs); extending the lifespan of processing equipment; improving production quality (e.g., processing results of substrates); improving processing efficiency (e.g., increasing the time tools are in production mode and reducing processing equipment downtime); and / or maintaining the safety and integrity of the equipment.

[0057] In some embodiments, preventive maintenance 202 includes cleaning one or more processing chambers (e.g., wet cleaning maintenance) (e.g., removing deposits on the chamber walls resulting from processing procedures performed within the processing chamber). Preventive maintenance 202 may restore the conditions of one or more processing tools to an unseasoned state (e.g., a clean chamber that has not yet undergone a seasoning process). As a result of replacing and / or cleaning one or more processing tools, the processing tools may operate in an even cleaner (e.g., raw) state by performing a substrate processing procedure using those tools. For example, the processing chamber may be returned to an unseasoned state, e.g., a new or clean state after performing preventive maintenance 202.

[0058] As shown in Figure 2A, the chamber recovery process 200A includes chamber initialization 204. Chamber initialization 204 may include warming up the chamber or starting up one or more systems associated with the processing chamber. For example, chamber initialization 204 may be performed to prepare the processing chamber for receiving and processing substrates. Chamber initialization may be associated with one or more process steps (e.g., process step 106 in Figure 1) which may include one or more process parameters (e.g., temperature, pressure, vacuum conditions, etc.).

[0059] As shown in Figure 2A, the chamber recovery process 200A includes seasoning the processing chamber 206. Chamber seasoning 206 is a procedure that includes processing a series of substrates (e.g., blank silicon wafers) to restore chamber conditions suitable for production (e.g., high-volume substrate processing). For example, in some embodiments, after chamber seasoning, a thin silicon oxide layer covers the chamber walls.

[0060] In some embodiments, the chamber recovery process 200A includes control knob tuning 208. Control knob tuning 208 is a procedure for tuning (e.g., updating and / or adjusting) one or more chamber control knobs, and the process procedure performed according to the tuned (e.g., updated and / or adjusted) control knobs, as described below with respect to sample test 210 and block 212, results in a processed substrate that satisfies a threshold condition. In some embodiments, this threshold condition may be related to the limit dimensions of the substrate (e.g., limit thickness). The chamber recovery process 200A utilizes sample test 210 to determine whether the chamber control knob settings (e.g., process parameters reflecting the updated control knobs) satisfy the threshold condition.

[0061] In some embodiments, a sample test 210 is performed and evaluated after the control knob tuning 208 is performed, as shown in Figure 2A. The test sample may involve processing the substrate by a processing chamber under conditions after a first amount of seasoning of the substrate, and / or after tuning of one or more chamber control knobs. The sample is measured in this sample test and various process outcome criteria are evaluated. For example, the sample test 210 may be evaluated to determine whether the substrate contains processing results that meet threshold criteria such as limit dimensions, process uniformity thresholds, and thickness profile requirements. Processing and evaluating the sample test may include a duration of downtime for processing the substrate and evaluating the processed substrate (e.g., measuring process results such as thickness and uniformity conditions across the surface of the substrate). In some embodiments, as discussed with reference to Figure 2B, the control knob tuning 208 may use one or more machine learning models, and one or more outputs thereof may be used to tune one or more chamber control knobs.

[0062] In block 212, the chamber recovery process 200A may determine, based on the sample test 210, whether one or more threshold conditions for the substrate are met. A sample test 210 indicating that one or more threshold conditions for the substrate are met may indicate correct (e.g., substantially correct) chamber knob tuning. The threshold conditions may include criteria related to one or more process results of the sample test 210. As previously noted, the sample test 210 may be evaluated to determine whether the substrate contains processing results that meet threshold criteria such as limit dimensions, process uniformity thresholds, and thickness profile requirements, and any of these threshold criteria may be used as evaluation criteria to determine whether the chamber and / or related process parameters meet the conditions for transitioning to a production operating mode. The processing logic may determine that the process results of a test sample processed in the chamber using the first control knob setting do not meet the threshold condition, and may further perform additional control knob tuning 208 and another sample test 210 to determine whether the chamber control knobs have been properly tuned (e.g., to meet the threshold condition). The process of control knob tuning 208, performing sample tests 210, and evaluating whether the threshold condition is met based on the previous sample test (e.g., block 212) may be repeated until the relevant process parameters meet the threshold condition. Additional chamber knob tuning may be performed to produce substrates that meet the threshold condition. After determining that the threshold condition is met, the chamber recovery process 200A may proceed to final tuning 214. Final tuning 214 may include finally tuning one or more chamber control knobs before starting the production mode of the processing chamber.

[0063] Referring to Figure 2B, a process 200B for performing control knob tuning according to an aspect of the present disclosure is shown. Process 200B may include preventative maintenance 202, chamber initialization 204, and seasoning 206, which have been outlined previously. In some embodiments, process 200B tunes the chamber control knob using one or more machine learning models. Sensor data 224 is input to a machine learning engine (e.g., ML engine 226). The sensor data 224 may indicate the state of the processing chamber after a chamber recovery procedure (e.g., seasoning 206). In some embodiments, the sensor data 224 is received from one or more sensors in the processing chamber. These one or more sensors may include an OES sensor and / or a reflectometry spectral sensor. In some embodiments, the sensor data may include one or more values ​​from among temperature (e.g., heater temperature), spacing (SP), pressure, high-frequency radio frequency (HFRF), electrostatic chuck (ESC) voltage, current, flow rate, power, voltage, etc. Sensor data may relate to or indicate hardware parameters, such as processing chamber settings or components (e.g., size, type, etc.), or manufacturing parameters, such as process parameters of manufacturing equipment (e.g., control knobs). Sensor data may be provided while the processing chamber is performing the substrate process. Sensor data may vary for each processed substrate.

[0064] In some embodiments, the ML engine 226 includes the machine learning system 170 shown in Figure 1. The ML engine 226 may train a machine learning model (e.g., model 190 in Figure 1) based on sensor data 224. The training, verification, and testing of this machine learning model are discussed in detail with reference to Figures 6 and 7.

[0065] The control knob prediction 228 may be determined based on one or more outputs of the ML engine 226. In some embodiments, the outputs of the ML engine 226 are used to generate one or more predictions for one or more chamber control knob settings. The control knob prediction 228 may be related to tuning one or more chamber control knobs (e.g., control knob tuning 208). In some embodiments, the control knob prediction 228 is determined considering substrate processing that satisfies threshold conditions (e.g., threshold conditions discussed with respect to block 212).

[0066] The control knob prediction 228 may be used to update one or more chamber control knobs (e.g., control knob update 230). In some embodiments, after updating one or more control knobs according to the control knob update 230, the processing chamber may produce substrates according to the chamber control knob update that satisfies the process procedure and threshold conditions. By satisfying the threshold conditions, the processing chamber may be placed in a state associated with a process procedure that, when executed according to the control knob update 230, results in substrates having one or more process results that satisfy one or more target conditions (e.g., thickness, limit dimensions, etc.). In some embodiments, the production mode of the processing chamber may be started after the control knob update 230.

[0067] In some embodiments, process 200B returns the chamber to production mode using a shorter time (e.g., shorter processing chamber downtime). For example, the chamber recovery process 200B may not require the execution of sample tests (e.g., processing and evaluation of a test wafer) to determine that the processing chamber can produce substrates according to a process procedure using chamber control knob updates such that the substrates meet one or more threshold conditions. In another example, process 200B may determine a chamber control knob update (e.g., control knob update 230) that may immediately enable the processing chamber to start production mode without performing additional or final tuning of the chamber control knobs to meet one or more threshold conditions.

[0068] As will be discussed further in other embodiments, the chamber recovery process 200A may be used to generate training data for the ML engine 226. Sensor data from the processing chamber may be tracked throughout the seasoning process 206, and the output from the ML engine 226 may be verified using data related to the sample test 210. For example, the output from the ML engine 226 may be used to determine the update of the chamber control knob. The evaluation of the sample test processed according to the same process procedure and control knob update can be compared with the prediction by the ML engine 226 to verify and further improve the ML engine 226.

[0069] Figure 3 is a block diagram showing a chamber monitoring and maintenance system 300 in which embodiments of the present disclosure may function. The chamber monitoring and maintenance system 300 includes a substrate processing system 302 (e.g., manufacturing system 102 in Figure 1). The substrate processing system 302 may include a process tool 314 and chamber data 344 (e.g., acquired using the process tool 314). The process tool 314 may include a chamber sensor (for measuring, for example, radio frequency (RF), temperature, gas, pressure, etc.) and additional sensors (e.g., an optical emission spectroscopy (OES) sensor, a reflectometry spectral sensor, etc.). The chamber data 344 may include data indicating the state of the processing chamber. For example, the process tool 314 may measure and monitor parameters such as RF, temperature, gas, and spectral data.

[0070] In some embodiments, chamber data 344 is processed by a data preparation tool 316. The data preparation tool 316 can receive chamber data (e.g., raw sensor data) and generate synthetic data associated with various combinations, correlations, and / or artificial parameters. The data preparation tool 316 can reduce the dimensionality of the raw sensor data using groups or features. For example, the data preparation tool 316 may generate features that include parameter values ​​and / or rates of change of sets of parameter values. In some embodiments, the data preparation tool 316 performs one of the following: partial least squares analysis, principal component analysis, independent component analysis, multifactor dimensionality reduction, nonlinear dimensionality reduction, and / or any combination thereof. In some embodiments, the data preparation tool 316 is designed to detect edges in the chamber data. For example, the data preparation tool 316 includes a methodology aimed at identifying abruptly changing data points and / or data points having discontinuities. For example, the slope of one or more measured chamber parameters in the processing chamber.

[0071] In some embodiments, the data preparation tool 316 performs normalization across various sensor data associated with different process chamber conditions. Normalization may include processing incoming sensor data so that it looks similar across different chambers and / or sensors used to acquire the data. In some embodiments, the data preparation tool 316 may include embodiments and / or features of the data preparation tool 116 in Figure 1.

[0072] In some embodiments, a data preparation tool 316 is used to generate inputs for the chamber knob tool 326. As will be discussed further in other embodiments, the chamber knob tool 326 receives input data (e.g., raw sensor data, synthesized data, output of other models, etc.) to determine chamber control knob updates (e.g., updates to one or more process parameters related to the processing chamber). For example, these chamber control knob updates may include instructions to increase / decrease temperature, extend / shorten process time, etc. In some embodiments, the chamber knob tool 326 receives sensor and measurement data indicating the state of the processing chamber after processing a series of substrates (e.g., seasoning wafers). The chamber knob tool 326 may output a set of values ​​indicating chamber control knob updates that satisfy threshold conditions (e.g., substrates processed according to a process procedure including chamber control knob updates satisfy the target result). For example, the chamber knob tool 326 may output data indicating chamber control knob updates related to a substrate processing procedure. The chamber control knob update may also instruct an update to the associated process procedure, which, when the associated substrate is processed after the chamber control knob update, will result in a substrate having a process result that satisfies one or more threshold conditions associated with the substrate processing procedure (e.g., substrate process recipe). In some embodiments, as will be discussed further in other embodiments, the chamber knob tool 326 may include a machine learning model (e.g., trained using Figure 7 and / or implemented using Figure 6).

[0073] In some embodiments, the chamber knob tool 326 outputs data (e.g., chamber control knob update data) to the endpoint detection tool 324 (e.g., endpoint detection tool 124 in Figure 1). The endpoint detection tool 324 may process this data output by the chamber knob tool 326 to determine the relevant termination conditions (e.g., measurable / observable features) of the processing chamber associated with the chamber control knob update. In some embodiments, the endpoint detection tool 324 determines the endpoint of a process procedure operation (e.g., a step in the process procedure) based on the data received from the chamber knob tool 326. This endpoint may be a termination time (e.g., duration), termination temperature, termination pressure, or any other termination parameter related to the process procedure. For example, the data received from the chamber knob tool 326 may indicate an updated process duration related to the process procedure. The endpoint detection tool 324 may determine endpoint criteria, such as optical spectral measurements correlated with the process duration. In some embodiments, the endpoint detection tool 324 determines the correlation between substrate results (e.g., substrate thickness, substrate dimensions, substrate optical quality, etc.) and the endpoint. This correlation may be a deterministic correlation (for example, one input reduces to one output). In some embodiments, the endpoint detection tool 324 uses a machine learning model to determine the endpoint of a process procedure operation.

[0074] In some embodiments, the chamber knob tool 326 outputs a command to the process controller 308 to modify one or more operations related to the substrate processing system 302. For example, the chamber knob tool 326 may instruct a chamber control knob update, and the process controller may, based on this control knob update, modify the operating parameters of the substrate processing system to enter a production operation mode. In another example, the chamber knob tool 326 may instruct an update to one or more process parameters. In yet another example, the chamber knob tool 326 may instruct the process controller 308 that the chamber control knob tuning is insufficient (e.g., threshold conditions are not met), and may instruct the process controller 308 to perform additional chamber control knob tuning on the substrate processing system 302.

[0075] In some embodiments, the endpoint detection tool 324 outputs instructions to the process controller 308 to modify one or more operations related to the substrate processing system 302. For example, the endpoint detection tool 324 may instruct the process controller 308 that the chamber control knob tuning is insufficient (e.g., the substrates processed according to the chamber control knob update do not result in substrates that satisfy one or more target conditions), and may instruct the process controller 308 to perform additional chamber control knob tuning on the substrate processing system 302. In another example, the endpoint detection tool 324 may instruct the process controller 308 that the chamber control knob tuning is sufficient (e.g., the substrates processed according to the chamber control knob update result in substrates that satisfy one or more target conditions), and that the production mode of the substrate processing system should be initialized. The endpoint detection tool 324 may update the endpoint conditions based on the updated chamber control knob tuning. For example, the endpoint detection tool 324 may update the endpoint conditions based on feature-based data such as optical data, the detectable thickness of the substrate, and the plasma flux passing through the processing chamber of the substrate processed according to the process procedure and updated chamber control knob tuning.

[0076] As shown in Figure 3, the process controller 308 receives input from the chamber knob tool 326 and the endpoint detection tool 324 and provides instructions to the substrate processing system 302 when performing additional operations. For example, the process controller 308 may instruct the substrate processing system 302 to execute a process recipe in accordance with the chamber control knob update. In another example, the process controller 308 may instruct the substrate processing system 302 to update one or more process parameters in accordance with the chamber control knob update. In yet another example, the process controller 308 may instruct the substrate processing system to initialize a production mode.

[0077] As shown in Figure 3, the chamber knob tool 326 and the endpoint detection tool 324 may output instructions to be displayed on a graphical user interface (GUI). This graphical user interface may output chamber control knob predictions on the graphical user interface. This GUI may output instructions to initialize the production mode of the substrate processing system 302. In some embodiments, the chamber knob tool 326 and / or the endpoint detection tool 324 output instructions related to one or more operations to be performed on the substrate processing system for display on the GUI 328. For example, the GUI 328 may display a notification informing that one or more chamber control knobs (e.g., process parameters related to the substrate processing system 302) should be updated. In another example, the GUI 328 may display a notification informing that the production mode of the substrate processing system 302 should be initialized. In yet another example, the GUI 328 may display a notification informing that additional tuning of one or more chamber control knobs should be performed.

[0078] Figure 4A shows a block diagram illustrating an exemplary modeling architecture 400A for chamber control knob tuning according to several embodiments. Architecture 400A may use one or more machine learning models to determine chamber control knob predictions (e.g., updates to one or more process parameters associated with the processing chamber). In some embodiments, architecture 400A uses boosted modeling to determine the predictions.

[0079] In some embodiments, the input data 404 is associated with one or more instances that execute a process recipe, and these instances are sometimes called events. For example, the first event may be associated with a first instance that executes a process recipe. The input data 404 may contain multiple subsets (e.g., input data 404(1.1), input data 404(1.2), etc.). Each subset of the input data may be associated with an operation of the process recipe. For example, input data 404(1.1) may be associated with the first operation of the first event. As another example, input data 404(1.2) may be associated with the second operation of the first event. As yet another example, input data 404(NM) may be associated with the Mth operation of the Nth event. In some embodiments, the input data 404 corresponds to one or more processing events, and the subsets of the input data 404 (e.g., input data 404(1.1), input data 404(1.2), etc.) correspond to one or more operations (e.g., a trim step) of each event. In some embodiments, input data 404 is collected after each corresponding event. For example, input data 404(1.1) to input data 404(1.M) may be collected after the Mth operation of the first event. In another example, input data 404(N.1) to input data 404(NM) may be collected after the Mth operation of the Nth event. In some embodiments, the recipe operation is called the trim step.

[0080] In some embodiments, a data preparation tool (e.g., data preparation tool 316 in Figure 3) subdivides the input data 404 into multiple subsets corresponding to each corresponding operation of the associated process recipe. For each subset of the input data 404 corresponding to an operation, a set of process parameters (e.g., chamber control knobs) is predicted.

[0081] In some embodiments, input data 404 is provided as input to an output process tool 406. The output process tool 406 may include logic for performing filtering, splitting, processing, etc., of the input data 404. For example, the output process tool 406 may perform statistical analysis on the input data 404. The output process tool 406 may calculate one or more statistical values ​​(e.g., mean, median, minimum, maximum, standard deviation, etc.) for each corresponding subset of the input data 404. In some embodiments, the output process tool 406 outputs one or more sets of data. The output process tool 406 may provide output base data 408. The output base data 408 may indicate the base value for each corresponding subset of the input data 404. In some embodiments, the output base data 408 indicates the average value (e.g., mean value) for each corresponding subset of the input data 404. For example, the output base data 408 may indicate the mean of a subset of input data 404(1.1) to input data 404(N.1) (e.g., input data 404(1.1), input data 404(2.1), input data 404(3.1), etc.). Alternatively, the output base data 408 may indicate the mean of a subset of input data 404(1.M) to input data 404(NM) (e.g., input data 404(1.M), input data 404(2.M), input data 404(3.M), etc.). In some embodiments, the output base data 408 is provided as input to a machine learning model. This machine learning model may be part of a boosted machine learning model algorithm that uses two or more machine learning models.

[0082] The output processing tool 406 may provide output offset data 409. The output offset data 409 may indicate the offset value for each corresponding subset of the input data 404. In some embodiments, this offset value may be based on statistical deviation. This offset value may be a delta value (e.g., difference). This offset value may be an error value. For example, the offset value for each corresponding subset of the input data 404 may be based on the standard deviation associated with each corresponding subset of the input data 404. In some embodiments, the output offset data 409 indicates the delta value associated with each corresponding subset of the input data 404. The delta value associated with each corresponding subset of the input data may be based on the standard deviation value associated with each corresponding subset of the input data 404. For example, the output offset data 409 may indicate the delta value associated with subsets of input data 404(1.1) to input data 404(N.1) (e.g., input data 404(1.1), input data 404(2.1), input data 404(3.1), etc.). As another example, the output offset data 409 may indicate delta values ​​associated with a subset of input data 404(1.M) to input data 404(NM) (e.g., input data 404(1.M), input data 404(2.M), input data 404(3.M), etc.). In some embodiments, the output offset data 409 is provided as input to a machine learning model. This machine learning model may be part of a boosted machine learning model algorithm that uses two or more machine learning models.

[0083] In some embodiments, output base data 408 is provided to a predictive model 410(1). The predictive model 410(1) may be a trained machine learning model (e.g., trained and / or deployed using methods 700 and 800 in Figures 7 and 8, respectively). The predictive model 410(1) may be a machine learning model of a boosting machine learning modeling architecture. The predictive model 410(1) may be trained to determine at least a first portion of one or more chamber control knob updates (e.g., an update to at least one process parameter value of a set of process parameter values ​​associated with performing a substrate processing procedure). In some embodiments, the predictive model 410(1) predicts the base control knob updates. The base control knob updates may be a subset of control knob updates that constitute the majority of the total control knob updates. The base control knob updates may be insufficient by the amount of delta (e.g., offset control knob updates) described later. The base control knob updates determined by the predictive model 410(1) may be a first model output.

[0084] In some embodiments, output offset data 409 is provided to a predictive model 410(2). The predictive model 410(2) may be a trained machine learning model. The predictive model 410(2) may be a machine learning model of a boosting machine learning modeling architecture. The predictive model 410(2) may be trained to determine at least one or more second parts of chamber control knob updates. In some embodiments, the predictive model 410(2) predicts the offset control knob update. The offset control knob update may be a subset of control knob updates that constitute a part of the overall control knob update. The overall control knob update may include a base control knob update and an offset control knob update. Both the base control knob update and the offset control knob update may be weak learners (e.g., weak learning models), and they may be combined to produce a strong learner (e.g., a strong learning model; control knob update). The offset control knob update may be a delta control knob update (e.g., the difference between a first value associated with the control knob update and a second value associated with the control knob update). The offset control knob update may compensate for the error in the base control knob update between the value of the base control knob update and the value of the control knob update (e.g., the sum of the base control knob update and the offset control knob update). The offset control knob update determined by the prediction model 410(2) may be the second model output.

[0085] In some embodiments, a base control knob update determined by a predictive model 410(1) and an offset control knob update determined by a predictive model 410(2) are combined to form a final prediction 412. In some embodiments, the combination of the base control knob update and the offset control knob update is more accurate (e.g., more reliable, stronger) than one or either update alone. The combination of the base control knob update and the offset control knob update may be a stronger prediction than a prediction by a single model (e.g., a prediction that the update, when used, will result in a base control knob with greater confidence that it will result in a substrate that satisfies threshold conditions). In some embodiments, the final prediction 412 is a first update for one or more chamber control knobs based on the combination of at least a first model output and a second model output. The final prediction 412 may be based on one or more outputs of predictive model 410(1) and predictive model 410(2). The substrate processing procedure may be performed based on the final prediction 412. In some embodiments, performing a substrate processing procedure based on a final prediction 412 in a processing chamber having a first state environment (e.g., after a chamber recovery process) results in processing a substrate having a process result that satisfies threshold conditions (e.g., target thickness, limit dimensions, quality, etc.).

[0086] Figure 4B shows a block diagram illustrating an exemplary data preparation architecture according to several embodiments. Architecture 400B may use a data preparation tool 416 (e.g., data preparation tool 316 in Figure 3) to isolate one or more subsets of input data. In some embodiments, a predictive model 416 receives input data related to one or more instances of the execution of a chamber process recipe. This input data may be sensor data from one or more sensors associated with the processing chamber. In some embodiments, this sensor data may be collected after each instance of the execution of the chamber process recipe. For example, recipe data 413(1) may be collected after the first instance of the execution of the chamber process recipe. As another example, recipe data 413(N) may be collected after the Nth instance of the execution of the chamber process recipe. In some embodiments, each subset of recipe data includes one or more subsets of sensor data related to one or more operations of the chamber process recipe. Each subset of recipe data may include sensor data corresponding to each operation of the corresponding chamber process recipe and / or chamber seasoning recipe iteration. The operation data 414 may also be sensor data corresponding to each operation in the corresponding iteration of the chamber process recipe and / or chamber seasoning recipe. Operation data 414(1.1) may correspond to the first operation (e.g., step) of the first iteration of the chamber process / seasoning recipe. Operation data 414(NM) may correspond to the Mth operation (e.g., step) of the Nth iteration of the chamber process / seasoning recipe. Recipe data 413 may contain operation data 414. For example, recipe data 413(1) may contain operation data 414(1.1) to operation data 414(1.M) (e.g., operation data 414(1.1), operation data 414(1.2), operation data 414(1.3), etc.).As another example, recipe data 413(N) may include operation data 414(N.1) to operation data 414(NM) (e.g., operation data 414(N.1), operation data 414(N.2), operation data 414(N.3), etc.). Operation data 414 may be associated with one or more recipe operation predictions of the corresponding chamber process / seasoning recipe.

[0087] In some embodiments, the predictive model 416 predicts one or more process parameters associated with a chamber process recipe. A subset of the output of the predictive model 416 may correspond to a subset of the recipe data 413. The output data 418 may include predictions of one or more process parameters output by the predictive model 416. A subset of the output data 418 may correspond to a subset of the operation data 414. For example, output data 418(1.1) may correspond to operation data 414(1.1). As another example, output data 418(NM) may correspond to operation data 414(NM). The output data 418(1.1) may include predictions of one or more process parameters associated with a recipe operation (e.g., a recipe step) corresponding to the operation data 414(1.1). The output data 418 may be input to one or more tools or machine learning models of this disclosure.

[0088] Figure 5 shows a block diagram of a chamber knob prediction system 500 using stacked modeling according to an aspect of this disclosure. One or more of the models (e.g., machine learning models) described herein may include model stacking as described in relation to Figure 5. For example, model 190 of machine learning system 170 (e.g., Figure 1) may include one or more methodologies and / or processes shown in Figure 5. As another example, prediction model 410 in Figure 4A and prediction model 416 in Figure 4B may include one or more methodologies and / or processes shown in Figure 5.

[0089] As shown in Figure 5, the process result prediction system 500 may include a dataset containing a set of input data 502 (e.g., one or more of the input data 404 in Figure 4A or the recipe data 413 in Figure 4B) and a set of output data 504 corresponding to each of the input data 502 (e.g., the output data 418 in Figure 4B). The input data 502 and output data 504 may be received by a data processing tool 506. The data processing tool 506 may perform the task of dividing these input and output data into data groups 508. The data groups 508 may contain different combinations of groupings of the input data 502 and output data 504. In some embodiments, the data groups 508 are mutually exclusive, while in other embodiments, the data groups 508 contain common data points.

[0090] As shown in Figure 5, this chamber knob prediction system generates a stack of local models 510. Each local model may be generated and / or trained based on an individual associated data group 508. Each local model 510 may be trained to produce an output independent of other local models 510 based on the same input it receives. Each local model may receive new input data and provide new output data based on the trained model. Each model may identify different features, artificial parameters and / or principle components based on the difference in the data group 508 used to train the corresponding model 510 (e.g., by the difference in the training dataset).

[0091] Local models 510(M) may be used in conjunction with each other to generate and / or train the final model. In some embodiments, the final model includes a weighted mean ensemble. The weighted mean ensemble weights the contribution of each local model 510 by the credibility or confidence of the contribution (e.g., output) that its corresponding model receives. In some embodiments, these weights are equal across the local models 510 (e.g., each output from each local model 510 is treated equally across multiple models). In some embodiments, the final model is trained to determine various weights (e.g., contribution weights) of the local models (e.g., using a neural network or deep learning network). For example, one or more types of regression (e.g., gradient boosting regression, linear regression, logistic regression) may be performed to determine one or more contribution weights associated with the local models. The final model 512 may take outputs from the local models 510 as input and attempt to learn how best to combine the input predictions to perform improved output predictions. In some embodiments, the final model 512 is configurable. The final prediction model may receive output from only one local model 510 (for example, the final model 512 may be configured to receive output from only one local model 510).

[0092] Figures 6–7 show flowcharts illustrating exemplary methods 600–700 relating to training and / or using machine learning models related to chamber data, according to several embodiments of the present disclosure. For simplicity of explanation, methods 600–700 are presented and described as a series of operations. However, the operations according to the present disclosure may be performed in various orders and / or simultaneously, and may be performed in conjunction with other operations not presented and described herein. Furthermore, not all of the operations shown may be performed to carry out methods 600–700 according to the disclosed subject matter. In addition, those skilled in the art will understand and recognize that methods 600–700 may also be represented through a state diagram or events as a series of interrelated states.

[0093] Figure 6 shows a flowchart of a method 600 for updating a chamber control knob after a chamber preventive maintenance event, according to several embodiments of the present disclosure. The method 600 is performed by processing logic, which may include hardware (e.g., circuits, dedicated logic, etc.), software (e.g., software running on a general-purpose computer system or a dedicated machine), or any combination thereof. In one embodiment, the method is performed using the server 120 and trained machine learning model 190 of Figure 1, while in several other embodiments, one or more blocks of Figure 6 may be performed by one or more other machines not shown in the figure.

[0094] Method 600 may include receiving sensor data (e.g., related to a processing chamber on which a process procedure is performed) and processing this sensor data using a trained machine learning model 190. The trained machine learning model (MLM) may be configured to generate one or more outputs that, based on this sensor data, instruct a first update related to a substrate processing procedure. In some embodiments, the trained MLM is configured to generate chamber control knob updates related to the substrate processing procedure (e.g., updates to one or more process parameters).

[0095] In block 602, the processing logic receives sensor data indicating a first state of the environment in the processing chamber processing the substrate. This sensor data may be provided by one or more sensors, including an OES sensor, a reflectometry spectral sensor, and / or other sensors for measuring various parameters such as temperature, pressure, HFRF, ESC voltage, current, and power supply voltage. In some embodiments, the first state of the environment in the processing chamber processing the substrate occurs after the execution of a chamber recovery procedure. This chamber recovery procedure may include a seasoning procedure consisting of multiple seasoning runs (e.g., seasoning 206 in Figure 2A).

[0096] In some embodiments, this sensor data further includes synthetic data or data engineered from raw sensor data. For example, as described in previous embodiments, various engineering tools can perform feature extraction and / or generate combinations of artificial and / or virtual parameters. Feature extractors (e.g., data preparation tool 116 in Figure 1) can generate various features by performing various analyses on the raw sensor data, such as process control analysis (PCA), univariate / multivariate statistical process control (SPC), univariate limit violation analysis and / or multivariate limit violation analysis. In some embodiments, the sensor data is normalized across multiple processing chambers and / or process recipes to generate comparable datasets with a common base.

[0097] In block 604, the processing logic receives substrate process data that specifies a set of process parameter values ​​related to executing a substrate processing procedure. These process parameter values ​​may include one or more chamber control knob settings. In some embodiments, the substrate process data may include process duration (e.g., trim time). The substrate process data may include sequentially recorded process data. For example, the substrate process data may include a first process data piece, followed by a second process data piece, and so on. In some embodiments, the substrate process data includes sequentially recorded process trim times. In some embodiments, the substrate process data includes a set of data that specifies a process recipe (e.g., a set of data that specifies a set of process recipe operations) (see Figure 4B). In some embodiments, the substrate processing procedure is performed by a processing chamber having a second state environment. The second state of the processing chamber may specify a state that occurs before a chamber recovery procedure. The second state of the processing chamber may specify a state that occurs after a preventive maintenance event.

[0098] In block 606, one or more MLMs process sensor data and substrate process data to determine one or more outputs. In some embodiments, this one or more output includes a first update related to performing a substrate processing procedure. This first update may be an update to one or more process parameter values. This first update may be an update to one or more chamber control knob settings. The chamber control knob settings may include updates to the operation and / or configuration of hardware and / or software elements that control one or more processing chamber parameters related to processing one or more substrates in the processing chamber. A substrate processed using a substrate processing procedure including this first update may have a process result that satisfies threshold conditions related to the substrate processing procedure (e.g., uniformity requirements, limit dimension criteria, film thickness criteria, etc.).

[0099] In block 608, the processing logic prepares a notification instructing a first update. This notification may be prepared to be presented on a GUI. In some embodiments, this GUI is presented on a client device (e.g., client device 150 in Figure 1). In some embodiments, this notification is provided to the user via the GUI.

[0100] In block 610, the processing logic causes the processing chamber to perform a selection of a substrate processing procedure using a first update. In some embodiments, the selection of a substrate processing procedure may include a first update. In some embodiments, process parameters associated with the processing procedure are updated based on the first update. The chamber control knobs may be updated based on a chamber control knob update corresponding to the first update.

[0101] In some embodiments, in block 612, processing logic determines a first endpoint detection condition. This first endpoint detection condition may correspond to a first update. The first endpoint detection condition may indicate the end of a chamber control knob tuning procedure, corresponding to a chamber processing a substrate according to a first update that results in a substrate satisfying a process recipe and target results. In some embodiments, the first endpoint detection condition corresponds to an optical sensor measurement of the substrate processed according to the first update.

[0102] Determining a first endpoint condition may involve applying a deterministic model to the first update. This deterministic model may generate the first endpoint condition based on the first update. The first endpoint condition may correspond to a physical measurement of the substrate processed according to the first update, and the measurement value of the physical measurement itself may correspond to the first update. For example, a measurement of the substrate thickness may be performed, in which case the measured thickness value corresponds to one or more updated parameters of the first update. The first endpoint condition may be generated using one or more history process parameters (e.g., history chamber control knob settings) and one or more history endpoint detection conditions associated with the history process parameters. For example, a machine learning model may predict the first endpoint condition based on a first update given as input, and this machine learning model is trained using a training input containing history process parameters and a target output containing history endpoint detection conditions associated with the history process parameters.

[0103] In some embodiments, the deterministic model includes a mechanism model. This mechanism model may be generated by performing linear exponential regression using historical process parameters and historical endpoint detection conditions associated with the historical process parameters. In some embodiments, a correlation is established between the historical process parameters and the historical endpoint detection conditions that enables the generation of the mechanism model by performing linear exponential regression.

[0104] In some embodiments, in block 614, processing logic causes the processing chamber to select a substrate processing procedure using a first endpoint detection condition. The processing chamber may be caused to execute a substrate processing procedure based on a first update and / or the first endpoint detection condition. In some embodiments, this selection of the substrate processing procedure may include the first endpoint detection condition. In some embodiments, process parameters associated with the processing procedure are updated based on the first endpoint condition. The chamber control knobs may be updated based on the chamber control knob update corresponding to the first endpoint condition.

[0105] Figure 7 is a diagram illustrating the training phase of a machine learning system according to several embodiments of the present disclosure. The machine learning system 170 may use Method 700 to perform at least one of training, verification, or testing a machine learning model according to embodiments of the present disclosure. In some embodiments, one or more operations of Method 700 may be performed by a dataset generator 174 of the server machine 172 described with respect to Figure 1. Note that aspects of Figure 7 may be illustrated using the components described with respect to Figure 1.

[0106] In block 702, the processing logic initializes the training set. In block 704, the processing logic identifies a first data input (e.g., a first training input, a first confirmation input). This first data input may include first training input data indicating a first state of the environment of the first processing chamber processing the substrate after a chamber recovery procedure performed on the first processing chamber. The first data input may include first substrate process data indicating a first set of process parameter values ​​related to performing a first substrate processing procedure. The execution of the first substrate processing procedure may be by the first processing chamber having an environment in a second state. The second state may be a state that occurs before the chamber recovery procedure. In some embodiments, the first data input may include chamber environment parameters (e.g., temperature, pressure, vacuum conditions, plasma conditions, optical measurements, spectral conditions, etc.). The first target input may be identified from sensor data 144 of history data 142 (e.g., data store 140).

[0107] In block 706, the processing logic identifies a target output for a training input. In some embodiments, the target output includes second substrate process data that instructs a first update to at least one process parameter value from a first set of process parameter values. The first update may be an update to one or more chamber control knobs associated with a first substrate processing procedure. The first update may be associated with a first processing chamber having a first state environment in which the first substrate processing procedure is being performed. The first target input may be identified from sensor data 144 of historical data 142 (e.g., data store 140).

[0108] In block 708, the processing logic optionally generates mapping data that instructs input / output mapping. This input / output mapping (or mapping data) may relate to a data input (e.g., one or more of the data inputs described herein), a target output for the data input (e.g., one or more of the data inputs described herein), a target output for data (e.g., if the target output identifies the relevant chamber status), and the relationship between the data input and the target output. The input may be based on sensor data and first substrate process data. The output may be based on second substrate process data.

[0109] In block 710, the processing logic adds the mapping data generated in block 704 to dataset T.

[0110] In block 712, the processing logic branches based on whether the dataset T is sufficient for at least one of training, verification, or testing the machine learning model 190. If it is sufficient ("yes" branch), execution proceeds to block 714; otherwise, it returns to block 704 ("no" branch). It should be noted that in some embodiments, whether the dataset T is sufficient may be determined simply based on the number of input / output mappings in the dataset, while in some other embodiments, whether the dataset T is sufficient may be determined based on, in addition to or instead of, the number of input / output mappings, one or more other criteria (e.g., measures such as diversity or accuracy of data examples).

[0111] In block 714, the processing logic provides a dataset T for training, verifying, or testing a machine learning model 190. In some embodiments, dataset T is a training set and is provided to the training engine 182 of the server machine 180 to perform training. In some embodiments, dataset T is a verification set and is provided to the verification engine 184 of the server machine 180 to perform verification. In some embodiments, dataset T is a test set and is provided to the test engine 186 of the server machine 180 to perform testing. For example, in the case of a neural network, the input values ​​of a given input / output mapping (e.g., numerical values ​​associated with data inputs) are input to the neural network, and the output values ​​of the input / output mapping (e.g., numerical values ​​associated with target outputs) are stored in the output nodes of the neural network. The connection weights of the neural network are then adjusted according to a learning algorithm (e.g., backpropagation), and this procedure is repeated for the remaining input / output mappings in dataset T. After block 714, at least one of the following can be performed: training the machine learning model (e.g., machine learning model 190) using the training engine 182 of the server machine 180, verifying it using the verification engine 184 of the server machine 180, or testing it using the test engine 186 of the server machine 180. The trained machine learning model may be implemented by the chamber status component 194 (of the server machine 192) to identify chamber control knob updates (e.g., updates to one or more process parameters related to the substrate process procedure) based on the received sensor data.

[0112] In some embodiments, the generated training dataset is used to train a machine learning model and / or a physical model. The model may be trained to accept chamber sensor data as input. The model may output chamber control knob updates (e.g., updates to one or more process parameters related to a substrate process procedure) based on the received sensor data. In embodiments, the model may be agnostic to individual processing chambers, performed preventive maintenance, and / or process recipes used in production modes. Thus, the model may be generated based on training data items generated based on the execution of a seasoning process on a first processing chamber, and then used for a second processing chamber without performing transfer learning to tune the model for the second processing chamber.

[0113] In one embodiment, the trained machine learning model is a regression model trained using regression. An example of a regression model is a regression model trained using linear / nonlinear regression or Gaussian regression. The regression model predicts a value of Y given a known value for the variable X. The regression model may be trained using regression analysis, which may include interpolation and / or extrapolation. In one embodiment, the parameters of the regression model are estimated using the least squares method. Alternatively, the regression model may be trained by performing Bayesian linear regression, partial least squares regression, support vector regression, percentage regression, minimum absolute deviation, nonparametric regression, regularized regression, scenario optimization and / or distance learning. In one embodiment, the trained machine learning model includes a support vector machine or other types of machine learning models.

[0114] In one embodiment, the trained machine learning model uses ensemble modeling. Ensemble modeling may include training multiple models (e.g., two or more related but different analytical models) and synthesizing the results of each model into a single score or spread (e.g., improving and reducing predictive accuracy from any one machine learning model). For example, the trained machine learning model may include a bootstrap aggregation ("bagging") model. Bagging may include an aggregation stage and a bootstrapping stage for each individual model. Bootstrapping includes sampling or generating a dataset (e.g., using method 700), where each sample (e.g., sensor data related to the execution of a seasoning procedure on a processing chamber) is selected (e.g., randomly) within a broader sample set and replaced. As previously noted, for each, multiple models may be generated based on different subsets of the broader overall set of training data (e.g., historical data 142). Those multiple models may output synthesized (e.g., aggregated) results to form an overall score (e.g., overall confidence) related to the identified output of the ensemble of models (e.g., predicted chamber status).

[0115] In another example, this machine learning model may include a random forest model. This machine learning model utilizes multiple decision trees or multiple types of analytical models designed to predict outputs based on different variables and rules. This random forest model may mix decision trees that analyze different sample data, evaluate different factors, or weight common variables by different amounts. The results of these various decision trees may be transformed into a combined result (e.g., an aggregated result through a simple average or additional weighting) to predict chamber control knob updates based on input sensor data. In another example, this machine learning model may include a gradient boosting regression (GBR) model. This machine learning model utilizes multiple (e.g., at least two) models to predict outputs based on different inputs. These inputs may be related. The gradient boosting regression model may use a first model to determine a first prediction, and a second model to determine the delta (e.g., difference) between the first prediction and the target value. The outputs of these models may be combined to predict chamber control knob updates based on input sensor data. In some embodiments, a first regression is performed to generate a first machine learning model using a first subset of the training data, and a second regression is performed to generate a second machine learning model using a second subset of the training data. The accuracy of the first machine learning model may be determined to be greater than the accuracy of the second machine learning model, based on a comparison of the first machine learning model, the second machine learning model, and the training data.

[0116] In one embodiment, the trained machine learning model is an artificial neural network (also simply called a neural network). This artificial neural network may be, for example, a convolutional neural network (CNN) or a deep neural network. In one embodiment, processing logic trains the neural network by performing supervised machine learning.

[0117] Artificial neural networks generally include feature representation components that have classifier or regression layers that map features to a target output space. For example, a convolutional neural network (CNN) acts as a host for multiple layers of convolutional filters. Pooling may be performed and nonlinearity may be addressed in the lower layers, and above the lower layers, a multilayer perceptron is typically added to map the upper layer features extracted by the convolutional layers to a decision (e.g., classification output). This neural network may be a deep network with multiple hidden layers, or a shallow network with no hidden layers or only a few (e.g., one or two) hidden layers. Deep learning is a type of machine learning algorithm that performs feature extraction and transformation using a cascade of multiple layers of nonlinear processing units. Each successive layer uses the output from the preceding layer as input. Neural networks may be trained in a supervised (e.g., classification) and / or unsupervised (e.g., pattern analysis) manner. Some neural networks (e.g., deep neural networks) have a hierarchical structure of layers, where different layers learn different levels of representation corresponding to different levels of abstraction. In deep learning, each level learns to transform its input data into a slightly more abstract and complex representation.

[0118] Training a neural network may be achieved through supervised learning, which involves feeding a training dataset consisting of labeled inputs through the network, observing its outputs, defining the error (by measuring the difference between the output and the labeled value), and tuning the network's weights across all its layers and nodes to minimize the error, using techniques such as deep gradient descent and backpropagation. In many applications, repeating this process across many labeled inputs of the training dataset gives a network that can produce the correct output when presented with inputs different from those present in the training dataset.

[0119] To achieve continuous learning and improvement of this trained machine learning model, the trained machine learning model may be retrained periodically or continuously. The model may generate an output based on an input, perform an action based on that output, and measure the result of that action. In some examples, the result of the action is measured within a few seconds or minutes, while in some examples, it takes longer to measure the result of the action. For example, one or more additional processes may be performed before it becomes possible to measure the result of the action. This action and its result may indicate whether the output was a correct output, and / or the difference between what the output should have been and what the output was. Thus, the action and its result may be used to determine a target output that can be used as a label for sensor measurements. After the result of the action has been determined, new training data items may be generated using the input (e.g., sensor data), the output of the trained machine learning model (e.g., chamber control knob update), the target result (e.g., target process result on a substrate), and the measured actual result (e.g., measured process result on a substrate). The trained machine learning model may then be further trained using the new training data items. In some embodiments, this retraining process may be performed on-tool on the manufacturing system (e.g., processing chamber).

[0120] Figure 8 shows a model training workflow 805 and a model application workflow 817 for chamber knob prediction and simulation according to an aspect of the present disclosure. In an embodiment, the model training workflow 805 may be executed on a server, which may or may not include an endpoint detector and / or chamber knob predictor application, which is provided with a trained model, and which endpoint detector and / or chamber knob predictor application may execute the model application workflow 817. The model training workflow 805 and the model application workflow 817 may be executed by processing logic executed by a processor of a computing device (e.g., server 120 in Figure 1). One or more of these workflows 805, 817 may be executed, for example, by one or more machine learning modules implemented on the processing device, and / or other software and / or firmware running on the processing device.

[0121] The model training workflow 805 is a workflow for training one or more machine learning models (e.g., deep learning models) to perform one or more tasks related to the endpoint detector and / or chamber knob predictor, such as judgment, prediction, and modification (e.g., a task to determine whether one or more chamber control knobs are properly tuned and ready to transition to production operating mode). The model application workflow 817 is a workflow for applying one or more trained machine learning models to perform tasks such as judgment and / or tuning on chamber data (e.g., raw sensor data, synthetic data indicating the state of the processing chamber). One or more of these machine learning models may receive process result data (e.g., endpoint detection conditions and / or chamber knob settings).

[0122] This specification describes various machine learning outputs. Certain numbers and configurations of machine learning models are described and illustrated. However, it should be understood that the same or similar final results can be achieved by changing the number and types of machine learning models used, as well as the configuration of such models. Therefore, the configurations of machine learning models described and illustrated are merely examples and should not be interpreted as limiting.

[0123] In some embodiments, one or more machine learning models are trained to perform one or more of the following tasks. Each task may be performed by a separate machine learning model. Alternatively, a single machine learning model may perform each or a subset of those tasks. In addition, or instead, different machine learning models may be trained to perform different combinations of these tasks. In one example, one or a few machine learning models may be trained, and the trained machine learning (ML) model is a single shared neural network having multiple shared layers and multiple separate higher-order output layers, each of which outputs a different prediction, classification, identification, etc. The tasks that one or more trained machine learning models may be trained to perform are as follows: a. Endpoint detector - As previously discussed, the endpoint detector may receive and process various input data, such as sensor data, pre-processed data, process data, and synthesis data, that indicate the state of the processing chamber during a seasoning process (e.g., processing a series of seasoning wafers). The endpoint detector may output various values ​​corresponding to various endpoint detection setpoints in the chamber control knob tuning procedure, indicating that the substrate processed according to the process procedure after chamber control knob tuning should result in a substrate having a process outcome that satisfies threshold conditions related to the substrate process (e.g., the substrate and / or manufacturing recipe). b. Chamber Knob Predictor - The chamber knob predictor may receive data and generate chamber control knob updates related to the recovery of the processing chamber after preventive maintenance and / or seasoning procedures. The chamber knob predictor may determine updates to one or more chamber control knobs related to a substrate processing recipe performed in the processing chamber. For example, the chamber knob predictor may identify updates to process parameters related to a processing chamber corresponding to a substrate process recipe, such that a substrate processed according to a substrate process recipe and / or process parameter update results in a substrate having a process result that satisfies threshold conditions related to one or more process procedures.

[0124] For the model training workflow 805, the training dataset should be formed using a training dataset that includes hundreds, thousands, tens of thousands, hundreds of thousands or more of chamber data 810 (e.g., sensor data, process data, synthetic data indicating the state of the associated processing chambers). In some embodiments, the chamber data 810 includes process recipe data. For example, the chamber data 810 may include a set of process recipe durations (e.g., trim times) corresponding to a set of process recipe operations. In some embodiments, the chamber data 810 includes trim times for sequentially recorded process recipe operations. In some embodiments, the chamber data includes a set of data indicating a process recipe (e.g., a set of data indicating a set of process recipe operations) (see Figure 4B). In some embodiments, the training dataset may further include relevant process result data 812 for forming the training dataset (e.g., measured parameters of the substrate (e.g., limit dimensions, uniformity requirements, film thickness results, etc.)), and each data point may include various labels or classifications of one or more types of useful information. In some embodiments, the training dataset may further include relevant endpoint data 814 (e.g., optical measurements related to the endpoint of a chamber control knob tuning procedure). Each case may include, for example, data indicating one or more processing chambers undergoing a seasoning process, as well as relevant process results of the substrate evaluated during and / or after the seasoning process. This data may be processed to generate one or more training datasets 836 for training one or more machine learning models. The machine learning models may be trained to automate, for example, chamber control knob predictions (e.g., updates to one or more parameters related to the process recipe of the processing chamber), and / or to predict chamber process knob settings after a chamber seasoning procedure and / or a chamber preventive maintenance procedure.

[0125] In one embodiment, generating one or more training datasets 836 includes performing a chamber recovery process 200A by performing preventive maintenance 202, seasoning 206, control knob tuning 208, and evaluating a sample test 210 to determine whether the tuning of one or more chamber control knobs meets a threshold condition (e.g., block 212). One or more labels may be used in different iterations of the chamber recovery process 200A to identify the sample test 208 as meeting a threshold condition for initiating a production mode for the processing chamber and / or initiating additional chamber control knob tuning corresponding to the processing chamber. The labels used may depend on how a particular machine learning model is to be trained. In some embodiments, the chamber data, process results and / or sample test results may be represented as vectors, as described in other embodiments, and the process rate may be represented as one or more matrices.

[0126] To perform training, the processing logic inputs the training dataset 836 into one or more untrained machine learning models. The machine learning models may be initialized before inputting the first input into them. The processing logic trains the untrained machine learning models based on this training dataset to produce one or more trained machine learning models that perform the various operations described above.

[0127] Training may be performed by inputting one or more of the chamber data 810 and process result data 812 into the machine learning model one by one.

[0128] A machine learning model processes this input to produce an output. An artificial neural network includes an input layer consisting of values ​​at data points. The next layer is called a hidden layer, and each node in the hidden layer receives one or more input values ​​from those input values. Each node contains parameters (e.g., weights) to apply to the input values. Thus, each node essentially inputs those input values ​​into a multivariate function (e.g., a nonlinear mathematical transformation) to produce an output value. The next layer may be another hidden layer, or it may be an output layer. In either case, the nodes in the next layer receive output values ​​from the nodes of the previous layer, and each node applies weights to those values ​​and then produces its own output value. This may be done in each layer. The final layer is the output layer, which has one node for each class, prediction, and / or output that the machine learning model can produce.

[0129] Therefore, the output may include one or more predictions or inferences. For example, the output prediction or inference may include one or more changes to the plasma process data (e.g., one or more changes to the plasma exposure time). The processing logic may cause the substrate to process using the updated recipe (e.g., with the identified changes) and the substrate to receive the updated thickness profile. The processing logic may compare the updated thickness profile with the target thickness profile and determine whether the threshold criteria are met (e.g., whether the thickness value measured across the surface of the wafer falls within the target threshold window). The processing logic determines an error (i.e., endpoint error or knob prediction error) based on the difference between the identified chamber endpoint or knob prediction and the evaluated process result (e.g., sample test result). Based on this error, the processing logic adjusts the weights of one or more nodes of the machine learning model. An error term or delta may be determined for each node of the artificial neural network. Based on this error, the artificial neural network adjusts one or more parameters of the artificial neural network parameters (weights for one or more inputs of one node) for one or more nodes of the artificial neural network. Parameters may be updated by backpropagation, starting with the topmost layer's nodes, followed by the nodes of the next layer, and so on. An artificial neural network contains multiple layers of "neurons," each layer receiving values ​​from the neurons of the previous layer as input. The parameters for each neuron include weights associated with the values ​​received from each of the neurons of the previous layer. Therefore, tuning parameters may involve adjusting the weights assigned to each of the inputs to one or more neurons in one or more layers of the artificial neural network.

[0130] After optimizing the model parameters, a model check may be performed to determine whether the model has improved and to determine the current accuracy of the deep learning model. After one or more training iterations, the processing logic may determine whether a stopping criterion has been met. The stopping criterion may be a target accuracy level, a target number of processed images from the training dataset, a target amount of change to the parameters across one or more previous data points, a combination thereof, and / or other criteria. In one embodiment, this stopping criterion is met when at least a minimum number of data points have been processed and at least a threshold accuracy has been achieved. This threshold accuracy may be, for example, 70%, 80%, or 90% accuracy. In one embodiment, this stopping criterion is met when the accuracy improvement of the machine learning model has stopped. If the stopping criterion is not met, additional training is performed. If the stopping criterion is met, training may be completed. After the machine learning model has been trained, the model may be tested using a reserve portion of the training dataset.

[0131] After one or more trained machine learning models 838 are generated, they may be stored in model storage 845 and added to a board process rate determination and / or process tuning application. The board process rate determination and / or process tuning application may then implement an automated mode using the one or more trained ML models 838 and additional processing logic. In the automated mode, user manual input of information is minimized, or in some cases, eliminated altogether.

[0132] In one embodiment of the model application workflow 817, input data 862 may be input to a chamber knob predictor 864, which may include a trained machine learning model. Based on the input data 862, the chamber knob predictor 864 may output a chamber control knob prediction. The chamber control knob prediction may be the chamber control knob settings required for the chamber to meet threshold conditions for operating in a production operating mode (e.g., related to a substrate processing recipe). The chamber knob data 866 may include this chamber control knob prediction.

[0133] According to one embodiment, chamber knob data 866 may be input to an endpoint detector 867, which may include a trained machine learning model. Based on the input data 862, the endpoint detector 867 outputs endpoint detection data 869. The endpoint detection data 869 may include information indicating an endpoint related to the chamber control knob tuning process. This endpoint may indicate that the chamber control knob tuning process is complete. This endpoint may indicate physical conditions of the substrate (e.g., measured values ​​or limit dimensions) corresponding to the completion of the chamber control knob tuning process (for example, the chamber control knob tuning process is complete when the substrate processed according to the results of the chamber control knob tuning process satisfies threshold measured values ​​or threshold limit dimensions). This endpoint may correspond to a condition including a first process result in which the associated substrate processed according to the chamber control knob tuning satisfies the threshold conditions of the substrate processing procedure.

[0134] Figure 9 shows a block diagram of an exemplary computing device 900 operating according to one or more aspects of the present disclosure. In various illustrative examples, various components of the computing device 900 may represent various components of the client device 150, measurement system 110, server 120, data store 140, and machine learning system 170 shown in Figure 1.

[0135] The exemplary computing device 900 may be connected to other computer devices on a LAN, intranet, extranet and / or the Internet. The computing device 900 may operate within the capacity of a server in a client-server network environment. The computing device 900 may be a personal computer (PC), a set-top box (STB), a server, a network router, a switch or bridge, or any device capable of executing (sequentially or otherwise) a set of instructions that specify the actions the device takes. Furthermore, although only a single exemplary computing device is shown, the term “computer” should be interpreted to 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 discussed herein.

[0136] An exemplary computing device 900 may include a processing unit 902 (also called a processor or CPU), main memory 904 (e.g., read-only memory (ROM), flash memory, dynamic random access memory (DRAM), e.g., synchronous DRAM (SDRAM), etc.), static memory 906 (e.g., flash memory, static random access memory (SRAM), etc.), and secondary memory (e.g., data storage device 918), which may communicate with each other via a bus 930.

[0137] The processing unit 902 represents one or more general-purpose processing units, such as a microprocessor, a central processing unit, or other similar devices. More specifically, the processing unit 902 may be a complex instruction set computing (CISC) microprocessor, a reduced instruction set computing (RISC) microprocessor, a very long instruction word (VLIW) microprocessor, a processor that implements other instruction sets, or a processor that implements a combination of instruction sets. The processing unit 902 may further be one or more dedicated processing units, such as an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), a digital signal processing processor (DSP), a network processor, or other similar devices. According to one or more aspects of this disclosure, the processing unit 902 may be configured to execute instructions that implement methods 600, 700 shown in Figures 6-7.

[0138] The exemplary computing device 900 may further include a network interface device 908, which may be communicatively coupled to a network 920. The exemplary computing device 900 may further include a video display 910 (e.g., a liquid crystal display (LCD), a touchscreen, or a cathode ray tube (CRT)), a character / number input device 912 (e.g., a keyboard), a cursor control device 914 (e.g., a mouse), and an acoustic signal generator 916 (e.g., a speaker).

[0139] The data storage device 918 may include a computer-readable storage medium (or more specifically, a non-transient machine-readable storage medium) 928 containing one or more sets of executable instructions 922. According to one or more aspects of the present disclosure, the executable instructions 922 may include executable instructions related to performing the methods 600, 700 shown in Figures 6-7.

[0140] The executable instruction 922 may also reside entirely or at least partially in the main memory 904 and / or the processing unit 902 while the executable instruction 922 is being executed by the exemplary computing device 900, and the main memory 904 and the processing unit 902 may also constitute a computer-readable storage medium. The executable instruction 922 may also be transmitted or received across a network via the network interface device 908.

[0141] Although Figure 9 shows the computer-readable storage medium 928 as a single medium, the term “computer-readable storage medium” should be interpreted to include a single medium or multiple mediums (e.g., a centralized or distributed database and / or associated caches and servers) that store one or more sets of operating instructions. The term “computer-readable storage medium” should further be interpreted to include any medium capable of storing or encoding a set of instructions for a machine to execute, causing that machine to execute one or more of the methods described herein. Thus, the term “computer-readable storage medium” should be interpreted to include, but are not limited to, solid memory and optical and magnetic media.

[0142] Some parts of the detailed explanation above are presented as algorithms and symbolic representations of operations on data bits in computer memory. These algorithmic descriptions and representations are the means by which data processing technicians most effectively communicate the essence of their work to other technicians. Here, an algorithm is generally considered to be a consistent set of steps leading to a desired result. These steps are steps that require the physical manipulation of physical quantities. While not always, these quantities usually take the form of electrical or magnetic signals that can be stored, transferred, combined, compared, and otherwise manipulated. Sometimes, primarily for reasons of general use, it is convenient to refer to these signals as bits, values, elements, symbols, characters, terms, numbers, or other homogeneous things.

[0143] However, it should be noted that these terms and similar terms are all associated with appropriate physical quantities and are merely convenient labels applied to those quantities. Unless otherwise specified, as will be evident from the following discussion, discussions throughout this explanation using terms such as “identify,” “determine,” “store,” “adjust,” “produce,” “return,” “comparise,” “generate,” “stop,” “load,” “copy,” “throw,” “exchange,” “execute,” or other similar terms are understood to refer to the operations and processes of a computer system or similar electronic computing device that manipulate data in the registers and memory of the computer system, expressed as physical (electronic) quantities, and convert such data into other data in the computer system memory or registers or other such information storage devices, transmitters or display devices, also expressed as physical quantities.

[0144] Examples of the present disclosure further relate to apparatus for performing the methods described herein. Such apparatus may be specifically constructed for a required purpose, or it may be a general-purpose computer system selectively programmed by a computer program stored in the computer system. Such computer programs may be stored in computer-readable storage media, including, but not limited to, any type of disk, including optical disks, compact disk read-only memory (CD-ROM) and magneto-optical disks, read-only memory (ROM), random access memory (RAM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), magnetic disk storage media, optical storage media, flash memory devices, other types of machine-accessible storage media, or any type of media suitable for storing electronic instructions, each coupled to a computer system bus.

[0145] The methods and displays described herein are not inherently related to any particular computer or other device. Various general-purpose systems may be used with the programs in accordance with the teachings herein, or it may be more convenient to construct a more specialized device to perform the necessary method steps. The structures required for these various systems are described below. In addition, the scope of this disclosure is not limited to any particular programming language. It will be understood that various programming languages ​​may be used to carry out the teachings of this disclosure.

[0146] It should be understood that the above description is intended to be illustrative and not limiting. Many other examples of embodiments will be obvious to those skilled in the art who have read and understood the above description. Although specific examples are described in this disclosure, it will be recognized that the systems and methods of this disclosure are not limited to the examples described herein and may be implemented with modifications included in the appended claims. Accordingly, this specification and drawings should be construed as illustrative and not limiting. Accordingly, the scope of this disclosure should be determined in relation to the full scope of the appended claims and the equivalents to which such claims are claimed.

Claims

1. The processing apparatus receives (i) sensor data indicating a first state of the coating of the processing chamber processing the substrate after a chamber recovery procedure performed on the processing chamber, and (ii) substrate process data indicating a set of process parameter values ​​related to performing a substrate processing procedure with the processing chamber having the coating in a second state prior to the chamber recovery procedure. The process involves using one or more machine learning models (MLMs) to process the sensor data and the substrate process data to determine one or more outputs, wherein the one or more outputs include a first update to at least one process parameter value from a set of process parameter values, and the first update is related to performing the substrate processing procedure by the processing chamber having the coating in the first state, and The processing apparatus performs at least one of the following: a) preparing a notification instructing the first update for presentation on a graphical user interface (GUI), or b) causing the processing chamber to perform the selection of the substrate processing procedure based on the first update. A method that includes this.

2. To determine the first endpoint detection condition corresponding to the first update, and The process chamber is made to execute the selection of the substrate processing procedure using the first endpoint detection condition. The method according to claim 1, further comprising:

3. The method according to claim 2, wherein the first endpoint detection condition corresponds to the optical sensor measurement value of the coating of the processing chamber.

4. Determining the first endpoint detection condition includes applying a deterministic model to the first update to generate the first endpoint detection condition, wherein the deterministic model is generated using one or more history process parameters and one or more history endpoint detection conditions associated with the one or more history process parameters. The method according to claim 2.

5. The method according to claim 4, wherein the deterministic model includes a mechanism model generated by performing linear exponential regression using at least one or more history process parameters and one or more history endpoint detection conditions.

6. The substrate processing procedure includes a plurality of process operations, The first value among the set of process parameter values ​​corresponds to the first process operation among the plurality of process operations. The first update corresponds to the first value, The second value of the set of process parameter values ​​corresponds to the second process operation among the plurality of process operations. The one or more outputs further include a second update corresponding to the second value, The method according to claim 1.

7. The first process operation and the first endpoint detection conditions corresponding to the processing chamber are determined based on the first update. The second process operation and the second endpoint detection conditions corresponding to the processing chamber are determined based on the second update, and The processing chamber is made to execute the selection of the substrate processing procedure using the first endpoint detection condition and the second endpoint detection condition. The method according to claim 6, further comprising:

8. Processing the aforementioned sensor data and the aforementioned substrate process data Using the first MLM among the one or more MLMs, the sensor data and the substrate process data are processed to obtain a first model output. To process the first model output using a second MLM from the one or more MLMs mentioned above to obtain a second model output, and The first update is determined based at least on the combination of the first model output and the second model output. The method according to claim 1, further comprising:

9. The method according to claim 1, wherein at least one of the one or more MLMs is trained using at least one of Least Absolute Shrinkage and Selection Operator (LASSO) regression or Ridge regression.

10. The method according to claim 1, wherein the chamber recovery procedure includes a chamber seasoning procedure performed after a preventive maintenance (PM) procedure.

11. The method according to claim 1, wherein the set of process parameter values ​​includes one or more values ​​indicating at least one of process temperature, process pressure, or process duration.

12. The processing apparatus receives training data comprising: (i) first sensor data indicating a first state of the coating of the first processing chamber processing the substrate after a chamber recovery procedure performed on the first processing chamber; (ii) first substrate process data indicating a first set of process parameter values ​​related to performing a first substrate processing procedure with the first processing chamber having the coating in a second state prior to the chamber recovery procedure; and (iii) second substrate process data indicating a first update to at least one of the first set of process parameter values, wherein the first update is related to performing the first substrate processing procedure with the first processing chamber having the coating in the first state, and The processing apparatus trains one or more machine learning models (MLMs) using (i) input data including the first sensor data and the first substrate process data, and (ii) a target output including the second substrate process data, wherein the trained one or more MLMs receive a new input having new sensor data indicating a third state of the coating in a new processing chamber processing a new substrate after the chamber recovery procedure, and third substrate process data indicating a second set of process parameter values ​​related to performing the first substrate processing procedure with the new processing chamber having the coating in a fourth state prior to the chamber recovery procedure, and generates a new output based on the new input, wherein the new output indicates a second update to at least one of the second set of process parameter values, and the second update is related to performing the first substrate processing procedure with the new processing chamber having the coating in a third state. A method that includes this.

13. Perform a first regression using a first subset of the training data to generate a first MLM. Perform a second regression using a second subset of the training data to generate a second MLM, and The determination is made based on a comparison of the first MLM, the second MLM, and the training data, that the first accuracy of the first MLM is greater than the second accuracy of the second MLM, and that the one or more trained MLMs include the first MLM. The method according to claim 12, further comprising:

14. The method according to claim 12, further comprising training the one or more MLMs by performing gradient boosting regression (GBR) using the training data.

15. The first substrate processing procedure includes a series of process operations, and each of the second set of process parameter values ​​is related to one of the process operations in the series of process operations. The second update instructs an update to the first process parameter value of the second set of process parameter values, the first process parameter value corresponds to the first process operation of the series of process operations, The new output further instructs a third update to the second process parameter value of the second set of process parameter values, wherein the second process parameter value corresponds to the second process operation of the series of process operations. The method according to claim 12.

16. A non-transient machine-readable storage medium containing instructions, wherein when the instructions are executed by a processing device, the processing device causes the processing device to perform an operation, and the operation is (i) receiving sensor data indicating a first state of the coating of the processing chamber processing the substrate after a chamber recovery procedure performed on the processing chamber, and (ii) receiving substrate process data indicating a set of process parameter values ​​related to performing a substrate processing procedure with the processing chamber having the coating in a second state prior to the chamber recovery procedure. The determination involves processing the sensor data and the substrate process data using one or more machine learning models to determine one or more outputs, wherein the one or more outputs include a first update to at least one process parameter value from a set of process parameter values, and the first update is related to performing the substrate processing procedure by the processing chamber having the coating in the first state, and a) providing a notification instructing the first update and preparing the notification for presentation on a graphical user interface (GUI), or b) causing the processing chamber to perform the selection of the substrate processing procedure based on the first update, at least one of the above. Non-transient, machine-readable storage media, including [specific data / information].

17. The aforementioned operation, To determine the first endpoint detection condition corresponding to the first update, and The process chamber is made to execute the selection of the substrate processing procedure using the first endpoint detection condition. A non-transient machine-readable storage medium according to claim 16, further comprising:

18. The substrate processing procedure includes a plurality of process operations, The first value among the set of process parameter values ​​corresponds to the first process operation among the plurality of process operations. The first update corresponds to the first value, The second value of the set of process parameter values ​​corresponds to the second process operation among the plurality of process operations. The one or more outputs further include a second update corresponding to the second value, The non-transient machine-readable storage medium according to claim 16.

19. The aforementioned operation, The first process operation and the first endpoint detection conditions corresponding to the processing chamber are determined based on the first update. The second process operation and the second endpoint detection conditions corresponding to the processing chamber are determined based on the second update, and The processing chamber is made to execute the selection of the substrate processing procedure using the first endpoint detection condition and the second endpoint detection condition. A non-transient machine-readable storage medium according to claim 18, further comprising:

20. The aforementioned operation, Using the first MLM among the one or more MLMs, the sensor data and the substrate process data are processed to obtain a first model output. To process the first model output using a second MLM from the one or more MLMs mentioned above to obtain a second model output, and The first update is determined based at least on the combination of the first model output and the second model output. A non-transient machine-readable storage medium according to claim 16, further comprising:

Images