Maintenance of modules for light sources used in semiconductor photolithography.

Abstract

A system for maintaining a light source for semiconductor photolithography, in which a module forming part of the light source is evaluated at various pulse counts to generate a binary prediction as to whether the module is sufficiently likely to operate without fault in a subsequent sequence of pulses. The binary prediction may be made by a machine learning model trained on metrics extracted from measurements made on the removed module. A group of models, each trained differently, may be made available according to a selection made by a user or according to the user's maintenance goals.

Description

Technical Field

[0001] Cross - Reference to Related Applications

[0001] This application claims priority to U.S. Patent Application No. 63 / 162,249, entitled "MAINTENANCE OF MODULES FOR LIGHT SOURCES USED IN SEMICONDUCTOR PHOTOLITHOGRAPHY," filed on March 17, 2021, and U.S. Patent Application No. 63 / 188,020, entitled "MAINTENANCE OF MODULES FOR LIGHT SOURCES USED IN SEMICONDUCTOR PHOTOLITHOGRAPHY," filed on May 13, 2021, the entire contents of both of which are incorporated herein by reference.

[0002]

[0002] The subject matter disclosed herein relates to the maintenance of light sources such as those used in integrated circuit photolithography manufacturing processes.

Background Art

[0003]

[0003] Laser radiation for semiconductor photolithography is typically supplied by a system referred to as a light source. These light sources generate radiation as a series of pulses, for example, at a specified repetition rate within the range of about 500 Hz to about 6 kHz. They are conventionally expected to have a useful life, typically measured in terms of the number of pulses that can be generated before repair or replacement is required, which is typically on the order of billions of pulses.

[0004]

[0004] One system for generating laser radiation at frequencies useful for semiconductor photolithography (deep ultraviolet (DUV) wavelengths) requires the use of a main oscillator power amplifier (MOPA) dual gas discharge chamber configuration. This configuration has two chambers: a main oscillator chamber (MO chamber) and a power amplifier chamber (PA chamber). These chambers and many other system components can be considered modules, and the entire light source can be considered a collection of modules. Each module will generally have a shorter lifespan than the entire system. Therefore, the health status of individual modules must be evaluated over the entire lifespan of the system in order to determine whether they should be repaired or replaced.

[0005]

[0005] The timing of module maintenance is determined according to the maintenance strategy. The earliest maintenance strategy is unplanned maintenance (performed in response to failure), in which maintenance is not performed until module failure occurs. While this strategy may increase the utilization rate of components to some extent, unplanned module failures can have a significant economic impact on the entire production line, resulting in unplanned downtime and costs.

[0006]

[0006] Another maintenance strategy is preventative maintenance, in which maintenance actions are carried out according to a planned schedule based on time series (i.e., time since service commencement) or machine time, and components are maintained in periodic increments to reduce unexpected machine failure. However, the practice of periodic inspection / maintenance can result in unnecessarily long downtime and high maintenance costs. Another maintenance strategy is condition-based maintenance, in which maintenance actions are carried out after the identification of one or more conditions that indicate deterioration in the operation of a module.

[0007]

[0007] Predictive maintenance (PdM) is a maintenance strategy designed to monitor the conditions of equipment in service in order to predict when the equipment will fail. The future behavior / conditions of the machine components are approximated, making it possible to optimize maintenance tasks (e.g., prognosis monitoring). Thus, machine downtime and maintenance costs can be significantly reduced while performing maintenance as infrequently as possible. The PdM system enables the advance detection of pending faults and allows for timely pre-fault intervention using predictive tools based on historical data.

[0008]

[0008] Typically, end users of such systems and field technicians of tool manufacturers manually monitor the tools using responsive reports and interfaces, meaning they are based on historical and current performance. Furthermore, these reports and interfaces tend to be global or universal, in the sense that they use global parameters as opposed to customer-specific parameters to evaluate the tool's ability to meet maintenance intervals. These global parameters and reports are generally not customizable on a per-customer basis.

[0009]

[0009] It is desirable to implement a maintenance strategy for light sources with module lifetimes measured in billions of pulses in a manner that produces the best possible availability without affecting the technical performance of the light source. [Overview of the project]

[0010]

[0010] Below, a simplified overview of one or more embodiments is provided to give a basic understanding of the present invention. This overview is not intended to be an extensive summary of all intended embodiments, nor to identify the main or important elements of all embodiments, nor to detail the scope of any or all embodiments. The sole purpose is to present some concepts of one or more embodiments in a simplified form as a preliminary step to the more detailed descriptions that will be presented later.

[0011]

[0011] According to one embodiment, the module is evaluated at various pulse milestones. In each evaluation, a model is used to represent a binary (true / false) prediction of whether the module is likely to operate without failure during a predicted increment consisting of a given number of pulses immediately following it. In other words, the model predicts whether the module is likely to survive another predicted increment or to fail before the end of the predicted increment. The model may be a mathematical model of its components or a model trained using machine learning.

[0012]

[0012] According to another aspect of one embodiment, a method for maintaining a light source for semiconductor photolithography is disclosed, the light source comprising one or more modules, the method comprising: determining whether one of the one or more modules is based on an initial evaluation based at least in part on a first number of pulses in which the module was involved in generating; performing the initial evaluation to determine whether the module has at least a minimum probability of fault-free operation in a predicted increment measured as a second number of pulses; and if the initial evaluation determines that the module does not have at least a minimum probability of fault-free operation in the predicted increment, generating a module failure alert for the module, or if the initial evaluation determines that the module has at least a minimum probability of fault-free operation in the predicted increment, leaving the module in service. The method further comprises removing the module if the initial evaluation determines that the module does not have at least a minimum probability of fault-free operation in the predicted increment. The first number may be about 10 billion pulses. The second number may be about 2 billion pulses.

[0013]

[0013] The method may further include the steps of: leaving the module in service; determining whether the module is due to an additional evaluation based at least in part on a third number of pulses the module has been involved in generating since the initial evaluation; performing an additional evaluation to determine whether the module has at least a minimum probability of fault-free operation in a second predicted increment measured as a fourth number of pulses; and, if the additional evaluation determines that the module does not have at least a minimum probability of fault-free operation in the second predicted increment, generating a module failure alert for the module; or, if the additional evaluation determines that the module has at least a minimum probability of fault-free operation in the second predicted increment, leaving the module in service. After these further steps, the method may further include removing the module if the additional evaluation determines that the module does not have at least a minimum probability of fault-free operation in the second predicted increment. The third number may be approximately 100 million pulses. The fourth number may be approximately 2 billion pulses.

[0014]

[0014] Performing an initial evaluation may involve using a model. The model may be an analytical model. The model may be a trained model developed via machine learning by supplying feature data to train a trained model, which makes an initial decision based on at least some of the feature data. The module may comprise a main oscillator chamber module, in which case the feature data may include several main oscillator-related energy batch quality events in the preceding 100 million pulses and / or the average main oscillator energy in the preceding 100 million pulses.

[0015]

[0015] In another aspect of one embodiment, a computer implementation method is disclosed, which includes: determining by a computing device whether a module, which is one of more modules of a photolithography light source, is based on an initial evaluation that is at least partially based on a first number of pulses in which the module was involved in generating; performing an initial evaluation by the computing device to determine whether the module has at least a minimum probability of fault-free operation in a predicted increment measured as a second number of pulses; and if the initial evaluation determines that the module does not have at least a minimum probability of fault-free operation in the predicted increment, providing an instruction by the computing device that the module should be removed, or if the initial evaluation determines that the module has at least a minimum probability of fault-free operation in the predicted increment, providing an instruction by the computing device that the module should be left in service.

[0016]

[0016] In another aspect of one embodiment, a non-temporary computer-readable storage medium is disclosed which includes executable instructions for causing a processor to perform an operation, the instructions including instructions for determining whether a module, which is one or more modules of a photolithography light source, is based at least in part on an initial evaluation based on a first number of pulses in which the module was involved in generating; instructions for performing an initial evaluation to determine whether the module has at least a minimum probability of fault-free operation in a predicted increment measured as a second number of pulses; and instructions for providing an instruction that the module should be removed if the initial evaluation determines that the module does not have at least a minimum probability of fault-free operation in the predicted increment, or instructions for providing an instruction that the module should be left in service if the initial evaluation determines that the module has at least a minimum probability of fault-free operation in the predicted increment.

[0017]

[0017] According to another aspect of one embodiment, a method for maintaining a light source for semiconductor photolithography is disclosed, the light source comprising one or more modules, the method comprising: identifying a module which is one or more of the modules for evaluation; performing an evaluation to determine whether the module has at least a minimum probability of fault-free operation in a predicted increment measured as a second number of pulses; and if the initial evaluation determines that the module does not have at least a minimum probability of fault-free operation in the predicted increment, generating a module failure alert for the module, or if the initial evaluation determines that the module has at least a minimum probability of fault-free operation in the predicted increment, leaving the module in service. The method may further comprise removing the module if the initial evaluation determines that the module does not have at least a minimum probability of fault-free operation in the predicted increment. The first number may be approximately 10 billion pulses. The second number may be approximately 2 billion pulses. Performing the evaluation may involve using a trained model developed via machine learning by supplying feature data to train the trained model, which then determines, based on at least some of the feature data, whether the module has at least the minimum probability of fault-free operation in a predicted increment measured as a second pulse count.

[0018]

[0018] According to another aspect of one embodiment, a method for maintaining a light source for semiconductor photolithography is disclosed, the light source comprising one or more modules, the method comprising: obtaining user information indicating the user's relative prioritization of two or more maintenance preferences; training at least two models, including a first model based on a first relative prioritization of two or more maintenance preferences and a second model based on a second relative prioritization of two or more maintenance preferences; performing an evaluation to determine whether a module failure alert should be generated, the evaluation being performed using one of the at least two models based on which model's relative prioritization is most closely aligned with the user's relative prioritization; and generating a module failure alert for the module if the evaluation determines that a module failure alert should be generated. Performing an evaluation using one of the at least two models based on which model's relative prioritization is most closely aligned with the user's relative prioritization may include performing an evaluation using one of the at least two models selected by the user. Maintenance preferences may include maximizing output and avoiding unexpected downtime. The method may further include indicating that the module should remain in service if the evaluation determines that a module failure alert should not be generated. The method may further include determining whether the module is in service based at least partially on the number of pulses the module was involved in generating. Performing an evaluation to determine whether a module failure alert should be generated may include selecting a model from at least two models based on the user's relative priorities to represent a binary (true / false) decision on whether a module failure alert should be generated. The method may further include performing maintenance actions on the module if the evaluation determines that a module failure alert should be generated. Performing maintenance actions may include removing the module. Performing maintenance actions may include repairing the module.The model can be a trained model developed via machine learning by supplying feature data to train a trained model, and the selected trained model makes decisions based on at least some of the feature data.

[0019]

[0019] According to another aspect of one embodiment, a computer implementation method is disclosed, which includes: using a computing device to store user information indicating a user's relative prioritization of two or more maintenance preferences; using a computing device to train at least two models, including a first model based on a first relative prioritization of two or more maintenance preferences and a second model based on a second relative prioritization of two or more maintenance preferences; using a computing device to select one of the models as the model selected based on the user information; using the selected model on the computing device to perform an evaluation to determine whether a module failure alert should be generated; and if the evaluation determines that a module failure alert should be generated, the computing device provides an instruction that the module should be removed, or if the evaluation determines that a module failure alert should not be generated, the computing device provides an instruction that the module should remain in service. Maintenance preferences may include output maximization and avoidance of unexpected downtime.

[0020]

[0020] According to another aspect of one embodiment, a non-temporary computer-readable storage medium is disclosed which includes executable instructions for causing a processor to perform an operation, the instructions including instructions for storing user information indicating a user's relative prioritization of two or more maintenance preferences; instructions for training at least two models, including a first model based on a first relative prioritization of two or more maintenance preferences and a second model based on a second relative prioritization of two or more maintenance preferences; instructions for selecting one of the models as a model selected based on the user information; instructions for performing an evaluation using the selected model to determine whether a module failure alert should be generated; and instructions for providing instructions that, if the evaluation determines that a module failure alert should be generated, the module should undergo maintenance procedures, or instructions for providing instructions that, if the evaluation determines that a module failure alert should not be generated, the module should remain in service. Maintenance preferences may include output maximization and avoidance of unexpected downtime.

[0021]

[0021] According to another aspect of one embodiment, a system for maintaining a light source for semiconductor photolithography is disclosed, the light source comprising one or more modules, the system comprising: a user preference data storage unit adapted to store user preference data relating to a user's relative prioritization of two or more maintenance preferences; an evaluation timing unit adapted to determine whether one of the one or more modules is based at least in part on an evaluation of a first number of pulses in which the module was involved in generating; a first model based on a first relative prioritization of two or more maintenance preferences, and a second model based on a second relative prioritization of two or more maintenance preferences The system comprises a model training unit adapted to train a model; a model selection unit adapted to select one of the models as the selected model based on user preference data; a binary prediction unit adapted to respond to an evaluation timing unit and a user preference input unit, and to perform an evaluation by determining whether a module failure alert should be generated using the selected model; and a module failure alert generation unit adapted to respond to the binary prediction unit, and to generate a module failure alert for the module if the evaluation determines that a module failure alert should be generated. Maintenance preferences may include output maximization and avoidance of unexpected downtime. The module failure alert generation unit may, in addition, be configured to generate a positive no-failure indication if the binary prediction unit determines that a module failure alert should not be generated.

[0022]

[0022] Further features and advantages of the present invention, as well as the structure and operation of various embodiments of the present invention, will be described in detail below while referring to the accompanying drawings. It should be noted that the present invention is not limited to the specific embodiments described herein. Such embodiments are presented herein only for illustrative purposes. Those skilled in the art will recognize additional embodiments based on the teachings contained herein.

[0023]

[0023] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate the present invention and, together with the description, serve to explain the principles of the present invention and enable those skilled in the art to make and use the present invention.

Brief Description of the Drawings

[0024] [Figure 1]

[0024] A schematic view, not to scale, of the general broad concept of a photolithography system according to one aspect of the disclosed subject matter is shown. [Figure 2]

[0025] A schematic view, not to scale, of the general broad concept of an illumination system according to one aspect of the disclosed subject matter is shown. [Figure 3]

[0026] A diagram showing maintenance procedures according to one aspect of one embodiment. [Figure 4]

[0027] A flowchart showing maintenance procedures according to one aspect of one embodiment. [Figure 5]

[0028] A diagram showing procedures for creating, training, and using a model to be used for maintenance according to one aspect of one embodiment. [Figure 6]

[0029] A functional block diagram showing a maintenance system according to one aspect of one embodiment. [Figure 7]

[0030] A functional block diagram showing a maintenance system according to another aspect of one embodiment. [Figure 8]

[0031] This is a graph showing various possible combinations of prioritization according to one aspect of one embodiment. [Figure 9]

[0032] This is a functional block diagram showing a possible arrangement for a maintenance system according to one aspect of one embodiment. [Figure 10]

[0033] This is a functional block diagram showing a possible arrangement for a maintenance system according to one aspect of one embodiment. [Modes for carrying out the invention]

[0025]

[0034] The features and advantages of the present invention will become more apparent from the detailed description below, which is made in conjunction with the drawings, and similar reference letters throughout will identify corresponding elements. In the drawings, similar reference numbers generally indicate identical, functionally similar, and / or structurally similar elements.

[0026]

[0035] Next, various embodiments will be described with reference to the drawings, and similar reference numerals will be used throughout to refer to similar elements. In the following description, numerous specific details are given for illustrative purposes to facilitate a full understanding of one or more embodiments. However, in some or all instances, it may be apparent that any embodiment described below can be realized without employing the specific design details described below. In other instances, well-known structures and devices are shown in block diagram form to facilitate the description of one or more embodiments.

[0027]

[0036] To provide a basic understanding of the embodiments, a simplified overview of one or more embodiments is provided below. This overview is not intended to be a comprehensive overview of all conceived embodiments, nor is it intended to identify the main or essential elements of all embodiments, nor to delineate the scope of any or all embodiments.

[0028]

[0037] The embodiments described herein, and any references such as “one embodiment,” “embodiment,” “exemplary embodiment,” or “exemplary embodiment” in this specification, indicate that the embodiments described may include certain features, structures, or characteristics, but not all embodiments may necessarily include certain features, structures, or characteristics. Furthermore, such terms do not necessarily refer to the same embodiment. Moreover, when certain features, structures, or characteristics are described in relation to an embodiment, it is understood that it is within the knowledge of those skilled in the art to produce such features, structures, or characteristics in relation to other embodiments, whether or not they are explicitly described.

[0029]

[0038] Embodiments of the present invention may be implemented in hardware, firmware, software, or any combination thereof. Embodiments of the present invention may also be implemented as instructions stored on a machine-readable medium that can be read and executed by one or more processors. The machine-readable medium may include any mechanism for storing or transmitting information in a form readable by a machine (e.g., a computing device). For example, the machine-readable medium may include read-only memory (ROM), random access memory (RAM), magnetic disk storage media, optical storage media, flash memory devices, electrical, optical, acoustic, or other forms of propagating signals (e.g., carrier waves, infrared signals, digital signals, etc.), and others. Furthermore, firmware, software, routines, and instructions may be described herein as performing specific actions. However, such descriptions are merely for convenience, and it should be understood that such actions actually originate from computing devices, processors, controllers, or other devices that perform firmware, software, routines, instructions, etc.

[0030]

[0039] Referring to Figure 1, the photolithography system 100 includes an illumination system 105. As will be described in more detail below, the illumination system 105 generates a pulsed light beam 110, which is directed to a photolithography exposure apparatus or scanner 115 that patterns microelectronic features onto a wafer 120. The wafer 120 is placed on a wafer table 125 connected to a positioner, which is constructed to hold the wafer 120 and configured to precisely position the wafer 120 according to specific parameters.

[0031]

[0040] The photolithography system 100 may use a light beam 110 having a wavelength in the deep ultraviolet (DUV) range, for example, 248 nanometers (nm) or 193 nm. The size of the microelectronic features patterned on the wafer 120 depends on the wavelength of the light beam 110, with lower wavelengths resulting in smaller minimum feature sizes. When the wavelength of the light beam 110 is 248 nm or 193 nm, the minimum size of the microelectronic features can be, for example, 50 nm or less. The bandwidth of the light beam 110 can be the actual instantaneous bandwidth of its light spectrum (or emission spectrum), and includes information on how the light energy of the light beam 110 is distributed across different wavelengths.

[0032]

[0041] The scanner 115 includes an optical configuration having, for example, one or more condenser lenses, a mask, and an objective system configuration. The mask is movable along one or more directions, such as along the optical axis of the light beam 110, or in a plane perpendicular to the optical axis. The objective system configuration includes a projection lens and can transfer an image so that it is produced from the mask onto the photoresist on the wafer 120. The illumination system 105 adjusts the angular range of the light beam 110 that strikes the mask. The illumination system 105 also homogenizes (makes uniform) the intensity distribution of the light beam 110 across the mask.

[0033]

[0042] The scanner 115 may include, among its features, a lithography controller 130, an air conditioning device, and a power supply for various electrical components. The lithography controller 130 controls how layers are printed on the wafer 120. The lithography controller 130 includes memory for storing information such as a processor recipe. The process program or recipe determines the exposure length on the wafer 120, the mask used, and other factors affecting exposure. During lithography, multiple pulses of the light beam 110 illuminate the same area on the wafer 120 to form an illumination dose together.

[0034]

[0043] The photolithography system 100 preferably also includes a control system 135. Generally, the control system 135 includes one or more of the following: digital electronic circuit elements, computer hardware, firmware, and software. The control system 135 also includes memory, which may be read-only memory and / or random-access memory. Suitable storage devices for tangibly embodying computer program instructions and data include, for example, semiconductor memory devices such as EPROM, EEPROM, and flash memory devices; magnetic disks such as internal hard disks and removable disks; magneto-optical disks; and CD-ROM disks, and all forms of non-volatile memory.

[0035]

[0044] The control system 135 may also include one or more input devices (such as a keyboard, touchscreen, microphone, mouse, or handheld input device) and one or more output devices (such as a speaker or monitor). The control system 135 may also include components to enable wireless communication, including Bluetooth, NFC, and Wi-Fi. In particular, the control system 105 may include components that allow the control system to exchange data, commands, etc., with the cloud.

[0036]

[0045] The control system 135 also includes one or more programmable processors and one or more computer program products tangibly embodied in a machine-readable memory device for execution by the one or more programmable processors. Each of the one or more programmable processors is capable of programming instructions to perform a desired function by operating on input data and generating appropriate outputs. Generally, the processors receive instructions and data from memory. Any of the above can be supplemented or incorporated by specially designed ASICs (Application-Specific Integrated Circuits). The control system 135 can be centralized or distributed partially or entirely throughout the photolithography system 100.

[0037]

[0046] Referring to Figure 2, the exemplary illumination system 105 is a pulsed laser source that generates a pulsed laser beam as an optical beam 110. Figure 2 generally shows one particular set of components or modules and optical paths solely to facilitate the explanation of the broader principles of the present invention, and it will become clear to those skilled in the art that the principles of the present invention can be advantageously applied to lasers having other modules and configurations.

[0038]

[0047] The illumination system 105 may include, for example, a solid-state or gas discharge seed laser system 140, a power amplification ("PA") stage, such as a power ring amplifier ("PRA") stage 145, a relay optical system 150, and a laser system output subsystem 160. The seed system 140 may include, for example, a main oscillator ("MO") chamber module 165, which discharges between electrodes (not shown) to generate relatively broadband radiation that can be a line narrowed to a relatively very narrow bandwidth, and a central wavelength selected in a line narrowing module ("LNM") 170, as known in the art, causing a lazing gas discharge in the lazing gas to create a population inversion of high-energy molecules, such as Ar, Kr, or Xe.

[0039]

[0048] The seed laser system 140 may also include a main oscillator output coupler ("MO OC") 175, which may comprise a partial reflective mirror formed with a reflective grating (not shown) in the LNM 170, and an oscillator cavity from which the seed laser 140 oscillates to form a seed laser output pulse, i.e., to form a main oscillator ("MO"). The system may also include a line center analysis module ("LAM") 180. The LAM 180 may include, for example, an etalon spectrometer for fine wavelength measurements and a coarse resolution grating spectrometer. The MO wavefront engineering box ("WEB") 185 may function to redirect the output of the MO seed laser system 140 toward the amplification stage 145, and may include, for example, a beam expander, and coherence busting in the form of, for example, an optical delay path (not shown), with, for example, a multiprism beam expander (not shown).

[0040]

[0049] The amplification stage 145 may be incorporated within the PRA WEB 210 and may also be an oscillator formed by, for example, a seed beam incidence and output coupling optical system (not shown), which can be inversely redirected by the beam reservoir 220 through the gain medium in the chamber 200, and may include, for example, a PRA racing chamber module 200. The PRA WEB 210 may incorporate a partial reflection input / output coupler (not shown), a maximum reflection mirror at the nominal operating wavelength (e.g., approximately 193 nm for an ArF system), and one or more prisms.

[0041]

[0050] The Bandwidth Analysis Module ("BAM") 230 at the output of the amplification stage 145 can receive the pulsed output laser beam from the amplification stage and may pick off a portion of the beam for metronome purposes, for example, to measure the output bandwidth and pulse energy. The pulsed laser output beam then passes through the Optical Pulse Stretcher ("OPuS") 240 and the Output Combination Auto-Shutter Metrology Module ("CASMM") 250, which may also be the location of the pulse energy meter. One purpose of the OPuS 240 may be, for example, to convert a single output laser pulse into a pulse train. Secondary pulses created from the original single output pulse may be delayed relative to each other. By distributing the original laser pulse energy within the train of secondary pulses, it is possible to extend the effective pulse length of the laser and, at the same time, reduce the peak pulse intensity. Thus, the OPuS 240 can receive the laser beam from the PRA WEB 210 via the BAM 230 and direct the output of the OPuS 240 to the CASMM 250.

[0042]

[0051] The overall availability of a light source (e.g., lighting system 105) is a direct result of the availability of each individual module that makes up the light source. In other words, a light source cannot be available unless all of the critical modules that make up the light source are available. This requires the use of several forms of maintenance strategies. In addition to the aforementioned maintenance strategies, one technique for maintaining a light source is called umbrella maintenance, in which a group of several modules, some of which may not be faulty, are all replaced simultaneously to optimize the availability of the light source and thereby the productivity of the manufacturing plant.

[0043]

[0052] Using an umbrella maintenance strategy, each module is assumed to have a minimum lifespan or a lifespan that is an integer multiple of the lifespan of another module. For example, module A has a nominal lifespan of 6 months, and module B has a nominal lifespan of 18 months. In this scenario, module B would be replaced every three times module A is replaced.

[0044]

[0053] If the actual module lifespan is shorter than the rated or expected minimum lifespan, the umbrella maintenance strategy can be blocked, potentially creating a cascading effect by disrupting synchronous maintenance schedules for other modules. Modules may also have a potential or actual lifespan that exceeds their rated minimum lifespan, in which case umbrella maintenance may involve removing modules that could still provide additional satisfactory operation. System maintenance events require light sources to be removed from production. Therefore, umbrella maintenance can result in unnecessary disruptions to productivity when the manufacturing plant cannot continue operations in any other way.

[0045]

[0054] To supplement existing field service operations, that is, to provide field technicians with final decisions regardless of any model outcomes, there is a need for methods of maintaining light sources based on the use of systems to provide effective module failure alerts. In some applications, it would be beneficial if these systems were fully automated and provided with real-time assessments.

[0046]

[0055] According to one aspect of one embodiment, as shown in Figure 3, the module is first evaluated after an increment with respect to a number of pulses N1, which is measured in length. The number N1 is selected to be within a range of the module's reasonably expected lifespan with high reliability. In evaluation E1, a decision is made as to whether the probability P1 of the module failing during the predicted increment I1 of additional pulses is less than some predetermined value P0. Or, equivalently, a decision is made as to whether the probability P1 of the module surviving during the predicted increment I1 is greater than some predetermined value P0.

[0047]

[0056] It should be noted that this is a binary, yes / no, true / false determination. The determination does not result in a numerical value of a continuous variable. It simply determines whether the value is within a given range. However, a PdM system typically returns results such as the remaining useful life (RUL) for continuous variables. This form of result may not be very useful in a system designed to perform maintenance on a semiconductor photolithography system. The binary outcome can provide a higher degree of accuracy because any one of the possible values within the space of possible outcomes, that is, within a given range for either true or false, is smaller.

[0048]

[0057] When P1 > P0, a module failure alert is generated and the module can be maintained, i.e., replaced or sustained. If the determination made as a result of the evaluation E1 indicates P1 < P0, the module remains in service until at least the next evaluation E2 after the additional pulse of N2.

[0049]

[0058] In evaluation E2, a determination is made as to whether the probability P2 of the module failing during the measured increment I2 is less than some predetermined value that can be P0. If P0 < P2, the module is replaced. If the determination made as a result of the evaluation E2 indicates P2 < P0, the module remains in service until at least the next evaluation E3 after the additional pulse of N3. This process is repeated until the module is replaced, at which point the process is restarted for the replacement module.

[0050]

[0059] For example, if the predicted increment is 2 billion pulses, the process may first evaluate the module at approximately 10 billion pulses (Bp), and the module may be predicted to either survive or fail at a predicted increment of approximately 2 Bp (i.e., up to approximately 12 Bp). The same module can be evaluated at approximately 10.1 Bp to make another binary prediction regarding failure or survival, such as until up to approximately 12.1 Bp or until the module requires maintenance operations, e.g., removal, repair, or service.

[0051]

[0060] It will be noted that in this and other parts of this specification, including the claims, pulse counts are specified as "approximately" a specific value. Those skilled in the art will understand the degree of discretionary freedom permitted when performing an operation based on a predetermined number of pulse counts, between that predetermined number and the actual number of pulse counts on which the operation is performed. In other words, an operation specified to be performed at 10 Bp can be sufficiently performed at actual pulse counts between, for example, 9.9 Bp and 10.1 Bp. Therefore, for the purposes of this description, the adverb "approximately" should be interpreted as meaning sufficiently close to a nominal value indicating that maintenance will not have adverse effects.

[0052]

[0061] Note that in the example above, the initial evaluation increment is considerably longer than the subsequent evaluation increments. Understand that this is an optional design choice. Also, in the example above, all subsequent evaluation increments are the same length. Similarly, in the example above, all predictive increments are the same length. Generally, one or both of these increments are configurable by the equipment manufacturer or end-user and can vary as a function of, for example, time-series age, cumulative pulse count (i.e., pulse count since service commencement), or a feature indicating module health. Therefore, depending on the cumulative pulse count, the evaluation increment can be shorter or longer, and the predictive increment can be shorter or longer. For example, as the module ages (time-series time or pulse count), the evaluation increment or predictive increment, or both, can be shortened to reflect the increasing probability of module failure with age. Understand that evaluation increments and predictive increments can be of different lengths. Also, in the example above, the failure threshold probability is the same for all predictive increments. It will become clear that different threshold probabilities are available. Furthermore, in the example above, all predicted increments are the same length. Again, it will become clear that predicted increments of different lengths can be used. Also, the process described above relates to performing evaluations at milestones in the pulse count, i.e., at a predetermined number of pulse counts. However, there may be cases where it is desirable to perform evaluations at a time or pulse count other than a predetermined time or pulse count. Therefore, the first step in such a process may be to perform the evaluation "on demand," that is, when it is more desirable than waiting for a specific time, event, or pulse count.

[0053]

[0062] Furthermore, in the example above, only one binary decision is made. However, in some implementations, it may be advantageous to have the ability to make at least one additional binary prediction. This is another advantage of binary prediction results. It is possible to make different predictions for a single evaluation pulse count; for example, assuming the module survives until 22.0 Bp, the first binary prediction would be whether it survives until 25.0 Bp, and the second binary prediction would be whether it survives until 23.0 Bp. A model that produces binary results can be tuned to make both of these predictions. Techniques that predict one of many values ​​in a continuum, such as predicting RUL, can only make one prediction.

[0054]

[0063] Figure 4 illustrates the process using a flowchart. In step S10, it is determined whether the module is in the evaluation pulse count based on the number of pulses involved in its generation. In step S20, the module is evaluated as described herein. This evaluation results in a determination in step S30 as to whether the probability of the module failing in the next predicted increment is greater than a specific threshold probability P. If the probability is greater than the threshold, a module failure alert is generated in step S40. This step is followed by a step to replace the module in response to the module failure alert. If the probability of module failure in the next predicted increment is not greater than the threshold probability P, the process returns and waits for the next evaluation point.

[0055]

[0064] In one embodiment, one feature of a module failure prediction model according to one embodiment is similar to that of a binary classification model. In a binary classification model, each data instance represents an evaluation of the module at a particular pulse count. In each evaluation, the objective of the model is to predict, with a yes or no, whether the module is likely to survive another predicted increment or to fail before the end of the predicted increment.

[0056]

[0065] The choice of threshold probability or the size of the probability limit P is a design choice. The probability limit P separates yes / no predictions. The machine learning model generates the "probability of belonging to the positive class," and the probability limit separates the classes. The probability limit can be, for example, 50%. However, other values ​​such as 90% are also possible. The probability limit is adjustable to obtain the best modeling performance, i.e., it can be changed to achieve the best fit between model predictions and historical data.

[0057]

[0066] The method described above offers several advantages. For example, the model returns a binary outcome (true or false) compared to a continuous outcome (for instance, if the model attempts to predict the remaining useful lifetime in terms of the number of pulses—a continuous variable). This is important because the accuracy of a data model in predicting binary outcomes is generally likely to be more accurate than that of a data model in predicting continuous outcomes.

[0058]

[0067] The evaluation step may be carried out, for example, using an analytical model-based method, where the analytical model is based on a physical understanding of the module, relying on the analytical model to represent the system's behavior. According to one aspect of one embodiment, a data-driven method, also called a machine learning method, is employed. This method uses historical data to train a model of the system's behavior. Machine learning, or ML, can be defined as a technique whose outcomes are predictable, based on a model prepared and trained with respect to past or historical input data.

[0059]

[0068] These learning models are derived from machine learning algorithms such as neural networks, decision trees, or regression analysis to generate conclusions. For example, SVM is a supervised learning method that can be used for classification and regression analysis. It generally involves a training phase that requires a machine condition (health) indicator with corresponding labels or equipment states such as good, bad, or failed.

[0060]

[0069] Figure 5 is a block diagram illustrating the model construction in relation to modules for light sources. Block 510 represents data acquisition. Over time, many instances of the same module are installed and subsequently removed within light sources deployed in global semiconductor manufacturing plants, and the data collected during the lifespan of each individual module is centrally stored. This historical data for removed modules can be used to develop a model capable of providing advance notice of module failure for currently installed modules.

[0061]

[0070] In another embodiment, the model uses many instances for every module within the distribution of removed modules. For a strategy that evaluates modules with a lifetime starting at 10 billion pulses and ending in 100 million pulse increments, for example, a module that fails at 32.55 billion pulses would give the model 226 instances. Thus, a distribution of several hundred modules can be used to create a training dataset with tens of thousands of instances, which is preferable from a machine learning perspective in that any of many different machine learning models can be used for implementation. This is also preferable because it enhances accuracy on larger training datasets.

[0062]

[0071] The raw data can be preprocessed to make it suitable for extracting state indicators.

[0063]

[0072] Block 520 is feature identification or extraction. This involves identifying features (or metrics) whose behavior changes predictably as the components age and degrade.

[0064]

[0073] In block 530, the identified features are used to train a machine learning (ML) model to provide binary predictive incremental estimates. The ML predicts the outcome based on a model prepared and trained on past or historical input data and its output behavior. Here, training involves aligning the estimates with real-world data, as described below.

[0065]

[0074] As a specific example, the main oscillator chamber is used below as an example of a module, but those skilled in the art will understand that this is merely one example of a module to which the teachings presented herein may apply, and that the teachings are equally and fully applicable to other modules. Other examples include PA chambers, LNM, LAM, BAM, and OPuS modules, as well as any module to which data is generally collected and which relates to the performance of the module. There exists a particular distribution of these modules that have been removed as part of the implementation of umbrella maintenance. A subset of these modules will fail. A complementary subset of these modules will be removed not because they have failed (they have not failed), but because they will be replaced when another module is replaced.

[0066]

[0075] Each of these removed modules is associated with an array of stored historical data in the form of metrics collected at various pulse counts. These metrics can be extracted and used by predictive logic for all modules on the pulse count grid (e.g., from 10Bp to the end of life in 0.1Bp increments). The model is then trained by evaluating how well its predictions align with the actual historical field outcomes. Ideally, the model should predict failures within the predictive increment; otherwise, it should not.

[0067]

[0076] For example, each of approximately 1000 removed modules can be evaluated first at 10Bp, and then in increments of 0.1Bp up to the pulse count at the time of removal. Assuming an average removal pulse count of about 30Bp, this means that each module gave an average of about 200 evaluations ((30Bp-10Bp) / .1), and the entire group of removed modules would result in 200,000 evaluations.

[0068]

[0077] There is a group of measurement metrics collected in each evaluation. Each of these metrics is extracted for all evaluation pulse counts. There are several different metrics that can be extracted and used as features. For example, if the module is an MO chamber module, the metrics are: (1) For all modules, the number of MO-related energy batch quality ("BQ") events within the previous 100Mp at 10Bp, 10.1Bp, 10.2Bp, ... (2) The number of very low energy detections that caused system shutdowns in the previous week, (3) The average MO energy and voltage within the previous 100 Mp, It can include...

[0069]

[0078] The metrics may also include additional parameters such as laser attributes, including model, deployed area, account associated with the module, customer type, power level employed, temperature and blower speed, bandwidth and wavelength stability, bearing time, neon reduction settings and failures, efficiency metrics, etc. The above are captured as average wavelength sigma in a fluctuating lookback window, or at ratings, e.g., at the previous 1 Bp and the previous 5 Bp, or at the previous 1 Bp with respect to the first 5 Bp of the lifetime.

[0070]

[0079] The metrics may also include a consistent set of features, such as counts of individual fault signatures in the previous 100Mp, 1Bp, and 2B, or data derived from individual sensors such as voltage and MO energy.

[0071]

[0080] Assuming at least 4 metrics for each of the 200,000 evaluations, this results in approximately 800,000 metrics.

[0072]

[0081] These metrics are then fed into the model. During training, the model makes predictions, compares them to field results, then adjusts the application of the metrics and compares again, focusing on a predictive logic that aligns with field results.

[0073]

[0082] For example, the prediction logic is feasible to derive approximately 200,000 predictions. These 200,000 predictions can then be compared to a prediction target, for instance, predicting failures within 1 Bp or less in terms of the removal pulse count for a module that has actually failed, while not predicting any other failures.

[0074]

[0083] According to one aspect of one embodiment, the system is designed to predict binary module failures in a predicted increment, for example, predicting module failures within the predicted increment before removal due to failure, while never predicting module failures before removal of modules that do not cause technical problems. The system can work for any module and is designed and tested using historical data.

[0075]

[0084] The operation described can be performed using a control system 600 as shown in Figure 6. The control system 600 includes a data generation unit 620, a memory 630, and a processing unit 610 connected to a communication interface 640, the communication interface 640 being connected to a user interface 650. This control system 600 can be the same as, additional to, or share components with the control system 135 in Figure 1.

[0076]

[0085] The data generation unit 620 may include sensors arranged to measure various metrics. For example, if the module of interest is an MO chamber module, the sensors may include sensors for measuring voltage, energy, gas pressure, temperature, pulse length, etc., across electrodes within the module. The data generation unit 620 may also include counters for measuring total pulses and repetition rate. The data generation unit 620 may include a clock for measuring total runtime. Of course, these are just examples, and it will be understood that other data or additional data can be collected. The data generation unit 620 may include means for data input, either manually or via electronic data transfer.

[0077]

[0086] The processing unit 610 may store this data in a data storage module. This data can then be accessed wirelessly, locally, or remotely, either directly using a hardwired user interface or via the communication interface 640. For example, the communication interface 640 could include a Wi-Fi interface, a cellular interface such as CDMA, a Bluetooth interface, a wireless interface, a USB interface, and / or a near-field communication interface. Of course, these are merely examples, and it will become clear that other or additional communication modules, such as an RFID interface, may be used. The communication interface 640 could, in essence, be any device for communicating data between the processing unit 610 and one or more user interface devices.

[0078]

[0087] The user interface 650 may include any device or system hardware and software through which users can provide or receive information from each other. Therefore, the user interface 650 may include a keypad, a barcode reader, a mobile device such as a smartphone with appropriate applications, a general-purpose computer interface, a hardwired connection, and a display. Of course, these are merely examples, and it will become clear that other or additional user interface devices may be used. The user interface may include various controls, such as a touchscreen and an array of indicators. The user interface may include a digital display.

[0079]

[0088] As described above, the user interface 650 can be implemented as software running on a computer, or as an application on a smartphone, tablet, or other wireless communication device. In such an environment, the communication interface 640 can be configured to interface with an external device, such as a wireless-enabled device like a computer, tablet, or mobile phone. A user can use the application on a mobile device to control the operation of the maintenance data acquisition system. If the external device is a wireless-enabled device such as a computer, tablet, or mobile phone, the application can be installed on the external device, and the user interface for the application may be, for example, a visual representation of a display with controls.

[0080]

[0089] According to one embodiment of the present invention, an embodiment of the present invention provides advance notification of module failures likely to occur in a predicted increment. This notification allows field teams to prepare for maintenance events, thereby preventing unplanned events caused by failures before the umbrella service increment. By reducing prolonged downtime events associated with unplanned maintenance, manufacturing plant productivity is increased. The same functionality may be useful for prioritizing multiple maintenance events in order of their likelihood of failure in a predicted increment.

[0081]

[0090] Figure 7 is a functional block diagram of a maintenance system 700 according to one aspect of one embodiment. The maintenance system 700 includes an evaluation timing unit 710 that determines whether an evaluation point has been reached for a module. This determination may be based, for example, on a pulse count. If the evaluation timing unit 710 determines that an evaluation point has been reached, it instructs a binary prediction unit 720 to make a binary determination of whether the module being evaluated is likely to operate satisfactorily over the entire number of subsequent pulses that make up the predicted increment. The binary prediction unit 720 may include a model of the module that has been learned through training on a set of features. The binary prediction unit 720 uses a set of features that may be supplied by the source of evaluation feature data 730 as input to the model to generate a binary determination. If the binary determination is that the module being evaluated is unlikely to operate satisfactorily over the entire predicted increment, the binary prediction unit 720 instructs a module failure alert generation unit 740 to generate a failure alert. Alternatively and equivalently, the binary prediction unit 720 supplies a binary determination to the module failure alert generation unit 740, which generates a failure alert if the binary determination indicates that the module being evaluated is unlikely to operate satisfactorily over the entire predicted increment. The module failure alert generation unit 740 may also be configured to generate a positive "no failures" indication if the module being evaluated is likely to operate satisfactorily over the entire predicted increment.

[0082]

[0091] According to another aspect of one embodiment, multiple models are each trained according to different relative priorities of specific maintenance parameters, such as tool availability to avoid the risk of the tool unexpectedly failing, as described below. Differentiation between customers or sectors, for example, between DRAM memory customers and another class of customers, such as semiconductor manufacturing plant customers, can increase the usefulness of these interfaces and reports for a single customer or sector. Similarly, such differentiation can improve the usefulness of interfaces and reports for current generation tools with tight tolerances and legacy technology tools.

[0083]

[0092] One advantage of machine learning models is that they are selectable and / or tunable to provide predictions and recommendations that best match a particular customer's preferences and the needs determined by that customer. For example, one customer may determine, given the specific conditions of their technology and supply environment, that they would prefer to prioritize tool availability over avoiding the risk of the tool unexpectedly failing. Another customer may reach the opposite conclusion, namely, that they would prefer to avoid unexpected downtime at the expense of accepting some additional lack of tool availability. The first customer would prefer to maximize their wafer output by accepting higher risk. The second customer would prefer to receive early notification of module failures at the expense of availability. This can translate into an assessment of the relative importance of cost metrics in the form of susceptibility (true positive accuracy), specificity (true negative accuracy), and loss pulses. The term "loss pulse" as used herein refers to the number of pulses lost if a module is removed in the first failure prediction. Using this metric, the predictive model has a way of quantifying false predictions, thereby being trained to avoid premature false predictions. Assigning these relative priorities inevitably involves trade-offs. In some implementations, it may be advantageous to allow customers to weigh these trade-offs and optimize the model for their specific use case.

[0084]

[0093] For example, customers may be offered a maintenance option under a “maximum output” model, which seeks to achieve increased availability along with a higher risk assumption of failure indicators being lost. These customers may prioritize several targets, including external time-based output targets for financial reporting, internal production targets, or high-pricing windows. Alternatively, customers may be offered a maintenance option under an “early warning” model, which targets advanced planning for their toolset. The “early warning” model can be defined as a model that achieves to reduce the risk of unscheduled maintenance with the potential for loss of availability. This provides, through information, values ​​that enable optimal manufacturing plant routines and scheduling.

[0085]

[0094] As a specific example, in modern manufacturing plants, when a lithocell (generally considered a combination of laser, scanner, and tracker) is down for maintenance, all products planned for that lithocell must be rerouted until the lithocell is restored. With state-of-the-art immersion tools, customers can choose to use a single tool for multiple wafer passes and benefit from tighter, dedicated chuck overlay specifications from scanner manufacturers. With improved knowledge of where failures may occur, manufacturing plant operation management can better optimize the overall volume of the plant and ensure high productivity. Furthermore, lithocell systems can combine planning challenges and have extended reserve time (days), increasing the plant's uptime, i.e., the value between green times.

[0086]

[0095] Therefore, customers may find it advantageous to be able to choose from different models based on various maintenance objectives, some of which may be contradictory, such as output maximization and early warning.

[0087]

[0096] For example, one trade-off can be susceptibility versus specificity. This is illustrated graphically in Figure 8. The x-axis represents the percentage of correctly predicted failures, i.e., the specificity or accuracy of true positives. The y-axis represents the percentage of correctly predicted survivals, or in other words, the y-axis represents susceptibility, i.e., the accuracy of true negatives. The area indicated by box 800 is selected by customer choice to pursue a “maximum output” strategy optimized for low extension risk, i.e., prioritizing correctly predicting survival over predicting failures. Priorities for customers adopting this strategy are true negatives and continued availability. On the other hand, customers who choose a combination of specificity and susceptibility, shown within the area marked by box 810, will adopt an “early warning” strategy. Priorities for these customers are the number of true positives and avoiding unplanned downtime. Box 820 shows the area in the susceptibility / selectivity space that represents a balanced approach where susceptibility and specificity are given equal weight.

[0088]

[0097] As an example, we can consider a model definition where evaluations are performed every 0.1 Bp, with a 2 Bp prediction increment. Otherwise, if the model is supposed to be removed at 24.8 Bp due to a failure, the ideal case predicts the failure at 22.8 Bp. An early warning model that ensures the possibility of failure is detected before an actual failure will show a false positive at some point, at or before 22.8 Bp. A less conservative, maximum output model that allows for maximum output will show a true positive after 22.8 Bp.

[0089]

[0098] The selection of a model is based, at least in part, on information indicating the user's maintenance goals or preferences. Achieving these preferences simultaneously may or may not involve trade-offs. For example, as mentioned above, one maintenance preference might be output maximization, which can be achieved, for instance, by reducing the frequency of evaluations, through limiting the amount of time used for equipment offline use and evaluation. Another preference might be avoiding unexpected downtime, which can be achieved through more frequent and / or longer evaluations. A model may be selected based on information indicating the relative importance or prioritization to which the user assigns these two preferences. Other preferences may include reducing overall running costs, the total lifespan of the equipment, and any other considerations that generally affect how maintenance operations are performed. All of these are intended to be incorporated under the term “maintenance preferences.”

[0090]

[0099] Therefore, there are various models that differ from one another in terms of how they are trained, particularly in terms of prioritizing specific maintenance goals or parameters.

[0091]

[0100] Figure 9 shows a functional block diagram conceptually illustrating the operation of a system that allows customers to select priorities. In Figure 9, the model selection module 910 receives evaluation data acquired during evaluation events, and also receives prioritization data indicating how a particular customer would like to assign relative priorities, for example, on sensitivity and selectivity. This prioritization data could simply be the customer selecting a particular model from a menu of candidate models. The model selection module 910 determines which of a set of differently trained models, e.g., Model 1 (specified by the digit 920), Model 2 (specified by the digit 930), and Model 3 (specified by the digit 940), has been selected by the customer or best matches the prioritization specified by the customer. For example, Model 1 could be a maximum output model, Model 2 a balanced model, and Model 3 an early warning model as described above. The arrangement in Figure 9 shows three candidate models, but it will be obvious to those skilled in the art that the number of models can be two, three, or more. The selected model operates on evaluation data and feeds the results of its operation to the binary prediction model 950, which then expresses a binary prediction about whether events such as removal should occur before the next evaluation, according to the customer's priorities.

[0092]

[0101] The arrangement in Figure 9 is merely one example of how the system can be conceptually divided into separate modules. Functionality will become clear in conceptually different ways. For example, the models can directly represent predictions. Also, prioritization data can be the direct selection of one of the three models, which can then directly receive evaluation data.

[0093]

[0102] For example, Figure 10 is a functional block diagram of an arrangement according to another aspect of one embodiment. In Figure 10, the model training unit 1000 trains multiple models using different weightings for various maintenance objectives. The models are stored in the model storage unit 1010 within the binary prediction unit 1020. The user preference data storage unit 1030 contains information about the user's relative preferences for maintenance objectives, or explicit customer selection of models, and supplies this information to the model selection unit 1040, also located within the binary prediction unit 1020. Thus, when the evaluation timing unit 1050 determines that the time has come for an evaluation based on the passage of time series, or time or machine time based on the number of pulses, the binary prediction unit 1020 generates a signal to the model module failure alert generation unit 1060, which will operate on the evaluation data using the selected model and generate a module failure alert. The binary prediction unit 1020 can also generate an affirmative signal indicating that no maintenance action is required.

[0094]

[0103] The above description includes examples of multiple embodiments. Of course, it is impossible to describe every conceivable combination of components or method systems in order to illustrate the above embodiments, but those skilled in the art will recognize that many further combinations and rearrangements of various embodiments are possible. Accordingly, the embodiments described are intended to encompass all such changes, modifications, and variations that fall within the spirit and scope of the appended claims. Furthermore, insofar as the term “includes” is used in either the detailed description or the claims, such terms are intended to be as comprehensive as the term “comprising,” and “comprising,” when adopted, is interpreted as a transitional term within the claims. Furthermore, elements of the described aspects and / or embodiments may be described or claimed in the singular, but the plural is intended unless a limitation to the singular is explicitly indicated. In addition, all or part of any aspect and / or embodiment may be used in conjunction with all or part of any other aspect and / or embodiment unless otherwise specifically indicated.

[0095]

[0104] The embodiments can be further described using the following clauses. 1. A method for maintaining a light source for semiconductor photolithography, wherein the light source comprises one or more modules, To determine whether one of more modules is based on an initial evaluation that is at least partially based on the number of first pulses that the module was involved in generating, To determine whether the module has at least the minimum probability of fault-free operation in the predicted increment, which is measured as the second pulse count, an initial evaluation is performed, and If the initial assessment determines that the module does not have at least the minimum probability of failure-free operation in the predicted increment, a module failure alert will be generated for the module. Methods that include... 2. The method of Clause 1, further comprising: if the initial assessment determines that the module has at least the minimum probability of failure-free operation in the predicted increment, leaving the module in service. 3. The method according to Clause 1, wherein determining whether a module, which is one of one or more modules, is based on an initial assessment that is at least partially based on a first number of pulses in which the module was involved in generating, includes evaluating whether the first number corresponds to a pulse milestone of a predetermined number of pulses. 4. The method of Clause 3, which includes using a model to represent a binary (true / false) prediction of whether the module is likely to operate without failure during the predicted increment, if the initial assessment determines that the module does not have at least the minimum probability of failure-free operation in the predicted increment, to generate a module failure alert for the module. 5. The method according to Clause 1, further comprising performing a maintenance operation on the module if the initial assessment determines that the module does not have at least the minimum probability of failure-free operation in the predicted increment. 6. Performing maintenance operations includes removing modules, as described in Clause 5. 7. Performing maintenance operations includes repairing modules, as described in Clause 5. 8. The first number is approximately 10 billion pulses, as described in Clause 1. 9. The method according to Clause 1, wherein the second number is approximately 2 billion pulses. 10. The method according to Clause 1, wherein the second number is determined at least in part on the age of the module with respect to the time-series age or cumulative pulse count. 11. After leaving the module in service, A step of determining whether the module is the result of an additional evaluation that is at least partially based on the number of third pulses that the module has been involved in generating since the initial evaluation. The steps include performing an additional evaluation to determine whether the module has at least the minimum probability of fault-free operation in a second predicted increment, which is measured as a fourth pulse count, and If the additional evaluation determines that the module does not have at least the minimum probability of failure-free operation in the second predicted increment, the step is to generate a module failure alert for the module, or If the additional evaluation determines that the module has at least the minimum probability of failure-free operation in the second predicted increment, the step is to keep the module in service. The method described in Clause 1, further including the following: 12. The method according to Clause 11, further comprising removing the module if an additional evaluation determines that the module does not have at least the minimum probability of failure-free operation in the second predicted increment. 13. The third number is approximately 100 million pulses, as described in Clause 11. 14. The fourth number is approximately 2 billion pulses, as described in Clause 11. 15. The method according to Clause 1, wherein at least one of the third and fourth numbers is determined at least in part on the age of the module with respect to the time-series age or cumulative pulse count. 16. Performing an initial evaluation is the method described in Clause 1, which includes using a model. 17. The model is an analytical model, as described in Clause 16. 18. The method according to Clause 16, wherein the model is a trained model developed via machine learning by supplying feature data to train the trained model, and the trained model makes an initial decision based on at least some of the feature data. 19. The module comprises a main oscillator chamber module, and the feature data includes several main oscillator-related energy batch quality events in the preceding 100 million pulses, as described in Clause 18. 20. The module comprises a main oscillator chamber module, and the characteristic data includes the average main oscillator energy in the preceding 100 million pulses, as described in Clause 18. 21. Computer implementation method, A computing device determines whether one of the modules of a photolithography light source is based on an initial evaluation that is at least partially based on the number of first pulses that the module was involved in generating. To determine whether the module has at least the minimum probability of fault-free operation in a predicted increment measured as a second pulse count, an initial evaluation is performed by a computing device, and If the initial assessment determines that the module does not have at least the minimum probability of failure-free operation in the predicted increment, the computing device provides an instruction that the module should be removed, or If the initial assessment determines that the module has at least the minimum probability of failure-free operation in the predicted increment, the computing device shall provide an instruction that the module should remain in service. Computer implementation methods, including those mentioned above. 22. A non-temporary computer-readable storage medium containing executable instructions for causing a processor to perform an operation, wherein the instructions are: Instructions for determining whether a module, which is one or more modules of a photolithography light source, is based on an initial evaluation that is at least partially based on the number of first pulses that the module was involved in generating, Instructions for performing an initial evaluation to determine whether the module has at least the minimum probability of fault-free operation in a predicted increment measured as a second pulse count, and If the initial assessment determines that the module does not have at least the minimum probability of failure-free operation in the predicted increment, an instruction to remove the module is provided, or If the initial assessment determines that the module has at least the minimum probability of failure-free operation in the predicted increment, an instruction to provide that the module should be kept in service. Non-temporary computer-readable storage media, including [specific type of storage medium]. 23. A method for maintaining a light source for semiconductor photolithography, wherein the light source comprises one or more modules, and the method is Identifying a module that is one of one or more modules for evaluation. To determine whether the module has at least the minimum probability of fault-free operation in the predicted increment, which is measured as the second pulse count, an evaluation is performed, and If the initial assessment determines that the module does not have at least the minimum probability of failure-free operation in the predicted increment, a module failure alert will be generated for the module, or If the initial assessment determines that the module has at least the minimum probability of failure-free operation in the predicted increment, the module should be left in service. Methods that include... 24. The method of Clause 23, further comprising removing the module if the initial assessment determines that the module does not have at least the minimum probability of failure-free operation in the predicted increment. 25. The first number is approximately 10 billion pulses, as described in Clause 23. 26. The method described in Clause 23, where the second number is approximately 2 billion pulses. 27. Performing an evaluation involves using a trained model developed via machine learning by supplying feature data to train the trained model, and the trained model determines, based on at least some of the feature data, whether the module has at least the minimum probability of fault-free operation in a predicted increment measured as a second pulse count, as described in Clause 23. 28. A system for maintaining a light source for semiconductor photolithography, wherein the light source comprises one or more modules, and the system is An evaluation timing unit adapted to determine whether one of one or more modules is evaluated based at least partially on a first number of pulses in which the module was involved in generating, A binary prediction unit, positioned to respond to an evaluation timing unit and adapted to perform an evaluation by determining whether the module has at least the minimum probability of fault-free operation in a measured prediction increment as a second pulse count, A module failure alert generation unit is configured to respond to a binary prediction unit and is adapted to generate a module failure alert for a module if the evaluation determines that the module does not have at least the minimum probability of failure-free operation in the prediction increment. A system that includes these features. 29. The module failure alert generation unit is further configured to generate a positive failure-free indication if the module being evaluated is likely to operate satisfactorily throughout the entire predicted increment, as described in Clause 28. 30.2 The value prediction unit is a system as described in Clause 28, comprising a model generated by machine learning from a feature set. 31.2 The system as described in Clause 28, wherein the value prediction unit is arranged to receive feature data and uses the feature data to determine whether the module has at least the minimum probability of failure-free operation in the prediction increment. 32. A system for maintaining a light source for semiconductor photolithography, wherein the light source comprises one or more modules, and the system is An evaluation timing unit, adapted to generate an evaluation signal when one of one or more modules reaches an evaluation point based on pulse counting, A binary prediction unit is positioned to receive an evaluation signal and adapted to make a binary determination in response to the evaluation signal of whether the module being evaluated is likely to operate satisfactorily over the entire number of subsequent pulses that constitute the prediction increment. A module failure alert generation unit is configured to respond to a binary prediction unit and is adapted to generate a failure alert if the binary determination indicates that the module being evaluated is unlikely to operate satisfactorily over the entire prediction increment. A system that includes these features. 33.2 The value prediction unit is a system as described in Clause 32, comprising a model generated by machine learning from a feature set. 34.2 The system as described in Clause 32, wherein the value prediction unit is arranged to receive feature data and uses the feature data to determine whether the module has at least the minimum probability of failure-free operation in the prediction increment. 35. A method for maintaining a light source for semiconductor photolithography, wherein the light source comprises one or more modules, and the method is Obtain user information that shows the relative prioritization of a user for two or more maintenance preferences. Training at least two models, including a first model based on a first relative prioritization of two or more maintenance preferences, and a second model based on a second relative prioritization of two or more maintenance preferences. The evaluation is performed to determine whether a module failure alert should be generated, and the evaluation is performed using at least one of two models based on which model's relative prioritization is most closely aligned with the user's relative prioritization, and If the evaluation determines that a module failure alert should be generated, then a module failure alert will be generated for the module. Methods that include... 36. The method of Clause 35, which includes performing an evaluation using at least one of two models selected by the user, based on which model's relative prioritization is most closely aligned with the user's relative prioritization. 37. Maintenance preferences are to maximize output and avoid unexpected downtime, as described in Clause 35. 38. The method as described in Clause 35, further including indicating that the module will remain in service if the evaluation determines that a module failure alert should not be generated. 39. The method of Clause 35, further comprising determining whether the module is based on an evaluation that is at least partially based on the number of pulses that the module was involved in generating. 40. The method of Clause 35, wherein performing an evaluation to determine whether a module failure alert should be generated includes selecting a model from at least two models based on the user's relative prioritization to represent a binary (true / false) determination regarding whether a module failure alert should be generated. 41. The method of Clause 35, further comprising performing maintenance actions on the module if the evaluation determines that a module failure alert should be generated. 42. Performing maintenance operations includes removing modules, as described in Clause 41. 43. Performing maintenance operations includes repairing modules, as described in Clause 41. 44. The method according to Clause 35, wherein the model is a trained model developed via machine learning by supplying feature data to train the trained model, and the selected trained model makes decisions based on at least some of the feature data. 45. Computer implementation method, Using a computing device to store user information indicating the relative prioritization of a user for two or more maintenance preferences, Using a computing device, train at least two models, including a first model based on a first relative prioritization of two or more maintenance preferences, and a second model based on a second relative prioritization of two or more maintenance preferences. Using a computing device, select one of the models as the model chosen based on user information. Using a selected model on the computing device, perform an evaluation to determine whether a module failure alert should be generated, and If the evaluation determines that a module failure alert should be generated, the computing device should provide instructions to remove the module, or If the assessment determines that a module failure alert should not be generated, the computing device should provide instructions that the module should remain in service. Computer implementation methods, including those mentioned above. 46. ​​Maintenance preferences are to maximize output and avoid unexpected downtime, as described in the computer implementation method in Clause 45. 47. A non-temporary computer-readable storage medium containing executable instructions for causing a processor to perform an action, wherein the instructions are: Instructions for storing user information that indicates the relative prioritization of a user for two or more maintenance preferences. Instructions for training at least two models, including a first model based on a first relative prioritization of two or more maintenance preferences, and a second model based on a second relative prioritization of two or more maintenance preferences. A command to select one of the models as the model selected based on user information. Instructions to perform an evaluation using the selected model to determine whether a module failure alert should be generated, and If the assessment determines that a module failure alert should be generated, it will provide instructions that the module should undergo maintenance procedures, or If the assessment determines that a module failure alert should not be generated, an instruction should be given to leave the module in service. Non-temporary computer-readable storage media, including [specific type of storage medium]. 48. Maintenance preferences are to maximize output and avoid unexpected downtime for non-transient computer-readable storage media as described in Clause 47. 49. A system for maintaining a light source for semiconductor photolithography, wherein the light source comprises one or more modules, and the system is A user preference data storage unit adapted to store user preference data relating to the relative prioritization of a user of two or more maintenance preferences, An evaluation timing unit adapted to determine whether one of one or more modules is evaluated based at least partially on a first number of pulses in which the module was involved in generating, A model training unit adapted to train a first model based on a first relative prioritization of two or more maintenance preferences, and a second model based on a second relative prioritization of two or more maintenance preferences, A model selection unit adapted to select one of the models as the model selected based on user preference data, A binary prediction unit, positioned to respond to an evaluation timing unit and a user preference input unit, is adapted to perform an evaluation by determining whether a module failure alert should be generated using a selected model. A module failure alert generation unit is configured to respond to a binary prediction unit and, if the evaluation determines that a module failure alert should be generated, is adapted to generate a module failure alert for the module. A system that includes these features. 50. Maintenance preferences are to maximize output and avoid unexpected downtime, as described in Clause 49 for the system. 51. The module failure alert generation unit is further configured to generate a positive no-failure indication if the binary prediction unit determines that a module failure alert should not be generated, as described in Clause 49 of the system.

[0096]

[0105] Other implementations are within the scope of the claims.

Claims

1. A method for maintaining a light source for semiconductor photolithography, wherein the light source comprises one or more modules. To determine whether one of the aforementioned modules is based on an initial evaluation that is at least partially based on a first number of pulses in which the module was involved in generating, Performing the initial evaluation to determine whether the module has at least a threshold probability of fault-free operation in a predicted increment measured as a second number of pulses, and If the initial evaluation determines that the module does not have at least the threshold probability of failure-free operation in the predicted increment, a module failure alert is generated for the module. Methods that include...

2. The method according to claim 1, wherein determining whether a module, which is one or more of the modules, is based on an initial evaluation that is at least partially based on a first number of pulses in which the module was involved in generating, includes evaluating whether the first number corresponds to a predetermined number of pulse milestones.

3. The method of claim 2, wherein if the initial assessment determines that the module does not have at least the threshold probability of failure-free operation in the predicted increment, generating a module failure alert for the module includes using a model to represent a binary (true / false) prediction of whether the module is likely to operate without failure during the predicted increment.

4. The method according to claim 1, further comprising performing a maintenance operation on the module if the initial evaluation determines that the module does not have at least the threshold probability of fault-free operation in the predicted increment.

5. The method according to claim 1, wherein the second number is determined at least in part on the age of the module with respect to a time-series age or cumulative pulse count.

6. After leaving the aforementioned module in service, A step of determining whether the module is based on an additional evaluation that is at least partially based on a third number of pulses that the module has been involved in generating since the initial evaluation. The steps include performing the additional evaluation to determine whether the module has at least a threshold probability of fault-free operation in a second predicted increment, which is measured as a fourth number of pulses, and If the additional evaluation determines that the module does not have at least the threshold probability of failure-free operation in the second predicted increment, the step of generating a module failure alert for the module, or If the additional evaluation determines that the module has at least the threshold probability of failure-free operation in the second predicted increment, the step of keeping the module in service, The method according to claim 1, further comprising:

7. The method according to claim 6, wherein at least one of the third number and the fourth number is determined at least in part on the age of the module with respect to time-series age or cumulative pulse count.

8. The method according to claim 1, wherein performing the initial evaluation includes using a model.

9. The method according to claim 8, wherein the model is an analytical model.

10. The method according to claim 8, wherein the model is a trained model developed via machine learning by supplying feature data to train the trained model, and the trained model performs the initial evaluation based on at least some of the feature data.

11. A method for maintaining a light source for semiconductor photolithography, wherein the light source comprises one or more modules, and the method is Identifying one of the aforementioned one or more modules for evaluation purposes. To determine whether the module has at least a threshold probability of fault-free operation in a predicted increment measured as a second number of pulses, an initial evaluation is performed, and If the initial evaluation determines that the module does not have at least the threshold probability of failure-free operation in the predicted increment, a module failure alert is generated for the module, or If the initial evaluation determines that the module has at least the threshold probability of failure-free operation in the predicted increment, the module will remain in service. Methods that include...

12. The method according to claim 11, further comprising removing the module if the initial evaluation determines that the module does not have at least the threshold probability of fault-free operation in the predicted increment.

13. The method according to claim 11, wherein performing the evaluation includes using a trained model developed via machine learning by supplying feature data to train the trained model, the trained model determines, based on at least some of the feature data, whether the module has at least a threshold probability of fault-free operation in a predicted increment measured as a second number of pulses.

14. A system for maintaining a light source for semiconductor photolithography, wherein the light source comprises one or more modules, and the system is An evaluation timing unit adapted to determine whether one of the one or more modules is evaluated based at least partially on a first number of pulses in which the module was involved in generating, A binary prediction unit, arranged to respond to the evaluation timing unit and adapted to perform the evaluation by determining whether the module has at least a threshold probability of fault-free operation in a measured predicted increment as a second number of pulses, A module failure alert generation unit is provided, which is configured to respond to the binary prediction unit and is adapted to generate a module failure alert for the module if the evaluation determines that the module does not have at least the threshold probability of failure-free operation in the prediction increment. A system that includes these features.

15. The system according to claim 14, wherein the module failure alert generation unit is further configured to generate a positive no-failure indication if the module being evaluated is likely to operate satisfactorily over the entire predicted increment.

16. The system according to claim 14, wherein the binary prediction unit comprises a model generated from a feature set by machine learning.

17. The system according to claim 14, wherein the binary prediction unit is arranged to receive feature data and uses the feature data to determine whether the module has at least a threshold probability of fault-free operation in the prediction increment.

18. A system for maintaining a light source for semiconductor photolithography, wherein the light source comprises one or more modules, and the system is An evaluation timing unit, which is configured to generate an evaluation signal when one of the one or more modules reaches an evaluation point based on pulse counting, A binary prediction unit is provided, which is arranged to receive the evaluation signal and is adapted to make a binary determination in response to the evaluation signal as to whether the module being evaluated is likely to operate satisfactorily over the entire number of subsequent pulses that constitute the prediction increment. A module failure alert generation unit, configured to respond to the binary prediction unit and adapted to generate a failure alert if the binary determination indicates that the module being evaluated is unlikely to operate satisfactorily over the entire prediction increment, A system that includes these features.

19. The system according to claim 18, wherein the binary prediction unit comprises a model generated from a feature set by machine learning.

20. The system according to claim 18, wherein the binary prediction unit is arranged to receive feature data and uses the feature data to determine whether the module has at least a threshold probability of fault-free operation in the prediction increment.

Images