Method for creating training model, laser device, and method for manufacturing electronic device

Abstract

This method for creating a trained model comprises: creating a database by breaking a part in a laser apparatus on a digital twin that models the electric component-system hardware and software of the laser apparatus, and accumulating data in which the part that has been destroyed and a trouble phenomenon that is output from the digital twin are associated with each other; and training a trained model using the data in the database as training data for machine learning, so that, in response to an input of information about a trouble phenomenon, the trained model outputs information about a failure location corresponding to the input.

Description

[Technical Field]

[0001] The present disclosure relates to a method for creating a training model, a laser apparatus, and a method for manufacturing an electronic device. [Background technology]

[0002] In recent years, semiconductor exposure devices have been required to improve their resolution in response to the miniaturization and high integration of semiconductor integrated circuits. To this end, the wavelength of light emitted from exposure light sources has been shortened. For example, KrF excimer laser devices that output laser light with a wavelength of approximately 248 nm and ArF excimer laser devices that output laser light with a wavelength of approximately 193 nm are used as gas laser devices for exposure.

[0003] The spectral linewidth of the spontaneously oscillating light from KrF excimer laser devices and ArF excimer laser devices is as wide as 350 to 400 pm. Therefore, if a projection lens is constructed using a material that transmits ultraviolet light, such as KrF and ArF laser light, chromatic aberration may occur. As a result, resolution may decrease. Therefore, it is necessary to narrow the spectral linewidth of the laser light output from the gas laser device to a level where chromatic aberration is negligible. Therefore, a line narrowing module (LNM) containing a line narrowing element (e.g., an etalon or grating) may be installed inside the laser resonator of the gas laser device to narrow the spectral linewidth. Hereinafter, a gas laser device with a narrowed spectral linewidth is referred to as a line narrowing gas laser device. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] U.S. Patent No. 10,990,089 [Patent Document 2] Japanese Patent Publication No. 2020-177276 [Patent Document 3] Summary of U.S. Patent Application Publication No. 2020 / 0103842

[0005] A method for creating a training model according to one aspect of the present disclosure includes creating a database by destroying components within a laser device on a digital twin that models the electrical hardware and software of the laser device and accumulating data that associates the destroyed components with malfunction phenomena output by the digital twin, and training the training model using the data in the database as machine learning training data so that the training model receives input of information on malfunction phenomena and outputs information on the fault location corresponding to the input.

[0006] A laser device according to another aspect of the present disclosure comprises a processor, electrical hardware including a monitored object including a sensor whose status is monitored by the processor, and equipment including wiring connecting the processor and the monitored object, and a training model trained by machine learning to receive input of information on a malfunction phenomenon and output information on the fault location corresponding to the input, wherein the training model is a model trained using data in a database created by destroying parts within the laser device on a digital twin that models the electrical hardware and software of the laser device, and accumulating data that associates the destroyed parts with the malfunction phenomenon output by the digital twin as training data.

[0007] Another aspect of the present disclosure provides a method for manufacturing an electronic device, the method comprising: a laser apparatus including a processor; electrical hardware including a monitored object including a sensor whose state is monitored by the processor; and equipment including wiring connecting the processor and the monitored object; and a training model trained by machine learning to receive input of information on a malfunction phenomenon and output information on the fault location corresponding to the input; the training model is a model trained using, as training data, data in a database created by destroying components within the laser apparatus on a digital twin that models the electrical hardware and software of the laser apparatus and accumulating data correlating the destroyed components with the malfunction phenomenon output by the digital twin; and the method includes generating laser light by the laser apparatus, outputting the laser light to an exposure apparatus, and exposing the laser light to a photosensitive substrate in the exposure apparatus to manufacture an electronic device. [Brief explanation of the drawings]

[0008] Some embodiments of the present disclosure will now be described, by way of example only, with reference to the accompanying drawings, in which: [Figure 1] FIG. 1 shows a schematic diagram of an exemplary laser device configuration. [Figure 2] FIG. 2 shows a schematic configuration for notifying error information based on information obtained from a monitor module. [Figure 3] FIG. 3 shows a schematic diagram of a service engineer's response to a malfunction of a laser device. [Figure 4] FIG. 4 is an explanatory diagram illustrating an overview of a system that implements the training model creation method according to the first embodiment. [Figure 5] FIG. 5 is a diagram showing an example of data stored in the database. [Figure 6] FIG. 6 is an explanatory diagram schematically illustrating an example of how to use a trained model created by implementing the training model creation method according to the first embodiment. [Figure 7] FIG. 7 is a chart showing an example of an estimated list of fault locations estimated using the training model. [Figure 8]FIG. 8 is an explanatory diagram illustrating an outline of a method for creating a training model according to the second embodiment. [Figure 9] FIG. 9 is an explanatory diagram showing an overview of a laser device including a training model according to the third embodiment and an example of its use. [Figure 10] FIG. 10 shows a schematic configuration example of an exposure apparatus. Embodiment

[0009] -table of contents- 1. Overview of the laser device 1.1 Configuration 1.2 Operation 1.3 Example of error information notification 1.4 Challenges 2. Embodiment 1 2.1 Configuration 2.2 Operation 2.2.1 Training data generation and training phase 2.2.2 Inference Phase 2.3 Actions and Effects 3. Embodiment 2 3.1 Configuration 3.2 Operation 3.3 Actions and Effects 4. Embodiment 3 4.1 Configuration 4.2 Operation 4.3 Actions and Effects 5. Other forms of laser devices 6. Functional Roles of Information Processing Systems 7. Manufacturing methods for electronic devices 8.Other Hereinafter, embodiments of the present disclosure will be described in detail with reference to the drawings. The embodiments described below show some examples of the present disclosure and do not limit the content of the present disclosure. Furthermore, not all of the configurations and operations described in each embodiment are necessarily essential as the configurations and operations of the present disclosure. Note that the same components are given the same reference symbols, and redundant explanations will be omitted.

[0010] 1. Overview of the laser device 1.1 Configuration 1 shows a schematic configuration of an exemplary laser apparatus 10. The laser apparatus 10 is a KrF excimer laser apparatus and includes a line narrowing module (LNM) 12, a chamber 14, an inverter 16, a front mirror (output coupling mirror) 18, a monitor module 20, a charger 22, a pulsed power module (PPM) 24, a processor 26, a gas supply device 28, a gas exhaust device 30, and an exit shutter 32.

[0011] The chamber 14 includes windows 34, 36, a cross-flow fan (CFF) 38, a motor 40 that rotates the CFF 38, a pair of electrodes 42a, 42b, an electrical insulator 44, a pressure sensor 46, and a heat exchanger (not shown).

[0012] The PPM 24 is connected to the electrode 42a via a feedthrough in the electrical insulator 44 of the chamber 14. The PPM 24 includes a semiconductor switch 48, a charging capacitor (not shown), a pulse transformer, and a pulse compression circuit.

[0013] The front mirror 18 is a partially reflective mirror and is arranged to form an optical resonator together with the LNM 12. The chamber 14 is arranged on the optical path of this optical resonator. The LNM 12 includes a beam expander consisting of two prisms 50 and 52, a rotation stage 54, and a grating 56. The prisms 50 and 52 are arranged to expand the beam of light emitted from the window 34 of the chamber 14 in the Y direction and cause the beam to enter the grating 56.

[0014] The grating 56 is arranged in a Littrow configuration so that the incident angle and diffraction angle of the laser light coincide with each other. The prism 52 is also arranged on the rotation stage 54 so that the incident angle and diffraction angle of the laser light on the grating 56 change when the rotation stage 54 rotates.

[0015] The monitor module 20 includes beam splitters 60 and 62, a pulse energy detector 64, and a spectrum detector 66. The beam splitter 60 is located on the optical path of the laser light output from the front mirror 18 and is arranged to reflect a portion of the incident laser light so that the reflected portion is incident on the beam splitter 62.

[0016] The pulse energy detector 64 is disposed so that the laser light transmitted through the beam splitter 62 is incident thereon. The pulse energy detector 64 may be, for example, a photodiode that measures the intensity of ultraviolet light. The beam splitter 62 is disposed so that a portion of the incident laser light is reflected and made incident on the spectrum detector 66. The spectrum detector 66 is, for example, a monitor etalon measurement device that measures interference fringes generated by an etalon using an image sensor. The center wavelength and spectral linewidth of the laser light are measured based on the generated interference fringes.

[0017] Gas supply device 28 is connected to an inert gas supply source 70, which is a supply source of inert laser gas, and a halogen gas supply source 72, which is a supply source of halogen-containing laser gas, via pipes 74 and 76. In the case of a KrF excimer laser device, the inert laser gas is a mixture of Kr gas and Ne gas, and the halogen-containing laser gas is a mixture of F gas, Kr gas, and Ne gas. Gas supply device 28 is connected to chamber 14 via pipe 78. Gas supply device 28 includes an automatic valve and a mass flow controller (not shown) for supplying predetermined amounts of the inert laser gas and the halogen-containing laser gas to chamber 14, respectively.

[0018] The gas exhaust device 30 is connected to the chamber 14 via piping 80, includes a halogen filter (not shown) that removes halogen, and an exhaust pump (not shown), and is configured to exhaust the laser gas from which the halogen has been removed to the outside.

[0019] The exit shutter 32 is disposed on the optical path of the laser light output from the laser device 10 to the outside.

[0020] The inverter 16 is a power supply device for the motor 40 that drives the CFF 38, and is configured to receive the frequency of the power supplied to the motor 40 from the processor 26. In the present disclosure, the processor 26 is a processing device including a storage device in which a control program is stored and a CPU (Central Processing Unit) that executes the control program. The processor 26 is specially configured or programmed to execute various processes included in the present disclosure. The processor 26 is electrically connected to multiple components of the laser device 10 and is configured to communicate with and control them. Components connected to the processor 26 also include components not shown. The processor 26 is also connected to the exposure device 90.

[0021] 1.2 Operation The processor 26 exhausts the gases in the chamber 14 via the gas exhaust device 30, and then fills the chamber 14 with laser gases via the gas supply device 28, including a mixed gas of Kr and Ne and a mixed gas of F2, Kr and Ne, so as to achieve the desired gas composition and total gas pressure.

[0022] The processor 26 rotates the motor 40 at a predetermined rotation speed via the inverter 16 to rotate the CFF 38. As a result, laser gas flows between the electrodes 42a and 42b. The processor 26 receives the target pulse energy Et from the exposure control unit 92 of the exposure tool 90, and transmits data on the charging voltage Vhv to the charger 22 so that the pulse energy becomes Et. The charger 22 charges the charging capacitor of the PPM 24 to the charging voltage Vhv.

[0023] When the exposure device 90 outputs a light emission trigger signal Tr1, the processor 26 inputs a trigger signal Tr2 to the semiconductor switch 48 of the PPM 24 in synchronization with the light emission trigger signal Tr1. When the semiconductor switch 48 operates, a current flows from the charging capacitor of the PPM 24, and the current is pulse-compressed by the magnetic compression circuit, and a high voltage is applied between the electrodes 42a and 42b. As a result, a discharge occurs between the electrodes 42a and 42b, exciting the laser gas in the discharge space.

[0024] When the excited laser gas in the discharge space returns to its ground state, excimer light is generated. This excimer light travels back and forth between the front mirror 18 and the LNM 12 and is amplified, resulting in laser oscillation. As a result, narrowband pulsed laser light is output from the front mirror 18. The pulsed laser light output from the front mirror 18 enters the monitor module 20.

[0025] In the monitor module 20, a portion of the pulsed laser beam is sampled by a beam splitter 60 and is incident on a pulse energy detector 64 and a spectrum detector 66 via a beam splitter 62. The pulse energy detector 64 measures the pulse energy E of the pulsed laser beam, and this data is sent to the processor 26. The spectrum detector 66 measures the central wavelength λ and the spectral linewidth Δλ of the pulsed laser beam, and this data is sent to the processor 26.

[0026] The processor 26 receives the target pulse energy Et and the target wavelength λt from the exposure tool 90. The processor 26 performs various controls, including pulse energy control and wavelength control. The pulse energy control is performed by controlling the charging voltage Vhv so that the difference ΔE between the pulse energy E measured by the pulse energy detector 64 and the target pulse energy Et approaches zero. The wavelength control is performed by controlling the rotation angle of the rotation stage 54 so that the difference δλ between the center wavelength λ measured by the spectrum detector 66 and the target wavelength λt approaches zero.

[0027] As described above, the processor 26 receives the target pulse energy Et and the target wavelength λt from the exposure device 90, and outputs a pulsed laser beam in synchronization with the light emission trigger signal Tr1 every time the light emission trigger signal Tr1 is input.

[0028] When the laser device 10 repeatedly discharges, the electrodes 42a and 42b wear down, the halogen gas in the laser gas is consumed, and impurity gas is generated. A decrease in the halogen gas concentration and an increase in the impurity gas concentration in the chamber 14 adversely affect the decrease in the pulse energy E of the pulsed laser light and the pulse energy stability. To suppress these adverse effects, the processor 26 executes the following gas control ([1] to [3]).

[0029] [1] Halogen injection control Halogen injection control is gas control that replenishes the halogen gas consumed mainly by discharge in chamber 14 during laser oscillation by injecting gas containing halogen gas at a higher concentration than the halogen gas in chamber 14.

[0030] [2] Partial gas exchange control Partial gas replacement control is gas control in which part of the laser gas in chamber 14 is replaced with new laser gas so as to suppress an increase in the concentration of impurity gas in chamber 14 during laser oscillation.

[0031] [3] Gas pressure control Gas pressure control is gas control that controls the pulse energy E by injecting laser gas into chamber 14 and changing the total pressure of the laser gas when it is difficult to improve the decrease in pulse energy E of the pulse laser light output from laser device 10 within the control range of charging voltage Vhv.

[0032] When exhausting the laser gas from the chamber 14, the processor 26 controls the gas exhaust device 30. The laser gas exhausted from the chamber 14 has halogen gas removed by a halogen filter and is then exhausted to the outside of the laser device 10.

[0033] The processor 26 transmits parameter data such as the number of oscillation pulses, the charging voltage Vhv, the gas pressure Pch in the chamber 14, the pulse energy E of the laser light, etc. to a laser apparatus management system (not shown).

[0034] 1.3 Example of error information notification 2 shows a schematic configuration for notifying error information based on information obtained from the monitor module 20. The laser device 10 can output information from multiple sensors to a display 84 or a computer network. For example, the pulse energy detector 64, spectrum detector 66, and temperature sensor 68 that measures the temperature inside the module of the monitor module 20 are connected to the processor 26. The processor 26 monitors the status quantities of these sensors, and when it detects an abnormal value, it can output a corresponding error code and each status quantity.

[0035] Other modules and devices of the laser system 10 are similarly connected to the processor 26 and have their operation monitored by the processor 26 .

[0036] The various sensors connected to the processor 26 and used to monitor state quantities, the signal lines, interfaces, or relay devices that electrically connect the various sensors to the processor 26, and the electrical and electronic elements contained in the processor 26 itself are collectively referred to as electrical hardware.

[0037] The electrical hardware is composed of the processor 26, monitored objects including sensors whose states are monitored by the processor 26 within the laser device 10, and devices including wiring that connect the processor 26 to the monitored objects. Therefore, the electrical hardware also includes actuators with encoders, proximity switches, Ethernet repeaters, sequencers, AD converters, DA converters, programmable controllers, etc.

[0038] 1.4 Challenges When the processor 26 detects an abnormality in the output from a module, sensor, or the like under monitoring, it notifies the service engineer (FSE) of the error or status quantity via the display 84, etc. Depending on the location of the failure or malfunction, multiple errors or status quantities may be displayed, making it difficult to identify the cause.

[0039] In such cases, the service engineer FSE checks the laser device 10 for errors or abnormal behavior against a malfunction list 88 that lists previously occurring malfunctions, considers and determines the appropriate response and replacement parts, and then performs adjustments or repairs (see Figure 3). The malfunction list 88 is a list of malfunction causes and / or replacement parts corresponding to previously occurring errors and status quantities. Note that part replacement is included in the concept of "repair." "Replacing" a part not only includes replacing a part with a new one, but also cleaning the part to maintain and / or restore its functionality and then relocating the same part.

[0040] 3 is a diagram showing the response of a service engineer FSE in dealing with a malfunction of the laser device 10. The service engineer FSE checks the malfunction, such as an error code or abnormal behavior of the laser device 10 notified on the display 84 or the like, and if it is difficult to identify the cause, refers to a malfunction list 88 for a solution or replacement part, and adjusts or repairs the equipment, parts, etc.

[0041] For this reason, when a new problem not on the problem list 88 occurred, it was sometimes unclear how to deal with it. In such cases, they were forced to perform partial operations or collect specific data to identify the faulty part, which took time. If they still could not identify the problem, a member of the development department had to inspect the actual machine, which took time to resolve.

[0042] 2. Embodiment 1 2.1 Configuration FIG. 4 is an explanatory diagram showing an overview of a system that implements a method for creating a training model 106 according to the first embodiment. A digital twin 100 that models electrical hardware, including various sensors of the laser device 10, and software, is constructed in a digital space DGS such as a server or cloud. The digital twin 100 is a digital replica of the laser device 10. This digital twin 100 may also include physical models of the electrical hardware and some components of the laser device 10. When information that a certain component has broken (failed) (hereinafter referred to as broken component information) is input, the digital twin 100 is configured to output an error code and state quantity as a malfunction phenomenon in the same way as the laser device 10 in real space.

[0043] Information on the sequential destruction of components within the laser device 10, either one by one or in multiple combinations, is input into the digital twin 100, which outputs the malfunction phenomenon, and the destroyed components are associated with the resulting malfunction phenomenon (behavior of the laser device 10, such as error codes and state quantities) and stored in the database 104. The concept of "components" when sequentially destroying components within the laser device 10 also includes wiring for transmission systems, such as signal lines and power supply lines. "Destroying" includes disconnecting wiring for transmission systems. The expressions "destroying" or "destroying" a component include the concepts of "causing a breakdown" or "failure."

[0044] The broken part information input to the digital twin 100 may be automatically generated by a program. By breaking parts in the laser device 10 one by one on the digital twin 100, or by combining multiple parts and breaking them in different combinations, data on malfunction phenomena corresponding to the broken parts can be obtained from the digital twin 100. The data stored in the database 104 may be, for example, as shown in FIG. 5.

[0045] FIG. 5 is a diagram showing an example of data stored in database 104. The destructible parts may be data linked to the module or assembly to which the destructible parts belong. An assembly is a unit made up of multiple parts. The "belonging module" item shown in FIG. 5 may be the assembly to which the destructible part belongs. The "belonging module" is a collection of parts that serve as the replacement unit when replacing parts during maintenance, and includes concepts such as assemblies, equipment, and detectors that serve as the replacement unit.

[0046] The data stored in the database 104 is used as training data for machine learning. That is, the training model 106 is trained by performing machine learning using the data stored in the database 104 (data created by the digital twin 100), and the training model 106 is created so that it receives input of a malfunction phenomenon and outputs the location of the failure.

[0047] The training model 106 may output the module or assembly to which the faulty part belongs along with the faulty part. The training model 106 may be configured with artificial intelligence (AI) such as an expert system or a Bayesian network, or may use a neural network.

[0048] The database 104 and the training model 106 are configured to be accessible from a network on the digital space DGS, which may be a wide area network such as the Internet.

[0049] The digital twin 100 and the database 104 may be constructed for each model of laser device 10. A training model for each model may be created using the database 104 constructed for each model of laser device 10. Furthermore, the training model 106 may be trained using databases for multiple laser models.

[0050] The digital space DGS is realized using a computer system including one or more processors (not shown) and one or more storage devices (not shown). The storage device is a computer-readable medium that is a non-transitory tangible entity, and includes, for example, a memory as a main storage device and a storage device as an auxiliary storage device. The computer that realizes the functions of the digital twin 100, the computer that stores and manages the database 104, and the computer that realizes the machine learning processing function of training the training model 106 may each be configured with separate hardware, or the computers that realize some or all of these processing functions may be configured with common hardware.

[0051] 2.2 Operation 2.2.1 Training data generation and training phase The digital twin 100 functions as a simulation model that can virtually reproduce the behavior of the laser device 10 in real space in the digital space DGS. By using the digital twin 100, it is possible to artificially create data about malfunctions for which it is difficult to actually collect data from the laser device 10 in real space, malfunctions for which it would take a great deal of time to actually collect data, and even extremely rare or unexpected malfunctions for which there have been no previous cases.

[0052] In this way, by performing supervised learning using the data set generated using the digital twin 100 as training data, a training model 106 is obtained that is trained to receive input of a malfunction phenomenon and output a fault location. The input to the training model 106 in the training phase may include an error code indicating the malfunction phenomenon. The input to the training model 106 may also include an output value of a sensor. The training model 106 performs a classification process to estimate the fault location based on the input malfunction information, and outputs the estimated result (classification result). The output from the training model 106 may be a classification score indicating the certainty of the fault location.

[0053] A broken part associated with a malfunction phenomenon in the training data is used as the correct answer data for the failure location corresponding to the malfunction phenomenon used as input. As an output from the training model 106, together with or instead of the failure location, an estimation result for the module or assembly to which the failure location belongs may be output. In this case, the module to which the broken part associated with a malfunction phenomenon in the training data belongs is used as the correct answer data for the module or assembly to which the failure location corresponding to the malfunction phenomenon used as input belongs.

[0054] By training the training model 106 using a large amount of data stored in the database 104, the parameters of the training model 106 can be updated to appropriate values, and the training model 106 can achieve the target inference performance. The trained training model 106 thus created can be used as a tool in place of the defect list 88 described in FIG. 3.

[0055] 2.2.2 Inference Phase Fig. 6 is an explanatory diagram schematically illustrating an example of how to use a trained model created by implementing the method for creating the training model 120 according to embodiment 1. The training model 120 shown in Fig. 6 is a trained model that has been trained by the method described in Fig. 4 and has acquired a reasonable inference performance.

[0056] The training model 120 is incorporated into a computer such as a server that can be accessed via a network. When the laser device 10 outputs information on a malfunction phenomenon (hereinafter referred to as malfunction information), the malfunction information is input to the training model 120 via the network. At this time, the laser device 10 may be connected to the network. Alternatively, a service engineer FSE may input the malfunction information to the training model 120 via the network using a terminal (not shown) that can be connected to the network. The terminal operated by the service engineer FSE may be a laptop personal computer, a tablet terminal, or the like.

[0057] The training model 120 provides a list of estimated fault locations based on the malfunction information to the on-site service engineer FSE via a network. At this time, the estimated list may be displayed on the display 84 of the laser device 10, or on the display of a terminal carried by the service engineer FSE. At this time, at least one fault location estimated by the training model 120 is displayed. Furthermore, the module to which the fault location belongs may be displayed.

[0058] FIG. 7 is a diagram showing an example of an estimated list of fault locations estimated using the training model 120. The display order of the estimated list may be set so that modules containing more fault locations are given priority (see FIG. 7). In other words, the estimated list is configured so that modules containing more fault locations are displayed with a higher priority. The "estimated ranking" in FIG. 7 corresponds to the display priority. A service engineer FSE can quickly perform part replacement or adjustment work by referring to the estimated list. If the estimated list is provided before the service engineer FSE goes to the site, he or she can head to the site with replacement parts in advance.

[0059] 2.3 Actions and Effects According to the method of the first embodiment, the training model 120 provides an estimation list, which significantly reduces the work required to identify the fault location. Furthermore, the malfunction phenomenon that occurs when all expected parts are broken can be grasped in advance on the digital twin 100. Therefore, it is possible to immediately derive appropriate countermeasures for errors that have never occurred in the real laser device 10 before or malfunctions that occur very rarely.

[0060] The method for creating the training models 106, 120 (creation method) described in embodiment 1 can be understood as a method for producing the training models 106, 120 (manufacturing method), as well as a method for producing a computer-readable medium on which the training models 106, 120 are recorded.

[0061] 3. Embodiment 2 3.1 Configuration Fig. 8 is an explanatory diagram showing an outline of a method for creating a training model 106 according to embodiment 2. The database 104 shown in Fig. 8 is an accumulation of artificial defect data created using the digital twin 100 as described in Fig. 4. Note that the digital twin 100 is not shown in Fig. 8. Differences between Fig. 8 and Fig. 4 will be described below.

[0062] In addition to training using the data in the database 104, the training model 106 may also be trained using data (actual data based on actual examples) showing the correspondence between actual malfunction phenomena obtained from the actual laser apparatus 10 and the corresponding actual failure locations. The information on the actual failure location preferably includes information on the module (failure module) to which the failure location belongs. Such data based on malfunction events actually confirmed in the laser apparatus 10 may be provided to the training model 106 via a network.

[0063] The data correlating actual malfunction phenomena with actual failure locations is not limited to data obtained from malfunction events that occurred at the operation site of the laser device 10, but may be data obtained from malfunction events that occurred experimentally in a development department, etc., and may also include data from a malfunction list accumulated as malfunction cases for the same type of laser model. Other system configurations may be similar to those shown in FIG.

[0064] 3.2 Operation As a method of utilizing actual data as training data, for example, there is a method of training the training model 106 using data in the database 104 to create a training model 120 that has inference performance at an acceptable practical level, and then further training the training model 120 using actual data as additional training data to enhance the inference performance. In this case, while the training model 120 that already has inference performance at a practical level is utilized at the service site, it is possible to further improve and enhance the performance of the training model 120 and update the training model 120 based on actual data obtained from the laser device 10 at the site.

[0065] Another method is to use actual data as training data, along with the data in the database 104, to train the training model 106 during the training process to achieve the target practical level of inference performance.

[0066] 3.3 Actions and Effects According to the method of the second embodiment, the inference of the training model 120 can be gradually strengthened by sequentially inputting actual data based on examples into the training model 120.

[0067] Furthermore, according to the method of embodiment 2, the inference accuracy of the training models 106, 120 can be improved by training the training models 106, 120 using actual data based on real examples in addition to data artificially created using the digital twin 100.

[0068] 4. Embodiment 3 4.1 Configuration FIG. 9 is an explanatory diagram showing an overview of laser devices 10 and 10B equipped with a training model 120 according to a third embodiment and examples of its use. In the configuration shown in FIG. 9, elements common to those in FIGS. 4 and 8 are designated by the same reference numerals. The training model 120 trained using data in the database 104 may be mounted on the laser device 10. The training model 120 may be mounted on a laser device 10B other than the laser device 10. If the training model 120 is created for each laser model, the laser devices 10 and 10B are the same model. If the training model 120 is created as a versatile model capable of inference for multiple laser models, the laser devices 10 and 10B may be different models. The laser device 10B includes a display 84B that displays various information.

[0069] The laser device 10, 10B may be shipped with the training model 120 installed, or the training model 120 may be downloaded into the device via a network after shipping and installation. The training model 120 may also be incorporated into the software of the laser device 10, 10B.

[0070] The training model 120 may be trained in a digital space DGS such as a server or cloud, as in the first or second embodiment.

[0071] 4.2 Operation The training model 120 in the laser device 10 may be connected to the control unit, software, memory, etc. of the laser device 10 so as to be accessible, and may receive data relating to malfunction phenomena from the processor 26.

[0072] When the training model 120 in the laser device 10 receives data related to the malfunction phenomenon, it presents the data to the on-site service engineer FSE by, for example, displaying an estimation list (see FIG. 5) including the estimation results of the malfunction cause and the fault location on the display 84, as in the first embodiment. Alternatively, the laser device 10 may transmit the estimation list of the malfunction cause and the fault location to a terminal of the laser device manufacturer or the service engineer FSE via a network, or may output it to an FDC (Fault Detection and Classification) system. The same applies to the operation of the training model 120 in the other laser device 10B.

[0073] The training model 120 may be updated as needed via a network. For example, data on actually occurring malfunctions and corresponding fault locations may be collected from each of the laser devices 10, 10B as needed, and the training model 120 may be updated by the training model 106, which is trained as needed in the digital space DGS using the collected actual data as additional training data. The model may be updated periodically, or the operator may specify the update time.

[0074] 4.3 Actions and Effects According to the configuration of the third embodiment, the service engineer (FSE) can quickly obtain a solution on-site or in advance, further shortening the time required to resolve the problem. Since the location of the failure can be estimated even with the laser device alone, appropriate countermeasures can be presented immediately, for example, even if a malfunction occurs in the communication function or the communication line conditions are poor. Furthermore, by performing additional training on the training model 106 using actual data, the inference accuracy of the training model 120 can be further improved, and the training model 120 can be kept up to date via the network.

[0075] 5. Other forms of laser devices 1 shows an example of a narrow-band KrF excimer laser device, but is not limited to this example and may be a narrow-band ArF excimer laser device. Also, while Fig. 1 shows an example of a single-chamber laser device 10, is not limited to this example and may be a laser device including a master oscillator that outputs narrow-band pulsed laser light and an amplifier that amplifies the pulsed laser light output from the master oscillator by a chamber containing excimer laser gas.

[0076] Furthermore, in a laser device including a master oscillator and an amplifier, the master oscillator may be a solid-state laser device that combines a solid-state laser with a nonlinear crystal and outputs narrowband laser light in the amplifiable wavelength range of an ArF laser or a KrF laser.

[0077] 6. Functional Roles of Information Processing Systems As described in FIG. 4, an information processing system that artificially creates data related to various malfunction phenomena using the digital twin 100 can function as a data generation device that automatically generates malfunction data. This information processing system can also function as a database creation device in that it can create the database 104 by organizing the automatically generated data into a database. The database 104 created using the digital twin 100 can be a malfunction data group that covers all malfunction phenomena that can occur in the laser device 10. While it is not possible to strictly cover all malfunction phenomena, it is preferable that the malfunction data group covers approximately all expected malfunction phenomena.

[0078] In addition to providing training data to the training model 106, the database 104 can also be used to, for example, accept input of a search key and return search results for data in the database 104.

[0079] As described with reference to FIGS. 4 and 8 , a machine learning system (machine learning device) that performs machine learning using the data in the database 104 as training data and trains the training model 106 functions as a training model creation device. An information processing system that estimates a failure location from a malfunction phenomenon using a trained training model 120 functions as a failure location estimation device. Furthermore, an information processing system that estimates a module, which is a replacement part to which the failure location belongs, from a malfunction phenomenon using the training model 120 functions as a replacement part estimation device or a malfunction countermeasure support device. The training model 120 is not limited to being incorporated into the laser device 10, 10B or a terminal carried by a service engineer (FSE), but may also be deployed on a cloud server or the like and applied as SaaS (Software as a Service) that accepts input of information on the malfunction phenomenon via a network and returns an estimation result of the failure location.

[0080] It is also possible to record a program that causes a computer to realize some or all of the processing functions of a data generation device, a machine learning device, a replacement part estimation device, or a defect countermeasure support device on a non-transitory, tangible computer-readable medium and distribute the program.

[0081] 7. Manufacturing methods for electronic devices FIG. 10 schematically illustrates an exemplary configuration of an exposure apparatus 90. The exposure apparatus 90 includes an illumination optical system 804 and a projection optical system 806. As described in FIG. 8, the laser apparatus 10 may be configured to include the training model 120. The laser apparatus 10 generates pulsed laser light and outputs it to the exposure apparatus 90. The illumination optical system 804 illuminates a reticle pattern of a reticle (not shown) placed on a reticle stage RT with the laser light incident from the laser apparatus 10. The projection optical system 806 reduces and projects the laser light that has passed through the reticle, forming an image on a workpiece (not shown) placed on a workpiece table WT. The workpiece is a photosensitive substrate such as a semiconductor wafer coated with photoresist.

[0082] The exposure apparatus 90 exposes a workpiece with laser light reflecting a reticle pattern by synchronously translating the reticle stage RT and the workpiece table WT. After the reticle pattern is transferred to the semiconductor wafer through the exposure process described above, a semiconductor device can be manufactured through multiple processes. A semiconductor device is an example of an "electronic device" in this disclosure. Instead of the laser apparatus 10, another laser apparatus 10B shown in FIG. 9 may be used, or a narrow-band ArF excimer laser apparatus may be used, or a laser apparatus including a master oscillator and an amplifier may be used.

[0083] 8.Other The above description is intended to be illustrative rather than limiting. Thus, it will be apparent to one skilled in the art that modifications can be made to the disclosed embodiments without departing from the scope of the claims. It will also be apparent to one skilled in the art that the disclosed embodiments can be used in combination.

[0084] Terms used throughout this specification and claims should be construed as "open ended" unless expressly stated otherwise. For example, terms such as "comprise," "have," "comprise," and "equip" should be construed as meaning "without excluding the presence of elements other than those listed." In addition, the modifier "a" should be construed as meaning "at least one" or "one or more." In addition, the term "at least one of A, B, and C" should be construed as "A," "B," "C," "A+B," "A+C," "B+C," or "A+B+C." Furthermore, it should be construed as including combinations of these with elements other than "A," "B," and "C."

Claims

1. Creating a database by destroying components within the laser device on a digital twin that models the electrical hardware and software of the laser device and accumulating data that associates the destroyed components with malfunction phenomena output by the digital twin; training a training model using the data in the database as training data for machine learning so as to receive input of information on a malfunction phenomenon and output information on a failure location corresponding to the input; How to create a training model, including:

2. 2. The method for creating a training model according to claim 1, The electrical hardware includes: a processor; a monitored object including a sensor whose state is monitored by the processor; a device including wiring connecting the processor and the monitored object; How to create a training model, including:

3. 2. The method for creating a training model according to claim 1, When information indicating that one or a combination of the components in the laser device has broken is input, the digital twin outputs at least one of an error code and a state quantity as the malfunction phenomenon. How to create a training model.

4. 2. The method for creating a training model according to claim 1, By destroying components in the laser device one by one on the digital twin, the data correlating the destroyed components with the malfunction phenomenon is accumulated in the database. How to create a training model.

5. 2. The method for creating a training model according to claim 1, By combining and destroying a plurality of parts in the laser device on the digital twin and changing the combination of the plurality of parts to be destroyed, the data correlating the destroyed parts with the malfunction phenomenon is accumulated in the database. How to create a training model.

6. 2. The method for creating a training model according to claim 1, The data stored in the database includes information on the module to which the destroyed part belongs. How to create a training model.

7. 7. A method for creating a training model according to claim 6, comprising: The information on the fault location output by the training model includes information on the module to which the fault location belongs. How to create a training model.

8. 7. A method for creating a training model according to claim 6, comprising: The training model outputs an estimation list including information on a module to which the fault location estimated from the input information on the malfunction phenomenon belongs, The estimation list is configured so that a module including a larger number of fault locations can be displayed with a higher priority. How to create a training model, including:

9. 2. The method for creating a training model according to claim 1, The digital twin and the database are constructed for each model of the laser device. How to create a training model.

10. 10. The method for creating a training model according to claim 9, The training model is trained using data contained in the database of multiple laser models constructed for each model. How to create a training model.

11. The method for creating a training model according to claim 1, further comprising: training the training model using actual data in which actual fault locations and malfunction phenomena occurring in the laser device are associated with each other; How to create a training model.

12. 12. The method for creating a training model according to claim 11, performing additional training on the training model using the actual data as additional training data. How to create a training model.

13. 13. The method for creating a training model according to claim 12, comprising: acquiring the actual data via a network to perform the additional training; How to create a training model.

14. 13. The method for creating a training model according to claim 12, comprising: updating the training model by performing the additional training using the actual data, and providing the updated training model over a network. How to create a training model.

15. 12. The method for creating a training model according to claim 11, The training model is trained using the actual data obtained from a plurality of laser devices as training data. How to create a training model.

16. 16. A method for creating a training model according to claim 15, comprising: performing additional training on the training model using the actual data obtained from the plurality of laser devices as additional training data, and updating the training model; How to create a training model.

17. 2. The method for creating a training model according to claim 1, loading the training model, trained using the data in the database, into a laser device; How to create a training model.

18. 1. A laser device, comprising: a processor; a monitored object including a sensor whose state is monitored by the processor; electrical hardware including a device including wiring connecting the processor and the monitored object; a training model trained by machine learning to receive input of information on a malfunction phenomenon and output information on a failure location corresponding to the input; Equipped with The training model is a model trained using, as training data, data in a database created by destroying components in the laser device on a digital twin that models the electrical hardware and software of the laser device and accumulating data correlating the destroyed components with malfunction phenomena output by the digital twin. Laser device.

19. A method for manufacturing an electronic device, comprising: a processor; a monitored object including a sensor whose state is monitored by the processor; electrical hardware including a device including wiring connecting the processor and the monitored object; a training model trained by machine learning to receive input of information on a malfunction phenomenon and output information on a failure location corresponding to the input; a laser device comprising: a laser beam generating unit configured to generate a laser beam from a laser beam source, the laser beam generating unit generating a laser beam from a laser beam source, a ... outputting the laser light to an exposure device; A method for manufacturing an electronic device, comprising exposing a photosensitive substrate to the laser light in the exposure apparatus to manufacture the electronic device.

Images