Training data generation method, machine learning method, consumable management device, and computer-readable medium

Abstract

A method for generating training data of one aspect of the present disclosure is a method for generating training data used in machine learning of a learning model for predicting the life of a consumable of a laser device. The method for generating training data includes the following steps: obtaining first life-related information, the first life-related information including data of at least one life-related parameter of the consumable recorded corresponding to different oscillation pulse numbers during the period from the start of use of the consumable to its replacement; determining a first degree of deterioration of the consumable according to the oscillation pulse number; determining a second degree of deterioration of the consumable according to at least one life-related parameter; determining a third degree of deterioration of the consumable according to the first degree of deterioration and the second degree of deterioration; and generating training data associating the first life-related information with the third degree of deterioration.

Description

Technical Field

[0001] The present disclosure relates to a training data generation method, a machine learning method, a consumable management device, and a computer-readable medium. Background Art

[0002] With the miniaturization and high integration of semiconductor integrated circuits, an improvement in resolution is required in semiconductor exposure apparatuses. Hereinafter, a semiconductor exposure apparatus will be simply referred to as an "exposure apparatus". Therefore, the shortening of the wavelength of light output from an exposure light source has been developed. In the exposure light source, a gas laser device is used instead of an existing mercury lamp. Currently, as the gas laser device for exposure, a KrF excimer laser device that outputs ultraviolet light with a wavelength of 248 nm and an ArF excimer laser device that outputs ultraviolet light with a wavelength of 193 nm are used.

[0003] As current exposure technology, the following immersion exposure has been put into practical use: a liquid is filled in the gap between the projection lens on the exposure apparatus side and the wafer, and by changing the refractive index of this gap, the apparent wavelength of the exposure light source is made shorter. In the case of performing immersion exposure using an ArF excimer laser device as the exposure light source, ultraviolet light with a wavelength of 134 nm in water is irradiated onto the wafer. This technology is called ArF immersion exposure. ArF immersion exposure is also called ArF immersion lithography.

[0004] The line width in the natural oscillation of a KrF or ArF excimer laser device is relatively wide, about 350 to 400 pm. Therefore, chromatic aberration occurs in the laser (ultraviolet light) that is reduced and projected onto the wafer by the projection lens on the exposure apparatus side, and the resolution is lowered. Therefore, it is necessary to narrow the line width of the laser output from the gas laser device to a degree where chromatic aberration can be ignored. The line width is also called the spectral width. Therefore, a line narrowing unit (Line Narrow Module) having a line narrowing element is provided in the laser resonator of the gas laser device, and the spectral width is narrowed by this line narrowing unit. In addition, the line narrowing element may be an etalon or a grating, etc. A laser device with a narrowed spectral width is called a line-narrowed laser device.

[0005] Prior Art Documents

[0006] Patent Documents

[0007] Patent Document 1: U.S. Patent Application Publication No. 2018 / 0246494

[0008] Patent Document 2: U.S. Patent No. 6219367

[0009] Patent Document 3: U.S. Patent No. 6697695 Summary of the Invention

[0010] A training data generation method according to one aspect of the present disclosure is a method for generating training data used in machine learning of a learning model for predicting the life of a consumable of a laser device. The training data generation method includes the following steps: obtaining first life-related information including data of at least one life-related parameter of the consumable recorded corresponding to different oscillation pulse numbers during the period from the start of use of the consumable to its replacement; determining a first deterioration degree of the consumable based on the oscillation pulse number; determining a second deterioration degree of the consumable based on at least one life-related parameter; determining a third deterioration degree of the consumable based on the first deterioration degree and the second deterioration degree; and generating training data that associates the first life-related information with the third deterioration degree.

[0011] A machine learning method according to another aspect of the present disclosure generates a learning model for predicting the life of a consumable of a laser device. The machine learning method includes the following steps: obtaining first life-related information including data of at least one life-related parameter of the consumable recorded corresponding to different oscillation pulse numbers during the period from the start of use of the consumable to its replacement; determining a first deterioration degree of the consumable based on the oscillation pulse number; determining a second deterioration degree of the consumable based on at least one life-related parameter; determining a third deterioration degree of the consumable based on the first deterioration degree and the second deterioration degree; generating training data that associates the first life-related information with the third deterioration degree; performing machine learning using the training data to thereby generate a learning model for predicting the deterioration degree of the consumable based on the data of the life-related parameters included in the first life-related information; and saving the generated learning model.

[0012] A computer-readable medium according to another aspect of the present disclosure is a non-volatile computer-readable medium recording a program. The program is a program that, when executed by a computer, implements a training data generation function used in machine learning of a learning model for predicting the life of a consumable of a laser device. The program includes commands for causing the computer to implement the following functions: obtaining first life-related information including data of at least one life-related parameter of the consumable recorded corresponding to different oscillation pulse numbers during the period from the start of use of the consumable to its replacement; determining a first deterioration degree of the consumable based on the oscillation pulse number; determining a second deterioration degree of the consumable based on at least one life-related parameter; determining a third deterioration degree of the consumable based on the first deterioration degree and the second deterioration degree; and generating training data that associates the first life-related information with the third deterioration degree. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Hereinafter, several embodiments of the present disclosure will be described as simple examples with reference to the drawings.

[0014] Figure 1 is a diagram schematically showing the structure of an exemplary laser device.

[0015] Figure 2 is a diagram schematically showing an example of the structure of a laser management system in a semiconductor factory.

[0016] Figure 3 is a graph showing an example of the relationship between the air pressure in a typical laser cavity and the number of oscillation pulses.

[0017] Figure 4 is a diagram showing the structure of the laser management system in the semiconductor factory of Embodiment 1.

[0018] Figure 5 is a block diagram showing the functions of a consumable management server.

[0019] Figure 6 is a flowchart showing an example of the processing content in a data acquisition unit.

[0020] Figure 7 is a flowchart showing an example of the processing content in a learning model generation unit.

[0021] Figure 8 is a graph showing an example of the relationship between the voltage in a laser cavity and the number of oscillation pulses, and showing an example of the degree of degradation up to the lifetime of the laser cavity given according to the number of oscillation pulses and the voltage.

[0022] Figure 9 is a graph showing an example of the relationship between the voltage in a laser cavity and the number of oscillation pulses, and showing another example of the degree of degradation up to the lifetime of the laser cavity given according to the number of oscillation pulses and the voltage.

[0023] Figure 10 is a graph showing an example of the relationship between the air pressure in a laser cavity and the number of oscillation pulses, and showing an example of the degree of degradation up to the lifetime of the laser cavity given according to the number of oscillation pulses, the air pressure, and the voltage.

[0024] Figure 11 is shown to be applied to Figure 7 Flowchart of Example 1 of the processing content of step S48.

[0025] Figure 12 is shown to be applied to Figure 11 Flowchart of Example 1 of the processing content of step S104.

[0026] Figure 13 is shown to be applied to Figure 11 Flowchart of Example 2 of the processing content of step S104.

[0027] Figure 14is a flowchart of Example 3 showing the processing content of step S104 applied to Figure 11 .

[0028] Figure 15 is a diagram showing an example of assigning a degradation degree based on a feature quantity derived from two values of voltage and air pressure.

[0029] Figure 16 is a diagram showing another example of assigning a degradation degree based on a feature quantity derived from two values of voltage and air pressure.

[0030] Figure 17 is a graph showing an image in which multiple data are included in one degradation degree interval after classification using the number of oscillation pulses.

[0031] Figure 18 is a flowchart of Example 2 showing the processing content of step S48 applied to Figure 7 .

[0032] Figure 19 is a flowchart showing an example of the processing content of step S104 applied to Figure 18 .

[0033] Figure 20 is a schematic diagram showing an example of a neural network model.

[0034] Figure 21 is an example of a model of a neural network when generating a learning model.

[0035] Figure 22 is a flowchart showing an example of the processing content in the consumable life prediction unit.

[0036] Figure 23 is a flowchart showing an example of the processing content of step S70 applied to Figure 22 .

[0037] Figure 24 is a graph showing an example of calculating the life and remaining life of a laser cavity using the generated learning model.

[0038] Figure 25 is a chart showing an example of the probability of each degradation degree classified into 10 levels.

[0039] Figure 26 is a diagram showing an example of the process of predicting the life of a consumable by a learned neural network model.

[0040] Figure 27 is a flowchart showing an example of the processing content in the data output unit.

[0041] Figure 28A chart showing an example of lifetime-related information of a laser cavity.

[0042] Figure 29 A chart showing an example of lifetime-related information of a laser cavity.

[0043] Figure 30 A chart showing an example of lifetime-related information of a laser cavity.

[0044] Figure 31 A chart showing an example of lifetime-related information of a monitor module.

[0045] Figure 32 A chart showing an example of lifetime-related information of a narrowbanding module. Detailed implementation

[0046] -Table of Contents-

[0047] 1. Explanation of terms

[0048] 2. Explanation of the laser device

[0049] 2.1 Structure

[0050] 2.2 Operation

[0051] 2.3 Maintenance of the main consumables of the laser device

[0052] 2.4 Others

[0053] 3. Example of the laser management system in a semiconductor factory

[0054] 3.1 Structure

[0055] 3.2 Operation

[0056] 4. Problems

[0057] 5. Embodiment 1

[0058] 5.1 Structure

[0059] 5.2 Operation

[0060] 5.2.1 Outline of the operation of machine learning in the consumable management server

[0061] 5.2.2 Outline of the operation of predicting the lifetime of consumables in the consumable management server

[0062] 5.2.3 Example of the processing of the data acquisition unit

[0063] 5.2.4 Example of the processing of the learning model generation unit

[0064] 5.2.5 Example 1 of the generation of the learning model used in predicting the lifetime of the laser cavity

[0065] 5.2.6 Generation Example 2 of Learning Model Used in Lifetime Prediction of Laser Cavity

[0066] 5.2.7 Example of Converting Multiple Parameters into One Parameter

[0067] 5.2.8 Explanation in the Case where Data D(s) Contains Multiple Data Counts

[0068] 5.2.9 Example of Neural Network Model

[0069] 5.2.10 Learning Mode of Neural Network Model

[0070] 5.2.11 Processing Example of Lifetime Prediction Unit of Consumable

[0071] 5.2.12 Example of Processing for Calculating Lifetime of Consumable Using Learning Model

[0072] 5.2.13 Lifetime Prediction Mode of Neural Network Model

[0073] 5.2.14 Others

[0074] 5.2.15 Processing Example of Data Output Unit

[0075] 5.3 Lifetime-Related Information of Laser Cavity

[0076] 5.4 Example of Lifetime-Related Information of Monitor Module

[0077] 5.5 Example of Lifetime-Related Information of Narrowbanding Module

[0078] 5.6 Function / Effect

[0079] 5.7 Others

[0080] 6. Variation Example

[0081] 7. Regarding Computer-Readable Medium Recorded with Program

[0082] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the accompanying drawings. The embodiments described below show several examples of the present disclosure and do not limit the content of the present disclosure. In addition, the structures and operations described in each embodiment are not necessarily all essential to the structures and operations of the present disclosure. In addition, the same reference numerals are assigned to the same structural elements and repeated descriptions are omitted.

[0083] 1. Explanation of Terms

[0084] "Consumables" is a term used to collectively represent items such as components and modules that require regular maintenance. Replacing components and replacing modules are included in the concept of "consumables". A module can be understood as a form of a component. In this specification, the term "consumables" is sometimes used synonymously with "replacement module or replacement component". Maintenance includes replacing consumables. In the concept of "replacement", in addition to replacing consumables with new ones, it also includes cases such as cleaning consumables to maintain and / or restore the functions of components and reconfiguring the same consumables.

[0085] "Burst operation" means an operation that alternately repeats a burst period and an oscillation pause period. During the burst period, a pulsed laser that is narrowed in bandwidth oscillates continuously in response to exposure, and during the oscillation pause period, the oscillation pauses in response to the movement of the stage.

[0086] 2. Description of the laser device

[0087] 2.1 Structure

[0088] In Figure 1 the structure of an exemplary laser device 10 is schematically shown. The laser device 10 is, for example, a KrF excimer laser device and includes a laser control unit 90, a laser cavity 100, an inverter 102, an output coupler 104, a line narrow module (LNM) 106, a monitor module 108, a charger 110, a pulse power module (PPM) 112, a gas supply device 114, a gas exhaust device 116, and an exit port shutter 118.

[0089] The laser cavity 100 includes a first window 121, a second window 122, a cross-flow fan (CFF) 123, a motor 124 that rotates the CFF 123, a pair of electrodes 125, 126, an electrical insulator 127, a pressure sensor 128, and a heat exchanger (not shown).

[0090] The inverter 102 is a power supply device for the motor 124. The inverter 102 receives a command signal from the laser control unit 90, and this command signal determines the frequency of the power supplied to the motor 124.

[0091] The PPM 112 is connected to the electrode 125 through a feed-through hole in the electrical insulator 127 of the laser cavity 100. The PPM 112 includes a semiconductor switch 129, charging capacitors (not shown), a pulse transformer, and a pulse compression circuit.

[0092] The output coupler 104 is a partially reflecting mirror and is configured to form an optical resonator together with the line narrow module 106. The laser cavity 100 is disposed on the optical path of this optical resonator.

[0093] The narrowbanding module 106 includes a beam expander using a first prism 131 and a second prism 132, a rotating stage 134, and a grating 136. The first prism 131 and the second prism 132 are configured to expand the beam of light exiting from the second window 122 of the laser cavity 100 so that it is incident on the grating 136.

[0094] Here, the grating 136 is in a Littrow configuration such that the incident angle and the diffraction angle of the laser are the same. The second prism 132 is disposed on the rotating stage 134 so that when the rotating stage 134 rotates, the incident angle and the diffraction angle of the laser with respect to the grating 136 change.

[0095] The monitor module 108 includes a first beam splitter 141 and a second beam splitter 142, a pulse energy detector 144, and a spectrum detector 146. The first beam splitter 141 is disposed on the optical path of the laser output from the output coupling mirror 104 and is configured such that a part of the laser is reflected and incident on the second beam splitter 142.

[0096] The pulse energy detector 144 is configured to receive the laser passing through the second beam splitter 142. The pulse energy detector 144 can be, for example, a photodiode that measures the light intensity of ultraviolet light. The second beam splitter 142 is configured such that a part of the laser is reflected and incident on the spectrum detector 146.

[0097] The spectrum detector 146 is, for example, a monitoring etalon measurement device that uses an image sensor to measure the interference fringes generated by an etalon. The central wavelength and the spectral line width of the laser are measured based on the generated interference fringes.

[0098] In the case of a KrF excimer laser device, the gas supply device 114 is connected via pipes to an inert gas supply source 152, which is a supply source of an inert laser gas, and a halogen gas supply source 153, which is a supply source of a laser gas containing a halogen, respectively. The inert laser gas is a mixed gas of Kr gas and Ne gas. The laser gas containing a halogen is a mixed gas of F 2 gas, Kr gas, and Ne gas. The gas supply device 114 is connected to the laser cavity 100 via a pipe.

[0099] The gas supply device 114 respectively includes an automatic valve and a mass flow controller (not shown) for supplying a predetermined amount of the inert laser gas or the laser gas containing a halogen to the laser cavity 100, respectively.

[0100] The gas exhaust device 116 is connected to the laser cavity 100 via a pipe. The gas exhaust device 116 includes a halogen filter (not shown) for removing halogen and an exhaust pump, and is configured to discharge the laser gas from which halogen has been removed to the outside.

[0101] The output port shutter 118 is disposed on the optical path of the laser output from the laser device 10 to the outside.

[0102] The laser device 10 is configured such that the laser output from the laser device 10 via the output port shutter 118 is incident on the exposure device 14.

[0103] 2.2 Operation

[0104] The operation of the laser device 10 will be described. After the laser control unit 90 discharges the gas in the laser cavity 100 via the gas exhaust device 116, the laser cavity 100 is filled with an inert laser gas and a laser gas containing a halogen via the gas supply device 114 to achieve a desired gas composition and total gas pressure.

[0105] The laser control unit 90 rotates the motor 124 at a specified rotational speed by means of the inverter 102 to rotate the CFF 123. As a result, the laser gas flows between the electrodes 125 and 126.

[0106] The laser control unit 90 receives the target pulse energy Et from the exposure control unit 50 of the exposure device 14 and sends data of the charging voltage Vhv to the charger 110 so that the pulse energy becomes Et.

[0107] The charger 110 charges the charging capacitor of the PPM 112 to the charging voltage Vhv. When the light emission trigger signal Tr1 is output from the exposure device 14, the trigger signal Tr2 is input from the laser control unit 90 to the semiconductor switch 129 of the PPM 112 synchronously with the light emission trigger signal Tr1. After the semiconductor switch 129 operates, the current pulse is compressed by the magnetic compression circuit of the PPM 112, and a high voltage is applied between the electrodes 125 and 126. As a result, a discharge occurs between the electrodes 125 and 126, and the laser gas is excited in the discharge space. The electrodes 125 and 126 are an example of the "discharge electrodes" in the present disclosure.

[0108] When the excited laser gas in the discharge space returns to the ground state, excimer light is generated. The excimer light reciprocates between the output coupling mirror 104 and the narrowing module 106 and is amplified, thereby performing laser oscillation. As a result, a pulse laser that has been narrowed is output from the output coupling mirror 104.

[0109] The pulse laser output from the output coupling mirror 104 is incident on the monitor module 108. In the monitor module 108, a part of the laser is sampled by the first beam splitter 141 and made incident on the second beam splitter 142. The second beam splitter 142 transmits a part of the incident laser and makes it incident on the pulse energy detector 144, and reflects the other part and makes it incident on the spectrum detector 146.

[0110] The pulse energy E of the pulsed laser output from the laser device 10 is measured by the pulse energy detector 144, and the data of the measured pulse energy E is sent from the pulse energy detector 144 to the laser control unit 90.

[0111] In addition, the center wavelength λ and the spectral line width Δλ are measured by the spectral detector 146, and the data of the measured center wavelength λ and spectral line width Δλ are sent from the spectral detector 146 to the laser control unit 90.

[0112] The laser control unit 90 receives the data of the target pulse energy Et and the target wavelength λt from the exposure device 14. The laser control unit 90 controls the pulse energy based on the pulse energy E measured by the pulse energy detector 144 and the target pulse energy Et. The control of the pulse energy includes controlling the charging voltage Vhv so that the difference ΔE = E - Et between the pulse energy E measured by the pulse energy detector 144 and the target pulse energy Et approaches 0.

[0113] The laser control unit 90 controls the wavelength based on the center wavelength λ measured by the spectral detector 146 and the target wavelength λt. The control of the wavelength includes controlling the rotation angle of the turntable 134 so that the difference δλ = λ - λt between the center wavelength λ measured by the spectral detector 146 and the target wavelength λt approaches 0.

[0114] As described above, the laser control unit 90 receives the target pulse energy Et and the target wavelength λt from the exposure device 14, and whenever the light emission trigger signal Tr1 is input, the laser device 10 outputs a pulsed laser synchronously with the light emission trigger signal Tr1.

[0115] After the laser device 10 repeatedly discharges, the electrodes 125 and 126 are consumed, the halogen gas in the laser gas is consumed, and impurity gases are generated. The decrease in the halogen gas concentration and the increase in the impurity gases in the laser cavity 100 cause a decrease in the pulse energy of the pulsed laser, which has an adverse effect on the stability of the pulse energy. The laser control unit 90 performs the following gas control, for example, to suppress these adverse effects.

[0116] [1] Halogen injection control

[0117] The halogen injection control is the following gas control: during laser oscillation, a gas containing halogen is injected at a concentration higher than that of the halogen gas in the laser cavity 100, thereby replenishing the halogen gas in the laser cavity 100 in an amount mainly consumed by discharge.

[0118] [2] Partial gas replacement control

[0119] Partial gas replacement control is the following gas control: during laser oscillation, a part of the laser gas in the laser cavity 100 is replaced with new laser gas to suppress an increase in the concentration of impurity gas in the laser cavity 100.

[0120] [3] Gas pressure control

[0121] Gas pressure control is the following gas control: laser gas is injected into the laser cavity 100 to change the gas pressure P of the laser gas, thereby controlling the pulse energy. Generally, the pulse energy is controlled by controlling the charging voltage Vhv. However, when the reduction in the pulse energy of the pulsed laser output from the laser device 10 cannot be compensated within the control range of the charging voltage Vhv, gas pressure control is executed.

[0122] When discharging the laser gas from the laser cavity 100, the laser control unit 90 controls the gas exhaust device 116. The laser gas discharged from the laser cavity 100 has the halogen gas removed by an unillustrated halogen filter and is discharged to the outside of the laser device 10.

[0123] The laser control unit 90 sends data of various parameters such as the number of oscillation pulses, charging voltage Vhv, gas pressure P in the laser cavity 100, pulse energy E of the laser, spectral line width Δλ, etc. to the laser device management system 206 (refer to Figure 2 ).

[0124] 2.3 Maintenance of main consumables of the laser device

[0125] The replacement operations of the main consumables performed by the field service engineer (FSE) are the replacement operations of the laser cavity 100, the narrowbanding module 106, and the monitor module 108.

[0126] The replacement periods of these main consumables are generally not managed by time but by the number of oscillation pulses of the laser device 10. The replacement operations of these main consumables sometimes require a replacement time of 3 hours to 10 hours. Among these main consumables, the consumable with the longest replacement time is the laser cavity 100.

[0127] 2.4 Others

[0128] In Figure 1 the example shown, as the laser device 10, an example of a KrF excimer laser device is shown, but it is not limited to this example and can also be applied to other laser devices. For example, the laser device 10 can also be an ArF excimer laser device or an XeCl excimer laser device.

[0129] In Figure 1In the example shown, regarding the gas control of the laser device 10, the implementation of halogen injection control, partial gas replacement control, and air pressure control is shown, but it is not limited to this example. For example, air pressure control may not be implemented.

[0130] 3. Example of a Laser Management System in a Semiconductor Factory

[0131] 3.1 Structure

[0132] Figure 2 A structural example of the laser management system 200 in a semiconductor factory is schematically shown. The laser management system 200 includes a plurality of laser devices 10, a management system 206 for laser devices, and a semiconductor factory management system 208.

[0133] The management system 206 for laser devices and the semiconductor factory management system 208 are each configured using a computer. The management system 206 for laser devices and the semiconductor factory management system 208 can each be a computer system configured using a plurality of computers. The semiconductor factory management system 208 is connected to the management system 206 for laser devices via the network 210.

[0134] The network 210 is a communication line capable of information transfer based on wired, wireless, or a combination thereof. The network 210 can be a wide area network or a local area network.

[0135] To separately identify the plurality of laser devices 10, the laser device identification labels #1, #2,... #k,... #w are used here. w is the number of laser devices 10 included in the laser management system 200 in the semiconductor factory. w is an integer of 1 or more. k is an integer in the range of 1 or more and w or less. Hereinafter, for the sake of convenience of explanation, it is sometimes referred to as the laser device #k. In addition, the laser devices #1 to #w may have the same device structure, and a part or all of the laser devices #1 to #w may have different device structures from each other.

[0136] The laser devices #1 to #w and the management system 206 for laser devices are each connected to the local area network 213. In Figure 2 it, the local area network 213 is shown as "LAN".

[0137] 3.2 Operation

[0138] The management system 206 for laser devices mainly manages the replacement timing of the main consumables of each of the laser devices #1 to #w according to the number of pulses of laser oscillation (oscillation pulse number) Np.

[0139] The management system 206 for laser devices can display the management information of maintenance on a display terminal, or can send it to the semiconductor factory management system 208 via the network 210.

[0140] Regarding the management lines for managing laser devices #1 to #w by the management system 206 using a laser device, each management line is independent, and the manager of the semiconductor factory determines the replacement timing of the main consumables of each laser device #1 to #w based on the maintenance management information output from each laser device #1 to #w.

[0141] Figure 3 It is a graph showing an example of the relationship between the gas pressure P and the number of oscillation pulses Np in a typical laser cavity 100. After the excimer laser device repeatedly discharges, the electrodes 125 and 126 are consumed, the halogen gas in the laser gas is consumed, and impurity gases are generated. The decrease in the halogen gas concentration and the increase in impurity gases in the laser cavity 100 cause a decrease in the pulse energy of the pulsed laser, which has an adverse effect on the pulse energy stability.

[0142] Therefore, in order to maintain the performance of the excimer laser device, halogen injection control, partial gas replacement control, gas pressure control, or full gas replacement is performed according to the situation. In Figure 3 the upward arrow is used to indicate the replacement timing of the laser cavity 100. The operation after replacing the laser cavity 100 is as described below.

[0143] [Step 1] The gas pressure P immediately after replacing the laser cavity 100 maintains the laser performance at the initial gas pressure Pch.

[0144] [Step 2] When continuous laser oscillation is performed, due to the consumption of the discharge electrodes and the generation of impurity gases, in order to maintain the laser performance, the gas pressure P is increased by gas pressure control. Figure 3 The curve shown by the thick line in shows the change of the gas pressure P in this Step 2.

[0145] [Step 3] However, when the laser performance still cannot be maintained even by gas pressure control, the laser oscillation is stopped and full gas replacement is performed. In Figure 3 the downward arrow is used to indicate the timing of full gas replacement.

[0146] [Step 4] Adjustment oscillation is performed after full gas replacement. Gas pressure control is performed to restore the laser performance. The gas pressure P when the laser performance is restored is called the "initial gas pressure after full gas replacement" and is set as Pini.

[0147] [Step 5] Then, Steps 2 to 4 are repeated multiple times. The initial gas pressure Pini after full gas replacement gradually increases as the number of oscillation pulses Np increases. Figure 3 The curve shown by the thin line in shows the change of the initial gas pressure Pini.

[0148] [Step 6] Finally, when the gas pressure P reaches the maximum allowable gas pressure Pmax, the laser cavity life Nchlife is reached.

[0149] In Figure 3 In the example shown, for simplicity, regarding the lifetime of the laser cavity 100, the process from the change in the gas pressure P with respect to the number of oscillation pulses Np of the laser device 10 until the lifetime is reached has been described. However, other laser performances, such as pulse energy stability and line width, etc., also need to be satisfied. Therefore, it is sometimes impossible to simply predict the lifetime of the laser cavity 100.

[0150] 4. Problems

[0151] [Problem 1] Sometimes, the value of the number of oscillation pulses as the standard lifetime is determined for each main consumable of the laser device. However, due to individual differences in the consumables, the number of oscillation pulses until the lifetime is reached is not fixed and there are deviations. In the case where the lifetime of the consumable is longer than the standard lifetime, sometimes the consumable is replaced as regular maintenance at the time of the standard lifetime. In addition, in the case where the lifetime of the consumable is shorter than the standard lifetime, sometimes the planned replacement of the consumable cannot be carried out and the production line stops.

[0152] [Problem 2] Currently, the FSE views log data such as Figure 3 the transition of the gas pressure with respect to the number of oscillation pulses and other parameters related to the lifetime, and predicts the lifetime of each consumable according to experience and takes countermeasures. Therefore, the prediction of the lifetime of the consumable and the countermeasures until the replacement of the consumable sometimes depend on the ability of the individual FSE.

[0153] 5. Embodiment 1

[0154] 5.1 Structure

[0155] Figure 4 is a diagram showing the structure of the laser management system 300 of the semiconductor factory according to Embodiment 1. Regarding Figure 4 the structure shown, the differences from Figure 2 will be described. Figure 4 The laser management system 300 of the semiconductor factory shown is configured by adding a consumable management server 310 to the structure of the laser management system 200 of Figure 2 . The consumable management server 310 is connected to the laser device management system 206 and the semiconductor factory management system 208 via the network 210.

[0156] The consumable management server 310 is configured to be able to send and receive data and signals to and from the laser device management system 206 and the semiconductor factory management system 208, respectively.

[0157] Figure 5It is a block diagram showing the functions of the consumable management server 310. The consumable management server 310 includes a data acquisition unit 320, a life-related information storage unit 330 for consumables, a learning model generation unit 340 based on machine learning, a learning model storage unit 350, a life prediction unit 360 for consumables, and a data output unit 370.

[0158] The life-related information of the consumables includes File A, File B, and File C. File A is a file storing the life-related log data of the laser cavity 100. File B is a file storing the life-related log data of the monitor module 108. File C is a file storing the life-related log data of the narrowbanding module 106.

[0159] The life-related information storage unit 330 for consumables includes a storage unit 332 that pre-stores File A, a storage unit 334 that pre-stores File B, and a storage unit 336 that pre-stores File C.

[0160] The learning model generation unit 340 is a processing unit that generates a learning model through machine learning. The learning model storage unit 350 for consumables stores the learning model generated by the learning model generation unit 340. The learning model storage unit 350 for consumables includes a storage unit 352 that pre-stores File Am, a storage unit 354 that pre-stores File Bm, and a storage unit 356 that pre-stores File Cm.

[0161] File Am is a file storing the first learning model, and the first learning model performs processing for predicting the life of the laser cavity 100. File Bm is a file storing the second learning model, and the second learning model performs processing for predicting the life of the monitor module 108. File Cm is a file storing the third learning model, and the third learning model performs processing for predicting the life of the narrowbanding module 106.

[0162] The storage units 332, 334, 336, 352, 354, and 356 are constituted by storage devices such as hard disk devices and / or semiconductor memories. The storage units 332, 334, 336, 352, 354, and 356 can be respectively constituted by different storage devices, or can be constituted as a part of the storage areas in one or more storage devices.

[0163] In the present disclosure, the laser control unit 90, the exposure control unit 50, the laser device management system 206, the semiconductor factory management system 208, and the consumable management server 310 can be respectively implemented by a combination of hardware and software of one or more computers. Software is synonymous with a program. A programmable controller is included in the concept of a computer.

[0164] A computer can be configured to include, for example, a CPU (Central Processing Unit) and a storage device. A programmable controller is included in the concept of a computer. A computer may also include a GPU (Graphics Processing Unit). The CPU and GPU included in the computer are an example of a processor. The storage device is a non-transitory computer-readable medium as a tangible object, and includes, for example, a memory as a main storage device and a storage as an auxiliary storage device. The computer-readable medium may also be, for example, a semiconductor memory, a hard disk drive (HDD) device, or a solid state drive (SSD) device, or a combination of multiple thereof. The program executed by the processor is stored in the computer-readable medium. The processor may also be a structure including the computer-readable medium.

[0165] In addition, part or all of the functions of various control devices and processing devices such as the laser control unit 90, the exposure control unit 50, the laser device management system 206, the semiconductor factory management system 208, and the consumable management server 310 may also be implemented using integrated circuits represented by an FPGA (Field Programmable Gate Array) or an ASIC (Application Specific Integrated Circuit).

[0166] In addition, it is also possible to implement the functions of multiple control devices and processing devices using one device. Furthermore, in the present disclosure, multiple control devices and processing devices may be interconnected via a communication network such as a local area network or an Internet line. In a distributed computing environment, program units may also be stored in both local and remote storage devices. The processors applied to the laser control unit 90, the exposure control unit 50, the laser device management system 206, the semiconductor factory management system 208, the consumable management server 310, etc. are specifically configured or programmed to execute various processes included in the present disclosure.

[0167] 5.2 Operations

[0168] 5.2.1 Outline of the Operations of Machine Learning in the Consumable Management Server

[0169] Figure 5The consumable management server 310 shown has the following functions: performing machine learning for generating a learning model used in the process of predicting the life of the consumables of the laser device 10; and performing a process of predicting the life of the consumables using the generated learning model. The consumable management server 310 is an example of the "consumable management device" in the present disclosure. First, a machine learning method for generating a learning model used in the prediction of the life of the consumables and a method for generating training data used in the machine learning in the consumable management server 310 will be described.

[0170] When a consumable is replaced in each laser device 10, the data acquisition unit 320 acquires life-related information from the laser device management system 206. The life-related information includes all data of life-related parameters continuously recorded corresponding to the oscillation pulse number Np during all periods of the use period of the replaced consumable. The data acquisition unit 320 writes the data acquired from the laser device management system 206 into the consumable life-related information storage unit 330.

[0171] The data acquisition unit 320 determines the file to be written according to the type of the replaced consumable and writes the data. When the replaced consumable is the laser cavity 100, the data acquisition unit 320 writes life-related log data as the life-related information of the laser cavity 100 into file A. When the replaced consumable is the monitor module 108, the data acquisition unit 320 writes life-related log data as the life-related information of the monitor module 108 into file B. When the replaced consumable is the narrowbanding module 106, the data acquisition unit 320 writes life-related log data as the life-related information of the narrowbanding module 106 into file C. The log data in each of the files A, B, and C written is an example of the "first life-related information" in the present disclosure. The data acquisition unit 320 is an example of the "information acquisition unit" in the present disclosure.

[0172] When data of new life-related information related to the replaced consumable is stored in the consumable life-related information storage unit 330, the learning model generation unit 340 acquires the newly stored data of the life-related information. In addition, the learning model generation unit 340 retrieves the learning model corresponding to the replaced consumable from the consumable learning model storage unit 350.

[0173] For example, when the replaced consumable is the laser cavity 100, the learning model generation unit 340 retrieves file Am. When the replaced consumable is the monitor module 108, the learning model generation unit 340 retrieves file Bm. When the replaced consumable is the narrowbanding module 106, the learning model generation unit 340 retrieves file Cm.

[0174] The learning model generation unit 340 performs machine learning based on the data of the life-related parameters recorded during the period from the start of use of the replaced consumable to its replacement, and generates a new learning model. The details of the specific machine learning method will be described later. The new learning model generated by the learning model generation unit 340 is stored in the learning model storage unit 350 of the consumable. When a new learning model is generated by performing machine learning, the file in the learning model storage unit 350 is updated, and the file of the latest learning model is written in the learning model storage unit 350.

[0175] 5.2.2 Outline of the operation of predicting the life of consumables in the consumable management server

[0176] Next, the operation of predicting the life of consumables in the consumable management server 310 will be described. The data acquisition unit 320 can receive a request signal for predicting the life of a consumable to be replaced from an external device. The external device here can be the semiconductor factory management system 208 or a terminal device (not shown), etc. The "consumable to be replaced" is the consumable currently mounted on the laser device 10 and in use, and is a consumable that is a candidate object to be studied for future replacement.

[0177] After the data acquisition unit 320 receives a request for predicting the life of a consumable to be replaced, it acquires the data of the current life-related information of the consumable to be replaced and the data of the number of oscillation-predicted pulses Nday per day from the management system 206 for the laser device.

[0178] The data acquisition unit 320 sends the data of the current life-related information of the consumable to be replaced and the data of the number of oscillation-predicted pulses Nday per day to the life prediction unit 360 of the consumable.

[0179] The life prediction unit 360 of the consumable acquires the data of the current life-related information of the consumable to be replaced and the data of the number of oscillation-predicted pulses Nday per day, and retrieves the learning model corresponding to the consumable to be replaced from the learning model storage unit 350 of the consumable.

[0180] For example, when the consumable to be replaced is the laser cavity 100, the life prediction unit 360 of the consumable reads the file Am from the learning model storage unit 350 of the consumable.

[0181] The life prediction unit 360 of the consumable predicts the life of the consumable based on the data of the current life-related information by using the learning model.

[0182] The consumable life prediction unit 360 calculates the data of the number of oscillation pulses of the predicted life Nlife and the remaining life Nre of the consumable to be replaced, and the recommended maintenance date Drec, and sends this data to the data output unit 370.

[0183] The recommended maintenance date Drec can be calculated, for example, using the following formula.

[0184] Drec = Dpre + Nre / Nday

[0185] Dpre: Acquisition date of the current life-related data of the consumable

[0186] The data output unit 370 sends the data of the number of oscillation pulses of the predicted life Nlife and the remaining life Nre of the consumable to be replaced and the data indicating the recommended maintenance date Drec to the laser device management system 206 via the network 210. The data output unit 370 is an example of the "information output unit" in the present disclosure.

[0187] The laser device management system 206 can also notify the semiconductor factory management system 208, the operator, and the FSE of the information on the predicted life Nlife, the number of oscillation pulses of the remaining life Nre, and the recommended maintenance date Drec of the consumable to be replaced through a display, email, etc.

[0188] Regarding this notification, it can also be notified from the consumable management server 310 via the network 210.

[0189] 5.2.3 Processing example of the data acquisition unit

[0190] Figure 6 It is a flowchart showing an example of the processing content in the data acquisition unit 320. Figure 6 The processing and operations shown in the flowchart are realized, for example, by a processor functioning as the data acquisition unit 320 executing a program.

[0191] In step S12, the data acquisition unit 320 determines whether the consumable has been replaced. If the determination result in step S12 is "yes", the data acquisition unit 320 proceeds to step S14. Steps S14 and S16 are the processing flows when generating the learning model.

[0192] In step S14, the data acquisition unit 320 receives all the life-related information during the usage period of the replaced consumable. That is, when replacing the consumable of the laser device 10, the data acquisition unit 320 receives all the life-related information during the usage period of the replaced consumable from the laser device management system 206.

[0193] Next, in step S16, the data acquisition unit 320 writes all the life-related information during the usage period of the consumable to be replaced into the life-related information storage unit 330. That is, the data acquisition unit 320 writes data in the file corresponding to the consumable to be replaced. Here, the consumable to be replaced is the laser cavity 100, the monitor module 108, or the narrowbanding module 106, and the data acquisition unit 320 writes data in file A, file B, or file C according to the type of the consumable.

[0194] After step S16, the data acquisition unit 320 proceeds to step S30. In step S30, the data acquisition unit 320 determines whether to stop receiving information. If the determination result in step S30 is "no", the data acquisition unit 320 returns to step S12.

[0195] If the determination result in step S12 is "no", the data acquisition unit 320 proceeds to step S20. In step S20, the data acquisition unit 320 determines whether to calculate the life of the consumable to be replaced. For example, if the user inputs a request for life prediction related to the consumable to be replaced from an input device (not shown), the determination result in step S20 becomes "yes".

[0196] If the determination result in step S20 is "yes", the data acquisition unit 320 proceeds to step S22. Steps S22, S24, and S26 are the processing flows when calculating the predicted life of the consumable to be replaced. Calculating the predicted life of the consumable means predicting the life of the consumable.

[0197] In step S22, the data acquisition unit 320 receives the current life-related information of the consumable to be replaced from the laser device management system 206.

[0198] In step S24, the data acquisition unit 320 receives the operation-related information of the laser device 10 from the laser device management system 206. The operation-related information of the laser device 10 is the number of pulses Nday scheduled for oscillation per day. Specifically, it can be the number of pulses Nday scheduled for oscillation per day grasped based on past operation data. Or, future operation schedule information can also be obtained from the semiconductor factory management system 208 to calculate the number of pulses Nday scheduled for oscillation per day.

[0199] Then, in step S26, the data acquisition unit 320 sends the current life-related information and the operation-related information of the laser device 10 to the consumable life prediction unit 360.

[0200] After step S26, the data acquisition unit 320 proceeds to step S30. Further, in the case where the determination result in step S20 is "No", the data acquisition unit 320 skips steps S22 to S26 and proceeds to step S30.

[0201] In the case where the determination result in step S30 is "Yes", the data acquisition unit 320 ends Figure 6 the flowchart.

[0202] 5.2.4 Processing Example of the Learning Model Generation Unit

[0203] Figure 7 is a flowchart showing an example of the processing content in the learning model generation unit 340. Figure 7 The processing and operations shown in the flowchart are implemented, for example, by a processor functioning as the learning model generation unit 340 executing a program.

[0204] In step S42, the learning model generation unit 340 determines whether new data has been written in the consumable life-related information storage unit 330. In the case where the determination result in step S42 is "No", the learning model generation unit 340 repeats step S42. In the case where the determination result in step S42 is "Yes", the learning model generation unit 340 proceeds to step S44.

[0205] In step S44, the learning model generation unit 340 acquires all the life-related information during the usage period of the replaced consumable. The learning model generation unit 340 acquires the data written in the file (file A, file B, or file C) corresponding to the replaced consumable (laser cavity 100, monitor module 108, or narrowbanding module 106).

[0206] In step S46, the learning model generation unit 340 retrieves the learning model of the replaced consumable. That is, the learning model generation unit 340 retrieves the learning model stored in the file (file Am, file Bm, or file Cm) corresponding to the replaced consumable.

[0207] In step S48, the learning model generation unit 340 executes the processing of the learning model generation subroutine. The learning model generation unit 340 performs machine learning based on the learning model corresponding to the replaced consumable and the life-related information to generate a new learning model.

[0208] In step S50, the learning model generation unit 340 stores the newly generated learning model in the learning model storage unit 350 of the consumable. The learning model generation unit 340 stores the newly generated learning model in a file (file Am, file Bm, or file Cm) corresponding to the replaced consumable. The latest learning model is stored in the learning model storage unit 350 so that the new learning model can be used from the next time.

[0209] In step S52, the learning model generation unit 340 determines whether to abort the generation of the learning model. When the determination result in step S52 is "no", the learning model generation unit 340 returns to step S42 and repeats steps S42 to S52. When the determination result in step S52 is "yes", the learning model generation unit 340 ends Figure 7 the flowchart of.

[0210] In addition, in Figure 7 the flowchart of, when initially generating the learning model of each consumable, the parameters of each initial learning model stored in the learning model storage unit 350 can be set to any value before learning. By implementing the machine learning described later, the parameters of the learning model are changed to appropriate values, and a learning model with the processing function of predicting the life of the consumable is generated.

[0211] Of course, the initial learning model can also be a tentative learning model whose parameters have been adjusted to a certain extent by previously implementing the same method as the machine learning method of the present embodiment.

[0212] 5.2.5 Generation Example 1 of the Learning Model Used in the Life Prediction of the Laser Cavity

[0213] The learning model generated by the learning model generation unit 340 is learned to receive the input of life-related information and output the degradation degree of the consumable as a prediction (inference) result. The processing performed by the learning model generation unit 340 includes the processing of generating training data used in machine learning and the processing of implementing machine learning using the generated training data. First, an example of the training data generation method implemented by the learning model generation unit 340 will be described. In addition, the training data is synonymous with "learning data" or "learning data".

[0214] Figure 8 is a graph showing an example of the relationship between the voltage V of the laser cavity 100 and the number of oscillation pulses Np, showing an example of the degradation degree up to the life of the laser cavity 100 given according to the number of oscillation pulses Np and the voltage V. Figure 8 The horizontal axis of represents the number of oscillation pulses Np, and the vertical axis represents the voltage V applied between the electrodes 125 and 126. It can be read from the life-related log data of the laser cavity 100 stored in file A Figure 8Data of such voltage V associated with the number of oscillation pulses Np.

[0215] Set one cycle of consumable replacement as the life of each consumable, and define the degree of deterioration DLn based on the number of oscillation pulses Np in levels (for example, 10 levels). The degree of deterioration DLn is obtained by evaluating the level of deterioration of the laser cavity 100 according to the number of oscillation pulses Np. As the number of oscillation pulses Np increases, the value representing the level of the degree of deterioration DLn becomes a larger value. The higher the level of the degree of deterioration DLn, that is, the larger the value representing the level of the degree of deterioration DLn, the more developed the state of deterioration is indicated. When defining the degree of deterioration DLn in 10 levels according to the number of oscillation pulses Np, the minimum level value can be set to 1, and the maximum level value can be set to 10.

[0216] Here, in order to compensate for the energy reduction associated with the increase in the impurity concentration of the gas, the voltage V applied to the electrodes 125 and 126 has a tendency to gradually increase. Therefore, set the maximum allowable voltage Vmax as the life, and define the degree of deterioration DLv based on the voltage V in levels (for example, 10 levels). The maximum allowable voltage Vmax is, for example, a value in the range of 17.5 kV to 20.0 kV. The degree of deterioration DLv is obtained by evaluating the level of deterioration of the laser cavity 100 according to the voltage V. As the voltage V increases, the value representing the level of the degree of deterioration DLv becomes a larger value. When defining the degree of deterioration DLv in 10 levels according to the voltage V, the minimum level value can be set to 1, and the maximum level value can be set to 10. It is preferable to make the upper limit (maximum level value) of the degree of deterioration DLv based on the voltage V and the upper limit of the degree of deterioration DLn based on the number of oscillation pulses Np coincide, and it is preferable that the relative degree of deterioration with respect to the upper limit of the degree of deterioration based on each parameter is roughly the same in advance. Regarding the classification of the degree of deterioration DLv based on the voltage V, the correspondence relationship between the voltage value and the level value can be determined in advance according to test results or on-site data, etc.

[0217] Regarding the state of the laser cavity 100 represented by the combination (Np, V) of the parameters of the number of oscillation pulses Np and the voltage V, the degree of deterioration DL given as a label representing the actual degree of deterioration until the life is set as the degree of deterioration of the higher level (the larger level value) among the degree of deterioration DLn(Np) based on the number of oscillation pulses Np and the degree of deterioration DLv(V) based on the voltage V. According to Figure 8 the example of, for example, in the region where the level of the degree of deterioration DLn based on the number of oscillation pulses Np is "2", the level of the degree of deterioration DLv based on the voltage V is "4", so the actually given degree of deterioration DL for this region becomes "4". In addition, sometimes the deviation at the time of obtaining the voltage V is large, so the moving average value for a certain period (for example, 1 week) can also be used.

[0218] In Figure 8In the definition of the degradation degree DLV of data based on the voltage V, the range from the initial voltage Vch after replacement of the laser cavity 100 to the maximum allowable voltage Vmax is equally divided into 10 levels, levels 1 to 10. However, the classification of the degradation degree DLV based on the voltage V is not limited to this example. For example, as Figure 9 shown, a threshold voltage Vth may also be determined for the voltage V, and the degradation degree DLV at this threshold voltage Vth may be determined, such as level 6, and the range from the threshold voltage Vth to the maximum allowable voltage Vmax may be equally divided to set levels 6 to 10. When the maximum allowable voltage Vmax is, for example, 19 kV, the threshold voltage Vth may also be, for example, 17.5 kV. In this case, the degradation degree DLV based on a lower voltage V smaller than the threshold voltage Vth may also be set to level 0. Level 0 means that the degradation degree is not evaluated (no evaluation). The threshold voltage Vth is an example of the "prescribed threshold" in the present disclosure.

[0219] According to Figure 9 the example, for example, the value of the voltage V in the region where the degradation degree DLn based on the oscillation pulse number Np is "3" is lower than the threshold voltage Vth. Therefore, the degradation degree DLV based on the voltage V in this region is "0". Therefore, the evaluation of the degradation degree DLn based on the oscillation pulse number Np takes precedence, and the actually assigned degradation degree DL for this region becomes "3". On the other hand, in the region where the degradation degree DLn based on the oscillation pulse number Np is "4", the degradation degree DLV based on the voltage V is "7". Therefore, the actually assigned degradation degree DL for this region becomes a larger value "7".

[0220] According to this setting of the degradation degree, when the voltage V is lower than the threshold voltage Vth, the degradation degree DLn based on the oscillation pulse number Np is maintained as the actually assigned degradation degree DL. On the other hand, when the voltage V is higher, i.e., above the threshold voltage Vth, the degradation degree DLV based on the voltage V is evaluated as the level value in the latter half of the degradation level (levels 6 to 10).

[0221] Regarding the voltage V, based on the view that the degradation state becomes a problem when the voltage becomes higher than a certain voltage value, the following structure may also be adopted: For the region where the voltage V is higher than the value of the voltage V that should be noted (threshold voltage Vth), a label of the degradation degree DLV based on the voltage V is assigned.

[0222] By adopting Figure 9In this method of assigning the degree of deterioration, in a region where the voltage V is lower than the threshold voltage Vth, the influence of the evaluation of the degree of deterioration DLv based on the voltage V can be relatively reduced, and the influence of the evaluation of the degree of deterioration DLn based on the number of oscillation pulses Np can be relatively increased. Moreover, in a region where the voltage V is higher than the threshold voltage Vth, the evaluation of the degree of deterioration DLv based on the voltage V and the evaluation of the degree of deterioration DLn based on the number of oscillation pulses Np can be emphasized to roughly the same extent, and the two are compared to determine the actually assigned degree of deterioration DL. In this case, the actually assigned degree of deterioration DL becomes the degree of deterioration of the higher-ranking one among the degrees of deterioration based on the two parameters.

[0223] In Figure 9 regarding the degree of deterioration DLv based on a lower voltage V smaller than the threshold voltage Vth, it is determined as "level 0", but the same result is obtained by determining "level 1" instead of "level 0". That is, the degree of deterioration DLv based on the voltage V in the case of a voltage V lower than the threshold voltage Vth can also be set to a value below the minimum level value (level 1) of the degree of deterioration DLn based on the number of oscillation pulses Np.

[0224] As Figure 8 or Figure 9 illustrated in, the degree of deterioration DL is assigned to the combination (Np, V) of the number of oscillation pulses Np and the voltage V. The data set of the parameters of the combination (Np, V) of the generated number of oscillation pulses Np and the voltage V and the degree of deterioration DL is used as training data for machine learning. That is, the combined data of the number of oscillation pulses Np and the life-related parameter becomes the input data for the learning model, and the value representing the level of the degree of deterioration DL corresponds to the label (teaching data) of the correct solution of the degree of deterioration for this input data.

[0225] In Figure 8 In the case of the example shown, the data set of the parameters of the combination (Np, V) of the number of oscillation pulses Np and the voltage V is the input data for the learning model, and the data of the corresponding degree of deterioration DL is the label of the correct solution. The learning model generation unit 340 performs machine learning using the generated teaching data and generates a learning model that outputs a predicted value of the degree of deterioration for the input of the combination of the number of oscillation pulses Np and the voltage V. That is, the learning model generation unit 340 generates a learning model that performs a 10-class classification task of predicting (inferring) which level among 10 levels of the degree of deterioration (levels 1 to 10) corresponds to the input data based on the combination of multiple parameters. Here, an example of defining 10 levels of the degree of deterioration is shown, but the number of levels of the degree of deterioration is not limited to 10 levels and can be set to an appropriate number of 2 or more levels.

[0226] The degradation degree DLn based on the number of oscillation pulses Np is an example of the "first degradation degree" in the present disclosure. The degradation degree DLv based on the voltage V is an example of the "second degradation degree" in the present disclosure. The actually assigned degradation degree DL is an example of the "third degradation degree" in the present disclosure.

[0227] 5.2.6 Generation example 2 of the learning model used for predicting the lifetime of the laser cavity

[0228] In Figure 8 and Figure 9 examples of assigning the degradation degree DL according to the number of oscillation pulses Np and the voltage V were described. However, the parameters used to evaluate the lifetime of the laser cavity 100 are not limited to the number of oscillation pulses Np and the voltage V. As Figure 3 described, similar to the voltage V, when the number of oscillation pulses Np increases, the initial pressure Pini after replacing all the gas also increases (refer to Figure 3 ). Hereinafter, the initial pressure Pini after replacing all the gas will be referred to as "pressure Pini". Therefore, the degradation degree DLp can be determined based on the pressure Pini instead of the degradation degree DLv based on the voltage V.

[0229] Furthermore, it is more preferable to use both the voltage V and the pressure Pini to determine the degradation degree. In this case, regarding the actually assigned degradation degree DL, the degradation degree with the largest value among the degradation degree DLn based on the number of oscillation pulses Np, the degradation degree DLv based on the voltage V, and the degradation degree DLp based on the pressure Pini can be assigned. In addition, similar to the voltage V, there is sometimes a large deviation in obtaining the pressure Pini. Therefore, the moving average value over a certain period (for example, one week) can also be used.

[0230] Figure 10 is a graph showing an example of the relationship between the pressure Pini of the laser cavity 100 and the number of oscillation pulses Np, and shows an example of defining the degradation degree up to the lifetime of the laser cavity 100 according to the number of oscillation pulses Np, the pressure Pini, and the voltage V. Figure 10 The vertical axis of represents the pressure Pini, with the unit of [Pa]. The maximum allowable pressure Pmax is set as the lifetime, and the degradation degree DLp based on the pressure Pini is defined in levels (for example, 10 levels).

[0231] In Figure 10 regarding the definition of the degradation degree DLp based on the pressure Pini, the range from the initial pressure Pch after replacing the laser cavity 100 to the maximum allowable pressure Pmax is equally divided into 10 levels from level 1 to level 10. However, the setting of the classification of the degradation degree DLp based on the pressure Pini is not limited to this example.

[0232] For example, a threshold air pressure that becomes a threshold value can also be determined with respect to the air pressure Pini, and the degree of deterioration at this threshold air pressure can be set to, for example, level 3, and equally divided from level 3 to level 10 of the maximum allowable air pressure Pmax, and set to levels 3 to 10. For example, when Pmax is 3300 [Pa], the threshold air pressure can also be 2400 [Pa]. In this case, the degree of deterioration corresponding to an air pressure lower than level 3 can also be set to level 0 or level 1, etc.

[0233] The degree of deterioration DLnp based on the combination (Np, Pini) of the number of oscillation pulses Np and the air pressure Pini is set to the degree of deterioration of the higher level of deterioration between the degree of deterioration DLn(Np) based on the number of oscillation pulses Np and the degree of deterioration DLp(Pini) based on the air pressure Pini.

[0234] Furthermore, in the case of a method of determining the degree of deterioration using both the applied voltage V and the air pressure Pini, for the combination (Np, V, Pini) of the number of oscillation pulses Np, the applied voltage V, and the air pressure Pini, the actually assigned degree of deterioration DLnvp is set to the highest degree of deterioration among the degree of deterioration DLn(Np) based on the number of oscillation pulses Np, the degree of deterioration DLv(V) based on the applied voltage V, and the degree of deterioration DLp(Pini) based on the air pressure Pini.

[0235] According to Figure 10 the example of, in the region where the level of the degree of deterioration DLn based on the number of oscillation pulses Np is "5", the level of the degree of deterioration DLp based on the air pressure Pini is "6", so for this region, the degree of deterioration DLnp based on the comprehensive judgment of the number of oscillation pulses Np and the air pressure Pini becomes "6". In addition, in Figure 10 the region where the level of the degree of deterioration DLn based on the number of oscillation pulses Np is "5", the level of the degree of deterioration DLv based on the applied voltage V is "7" (refer to Figure 8 ), so for this region, the degree of deterioration DLnvp based on the comprehensive judgment of the number of oscillation pulses Np, the applied voltage V, and the air pressure Pini becomes "7".

[0236] As Figure 10 illustrated in, for the state of the consumable represented by the combination (Np, V, Pini) of the number of oscillation pulses Np, the applied voltage V, and the air pressure Pini, the degree of deterioration DLnvp is assigned according to the degrees of deterioration DLn, DLv, DLp based on each parameter. The data that associates the combination (Np, V, Pini) of the number of oscillation pulses Np, the applied voltage V, and the air pressure Pini generated in this way with the degree of deterioration DLnvp is used as training data for machine learning.

[0237] Machine learning is performed using such training data to generate the following learning model. For an input of a combination (Np, V, Pini) of the number of oscillation pulses Np, voltage V, and air pressure Pini, the learning model outputs a grade representing the degree of deterioration of the consumable (i.e., a class classification label for the degree of deterioration) as a prediction result. The voltage V and the air pressure Pini are examples of the "multiple life-related parameters" in the present disclosure. The degree of deterioration DLv based on the voltage V and the degree of deterioration DLp based on the air pressure Pini are each examples of the "second degree of deterioration" in the present disclosure. The comprehensive degree of deterioration DLnvp based on the number of oscillation pulses Np, voltage V, and air pressure Pini is an example of the "third degree of deterioration" in the present disclosure.

[0238] Not limited to the laser cavity 100, the same applies to other consumables such as the monitor module 108 and the narrowbanding module 106. For each consumable, the data of the life-related information during the entire period of one cycle from the start of use of the consumable to replacement is divided into grades of multiple degrees of deterioration, and training data that associates the data of the life-related parameters with the grades representing the degrees of deterioration is generated.

[0239] Then, for each type of consumable, machine learning is performed using the respective training data to generate respective learning models.

[0240] Figure 11 It shows being applied to Figure 7 The flowchart of Example 1 showing the processing content of step S48. That is, Figure 11 It shows Example 1 of the generation subroutine of the learning model.

[0241] In Figure 11 In step S102, the learning model generation unit 340 divides all the life-related information during the usage period of the consumable to be replaced into Smax levels according to the grades of the degree of deterioration. For example, as illustrated in Figure 8 It can be 10.

[0242] In step S104, the learning model generation unit 340 generates data D(s) of each level of the life-related information divided into Smax levels. Here, s is an integer representing the grade of the degree of deterioration. s can take values from 1 to Smax. In the example of Figure 8 Regarding the laser cavity 100, data D(s) of each level of the life-related information of the degree of deterioration DL classified into 10 levels is generated. The data D(s) is data that associates the life-related information with the grade s of the degree of deterioration DL and is used as training data. The method of generating training data by implementing step S102 and step S104 is an example of the "training data generation method" in the present disclosure.

[0243] Next, in step S106, the learning model generation unit 340 sets the value of the variable s representing the degree of deterioration to the initial value of "1". Then, in step S108, the learning model generation unit 340 inputs the data of D(s) into the learning model retrieved in Figure 7 step S46.

[0244] Next, in Figure 11 step S110, the learning model generation unit 340 changes the parameters of the learning model so that the output of the learning model corresponding to the input of the data D(s) becomes the level s.

[0245] The learning model can be, for example, a neural network model. The learning model generation unit 340 changes the parameters of the learning model by machine learning using the teaching data to generate a new learning model.

[0246] In step S112, the learning model generation unit 340 determines whether the variable s is equal to or greater than Smax. If the determination result in step S112 is "no", the learning model generation unit 340 proceeds to step S114, increases the value of the variable s, and returns to step S108. If the determination result in step S112 is "yes", the learning model generation unit 340 ends Figure 11 the flowchart of, and returns to Figure 7 the flowchart of. That is, when the determination result in step S112 becomes "yes", the learning model is updated to a new learning model that reflects the result of the consumable replaced this time.

[0247] In addition, in Figure 11 the processing of steps S106 to S112, an example of learning according to each level of the degree of deterioration is described. However, regarding the input of the training data for the learning model, it is preferable that the learning unit is not each level, but rather random samples are learned according to an arbitrary number of pieces (for example, 1000 pieces each). When learning according to each level, the parameters inside the learning model may be close to the data of the last learned level. Therefore, it is preferable that the learning data group that is the learning unit is sampled randomly as much as possible.

[0248] The learning model generation unit 340 is an example of a processing unit that implements the "training data generation method" and the "machine learning method" in the present disclosure.

[0249] Figure 12 is a flowchart showing Example 1 of the processing content applied to Figure 11 step S104. Figure 12 The flowchart of is as Figure 8 and Figure 9An example of "processing for generating data D(s) of life correlation information at each level" is considered in the case of rewriting (overwriting) the degradation degree DLn based on the number of oscillation pulses Np with the degradation degree DLv based on the voltage V as exemplified.

[0250] In step S201, the learning model generation unit 340 sets the value of the variable s to the initial value "1". Here, the variable s represents an interval defined by dividing the number of oscillation pulses Np into Smax levels, which corresponds to the level of the degradation degree DLn based on the number of oscillation pulses Np.

[0251] In step S202, the learning model generation unit 340 reads the data of the interval s. Here, the data of the voltage V in the interval s is read.

[0252] In step S204, the learning model generation unit 340 calculates the degradation degree DLv based on the value of the voltage V and saves it as the variable L. For example, when the value of the voltage V is 19 kV, the degradation degree DLv is calculated as "6" (refer to Figure 8 ).

[0253] Next, in step S210, the learning model generation unit 340 compares the degradation degree "s" based on the number of oscillation pulses Np with the degradation degree "L" based on the voltage V, and determines whether "L" is greater than "s".

[0254] When the determination result in step S210 is "yes" (L > s), the learning model generation unit 340 proceeds to step S211. In step S211, the learning model generation unit 340 saves the data of the life correlation information of the interval s as D(L) with the degradation degree "L".

[0255] On the other hand, when the determination result in step S210 is "no" (L ≤ s), the learning model generation unit 340 proceeds to step S212. In step S212, the learning model generation unit 340 saves the data of the life correlation information of the interval s as D(s) with the degradation degree "s".

[0256] After step S211 or S212, the learning model generation unit 340 proceeds to step S213.

[0257] In step S213, the learning model generation unit 340 determines whether the variable s is equal to or greater than Smax. When the determination result in step S112 is "no", the learning model generation unit 340 proceeds to step S214, increases the value of the variable s, and returns to step S202. When the determination result in step S213 is "yes", the learning model generation unit 340 ends Figure 12 the flowchart, and returns Figure 11 the flowchart.

[0258] Figure 13 is a flowchart of Example 2 showing the processing content of step S104 applied to Figure 11 . Figure 13 The flowchart is an example of "processing for generating data D(s) of life correlation information at each level" when considering rewriting the degradation degree DLn based on the oscillation pulse number Np in consideration of the degradation degree DLv based on the voltage V and the degradation degree DLp based on the initial pressure Pini. In Figure 13 , the steps of the processing that communicate with Figure 12 are labeled with the same step numbers and duplicate explanations are omitted.

[0259] After step S202, in step S204A, the learning model generation unit 340 calculates the degradation degree DLv based on the value of the voltage V and saves it as the variable L1.

[0260] In step S204B, the learning model generation unit 340 calculates the degradation degree DLp based on the initial pressure Pini and saves it as the variable L2.

[0261] In step S206, the learning model generation unit 340 sets the maximum value of L1 and L2 as L.

[0262] In step S209, the learning model generation unit 340 compares "s" and "L" and determines whether "L" is larger.

[0263] When the determination result in step S209 is a "yes" determination (L > s), the learning model generation unit 340 proceeds to step S211. For example, when s = 1, if L = 6, then in step S210, it becomes a "yes" determination and proceeds to step S211. On the other hand, when the determination result in step S209 is a "no" determination (L ≤ s), the learning model generation unit 340 proceeds to step S212. The subsequent steps S211 to step S214 are the same as Figure 12 .

[0264] Figure 14 is a flowchart of Example 3 showing the processing content of step S104 applied to Figure 11 . Figure 14 The flowchart is an example of "processing for generating data D(s) of life correlation information at each level" when considering rewriting the degradation degree DLn based on the oscillation pulse number Np in consideration of a total of n degradation parameters including the voltage V and the initial pressure Pini. In Figure 14 , the steps of the processing that communicate with Figure 13 are labeled with the same step numbers and duplicate explanations are omitted.

[0265] In Figure 14In step S205 after step S204B, n labels are generated by any single parameter value or combination of parameters, and label Ln is assigned. The process of step S205 is repeated any number of times according to the type of single parameter or combination of parameters for evaluating the degree of deterioration. Additionally, steps S204A and S204B may also be implemented in step S205. Voltage V and air pressure Pini are examples of single parameters for evaluating the degree of deterioration.

[0266] After step S205, in step S207, the learning model generation unit 340 sets the maximum value among L1 to Ln as L. Steps S209 and S211 - 214 thereafter are the same as Figure 13 the same.

[0267] 5.2.7 Example of Combining Multiple Parameters and Converting to One Parameter

[0268] Lifetime-related parameters such as voltage V and air pressure Pini can be used separately (as single parameters) for evaluating the degree of deterioration. However, multiple parameters can also be combined to define a new (different) parameter, and the degree of deterioration can be evaluated based on the value of this new parameter.

[0269] It is possible to derive one parameter by combining multiple r values (r - dimensional). As a method for converting from multiple parameters to one parameter, for example, methods using the following conversion formula, methods using coefficients in calculations, dimensionality reduction, etc. can be applied. The conversion formula calculates one value through arithmetic operations based on r values.

[0270] For example, convert from two values such as voltage V and air pressure Pini to values 1 - 100 representing a new characteristic quantity. At this time, consider dividing the converted values 1 - 100 into 10 levels and assigning degrees of deterioration 1 - 10 respectively. Figure 15 and Figure 16 shows an example of the relationship between two values and labels of 10 levels.

[0271] Figure 15 is an example in the case where the correlation between the two values of voltage V and air pressure Pini and the label is linear. Figure 15 In this, the horizontal axis represents the degree of deterioration based on the number of oscillation pulses Np or air pressure Pini, and the vertical axis represents the degree of deterioration based on voltage V. Figure 15 The numerical values shown in each cell of the 10 - row × 10 - column matrix indicate the 10 - level labels assigned to the new characteristic quantity derived from the combination of the values of two parameters.

[0272] Figure 16 is an example in the case where the influence of voltage V on the label is small and the influence of air pressure Pini is large. It can also be as Figure 16In that case, compared with the degradation degree based on the voltage V, the evaluation of the degradation degree based on the air pressure Pini is emphasized, and a new feature quantity is defined in such a way as to assign a label to the degradation degree based on the combination of these two parameters.

[0273] 5.2.8 Explanation when the data D(s) contains multiple data counts

[0274] As the acquisition conditions for the data used in learning, for example, assume the following situation.

[0275] [Condition 1] The data formula used in learning is acquired once a day. The "data formula" mentioned here includes data on various parameters such as the oscillation pulse number Np, voltage V, and air pressure Pini of the laser device 10.

[0276] [Condition 2] The laser device 10 is used with the same oscillation pulse number every day.

[0277] [Condition 3] The laser cavity 100 is replaced 500 days after it starts being used.

[0278] [Condition 4] Regarding the laser cavity 100 that has been working for 500 days until replacement, the working period is divided into 10 equal parts using the oscillation pulse number Np, and a label for the degradation degree based on the oscillation pulse number Np is assigned to each interval.

[0279] When conditions 1 to 4 are satisfied, the counting method for the number of data counts is as follows. That is, the total number of data counts in the data D as a whole is 500.

[0280] According to condition 2 and condition 4, the number of data counts included in each label s (s = 1 to 10) of the degradation degrees 1 to 10 is 50. 50 data counts are included in D(1), D(2)... D(10) respectively, and the whole of D(s), that is, the data D as a whole, becomes 500.

[0281] Figure 17 It is a curve graph showing an image in which multiple data are included in one interval of the degradation degree. Figure 17 The horizontal axis represents the oscillation pulse number, and the vertical axis is data on certain parameters associated with the lifespan. The horizontal axis is equally divided into 10 levels according to the oscillation pulse number to determine the intervals of the degradation degrees 1 to 10. In Figure 17 For simplicity of illustration, an example in which 10 data are included in one interval of the degradation degree is shown, but the number of data included in one interval is not limited to this example.

[0282] In the case of the assumed example that satisfies the above conditions 1 to 4, as described above, 50 data are included in one interval. Additionally, more data counts such as 100 or 1000 can also be included in one interval.

[0283] When learning, it is also possible to randomly sample any number of pieces from all the data D for learning. For example, when the total number of data pieces is 500, instead of learning sequentially for each degree of deterioration, a specified number of pieces (e.g., 50 pieces) are randomly and without repetition selected from all the data D for each degree of deterioration, and learning is performed 10 times.

[0284] In addition, when rewriting the labels of the degrees of deterioration, instead of uniformly rewriting multiple data within the same interval to the same label, it is preferable to rewrite the degrees of deterioration for each data within the interval, that is, for all the data D, respectively according to the voltage V and the air pressure Pini.

[0285] Figure 18 It shows being applied to Figure 7 Example 2 of the processing content of step S48. Figure 18 It shows Example 2 of the generation subroutine of the learning model. It can replace Figure 11 the flowchart described in Figure 18 and apply the flowchart shown in

[0286] Figure 18 The steps S102 and S104 shown in Figure 11 are the same. In addition, Figure 18 the data D(s) in step S104 of

[0287] In step S116 after step S104, the learning model generation unit 340 sets the variable n to "1" as the initial value and sets the variable m to "1000". The variable n here represents the number of loops of the subsequent processing. The variable m represents the number of data pieces extracted as the learning unit from all the data D. The case of m = 1000 means learning is performed uniformly for every 1000 pieces. That is, m can be understood as the batch size of a mini-batch. m can be any number less than the total number of data pieces. Here, an example of randomly extracting 1000 pieces of learning samples (data) from all the data D and performing learning in units of mini-batches of 1000 pieces is described.

[0288] In step S117, the learning model generation unit 340 randomly extracts m pieces from the data D. At this time, it is assumed that the data that has been extracted once will not be extracted repeatedly. Among the m pieces, there may be a mixture of data with various degrees of deterioration. In addition, the data D here is a set of all the data including D(1), D(2)... D(Smax).

[0289] In step S118, the learning model generation unit 340 retrieves the data of each D(s) and inputs it into the learning model.

[0290] In step S120, the learning model generation unit 340 changes the parameters of the learning model to increase the probability that the output of the learning model becomes level s. The learning model is a neural network model described later, and the parameters of the model are changed based on the teaching data to generate a new learning model.

[0291] In step S122, the learning model generation unit 340 determines whether the product of n and m is N or more. N is the total number of data items of data D. When the determination result in step S122 is "no", the learning model generation unit 340 proceeds to step S124, increases the value of n, and returns to step S117. When the determination result in step S122 is "yes", the learning model generation unit 340 ends Figure 18 the flowchart of, and returns Figure 7 the flowchart of. That is, when the determination result in step S112 becomes "yes", the learning model is updated to a new learning model that reflects the result of the consumable replaced this time.

[0292] In addition, in Figure 18 the flowchart of, it becomes a flowchart that ends learning in one epoch. However, the number of epochs can also be set to a value of 2 or more, and steps S116 to S122 can be repeated further multiple times.

[0293] Figure 19 is a flowchart showing an example of the processing content of step S104 applied to Figure 18 the flowchart of. Figure 19 the flowchart of is as Figure 8 and Figure 9 In the example of "processing of generating the data D(s) of the life correlation information for each level" when the degradation degree DLn based on the oscillation pulse number Np is rewritten in consideration of the degradation degree DLv based on the voltage V as exemplified in.

[0294] In Figure 19 step S251 of, the learning model generation unit 340 sets the index k indicating the data number to the initial value "1". The processing after step S252 is looped for all the data in the data set of data D.

[0295] In step S252, the learning model generation unit 340 reads the data with the data number "k".

[0296] In step S253, the learning model generation unit 340 calculates the degradation degree s based on the oscillation pulse number Np.

[0297] In step S254, the learning model generation unit 340 calculates the degradation degree DLv based on the value of the voltage V and saves it as the variable L.

[0298] In step S260, the learning model generation unit 340 compares "s" and "L" and determines whether "L" is greater than "s". If the determination result in step S260 is "yes", the learning model generation unit 340 proceeds to step S261. In step S261, the learning model generation unit 340 saves the data of the life correlation information with data number "k" as the data D(L) of the degradation degree "L".

[0299] On the other hand, if the determination result in step S260 is "no", the learning model generation unit 340 proceeds to step S262. In step S262, the learning model generation unit 340 saves the data of the life correlation information with data number "k" as the data D(s) of the degradation degree "s".

[0300] After step S261 or S262, the learning model generation unit 340 proceeds to step S263.

[0301] In step S263, the learning model generation unit 340 determines whether the value of k is equal to or greater than the total number of data pieces N of data D. If the determination result in step S263 is "no", the learning model generation unit 340 proceeds to step S264, increments the value of the index k, and returns to step S252. If the determination result in step S263 is "yes", the learning model generation unit 340 ends Figure 19 the flowchart of, and returns Figure 18 the flowchart of.

[0302] It can also be replaced with Figure 19 and applied with Figure 13 or Figure 14 the flowchart corresponding to the flowchart described in.

[0303] 5.2.9 Example of Neural Network Model

[0304] Figure 20 is a schematic diagram showing an example of a neural network model. In Figure 20 , the circles represent neurons, and the straight lines with arrows represent the signal flow. Starting from the left side of Figure 20 , there are neurons N 11 , N 12 , N 13 of the input layer 402, neurons N 21 of the hidden layer 404, and neurons N 31 of the output layer 406. Let the layer number of the neural network with a layer structure be i, the neuron number be j, and the intensity of the signal output from neuron N ij be X ij , expressed as signal X ij . Let the weight of the connection between neurons in layer i and layer (i + 1) be Wij .

[0305] The neurons N in the input layer 402 11 、N 12 、N 13 respectively output signals with intensities of X 11 、X 12 、X 13 . The neuron N in the hidden layer 404 21 when the input signals X 11 、X 12 、X 13 weighted signal sum (W 11 ×X 11 +W 12 ×X 12 +W 13 ×X 13 ) is greater than the threshold, it outputs signal X 21 . When setting the threshold here to b 21 , the neuron N 21 in W 11 ×X 11 +W 12 ×X 12 +W 13 ×X 13 -b 21 >0, it outputs signal X 21 . "-b 21 " is called the bias of the neuron N 21 .

[0306] The parameters of the neural network model include the weights and biases of the connections between neurons.

[0307] 5.2.10 Learning Mode of Neural Network Model

[0308] Figure 21 is an example of the neural network model when generating a learning model. The neural network model 400 includes an input layer 402, a hidden layer 404, and an output layer 406.

[0309] The input layer 402 contains n neurons N 11 ~N 1n , and for each neuron N 11 ~N 1n input the log data when the degree of deterioration is s in the life - related information of the replaced consumable.

[0310] The hidden layer 404 contains m neurons N 21 ~N 2m , and for each neuron in the hidden layer 404, input from the neurons N in the input layer 402 11 ~N1n The output signal. A parameter W1 that can set different weights for these input signals respectively. Here, regarding the weight parameter W1, the respective weights corresponding to the signals input to each neuron N 21 ~N 2m of the hidden layer 404 are collectively referred to as "the weight parameter W1".

[0311] The output layer 406 includes p neurons N 31 ~N 3p , and the signals output from the neurons N 21 ~N 2m of the hidden layer 404 are input to each neuron in the output layer 406. The number p of neurons in the output layer 406 can be the same as the number of levels (Smax) of the degradation degree levels. A parameter W2 that can set different weights for these input signals respectively. Here, regarding the weight parameter W2, the respective weights corresponding to the signals input to each neuron N 31 ~N 3p of the output layer 406 are collectively referred to as "the weight parameter W2".

[0312] The neurons N 31 ~N 3p of the output layer 406 output the probabilities of the degradation degrees Lv(1)~Lv(s)~Lv(Smax). The probability of the degradation degree here means a fraction, and this fraction represents the reliability corresponding to each level of the degradation degree.

[0313] When the degradation degree s is defined as Smax levels from 1 to Smax, the life-related information (log data) D1(s), D2(s)…Dn(s) of the replaced consumables are respectively input to the input layer 402.

[0314] Adjust the respective weights and biases between the neurons so that for the input of each degradation degree s, the output from the output layer 406 is a probability close to 1 for Lv(s) and close to 0 for other degradation degrees.

[0315] As described above, a learning model for predicting the life of consumables is generated by supervised machine learning. The machine learning method of Embodiment 1 is understood as a method for generating the following prediction model (learned model): This prediction model is learned to output the predicted value of the degradation degree of the consumable for the input of the life-related information of the consumable.

[0316] 5.2.11 Example of the processing of the consumable life prediction unit

[0317] Figure 22 It is a flowchart showing an example of the processing content in the consumable life prediction unit 360. Figure 22The processes and actions shown in the flowchart are implemented, for example, by a processor executing a program that functions as a consumable life prediction unit 360.

[0318] In Figure 22 step S62, the consumable life prediction unit 360 determines whether it has received the current life-related information of the consumable to be replaced as scheduled. If the determination result in step S62 is "no", the consumable life prediction unit 360 repeats step S62. If the determination result in step S62 is "yes", the consumable life prediction unit 360 proceeds to step S64.

[0319] In step S64, the consumable life prediction unit 360 obtains the current life-related information of the consumable to be replaced as scheduled. The current life-related information obtained in step S64 is an example of the "second life-related information" in the present disclosure.

[0320] In step S66, the consumable life prediction unit 360 obtains the operation-related information of the laser device 10. Then, in step S68, the consumable life prediction unit 360 retrieves the learning model of the consumable to be replaced as scheduled. Here, the learning model saved in the corresponding file (file Am, file Bm, or file Cm) of the consumable to be replaced (laser cavity 100, monitor module 108, or narrowbanding module 106) is retrieved.

[0321] Then, in step S70, the consumable life prediction unit 360 performs life calculation using the learning model. That is, the consumable life prediction unit 360 calculates the life, remaining life, and recommended maintenance date using the learning model based on the current life-related information of the consumable to be replaced as scheduled.

[0322] In step S72, the consumable life prediction unit 360 sends the data of the life, remaining life, and recommended maintenance date of the consumable to be replaced as scheduled to the data output unit 370.

[0323] In step S74, the consumable life prediction unit 360 determines whether to abort the calculation of the predicted life of the consumable. If the determination result in step S74 is "no", the consumable life prediction unit 360 returns to step S62 and repeats steps S62 to S74. If the determination result in step S74 is "yes", the consumable life prediction unit 360 ends Figure 22 the flowchart.

[0324] Figure 23 is a flowchart showing an example of a subroutine for the process of performing life calculation using a learning model. That is, Figure 23 is a flowchart showing Figure 22 an example of the processing content of step S70.

[0325] In Figure 23 In step S132, the consumable life prediction unit 360 inputs the current life correlation information of the consumable to be replaced into the learning model.

[0326] In step S134, the consumable life prediction unit 360 outputs the probabilities of the respective degradation levels Lv(1) to Lv(Smax) according to the learning model.

[0327] In step S136, the consumable life prediction unit 360 determines the current degradation level s of the consumable according to the probability distribution of the degradation level. As a first example of the method for determining the degradation level s, for example, the degradation level with the highest probability can be extracted. In addition, as a second example of the method for determining the degradation level s, an approximate curve can be obtained according to the probability distribution of the degradation level, and the degradation level s of the highest probability distribution can be obtained. In the case of the second example, the degradation level s is not an integer, and a value up to the decimal point is obtained. Compared with the case of the first example, the case of the second example can predict the life and remaining life of the consumable with higher accuracy.

[0328] In step S138, the consumable life prediction unit 360 further calculates the life Nches of the consumable according to the degradation level obtained in step S136. Using the current oscillation pulse number Nch, the life Nches of the consumable is calculated according to Nches = Nch·Smax / s. In addition, the "·" in the formula represents a multiplication operation.

[0329] In step S140, the consumable life prediction unit 360 further calculates the remaining life Nchre of the consumable according to the degradation level obtained in step S136. Using the current oscillation pulse number Nch, the remaining life Nchre of the consumable is calculated according to Nchre = Nches - Nch.

[0330] In step S142, the consumable life prediction unit 360 calculates the recommended maintenance date Drec of the consumable. The recommended maintenance date Drec is calculated according to Drec = Dpre + Nchre / Nday.

[0331] After step S142, the consumable life prediction unit 360 ends Figure 23 the flowchart of Figure 22 the flowchart of.

[0332] 5.2.12 Example of the process of calculating the life of a consumable using a learning model

[0333] Figure 24An example of calculating the lifetime and remaining lifetime of the laser cavity 100 using the generated learning model is shown. In the case of calculating the predicted lifetime of the currently operating laser cavity 100, the lifetime prediction unit 360 of the consumable obtains the current lifetime-related information of the laser cavity 100. Here, the data of the number of oscillation pulses Nch of the current laser cavity 100 is obtained.

[0334] Next, when the current lifetime-related information of the laser cavity 100 is input to the generated learning model, the probabilities of the respective degradation levels Lv(1) to Lv(10) of multiple levels are calculated. Figure 25 An example of the probability of each degradation level of 10 levels is shown. In Figure 25 the example, the probability of being determined as the degradation level 7 is the highest.

[0335] In this case, the predicted lifetime Nches of the currently operating laser cavity 100 is obtained using the following formula 1.

[0336] Nches = Nch · 10 / 7 (Formula 1)

[0337] The remaining lifetime Nchre is obtained using the following formula 2.

[0338] Nchre = Nches - Nch = Nch · 3 / 7 (Formula 2)

[0339] 5.2.13 Lifetime Prediction Mode of Neural Network Model

[0340] Figure 26 is an example of the process of predicting the lifetime of the consumable by the learned neural network model 400. The network structure of the neural network model 400 is the same as Figure 21 the same structure. In Figure 26 the parameters W1 and W2 of the weights between the neurons are set to the values optimized according to Figure 21 the learning mode described in

[0341] In the case of predicting the lifetime of the current consumable, the current lifetime-related information (log data) D1, D2... Dn of the consumable is respectively input to the input layer 402. As a result, the probabilities of the respective degradation levels Lv(1) to Lv(Smax) from the output layer 406 are output (refer to Figure 25 ).

[0342] 5.2.14 Others

[0343] In Figure 21 and Figure 26 an example of the case where the hidden layer 404 of the neural network model 400 is 1 layer is shown, but it is not limited to this, and the hidden layer 404 can also be multiple layers.

[0344] In this embodiment, an example of machine learning based on supervised learning is shown, but it is not limited to this example, and machine learning based on unsupervised learning can also be performed. For example, the dimension of the input data can be reduced so that data with similar features in these data sets can be clustered with each other. Using this result, a certain benchmark is set and the output is allocated to optimize it, thereby achieving output prediction.

[0345] 5.2.15 Processing example of data output unit

[0346] Figure 27 It is a flowchart showing an example of the processing content in the data output unit 370. Figure 27 The processing and operation shown in the flowchart are implemented, for example, by a processor functioning as the data output unit 370 executing a program.

[0347] In step S82, the data output unit 370 determines whether the life data of the consumables scheduled to be replaced is received. If the determination result of step S82 is "No", the data output unit 370 repeats step S82. If the determination result of step S82 is "Yes", the data output unit 370 proceeds to step S84.

[0348] In step S84 , the data output unit 370 reads data on the life, remaining life, and recommended maintenance date of the consumables scheduled to be replaced.

[0349] In step S86, the data output unit 370 sends data on the life, remaining life, and recommended maintenance date of the consumables to be replaced. The data may be sent to the laser device management system 206 and / or the semiconductor factory management system 208. In addition, the data may be sent to a terminal device (not shown) connected to the network 210.

[0350] In step S88, the data output unit 370 determines whether to stop sending data. If the result of the determination in step S88 is "No", the data output unit 370 returns to step S82 and repeats steps S82 to S88. If the result of the determination in step S88 is "Yes", the data output unit 370 ends. Figure 27 Flowchart of the process.

[0351] 5.3 Laser Cavity Lifetime Correlation Information

[0352] Figures 28 - 30 An example of the lifetime-related information of the laser cavity 100 is shown. The lifetime-related information of the laser cavity 100 includes, for example, electrode degradation parameters, pulse energy stability parameters, gas control parameters, operation load parameters, and degradation parameters of the optical elements of the laser resonator. Figure 30 The expression "OC" of the shown diagram indicates an output coupling mirror.

[0353] Among these lifetime-related parameters, the lifetime-related parameters that are at least required for highly accurate prediction of the lifetime of the laser cavity 100 are the electrode degradation parameter, the pulse energy stability parameter, and the gas control parameter. It is preferable to also use the operating load parameter, whereby the accuracy of lifetime prediction may be improved. This is because, when the operating load is high, the lifetime of the laser cavity 100 sometimes becomes short.

[0354] Furthermore, it is preferable to use the degradation parameter of the narrowbanding module 106, which is an index of the loss of the laser resonator, and the degradation parameter of the window of the laser cavity 100, whereby the accuracy of lifetime prediction may be further improved.

[0355] The electrode degradation parameter at least includes the number of discharges. The number of discharges is a value that is approximately equal to the number of oscillation pulses Np after replacement of the laser cavity 100. As the electrode degradation parameter, it is preferable to be able to further add the cumulative value of the input energy.

[0356] In Figure 1 In the case of the laser device 10 of this single cavity type shown, there is a correlation between the spectral line width and the discharge width. Therefore, as one of the electrode degradation parameters, the spectral line width can also be used.

[0357] The pulse energy stability parameter at least includes the deviation of the pulse energy. Furthermore, as the pulse energy stability parameter, the deviation of the cumulative value (exposure amount) of the pulse energy can also be added.

[0358] Regarding the gas control parameter, in the case of controlling the gas pressure of the laser cavity 100 so that the charging voltage falls within a specified range, this gas control parameter at least includes the gas pressure of the laser cavity 100 and the gas pressure of the laser cavity 100 after all the gas is replaced and the oscillation is adjusted.

[0359] In the case of fixedly controlling the gas pressure of the laser cavity 100 and controlling the charging voltage, as the gas control parameter, it at least includes the charging voltage and the charging voltage after all the gas is replaced and the oscillation is adjusted. It is preferable to be able to add the cumulative value of the injection amount of the halogen-containing gas or the cumulative value of the injection of the laser gas after replacement of the laser cavity 100. Thereby, the accuracy of lifetime prediction may be further improved.

[0360] In addition, it is preferable to be able to add the injection amount of the halogen-containing gas or the injection amount of the laser gas per unit oscillation pulse.

[0361] In the case where the average output or the target pulse energy of the laser output from the laser device 10 hardly changes, the operating load parameter can also be replaced by the duty ratio of the burst operation.

[0362] In particular, it is preferable to use the operation load parameters during the exposure operation. In a semiconductor manufacturing factory, when manufacturing memory elements, the operation load is sometimes high, and when manufacturing logic-related elements, the operation load is sometimes low.

[0363] The deterioration parameters of the optical elements of the laser resonator include the deterioration parameters of the window, the deterioration parameters of the narrowbanding module 106, and the deterioration parameters of the output coupling mirror 104, and at least include the number of oscillation pulses Np after replacement of each optical element. In the case where the pulse energy of the pulsed laser output from the laser device 10 changes significantly, the cumulative value of the pulse energy and the cumulative value of the square of the pulse energy, which are parameters for the deterioration of the optical element based on two-photon absorption, can be used.

[0364] 5.4 Examples of life-related information of the monitor module

[0365] Figure 31 An example of the life-related information of the monitor module 108 is shown. In most cases, the life of the monitor module 108 is determined by the deterioration of the optical element and the deterioration of the photosensor. The life-related information of the monitor module 108 includes at least one of the deterioration parameters of the optical element and the deterioration parameters of the photosensor configured in the monitor module 108.

[0366] The deterioration parameters of the optical element of the monitor module 108 at least include the number of oscillation pulses Np after replacement of the monitor module 108. In the case where the pulse energy of the pulsed laser output from the laser device 10 changes significantly, the cumulative value of the pulse energy and the cumulative value of the square of the pulse energy, which are parameters for the deterioration of the optical element based on two-photon absorption, can be used.

[0367] The deterioration parameters of the photosensor include the detected light intensity, spectral line width, pulse energy, and their cumulative values of the image sensor as the photosensor.

[0368] The deterioration parameter that is at least required for predicting the life of the monitor module 108 is the detected light intensity of the image sensor. The light intensity incident on the image sensor changes according to the spectral line width and the pulse energy. Therefore, the values of the spectral line width and the pulse energy can also be used assistively. The cumulative value of the pulse energy becomes a value close to the amount of light exposed to the image sensor. Therefore, this value can also be used.

[0369] [Other]

[0370] In Figure 31In the example shown, the optical sensors included in the pulse energy detector 144 of the monitor module 108 are, for example, photodiodes and pyroelectric elements. The degradation of these sensors can also be evaluated using the cumulative value of the pulse energy after replacing the monitor module 108. When the target pulse energy does not change significantly, it can be replaced by the number of oscillation pulses Np after replacing the monitor module.

[0371] 5.5 Examples of life-related information of the narrowbanding module

[0372] Figure 32 Examples of life-related information of the narrowbanding module 106 are shown. In most cases, the life of the narrowbanding module 106 is determined by the degradation of optical elements and the degradation of the wavelength actuator. The life-related information of the narrowbanding module 106 includes at least one of the degradation parameters of the optical elements (multiple prisms and gratings) configured in the narrowbanding module 106, the degradation parameters of the wavelength actuator, and the parameters of the wavefront degradation.

[0373] The degradation parameter that is at least required for predicting the life of the narrowbanding module 106 is the degradation parameter of the optical elements of the narrowbanding module 106. Preferably, the degradation parameters of the wavelength actuator and the parameters of the wavefront degradation can be added.

[0374] The degradation parameter of the optical elements of the narrowbanding module 106 includes at least the number of oscillation pulses after replacing the narrowbanding module 106. When the pulse energy of the pulsed laser output from the laser device 10 changes significantly, the cumulative value of the pulse energy and the cumulative value of the square of the pulse energy, which are parameters of the degradation of the optical elements based on two-photon absorption, can be used.

[0375] The degradation parameter of the wavelength actuator includes wavelength stability.

[0376] When the wavelength actuator degrades and the operation deteriorates, the wavelength control becomes unstable. Therefore, by using wavelength stability, it may be possible to evaluate the life.

[0377] The parameter of the wavefront degradation includes the spectral line width. The spectral line width of the pulsed laser output from the laser device 10 becomes wider due to the wavefront distortion of the narrowbanding module 106. Therefore, by using the spectral line width, it may be possible to evaluate the life. For example, when synthetic quartz is used for the prism, sometimes the transmitted wavefront of the prism is distorted due to compaction, resulting in a wider spectral line width.

[0378] 5.6 Function / Effect

[0379] The values of parameters such as voltage V and initial air pressure Pini have a clear correlation with the deterioration of the laser cavity 100. Assuming that only the number of oscillation pulses Np is used to evaluate the degree of deterioration of the laser cavity 100 and only labels with a degree of deterioration that increases step by step with respect to the number of oscillation pulses Np are assigned, the correlation with the above-mentioned voltage V and initial air pressure Pini cannot be learned. Therefore, in the case where the label of the degree of deterioration is assigned only by the number of oscillation pulses Np, it is difficult to generate a learning model that can predict an appropriate degree of deterioration, for example, for deterioration from the initial stage, rapid deterioration, or temporary deterioration.

[0380] Regarding this point, according to Embodiment 1, by using life-related parameters such as voltage V and / or initial air pressure Pini to rewrite the degree of deterioration DLn based on the number of oscillation pulses Np into a more appropriate label of the degree of deterioration, a learning model in which the value of the parameter and the label of the degree of deterioration have a correct correlation can be generated. According to the training data generation method of Embodiment 1, a data set capable of generating training data for a learning model with high prediction accuracy can be obtained.

[0381] According to the consumable management server 310 of Embodiment 1, for the consumables to be replaced as scheduled in the laser device 10, respectively according to the life-related information of the consumables and using the corresponding learning model, the life of each consumable to be replaced as scheduled can be predicted with high accuracy.

[0382] 5.7 Others

[0383] In Embodiment 1, as a method of using life-related parameters such as voltage V and / or initial air pressure Pini to rewrite the degree of deterioration DLn based on the number of oscillation pulses Np into a more appropriate label of the degree of deterioration, an example of assigning the degree of deterioration with the highest level of deterioration has been described. However, the method of determining one degree of deterioration from multiple degrees of deterioration based on different evaluation indexes is not limited to this example. For example, the average value of multiple degrees of deterioration can also be calculated and set as the label of the actually assigned degree of deterioration. As a specific example, when the degree of deterioration DLn based on the number of oscillation pulses Np is 2 and the degree of deterioration DLv based on voltage V is 6, the average value thereof, that is, "4", can also be set as the label of the actually assigned degree of deterioration.

[0384] In addition, weights can also be assigned to multiple parameters associated with the life of the consumable, the weighted average of the degrees of deterioration based on each parameter can be calculated, and this value can be set as the label of the actually assigned degree of deterioration.

[0385] In Embodiment 1, an example of a KrF excimer laser for an exposure apparatus in a semiconductor factory is shown, but it is not limited thereto. For example, it can also be applied to an excimer laser for annealing of a flat panel or an excimer laser for processing. In these cases, a rear mirror is arranged instead of the narrowbanding module 106, and the spectral detector 146 of the monitor module 108 may not be provided.

[0386] 6. Modification Example

[0387] The function of generating training data in the consumable management server 310 described in Embodiment 1, the function of generating a learning model by machine learning based on the generated training data, and the function of predicting the life of the consumable using the generated learning model may also be implemented by different devices (such as servers).

[0388] In addition, the generation process of training data and the learning process using the training data may be implemented in a series of processing flows, or each process may be implemented independently.

[0389] 7. Regarding a Computer-Readable Medium Recording a Program

[0390] As the consumable management server 310 described in the above embodiments, a program including commands for causing a computer to function can be recorded on an optical disc, a magnetic disk, or other computer-readable media (non-volatile information storage media as physical objects), and the program can be provided through the information storage media. By embedding the program in a computer and having a processor execute the commands of the program, the computer can implement the functions of the consumable management server 310.

[0391] The above description is not restrictive but a simple illustration. Therefore, those skilled in the art will understand that modifications can be made to the embodiments of the present disclosure without departing from the appended claims.

[0392] The terms used throughout this specification and the appended claims should be construed as "non-limiting" terms. For example, terms such as "comprising" or "comprised of" should be construed as "not limited to the parts described as being comprised of". The term "having" should be construed as "not limited to the parts described as having". In addition, the indefinite article "a" described in this specification and the appended claims 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 also including combinations with parts other than "A", "B", and "C".

Claims

1. A method for generating training data, which is a method for generating training data used in machine learning of a learning model for predicting the life of a consumable of a laser device. Among them, The training data generation method includes the following steps: Obtain first life-related information, where the first life-related information includes data of at least one life-related parameter of the consumable recorded corresponding to different oscillation pulse numbers during the period from the start of use of the consumable to its replacement. Determine the first degree of deterioration of the consumable according to the oscillation pulse number. Determine the second degree of deterioration of the consumable according to the at least one life-related parameter. Determine the third degree of deterioration of the consumable according to the first degree of deterioration and the second degree of deterioration; and Generate training data that correlates the first life-related information with the third degree of deterioration.

2. The training data generation method according to claim 1, Among them, Determine the degree of deterioration with the higher level of deterioration among the first degree of deterioration and the second degree of deterioration as the third degree of deterioration.

3. The training data generation method according to claim 1, Among them, The life-related parameter includes the voltage applied to the discharge electrode disposed in the laser cavity.

4. The training data generation method according to claim 1, Among them, The life-related parameter includes the initial air pressure after replacing the laser gas in the laser cavity.

5. The training data generation method according to claim 1, Among them, The first life-related information includes data of multiple life-related parameters. Respectively determine the second degree of deterioration according to the multiple life-related parameters, for each life-related parameter. Determine the degree of deterioration with the highest level of deterioration among the multiple second degrees of deterioration determined according to each life-related parameter and the first degree of deterioration as the third degree of deterioration.

6. The training data generation method according to claim 5, Among them, The multiple life-related parameters include the voltage applied to the discharge electrode disposed in the laser cavity and the initial air pressure after replacing the laser gas in the laser cavity.

7. The training data generation method according to claim 1, Among them, The first life-related information includes data of multiple life-related parameters. Determine the second degree of deterioration according to the combination of the multiple life-related parameters.

8. The training data generation method according to claim 1, Among them, The first life-related information includes data of multiple life-related parameters. Determine the second degree of deterioration according to the multiple life-related parameters, for each different type of parameter. Determine the degree of deterioration with the highest level of deterioration among the multiple second degrees of deterioration determined according to each different type of parameter and the first degree of deterioration as the third degree of deterioration.

9. The training data generation method according to claim 1, Among them, The first degree of deterioration is classified into multiple levels according to the oscillation pulse number, and it is determined that as the oscillation pulse number increases, the level of deterioration of the consumable becomes higher.

10. The training data generation method according to claim 9, wherein, the plurality of levels are 10 levels.

11. The training data generation method according to claim 9, wherein, the maximum level value of the second degradation degree is equal to the maximum level value of the first degradation degree.

12. The training data generation method according to claim 9, wherein, the second degradation degree is defined in the following manner: when the value of the at least one life-associated parameter is a value smaller than a specified threshold value, the second degradation degree becomes a value equal to or less than the minimum level value of the first degradation degree.

13. A machine learning method that generates a learning model for predicting the life of a consumable of a laser device, wherein, the machine learning method includes the following steps: acquiring first life-associated information, the first life-associated information including data of at least one life-associated parameter of the consumable recorded corresponding to different oscillation pulse numbers during the period from when the consumable starts to be used until it is replaced; determining the first degradation degree of the consumable according to the oscillation pulse number; determining the second degradation degree of the consumable according to the at least one life-associated parameter; determining the third degradation degree of the consumable according to the first degradation degree and the second degradation degree; generating training data that associates the first life-associated information with the third degradation degree; performing machine learning using the training data, thereby generating the learning model for predicting the degradation degree of the consumable according to the data of the life-associated parameter included in the first life-associated information; and saving the generated learning model.

14. The machine learning method according to claim 13, wherein, the learning model is a neural network model.

15. A consumable management device, wherein, the consumable management device has a storage device and a processor, the storage device pre-stores the learning model generated by implementing the machine learning method according to claim 13, the processor performs the following processing: receiving a request signal for a life prediction process related to a consumable to be replaced in the laser device, and acquiring current second life-associated information related to the consumable to be replaced; calculating the life and remaining life of the consumable to be replaced according to the learning model of the consumable to be replaced and the second life-associated information; notifying at least one of the information of the life and remaining life of the consumable to be replaced obtained by the calculation to an external device.

16. The consumable management device according to claim 15, wherein, the processor inputs the second life-associated information to the learning model, acquiring a score from the learning model, the score representing the reliability of the level of the degradation degree of the consumable corresponding to the second life-associated information, calculating the life and remaining life of the consumable to be replaced according to the current oscillation pulse number and the score included in the second life-associated information.

17. A computer-readable medium, which is a non-volatile computer-readable medium recording a program, wherein, The program is a program that, when executed by a computer, causes the computer to realize a function of generating training data used in machine learning of a learning model for predicting the life of consumables of a laser device. The program includes instructions for causing the computer to implement the following functions: Acquiring first life-related information, the first life-related information including data of at least one life-related parameter of the consumable, recorded corresponding to different numbers of oscillation pulses during a period from when the consumable starts to be used to when it is replaced; determining a first degradation degree of the consumable according to the number of oscillation pulses; determining a second degradation degree of the consumable according to the at least one life-related parameter; determining a third degree of degradation of the consumable product based on the first degree of degradation and the second degree of degradation; and Training data is generated in which the first life-related information and the third degree of degradation are associated with each other.

18. The computer readable medium of claim 17, in, The program also includes instructions for causing the computer to implement the following functions: Performing machine learning using the training data to generate the learning model for predicting the degree of degradation of the consumables based on the data included in the first life-related information; and The generated learning model is saved.

19. A non-volatile computer-readable medium having a program recorded thereon, The program is used to enable the computer to implement the following functions: saving the learning model generated by executing the program recorded in the computer-readable medium according to claim 18; receiving a request signal for life prediction processing related to a consumable scheduled to be replaced in the laser device; In response to receiving the request signal, obtaining current second life-related information related to the consumable scheduled to be replaced; calculating the life and remaining life of the consumable scheduled to be replaced based on the learning model of the consumable scheduled to be replaced and the second life-related information; and Information on the life and remaining life of the consumables scheduled to be replaced obtained through the calculation is notified to an external device.

Images