Performance prediction method and training method

Abstract

A performance prediction method according to one aspect of the present disclosure predicts the performance of a laser device having a cavity into which a laser gas is introduced and a pair of electrodes disposed in the cavity, the performance prediction method including the steps of: acquiring a target feature and a component replacement scene; the target feature includes at least one of air pressure and applied voltage between electrodes, and the component replacement scenario includes a replacement component and a replacement timing; obtaining a trained RNN model corresponding to the target feature; acquiring past data of the laser device corresponding to the RNN model; according to the component replacement scene, data of the future use pulse number of the replaced component is made; through the RNN model, predicting the performance of the target feature in the component replacement scene according to the past data and the data of the future use pulse number of component replacement; and outputting a prediction result.

Description

Technical Field

[0001] The present disclosure relates to performance prediction methods and training methods. Background Art

[0002] In recent years, semiconductor exposure equipment has been required to achieve higher resolution as semiconductor integrated circuits become increasingly miniaturized and highly integrated. Consequently, there has been a trend toward shorter wavelengths of light emitted from exposure light sources. For example, gas lasers used for exposure include KrF excimer lasers, which output laser light with a wavelength of approximately 248 nm, and ArF excimer lasers, which output laser light with a wavelength of approximately 193 nm.

[0003] The spectral line width of the natural oscillation light of KrF excimer laser devices and ArF excimer laser devices is relatively wide, ranging from 350 to 400 pm. Therefore, when a projection lens is constructed using a material that transmits ultraviolet light such as KrF and ArF lasers, chromatic aberration may sometimes occur. As a result, the resolution may be reduced. Therefore, it is necessary to narrow the spectral line width of the laser light output from the gas laser device to a level where chromatic aberration is invisible. Therefore, in order to narrow the spectral line width, a narrowing module (Line Narrowing Module: LNM) containing narrowing elements (etalon, grating, etc.) is sometimes included in the laser resonator of the gas laser device. Hereinafter, a gas laser device that narrows the spectral line width will be referred to as a narrowed gas laser device.

[0004] Prior art literature

[0005] Patent Literature

[0006] Patent Document 1: U.S. Patent Application Publication No. 2021 / 0333788

[0007] Patent Document 2: U.S. Patent Application Publication No. 2021 / 0117931

[0008] Patent Document 3: Japanese Patent Application Laid-Open No. 2021-177223 Summary of the Invention

[0009] A performance prediction method according to one aspect of the present disclosure is used to predict the performance of a laser device having a cavity into which laser gas is introduced and a pair of electrodes arranged in the cavity, wherein the performance prediction method includes the following steps: obtaining a target feature and a component replacement scenario, the target feature including at least one of the gas pressure in the cavity of the laser device and the applied voltage between the electrodes, and the component replacement scenario including component replacement and replacement timing; obtaining a trained recurrent neural network model corresponding to the target feature; obtaining past data of the laser device corresponding to the recurrent neural network model; generating data on the future number of usage pulses of the replacement component based on the component replacement scenario; predicting the performance of the target feature in the component replacement scenario based on the past data and the data on the future number of usage pulses of the replacement component using the recurrent neural network model; and outputting the predicted result.

[0010] A training method according to another aspect of the present disclosure is a training method for a recurrent neural network model, which predicts the performance of a first laser device, wherein the first laser device has a cavity into which laser gas is introduced and a pair of electrodes arranged in the cavity, wherein the training method includes the following steps: obtaining past data of a target feature, a replacement component, and a plurality of features, the target feature including at least one of the gas pressure in the cavity of the first laser device and the applied voltage between the electrodes; extracting an additional feature for predicting the target feature from the plurality of features; creating training data including data before and after replacement of the replacement component, the training data being data including the target feature, the number of usage pulses of the replacement component, and the additional feature; and training the recurrent neural network model using the training data.

[0011] A training method according to another aspect of the present disclosure is a training method for a recurrent neural network model, which predicts the performance of a first laser device, wherein the first laser device has a first cavity into which laser gas is introduced and a pair of first electrodes arranged in the first cavity, wherein the training method includes the following steps: obtaining past data of a target feature, a replacement component, and a plurality of features, the target feature including at least one of the gas pressure in a second cavity and the applied voltage between the second electrodes of a second laser device different from the first laser device, the second laser device having a second cavity into which laser gas is introduced and a pair of second electrodes arranged in the second cavity; extracting additional features for predicting the target feature from the plurality of features; creating training data including data before and after replacement of the replacement component, the training data being data including the target feature, the number of usage pulses of the replacement component, and the additional feature; and training the recurrent neural network model using the training data. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Several embodiments of the present disclosure are described below by way of example only with reference to the accompanying drawings.

[0013] Figure 11 is a diagram showing the structure of an exemplary laser device for exposure equipment.

[0014] Figure 2 FIG. 1 is a diagram showing the configuration of an exemplary light source management system.

[0015] Figure 3 The configuration of the light source management system according to the first embodiment is shown.

[0016] Figure 4 is a block diagram illustrating the functionality of a laser performance simulator.

[0017] Figure 5 This is a diagram showing the flow of processing for creating training data.

[0018] Figure 6 It is a diagram showing a specific example of target features, replacement parts, and additional features.

[0019] Figure 7 This is a diagram showing an example of the structure of an RNN model.

[0020] Figure 8 This is a diagram showing an example of data input to the RNN model.

[0021] Figure 9 This is a diagram showing the flow of a process for predicting future performance of a target feature.

[0022] Figure 10 1 is a diagram showing a specific example of target features and component replacement scenarios.

[0023] Figure 11 1 is a diagram showing a specific example of target features and component replacement scenarios.

[0024] Figure 12 It is a figure which shows the output example of a prediction result.

[0025] Figure 13 It is a figure which shows the output example of a prediction result.

[0026] Figure 14 It is a figure which shows the output example of a prediction result.

[0027] Figure 15 1 and 2 are diagrams illustrating specific examples of target features and component replacement scenarios according to a modified example.

[0028] Figure 16 It shows that the Figure 15 FIG2 shows an output example of target features and prediction results in the case of a parts replacement scenario.

[0029] Figure 17 This is a block diagram showing the functions of the laser performance simulator according to the second embodiment.

[0030] Figure 18 This is a diagram showing the flow of processing for creating training data.

[0031] Figure 19 This is a diagram showing the processing flow of the RNN model training unit.

[0032] Figure 20 This is a diagram showing the structure of a laser device for an exposure apparatus according to a fourth embodiment.

[0033] Figure 21 This is a block diagram showing the functions of the laser performance simulator according to the third embodiment.

[0034] Figure 22 This is a diagram showing the flow of processing for creating training data.

[0035] Figure 23 This is a diagram showing the flow of a process for predicting future performance of a target feature. DETAILED DESCRIPTION

[0036] -Table of contents-

[0037] 1. Explanation of terms

[0038] 2. Comparative Example Device

[0039] 2.1 Laser device

[0040] 2.1.1 Structure

[0041] 2.1.2 Action

[0042] 2.2 Light source management system

[0043] 2.2.1 Structure

[0044] 2.2.2 Action

[0045] 3.Topic

[0046] 4. Implementation Method 1

[0047] 4.1 Structure

[0048] 4.1.1 Light Source Management System

[0049] 4.1.2 Laser Performance Simulator

[0050] 4.2 Action

[0051] 4.2.1 RNN Model

[0052] 4.2.2 Preparation of training data

[0053] 4.2.3 Training of RNN Model

[0054] 4.2.4 Performance prediction of target features

[0055] 4.2.5 Output Example

[0056] 4.3 Effect

[0057] 4.4 Variations

[0058] 5. Implementation Method 2

[0059] 5.1 Structure

[0060] 5.2 Action

[0061] 5.2.1 Preparation of training data

[0062] 5.2.2 Training of RNN Model

[0063] 5.3 Effect

[0064] 6. Implementation Method 3

[0065] 6.1 Laser Device

[0066] 6.1.1 Structure

[0067] 6.1.2 Action

[0068] 6.2 Laser Performance Simulator

[0069] 6.2.1 Structure

[0070] 6.2.2 Action

[0071] 6.2.2.1 Preparation of training data

[0072] 6.2.2.2 Training of RNN Model

[0073] 6.2.2.3 Performance Prediction of Target Features

[0074] 6.3 Effect

[0075] 7. Others

[0076] Below, embodiments of the present disclosure are described in detail with reference to the accompanying drawings. The embodiments described below illustrate several examples of the present disclosure and do not limit the content of the present disclosure. In addition, the structures and actions described in each embodiment are not necessarily all required structures and actions of the present disclosure. In addition, the same reference numerals are given to the same structural elements, and repeated descriptions are omitted.

[0077] 1. Explanation of terms

[0078] The terms used in this specification are defined as follows.

[0079] “Characteristics” are numerically quantifiable characteristics of a laser device, such as the pulse energy, central wavelength, and linewidth of the pulsed laser output from the laser device, as well as the gas pressure within the cavity, the voltage applied between electrodes, and the number of pulses used by each component.

[0080] The "target feature" is a feature predicted by the Recurrent Neural Network (RNN) model. The "target feature" is also called an outcome variable, a response variable, or a dependent variable. The "target feature" is input from an external device.

[0081] Additional features are features selected by feature selection algorithms such as the random forest algorithm in order to predict target features.

[0082] The "parts replacement scenario" is data consisting of replacement parts (parts to be replaced during the simulation prediction period) and replacement timings (replacement timings) at which the replacement parts are replaced, and includes a laser identification number. The "parts replacement scenario" is input from an external device.

[0083] There are two different types of weights: model weights and sample weights. Model weights are parameters of the RNN model and are adjusted during training to minimize the difference between predicted values and actual data. Sample weights are scaling factors associated with each time step of the training and validation data. Scaling factors are calculated using a weighting scheme to improve the accuracy of predictions related to part replacements. Mathematically, scaling factors are applied to the loss term of the optimization algorithm.

[0084] The "weighting scheme" is an algorithm for assigning sample weights. For example, the algorithm generates unequal weights so that higher weights can be set in the initial steps of component replacement to improve the accuracy of component replacement prediction.

[0085] "Training data" is the data used to train the RNN model. It contains a subset of four permutations: target feature data, data on the number of pulses used for replacement parts, data on additional features, and sample weights.

[0086] "Validation data" is used to compare the effectiveness of trained RNN models. It includes a subset of four permutations: target feature data, pulse count data for replacement parts, additional feature data, and sample weights. "Validation data" is older than "training data." "Validation data" is data from a shorter period than "training data."

[0087] Examples of “hyperparameters” include the number of hidden layers, the number of hidden neurons, model weight initialization, learning rate, learning rate decay parameter, and momentum parameter.

[0088] 2. Comparative Example Device

[0089] 2.1 Laser device

[0090] 2.1.1 Structure

[0091] Figure 1 The structure of the laser device 10 for the exposure apparatus of a comparative example is shown. The comparative example disclosed in the present invention is a method that the applicant has recognized as only known to the applicant, and is not a publicly known example acknowledged by the applicant himself.

[0092] The laser device 10 is, for example, an excimer laser device, and includes an oscillator (OSC) 20 , a monitor module 22 , and a laser processor 24 .

[0093] The OSC 20 includes a narrowbanding module (LNM) 26 , a cavity 28 , an output coupler (OC) 30 , a charger 32 , and a pulsed power module (PPM) 34 . The PPM 34 includes a switch 35 .

[0094] The LNM 26 includes a first prism 36, a second prism 38, a rotating stage 40 for rotating the second prism 38, and a grating 42. The LNM 26 controls the center wavelength of the pulsed laser light by changing the angle of incidence on the grating 42 by rotating the second prism 38. The rotating stage 40 may also include a piezoelectric element.

[0095] Cavity 28 includes a pair of electrodes 44 and 46, an insulating member 47, and two windows 48 and 50 through which laser light passes. Excimer laser gas is introduced into cavity 28. Excimer laser gas includes, for example, an inert gas (Ar gas or Kr gas), a halogen gas (F2 gas), and a buffer gas (Ne gas). PPM 34 is connected to electrode 44 via a feedthrough in insulating member 47.

[0096] OC30 is a partial reflection mirror that reflects part of the pulsed laser light and transmits the other part.

[0097] The LNM 26 and the OC 30 constitute an optical resonator, and a cavity 28 may be arranged on the optical path of the optical resonator.

[0098] Monitor module 22 includes a first beam splitter 52, a second beam splitter 54, a spectrum detector 56 for measuring the wavelength and spectral line width of the pulsed laser light, and a photosensor 58 for detecting the pulse energy of the pulsed laser light. Spectrum detector 56 may be an etalon spectrometer, and photosensor 58 may be a photodiode.

[0099] 2.1.2 Action

[0100] The laser processor 24 receives a target central wavelength λt, a target line width Δλt, and a target pulse energy Et from an exposure device (not shown). The laser processor 24 sets the charging voltage V1 of the charger 32 so as to obtain pulsed laser light with the target pulse energy Et.

[0101] A charging capacitor (not shown) in the PPM 34 is charged at the charging voltage V1 .

[0102] Upon receiving a light emission trigger signal Tr1 from the exposure device, the laser processor 24 transmits the light emission trigger signal Tr1 to a switch 35 within the PPM 34. When the switch 35 is actuated, the charge stored in the charging capacitor is converted within the PPM 34 into a high-voltage pulse corresponding to the charge voltage V1 and applied between electrodes 44 and 46 within the cavity 28.

[0103] As a result, a discharge occurs between electrodes 44 and 46, exciting the excimer laser gas within cavity 28. Then, pulsed laser light with a narrowband wavelength of 150 nm to 380 nm, which is ultraviolet, is output from OSC 20 via the optical resonator formed by OC 30 and LNM 26. The wavelength of the pulsed laser light can be the oscillation wavelength of an ArF excimer laser or a KrF excimer laser.

[0104] The pulse laser light output from the OSC 20 enters the monitor module 22 .

[0105] The pulsed laser light incident on the monitor module 22 is partially reflected by the first beam splitter 52 and partially reflected by the second beam splitter 54 to be incident on the spectrum detector 56 . The pulsed laser light transmitted through the second beam splitter 54 is incident on the optical sensor 58 .

[0106] The central wavelength and spectral line width of the pulsed laser light are measured by the spectrum detector 56. The pulse energy of the pulsed laser light is measured by the optical sensor 58.

[0107] The laser processor 24 can control the rotation stage 40 in the LNM 26 so that the central wavelength measured by the spectrum detector 56 becomes the target central wavelength λt.

[0108] The laser processor 24 can control the charging voltage V1 output from the charger 32 so that the pulse energy measured by the optical sensor 58 becomes the target pulse energy Et.

[0109] 2.2 Light source management system

[0110] 2.2.1 Structure

[0111] Figure 2The structure of a light source management system 100 of a comparative example is shown. The light source management system 100 includes a plurality of laser devices 10 - 1 , 10 - 2 , . . . , 10 -S each of which outputs pulsed laser light, an external device 102 , and a database 104 .

[0112] The plurality of laser devices 10-1, 10-2, ..., 10-S may be all laser devices in a semiconductor factory. The laser devices may be excimer laser devices. The plurality of laser devices 10-1, 10-2, ..., 10-S each has a unique laser identification number.

[0113] The external device 102 may be a PC (Personal Computer), a display device such as an LCD (Liquid Crystal Display) or an organic electroluminescent display, or an input device such as a keyboard or a voice input device.

[0114] The database 104 may be located in a semiconductor factory or a laser device.

[0115] The plurality of laser devices 10 - 1 , 10 - 2 , . . . , 10 -S, the external device 102 , and the database 104 are connected via a communication network 106 .

[0116] The communication network 106 is a communication network capable of transmitting information via wired or wireless communication or a combination thereof. The communication network 106 may be a wide area network or a local area network.

[0117] 2.2.2 Action

[0118] Data from multiple laser devices 10-1, 10-2, ..., 10-S is continuously stored in database 104, associated with the total number of oscillation pulses of each laser device and the date and time. The data includes, for example, the gas pressure within cavity 28, the charging voltage V1, and the number of pulses used by LNM 26. The data may also include the voltage applied between electrodes 44 and 46, the number of pulses used in cavity 28, the number of pulses used by OC 30, pulse energy, spectral line width, central wavelength, pulse energy stability, and the partial pressure of the halogen gas within cavity 28.

[0119] The data in the database 104 can be accessed from the external device 102 via the communication network 106 .

[0120] 3.Topic

[0121] Field service engineers use external device 102 to access database 104, review the data within database 104, and determine which components to replace. However, individual components degrade at varying rates, and the characteristics of the laser system are also affected by the condition of other components. Therefore, component lifespan cannot be determined simply by the number of pulses used. Therefore, field service engineers use their experience to make the following estimates.

[0122] a. Changes (performance) in the number of pulses or time corresponding to the characteristics of future laser devices

[0123] b. Performance of the laser device after future component replacement

[0124] However, the above estimation is difficult even for experienced field service engineers.

[0125] An object of the present disclosure is to quantitatively predict the performance of a laser device in the future, including after component replacement.

[0126] 4. Implementation Method 1

[0127] 4.1 Structure

[0128] 4.1.1 Light Source Management System

[0129] Figure 3 The structure of light source management system 110 according to Embodiment 1 is shown. Light source management system 110 differs from light source management system 100 in that it includes laser performance simulator 120. Laser performance simulator 120 utilizes a computer. The computer can be a server, a PC, or a workstation.

[0130] The laser performance simulator 120 is connected to the plurality of laser devices 10 - 1 , 10 - 2 , . . . , 10 -S, the external device 102 , and the database 104 via the communication network 106 .

[0131] The laser performance simulator 120 includes a CPU (Central Processing Unit) 122, a main storage device 124, an auxiliary storage device 126, a network interface 128, and a device interface 130. The CPU 122, the main storage device 124, the auxiliary storage device 126, the network interface 128, and the device interface 130 are connected via a bus 132. There may be multiple CPUs 122, multiple main storage devices 124, multiple auxiliary storage devices 126, multiple network interfaces 128, and multiple device interfaces 130.

[0132] The main storage device 124 is a storage device that can be directly accessed by the CPU 122. The main storage device 124 temporarily stores programs and various data. The main storage device 124 can be a volatile memory or a non-volatile memory.

[0133] Auxiliary storage device 126 is a storage device that cannot be directly accessed by CPU 122. Auxiliary storage device 126 permanently stores programs and various data. Auxiliary storage device 126 can be an HDD (Hard Disk Drive), SSD (Solid State Drive), or USB (Universal Serial Bus) memory.

[0134] The network interface 128 is an interface for connecting to the communication network 106 via wired or wireless connections or a combination thereof.

[0135] The device interface 130 is an interface for connecting to a display device 134 and an input device 136 .

[0136] The laser performance simulator 120 may also be provided in the external device 102 .

[0137] 4.1.2 Laser Performance Simulator

[0138] Figure 4 1 is a block diagram illustrating the functions of the laser performance simulator 120. The laser performance simulator 120 includes a data acquisition unit 140, a training data creation unit 142, a training data storage unit 144, an RNN model training unit 146, an RNN model storage unit 148, a laser performance prediction unit 150, and a data output unit 152.

[0139] The training data generating unit 142 generates training data for training the RNN model.

[0140] The training data storage unit 144 includes a storage unit that stores a file A in advance. The file A stores the training data.

[0141] The RNN model training unit 146 is a processing unit that trains the RNN model through machine learning using training data.

[0142] The RNN model storage unit 148 includes a storage unit that pre-stores a file Am, which stores the RNN model trained in the RNN model training unit 146.

[0143] The storage unit of the training data storage unit 144 and the storage unit of the RNN model storage unit 148 are configured using the auxiliary storage device 126. The storage unit of the training data storage unit 144 and the storage unit of the RNN model storage unit 148 may be configured using separate auxiliary storage devices 126, or may be configured as part of a storage area in one or more auxiliary storage devices 126.

[0144] The laser performance prediction unit 150 is a processing unit that predicts the performance of the laser device 10 using the trained RNN model stored in the file Am.

[0145] The data output unit 152 is a processing unit that outputs the prediction result of the laser performance prediction unit 150 .

[0146] Figure 4 The illustrated laser performance simulator 120 includes an RNN model training unit 146 and a laser performance prediction unit 150. However, the structure of the laser performance simulator 120 is not limited to this example. For example, the laser performance simulator that includes the RNN model training unit 146 and trains the RNN model and the laser performance simulator that includes the laser performance prediction unit 150 and predicts the performance of the laser device 10 may be separate devices. The laser performance simulator that includes the RNN model training unit 146 and the laser performance prediction unit 150 may be deployed in separate factories.

[0147] 4.2 Action

[0148] 4.2.1 RNN Model

[0149] The auxiliary storage device 126 stores an RNN model for performing the following quantitative prediction.

[0150] a. Performance of future laser devices

[0151] b. Performance of the laser device after future component replacement

[0152] The RNN model is a neural network model designed to process a sequence of data. In other words, the laser performance simulator 120 can accept an input sequence and train the RNN model to generate a predicted output sequence.

[0153] The RNN model has a multi-head structure that imports multiple inputs and can process sequences of past and future data.

[0154] Furthermore, the RNN model has an encoder / decoder structure that can compress input sequences into output sequences of different lengths. In other words, the length of the past data sequence used for prediction does not need to be the same as the length of the future data sequence.

[0155] The number of layers, neurons, and hyperparameters of the RNN model are selected based on the training of the RNN model.

[0156] 4.2.2 Preparation of training data

[0157] Figure 5 The flow of the process of creating training data is shown.

[0158] Initially, in step S1, data acquisition unit 140 acquires target features and replacement parts from an external device. The target features are features used to predict future performance using a trained RNN model. The external device may be external device 102 connected to communication network 106 or input device 136 connected to device interface 130 of laser performance simulator 120.

[0159] In step S2, the data acquisition unit 140 further acquires past data of the laser device 10T from the database 104. This past data is past data of multiple characteristics that are continuously recorded in association with the total number of oscillation pulses and the date and time. The laser device 10T is an example of the "second laser device" of the present disclosure. The laser device 10T may also be any of the multiple laser devices 10-1, 10-2, ..., 10-S. The past data includes data before and after replacement of the replacement component. The data acquisition unit 140 may also acquire past data of multiple characteristics from the laser device 10T.

[0160] The data acquisition unit 140 transmits the target feature and replacement component acquired in step S1 and the past data of the plurality of features acquired in step S2 to the training data creation unit 142 .

[0161] Next, in step S3, the training data generator 142 extracts features necessary for predicting the target feature from the multiple features of the received past data as additional features. To perform this extraction, the training data generator 142 calculates the importance of each of the multiple features. This importance calculation is performed using, for example, a random forest algorithm. The training data generator 142 extracts features with relatively high importance as additional features.

[0162] Next, in step S4, the training data generator 142 generates training data that includes target feature data, data on the number of pulses used for the replacement component, additional feature data, and sample weights. The additional features vary depending on the target feature, so the training data also varies depending on the target feature.

[0163] Finally, in step S5, the training data generating unit 142 sends the file A storing the generated training data to the training data storing unit 144. The training data storing unit 144 writes the received file A into the storage unit.

[0164] Figure 6 Specific examples of target characteristics, replacement components, and additional characteristics are shown. Specific examples of target characteristics are the gas pressure within cavity 28 and the voltage applied between electrodes 44 and 46. Specific examples of replacement components are cavity 28, LNM 26, and OC 30. Additional characteristics are characteristics that differ from the target characteristics. Specific examples of additional characteristics are the pulse energy, linewidth, central wavelength, pulse energy stability of the output pulsed laser, and the partial pressure of the halogen gas within cavity 28.

[0165] Furthermore, the target characteristic only needs to include at least one of gas pressure and applied voltage. The replacement component only needs to include at least one of cavity 28, LNM 26, and OC 30. Additional characteristics only need to include at least one of pulse energy, linewidth, central wavelength, pulse energy stability, and halogen gas partial pressure.

[0166] 4.2.3 Training of RNN Model

[0167] An example of the "training method" of the present disclosure will be described. The RNN model training unit 146 reads the file A storing the training data from the training data storage unit 144 and trains the RNN model using the training data.

[0168] Figure 7 An example of the structure of the RNN model 160 is shown. The RNN model 160 includes an encoder 162 and a decoder 164. In the RNN model 160, the output of the encoder 162 is connected to the input of the decoder 164.

[0169] The encoder 162 has multiple neuron layers N n-j ~N n The decoder 164 has multiple neuron layers N n+1 ~N n+k In neuron layer N n-j ~N n+k There can be multiple neurons in each.

[0170] X p,n The past data including the characteristics of the pulse number n. The pulse number n may be, for example, the total number of oscillation pulses of the laser device 10T. s,n+1 Contains the past data of the number of pulses used for the replacement parts in pulse number (n+1). n+1 Contains the predicted value of the target feature in spike number (n+1).

[0171] The number of pulses increases at fixed intervals in each step from (nj) to (n+k). For example, if the number of pulses n is 20,000 × 1 million pulses and the number of pulses increases by 150 × 1 million pulses in each step, the number of pulses (n-2) becomes 19,700 × 1 million pulses, and the number of pulses (n-1) becomes 19,850 × 1 million pulses. Furthermore, the number of pulses (n+1) becomes 20,150 × 1 million pulses, and the number of pulses (n+2) becomes 20,300 × 1 million pulses.

[0172] The training data includes data for the period from (nj) to (n+k). The RNN model training unit 146 sets the number of pulses n in the training data as the step to start prediction. For example, the number of pulses n can be set to 2 / 3 of the total period of the training data.

[0173] The RNN model training unit 146 sends the neuron layer N of the encoder 162 of the RNN model 160 to the RNN model 160. n-j ~N n Input the target feature data of each step from the pulse number (nj) to the pulse number n at the start of prediction, the data of the number of pulses used for the replacement part, and the data of the additional feature X p,n-j ~X p,n In addition, the RNN model training unit 146 sends the neuron layer N of the decoder 164 of the RNN model 160 to the decoder 164. n+1 ~N n+k Input the data X of the number of pulses used for the replacement parts in each step from pulse number (n+1) to (n+k). s,n+1 ~X s,n+k For example, in the case where the target feature, replacement part, and additional feature are Figure 6 In the specific example shown, the RNN model training unit 146 inputs the RNN model 160 Figure 8 Xp and Xs shown.

[0174] That is, the input data X p,n-j ~X p,n The gas pressures P(nj) to P(n) in the chamber 28 and the applied voltages V(nj) to V(n) between the electrodes 44 and 46 are included as target features. p,n-j ~X p,n The pulse numbers C1p(nj) to C1p(n) of cavity 28, LNMp(nj) to LNMp(n) of LNM26, and OC1p(nj) to OC1p(n) of OC30 are included as the pulse numbers of replacement parts. p,n-j ~X p,nThe pulse energy E(nj)~E(n), the spectral line width SW(nj)~SW(n), the central wavelength W(nj)~W(n), the pulse energy stability ES(nj)~ES(n) and the partial pressure PF(nj)~PF(n) of the halogen gas in the cavity 28 are included as additional features.

[0175] In addition, the input data X s,n+1 ~X s,n+k It includes the number of used pulses C1p(n+1) to C1p(n+k) of cavity 28, the number of used pulses LNMp(n+1) to LNMp(n+k) of LNM26, and the number of used pulses OC1p(n+1) to OC1p(n+k) of OC30.

[0176] For these inputs, the RNN model 160 extracts the neuron layer N from the decoder 164. n+1 ~N n+k Output the predicted value Y of the target feature from the number of pulses (n+1) to (n+k) n+1 ~Y n+k For example, RNN model 160 outputs Figure 8 The output of the predicted value Y is n+1 ~Y n+k This includes the gas pressure within chamber 28 and the voltage applied between electrodes 44 and 46.

[0177] Then, the RNN model training unit 146 makes the predicted value Y of the target feature of each step with the number of pulses from (n+1) to (n+k) n+1 ~Y n+k The scaling factor associated with each step of the training data, that is, the model weight of the RNN model 160 is adjusted in such a way that the difference from the data of the target feature in the training data becomes smaller.

[0178] In this way, the RNN model training unit 146 uses the training data to adjust the model weights by using optimization algorithms such as stochastic gradient method (SGD), RMSprop (Root Mean Square Propagation) or Adam (Adaptive Moment Estimation) to reduce the difference between the predicted value and the true value.

[0179] In addition, the RNN model training unit 146 may also create an RNN model by performing a comprehensive grid search on all combinations of hyperparameters or by using a Bayesian optimization algorithm.

[0180] The RNN model training unit 146 sends the file Am storing the trained RNN model 160 to the RNN model storage unit 148. The RNN model storage unit 148 writes the received file Am into the storage unit.

[0181] Here, an example in which the RNN model training unit 146 generates a new RNN model is described. However, the RNN model training unit 146 may also retrain a trained RNN model and update the model weights.

[0182] 4.2.4 Performance prediction of target features

[0183] The “performance prediction method” of the present disclosure will be described. Figure 9 The flow of the process of predicting the future performance of the target characteristic of the laser device is shown. Here, the case where the replacement timing is the same as the timing when the prediction is started is described.

[0184] Initially, in step S11, the laser performance prediction unit 150 obtains target characteristics and a component replacement scenario from an external device. Here, it is assumed that the laser identification number included in the component replacement scenario is the identification number of the laser device 10I. The laser device 10I is an example of the "first laser device" of the present disclosure. The laser device 10I may also be any of the multiple laser devices 10-1, 10-2, ..., 10-S. The laser device 10I and the laser device 10T may also be devices of the same model, device of the same structure, device of the same application, or devices having the same parameters.

[0185] The laser device 10I and the laser device 10T may be devices of different models, devices of different structures, devices for different purposes, or devices having different parameters. In this case, the cavity 28 of the laser device 10I is an example of a "first cavity," and the electrodes 44 and 46 of the laser device 10I are an example of a "pair of first electrodes." Furthermore, the cavity 28 of the laser device 10T is an example of a "second cavity," and the electrodes 44 and 46 of the laser device 10T are an example of a "pair of second electrodes."

[0186] The cavity 28 of the laser device 10I and the cavity 28 of the laser device 10T may be components of the same model or different models. The electrodes 44, 46 of the laser device 10I and the electrodes 44, 46 of the laser device 10T may be components of the same model or different models.

[0187] The replacement timing included in the component replacement scenario can be represented by the total number of oscillation pulses of the laser device 10I or by date and time. The external device can be an external device 102 connected to the communication network 106 or an external device such as an input device 136 connected to the device interface 130 of the laser performance simulator 120.

[0188] Next, in step S12, the laser performance prediction unit 150 reads the file Am from the RNN model storage unit 148 to obtain the RNN model. The file Am stores the trained RNN model for predicting the target features obtained in step S11. If the device having the RNN model training unit 146 and the device having the laser performance prediction unit 150 are different devices, the device having the RNN model training unit 146 may also publish the file Am.

[0189] Furthermore, in step S13, the laser performance prediction unit 150 obtains past data of the laser device 10I from the database 104. This past data corresponds to the RNN model obtained in step S12. The past data corresponds to the total number of oscillation pulses and the date and time of the laser device 10I. The past data includes target feature data, data on the number of pulses used for replacement parts, and data on additional features. The past data is equivalent to Figure 8 XP.

[0190] In step S14, the laser performance prediction unit 150 generates data on the number of pulses used by the replacement component for each step of the laser apparatus 10I from (n+1) to (n+k). Specifically, the laser performance prediction unit 150 calculates the number of pulses used by the replacement component for each step based on the data on the number of pulses used by the replacement component at the start of the prediction, i.e., when the total number of pulses is n, and the amount of increase in the total number of pulses of the laser apparatus 10I between the steps from n to (n+k).

[0191] At this time, the laser performance prediction unit 150 initializes the number of used pulses to 0 at the time of replacement of the replacement component according to the component replacement scenario. Here, the replacement timing is the same as the timing when the prediction is started, so the number of used pulses of the replacement component when the total number of oscillation pulses is n is initialized to 0. The laser performance prediction unit 150 calculates the number of used pulses of the replacement component in the future based on the increase in the total number of oscillation pulses of the laser device 10I between steps. The laser performance prediction unit 150 creates data on the number of used pulses of the replacement component in the future based on the calculation results. The data on the number of used pulses of the replacement component is equivalent to Figure 8 Xs.

[0192] Next, in step S15, the laser performance prediction unit 150 inputs the past data of the laser device 10I obtained in step S13 and the data of the future number of pulses of the replacement component calculated in step S14 into the RNN model obtained in step S12. In step S16, the RNN model predicts the future performance of the target feature corresponding to the number of pulses or date and time in the component replacement scenario. The prediction result is equivalent to Figure 8 The RNN model can also further predict the future performance of the target feature without replacing the component.

[0193] Finally, in step S17 , the data output unit 152 outputs the prediction result predicted in step S16 . The data output unit 152 may display the prediction result on the display device 134 or notify the external device 102 of the prediction result by sending an e-mail or the like via the communication network 106 .

[0194] Figure 10 A specific example of the target feature and component replacement scenario acquired in step S11 is shown, in which the replacement timing is represented by the total number of oscillation pulses of the laser device 10I. Figure 10 In the example shown, specific examples of the target characteristics are the gas pressure within cavity 28 and the voltage applied between electrodes 44 and 46. Furthermore, specific examples of component replacement scenarios include four cases: replacing cavity 28 when the total number of oscillation pulses of laser apparatus 10I is 6.98 billion pulses, replacing LNM 26 when the total number of oscillation pulses of laser apparatus 10I is 6.98 billion pulses, replacing OC30 when the total number of oscillation pulses of laser apparatus 10I is 6.98 billion pulses, and simultaneously replacing cavity 28, LNM 26, and OC30 when the total number of oscillation pulses of laser apparatus 10I is 6.98 billion pulses.

[0195] Figure 11 A specific example of the target feature and component replacement scenario acquired in step S11 is shown, in which the replacement timing is represented by date and time. Figure 11 In the example shown, specific examples of the target characteristics are the gas pressure within cavity 28 and the voltage applied between electrodes 44 and 46. Furthermore, specific examples of component replacement scenarios include replacing cavity 28 at 9:00 AM on X-year-Y-month-Z-day, replacing LNM 26 at 9:00 AM on X-year-Y-month-Z-day, replacing OC30 at 9:00 AM on X-year-Y-month-Z-day, and simultaneously replacing cavity 28, LNM 26, and OC30 at 9:00 AM on X-year-Y-month-Z-day.

[0196] 4.2.5 Output Example

[0197] Figure 12 Shows the use of Figure 10 An example of the output of prediction results for the target features and parts replacement scenario is shown. Figure 12 This is a graph with the horizontal axis representing the total number of oscillation pulses and the vertical axis representing the value of the target characteristic. Figure 12 In the figure, the past data before the prediction is made and the prediction results after the start of the prediction are connected and shown in a time series. The past data is the data of the target feature until the total number of oscillation pulses of the laser device 10I is 69.8×10 billion pulses at the start of the prediction. The prediction result is the future performance of the target feature after the start of the prediction. In the prediction result, the future performance of the target feature under each component replacement scenario can be recorded in parallel. In the prediction result, the future performance of the target feature without replacing the component can also be recorded in parallel. The horizontal axis can also be shared and the prediction results can be displayed separately for each target feature. In the prediction result, for example, a line can be used to illustrate the start of the prediction.

[0198] The horizontal axis of the curve graph showing the forecast results can also be date and time. Figure 13 The total number of oscillation pulses and the date and time are written in parallel as shown. The date and time of the prediction period can also be calculated based on the average number of oscillation pulses and the total number of oscillation pulses per day calculated based on past data.

[0199] Figure 14 Shows the use of Figure 10 Another output example of prediction results for the target feature and part replacement scenario is shown. Figure 14 The bar graph shows the impact of the replacement of LNM26, OC30 and cavity 28 on the target features. Figure 14 In FIG. 1 , the effects on target features are shown separately on the left and right sides: the effect of reducing the gas pressure in cavity 28; and the effect of reducing the applied voltage between electrodes 44 and 46. The reduction effect is represented by the difference between the data when the total oscillation pulse number of laser device 10I is 69.8 × 1 billion pulses and the data when the total oscillation pulse number of laser device 10I is 70.0 × 1 billion pulses.

[0200] 4.3 Effect

[0201] According to the laser performance simulator 120 , even an inexperienced field service engineer can perform the following quantitative predictions.

[0202] a. Performance of future laser devices

[0203] b. Performance of the laser device after future component replacement

[0204] Furthermore, the field service engineer can use the laser performance simulator 120 to predict the future transition of target characteristics corresponding to the number of pulses or date and time in any component replacement scenario.

[0205] Furthermore, the field service engineer can confirm the effect of component replacement in advance based on the prediction results of the laser performance simulator 120 .

[0206] 4.4 Variations

[0207] Here, a case where there is a replacement opportunity after the start of prediction, that is, a case where the replacement opportunity is after (in the future) the start of prediction will be described.

[0208] Figure 15 The target features of the modified example and a specific example of a component replacement scenario are shown. Figure 15 In the example shown, specific examples of the target characteristics are the gas pressure within cavity 28 and the voltage applied between electrodes 44 and 46. Furthermore, specific examples of component replacement scenarios include replacing cavity 28 when the total number of oscillation pulses of laser device 10I is 6.98 billion pulses, replacing LNM 26 when the total number of oscillation pulses of laser device 10I is 7.1 billion pulses, replacing OC30 when the total number of oscillation pulses of laser device 10I is 6.98 billion pulses, and simultaneously replacing cavity 28, LNM 26, and OC30 when the total number of oscillation pulses of laser device 10I is 6.98 billion pulses.

[0209] Right now, Figure 15 The parts replacement scenario shown is similar to Figure 10 The component replacement scenario shown is different in that the LNM 26 is replaced when the total number of oscillation pulses of the laser device 10I is 71.0×1 billion pulses.

[0210] Figure 16 Shows the use of Figure 15 The output example of the prediction results for the target characteristics and component replacement scenario is shown. The prediction results for replacing LNM 26 in this scenario are the same as those for the case where component replacement is not performed, from the start of prediction (when the total oscillation pulse count is 69.8 billion pulses) to the point of replacement (when the total oscillation pulse count reaches 71.0 billion pulses). Furthermore, the prediction results for when LNM 26 is replaced are shown after the total oscillation pulse count reaches 71.0 billion pulses.

[0211] If the replacement timing is after the start of the forecast, Figure 16 As shown, the replacement timing after the start of prediction can be displayed using a line, for example.

[0212] Thus, the output example of the modified example of embodiment 1 produces the same effect as the output example of embodiment 1. Furthermore, the laser performance simulator 120 can predict the future performance of the target feature after the replacement timing even if the component is replaced after the prediction start time.

[0213] 5. Implementation Method 2

[0214] 5.1 Structure

[0215] Figure 17 This is a block diagram showing the functions of the laser performance simulator 120A according to Embodiment 2. The laser performance simulator 120A differs from the laser performance simulator 120 in the files stored in the training data storage unit 144 .

[0216] The training data storage unit 144 includes a storage unit that pre-stores a plurality of files At1, At2, At3, ..., which store a plurality of training data for training the RNN model for predicting the target feature.

[0217] The training data storage unit 144 further includes a storage unit that pre-stores a plurality of files Av1, Av2, Av3, ... storing a plurality of verification data having the same features as the respective training data.

[0218] 5.2 Action

[0219] 5.2.1 Preparation of training data

[0220] Figure 18 The flow of the process of creating training data is shown.

[0221] The processing of step S21 is the same as Figure 5 The steps S1 and S2 are the same. In addition, the processing of step S22 is the same as Figure 5 The same as step S3.

[0222] In the next step S23, the training data generation unit 142 generates training data, which includes target feature data, data on the number of pulses used for the replacement component, data on additional features, and sample weights. In this case, the training data generation unit 142 generates multiple training data sets with different numbers of additional features or different data periods. Here, multiple training data sets with different numbers of additional features and different data periods are generated. For example, the number of additional features in each training data set is 5, 20, 45, or 70. For example, the period of each training data set is 5, 10, 20, or 30 times 1 billion pulses.

[0223] Next, in step S24, the training data generator 142 generates multiple validation data sets with the same features as the training data. The validation data is used to compare the prediction accuracy of the trained RNN model. Here, the validation data includes a subset of the four permutations of the training data. The validation data is older than the training data. The validation data period is shorter than the training data period.

[0224] Finally, in step S25, the training data creation unit 142 sends the files storing the generated training data and verification data to the training data storage unit 144. For example, the file At1 storing the training data TD1 with 20 additional features, the file At2 storing the training data TD2 with 45 additional features, and the file At3 storing the training data TD3 with 70 additional features are sent to the training data storage unit 144. Furthermore, the training data creation unit 142 sends the file Av1 storing the verification data VD1 with the same features as the training data TD1, the file Av2 storing the verification data VD2 with the same features as the training data TD2, and the file Av3 storing the verification data VD3 with the same features as the training data TD3 to the training data storage unit 144.

[0225] The training data storage unit 144 writes the received plurality of files into the storage unit.

[0226] 5.2.2 Training of RNN Model

[0227] Figure 19 The following shows the processing flow of the RNN model training unit 146.

[0228] Initially, in step S31, the RNN model training unit 146 reads a file containing training data from the training data storage unit 144 and uses the training data to train the RNN model. Since there are multiple training data sets, a trained RNN model corresponding to each training data set is created. That is, each of the multiple RNN models is trained using each of the multiple training data sets. For example, a trained RNN model trained using training data TD1, a trained RNN model trained using training data TD2, and a trained RNN model trained using training data TD3 are created.

[0229] Next, in step S32, the RNN model training unit 146 calculates the prediction accuracy for each of the multiple trained RNN models using the verification data corresponding to each training data used for training. For example, the RNN model training unit 146 calculates the prediction accuracy for the trained RNN model trained using the training data TD1 using the verification data VD1. Furthermore, the RNN model training unit 146 calculates the prediction accuracy for the trained RNN model trained using the training data TD2 using the verification data VD2. Furthermore, the RNN model training unit 146 calculates the prediction accuracy for the trained RNN model trained using the training data TD3 using the verification data VD3.

[0230] Finally, in step S33, the RNN model training unit 146 selects a trained RNN model with relatively high prediction accuracy. For example, the RNN model training unit 146 selects the trained RNN model trained using training data TD3 as the RNN model with the highest prediction accuracy. The RNN model training unit 146 sends the file Am containing the selected RNN model to the RNN model storage unit 148.

[0231] When comparing RNN models, you can evaluate prediction accuracy and computational speed. This evaluation can be used to determine the RNN model. You can also compare model performance with validation data to determine the optimal settings for the input data period and the number of input features.

[0232] 5.3 Effect

[0233] According to the laser performance simulator 120A, the same effects as those of the laser performance simulator 120 are achieved.

[0234] Furthermore, according to the laser performance simulator 120A, the RNN model is trained using a plurality of training data, and a trained RNN model with better prediction accuracy is selected. Therefore, compared with the laser performance simulator 120 , the prediction accuracy can be improved.

[0235] 6. Implementation Method 3

[0236] 6.1 Laser Device

[0237] 6.1.1 Structure

[0238] Figure 20 The following shows the structure of a laser device 10A for an exposure apparatus according to Embodiment 3. The laser device 10A includes an OSC 200 , an amplifier (AMP) 202 , an optical pulse stretcher (OPS) 204 , a monitor module 206 , and a laser processor 208 .

[0239] The structure of OSC200 is the same as that of OSC20.

[0240] AMP 202 includes a rear mirror (RM) 210, a cavity 228, an OC 230, a charger 232, and a PPM 234. PPM 234 includes a switch 235. The structures of cavity 228, charger 232, and PPM 234 are the same as those of the corresponding elements of OSC 200.

[0241] RM210 is a partial reflector that partially reflects a portion of the pulsed laser light and transmits the remaining portion. The reflectivity of RM210 can be 80% to 90%.

[0242] The cavity 228 includes a pair of electrodes 244 and 246, an insulating member 247, and two windows 248 and 250 for transmitting laser light. Excimer laser gas is introduced into the cavity 228.

[0243] OC230 is a partial reflector that partially reflects a portion of the pulsed laser light and transmits the other portion. The reflectivity of OC230 can be 10% to 30%.

[0244] The RM 210 and the OC 230 constitute an optical resonator, and a cavity 228 may be arranged on the optical path of the optical resonator. The optical resonator may be a Fabry-Perot type optical resonator.

[0245] OPS 204 includes a beam splitter (BS) 260, a concave mirror (CM) 262, a CM 264, a CM 266, and a CM 268. The optical path length of the delayed optical path of OPS 204 is, for example, 2 to 14 meters. The reflectivity of BS 260 is, for example, 40% to 70%. Laser light reflected from BS 260 can be configured so that it is reflected by CM 262, CM 264, CM 266, and CM 268, and then the beam is imaged again by BS 260.

[0246] The structure of the monitor module 206 is the same as that of the monitor module 22 .

[0247] 6.1.2 Action

[0248] Laser processor 208 receives target central wavelength λt, target line width Δλt, and target pulse energy Et from the exposure device and sets charging voltage V1 of charger 32 and charging voltage V2 of charger 232 to obtain pulsed laser light of target pulse energy Et.

[0249] The charging capacitor of the PPM 34 is charged at the charging voltage V1 , and the charging capacitor (not shown) in the PPM 234 is charged at the charging voltage V2 .

[0250] Upon receiving the emission trigger signal Trt from the exposure device, the laser processor 208 transmits an emission trigger signal Tr1 to the switch 35 within the PPM 34. When the switch 35 is actuated, the charge stored in the charging capacitor is converted into a high-voltage pulse corresponding to the charge voltage V1 within the PPM 34 and applied between the electrodes 44 and 46 within the cavity 28.

[0251] As a result, a discharge occurs between electrodes 44 and 46, exciting the excimer laser gas in cavity 28. OSC 20 then outputs narrowband seed light with an ultraviolet wavelength of 150 to 380 nm. The seed light may also have the wavelength of an ArF excimer laser or a KrF excimer laser.

[0252] When the laser processor 208 receives the light emission trigger signal Trt from the exposure device, it sends a light emission trigger signal Tr2 to the switch 235 of the PPM234 in such a way that a discharge is generated between the electrodes 244 and 246 when the seed light output from the OSC20 enters the discharge space of the cavity 228 of the AMP202.

[0253] When the switch 235 is operated, the charge stored in the charging capacitor (not shown) in the PPM 234 is converted into a high voltage pulse corresponding to the charging voltage V2 in the PPM 234 and applied between the electrodes 244 and 246 in the cavity 228 .

[0254] As a result, a discharge occurs between electrodes 244 and 246, exciting the excimer laser gas within cavity 228. At this point, seed light output from OSC 20 passes through RM 210 and enters the discharge space within cavity 228. The incident seed light is amplified by the optical resonator formed by OC 230 and RM 210 and output from AMP 202.

[0255] The pulse laser output from the AMP 202 is incident on the OPS 204 .

[0256] The pulsed laser light incident on the OPS 204 is looped through the delay optical path of the OPS 204 to extend the pulse width.

[0257] The pulse laser light passing through the OPS 204 enters the monitor module 206 , and the central wavelength, line width, and pulse energy of the pulse laser light are measured.

[0258] The laser processor 208 can control the charging voltage V2 output from the charger 232 so that the pulse energy measured by the optical sensor 58 becomes the target pulse energy Et.

[0259] 6.2 Laser Performance Simulator

[0260] 6.2.1 Structure

[0261] Figure 21 This is a block diagram illustrating the functions of the laser performance simulator 120B according to Embodiment 3. The laser performance simulator 120B differs from the laser performance simulator 120 in the files stored in the training data storage unit 144 and the files stored in the RNN model storage unit 148 .

[0262] The type of laser device varies depending on the wavelength of the pulsed laser output from the laser device and the system structure of the laser device. Figure 1 The laser device 10 shown in FIG. 1 is a laser device of type K. Figure 20 The laser device 10A shown, which is composed of two cavities 28 and 228 , is a laser device of type T. Furthermore, the OSCs 20 and 200 and the AMP 202 are provided as separate modules.

[0263] The training data storage unit 144 includes a storage unit that pre-stores files A, B, C, . . . , which store different training data depending on the type and module of the laser device.

[0264] The RNN model storage unit 148 includes a storage unit that pre-stores files Am, Bm, Cm,..., which store RNN models trained by the RNN model training unit 146. These RNN models are different according to the type and module of the laser device.

[0265] 6.2.2 Action

[0266] 6.2.2.1 Preparation of training data

[0267] Figure 22 The flow of the process of creating training data is shown.

[0268] In step S41, in addition to the target characteristics and replacement parts, the data acquisition unit 140 also acquires information about the type of laser device from an external device. The information about the type of laser device includes the model of the laser device, the wavelength of the pulsed laser output by the laser device, the system configuration of the laser device, and the application of the laser device.

[0269] The training data generator 142 generates different training data according to the type and module of the laser device. Therefore, in step S42 , the data acquisition unit 140 acquires from the database 104 past data on multiple features of the same type of laser device as the one acquired in step S41 .

[0270] The processing of step S43 is the same as Figure 5 The same as step S3.

[0271] Next, in step S44, the training data generator 142 generates training data for the laser device type acquired in step S41. This training data includes target feature data, data on the number of pulses used for the replacement component, data on additional features, and sample weights. Training data TD11 for an RNN model that predicts the OSC feature of laser device type K as the target feature, training data TD12 for an RNN model that predicts the OSC feature of laser device type T as the target feature, and training data TD13 for an RNN model that predicts the AMP feature of laser device type T as the target feature are generated.

[0272] Finally, in step S45, the training data generating unit 142 sends the files storing the training data to the training data storing unit 144. Here, file A storing training data TD11, file B storing training data TD12, and file C storing training data TD13 are sent to the training data storing unit 144.

[0273] The training data storage unit 144 writes the received file A, file B, and file C into the storage unit.

[0274] In this way, different training data is generated depending on the type of laser device. Here, different training data is generated for training the RNN model that predicts the performance of a laser device composed of a single cavity and for training the RNN model that predicts the performance of a laser device composed of two cavities.

[0275] In addition, different training data is generated for each module. Here, different training data is generated for training an RNN model that predicts performance using OSC features as target features and for training an RNN model that predicts performance using AMP features as target features.

[0276] It is also possible to create different training data for the training of the RNN model that predicts the performance when the components constituting the OSC are replaced, and for the training of the RNN model that predicts the performance when the components constituting the AMP are replaced.

[0277] Different training data can also be produced in the training of an RNN model for predicting the performance of a laser device that outputs a pulsed laser with an oscillation wavelength of an ArF excimer laser and in the training of an RNN model for predicting the performance of a laser device that outputs a pulsed laser with an oscillation wavelength of a KrF excimer laser.

[0278] 6.2.2.2 Training of RNN Model

[0279] The RNN model training unit 146 reads the file storing the training data from the training data storage unit 144 and trains the RNN model. At this time, a trained RNN model corresponding to the training data is generated.

[0280] For example, the RNN model training unit 146 uses training data TD11 to train an RNN model that predicts the characteristics of the OSC of a laser device of type K. Furthermore, the RNN model training unit 146 uses training data TD12 to train an RNN model that predicts the characteristics of the OSC of a laser device of type T. Furthermore, the RNN model training unit 146 uses training data TD13 to train an RNN model that predicts the characteristics of the AMP of a laser device of type T.

[0281] The RNN model training unit 146 sends files storing the trained RNN models to the RNN model storage unit 148. For example, the file Am storing an RNN model for predicting the characteristics of the OSC of a type K laser device, the file Bm storing an RNN model for predicting the characteristics of the OSC of a type T laser device, and the file Cm storing an RNN model for predicting the characteristics of the AMP of a type T laser device are sent to the RNN model storage unit 148.

[0282] The RNN model storage unit 148 writes the received file Am, file Bm, and file Cm into the storage unit.

[0283] 6.2.2.3 Performance Prediction of Target Features

[0284] Figure 23 The flow of the process of predicting the future performance of a target feature is shown.

[0285] In step S51 , the laser performance prediction unit 150 obtains the type, target characteristics, and component replacement scenario of the laser device 10I to be predicted from the external device 102 .

[0286] Furthermore, in step S52, the laser performance prediction unit 150 reads a file from the RNN model storage unit 148, and obtains an RNN model. This file stores a trained RNN model for predicting target features of laser devices of the type to which the acquired laser device 10I belongs. For example, if the acquired laser device 10I is a laser device of type T and the acquired target features are OSC features, the RNN model storage unit 148 reads a file Bm storing an RNN model for predicting OSC features of the laser device of type T.

[0287] The processing of steps S53 to S57 is respectively Figure 9 Steps S13 to S17 are the same.

[0288] In this way, different trained RNN models are obtained depending on the type of laser device. Here, the trained RNN models obtained when predicting the performance of a laser device composed of a single cavity are different from those obtained when predicting the performance of a laser device composed of two cavities.

[0289] In addition, different trained RNN models are obtained depending on the module. Here, the trained RNN models obtained when the OSC feature is set as the target feature are different from those obtained when the AMP feature is set as the target feature.

[0290] The trained RNN models obtained when predicting the performance of a laser device that outputs pulsed laser light with an oscillation wavelength of ArF excimer laser and when predicting the performance of a laser device that outputs pulsed laser light with an oscillation wavelength of KrF excimer laser may be different.

[0291] The trained RNN model obtained may also be different depending on the replacement component. For example, the trained RNN model obtained when the replacement component is a component constituting an OSC may be different from the trained RNN model obtained when the replacement component is a component constituting an AMP.

[0292] 6.3 Effect

[0293] According to the laser performance simulator 120B, the same effect as that of the laser performance simulator 120 is achieved.

[0294] Furthermore, according to the laser performance simulator 120B, trained RNN models that differ according to the type and module of the laser device are created and used, thereby improving prediction accuracy compared to the laser performance simulator 120 .

[0295] 7. Others

[0296] The above description is not limiting, but merely illustrative. Therefore, those skilled in the art will appreciate that modifications can be made to the embodiments of the present disclosure without departing from the scope of the claims. Furthermore, those skilled in the art will appreciate that the embodiments of the present disclosure can be used in combination.

[0297] Unless otherwise expressly stated, the terms used in this specification and claims as a whole should be interpreted as “non-limiting” terms. For example, terms such as “including”, “having”, “having”, and “equipped” should be interpreted as “not excluding the presence of structural elements other than the structural elements described”. In addition, the modifier “one” should be interpreted as meaning “at least one” or “one or more”. In addition, terms such as “at least one of A, B, and C” should be interpreted as “A”, “B”, “C”, “A+B”, “A+C”, “B+C”, or “A+B+C”. Furthermore, it should be interpreted as also including combinations of these and parts other than “A”, “B”, and “C”.

Claims

1. A performance prediction method for predicting the performance of a laser device, wherein the laser device comprises a cavity into which a laser gas is introduced and a pair of electrodes arranged in the cavity, wherein: The performance prediction method comprises the following steps: Obtaining a target characteristic and a component replacement scenario, wherein the target characteristic includes at least one of the gas pressure in the cavity of the laser device and the applied voltage between the electrodes, and the component replacement scenario includes component replacement and replacement timing; Obtaining a trained recurrent neural network model corresponding to the target feature; Obtaining past data of the laser device corresponding to the recurrent neural network model; generating data on the number of future usage pulses of the replacement component according to the component replacement scenario; Predicting the performance of the target feature in the component replacement scenario based on the past data and the data on the number of future usage pulses of the replacement component using the recurrent neural network model; as well as Output the predicted result.

2. The performance prediction method according to claim 1, wherein: The replacement timing includes the total number of oscillation pulses of the laser device or date and time.

3. The performance prediction method according to claim 1, wherein: The past data corresponds to the total number of oscillation pulses of the laser device.

4. The performance prediction method according to claim 1, wherein: The replacement timing is after the start of prediction of the performance prediction.

5. The performance prediction method according to claim 1, wherein: When the laser device is a laser device that outputs a pulsed laser with an oscillation wavelength of ArF excimer laser and when the laser device is a laser device that outputs a pulsed laser with an oscillation wavelength of KrF excimer laser, different trained recurrent neural network models are obtained.

6. The performance prediction method according to claim 1, wherein: When the laser device is a laser device composed of one cavity and when the laser device is a laser device composed of two cavities, different trained recurrent neural network models are obtained.

7. The performance prediction method according to claim 1, wherein: The performance prediction method includes the following steps: obtaining different trained recurrent neural network models according to the replacement components.

8. The performance prediction method according to claim 7, wherein: The laser device has an oscillator and an amplifier. When the replacement component is a component constituting the oscillator and when the replacement component is a component constituting the amplifier, different trained recurrent neural network models are obtained.

9. The performance prediction method according to claim 1, wherein: The performance prediction method includes the following steps: further predicting the future performance of the target feature without replacing components through the recurrent neural network model.

10. The performance prediction method according to claim 1, wherein: The performance prediction method comprises the following steps: displaying a curve graph connecting the past data and the predicted result, In the graph, the vertical axis or the horizontal axis represents at least one of the total number of oscillation pulses of the laser device and the date and time.

11. The performance prediction method according to claim 1, wherein: The performance prediction method comprises the following steps: displaying the prediction result using a curve graph, The predicted result includes the reduction effect of the target feature.

12. A training method for a recurrent neural network model, wherein the recurrent neural network model predicts the performance of a first laser device, the first laser device having a cavity into which laser gas is introduced and a pair of electrodes disposed within the cavity, wherein: The training method comprises the following steps: acquiring past data of a target feature, a replacement component, and a plurality of features, wherein the target feature includes at least one of a gas pressure within the cavity of the first laser device and a voltage applied between the electrodes; Extracting additional features for predicting the target feature from the plurality of features; Creating training data including data before and after replacement of the replacement component, the training data including the target feature, the number of pulses used for the replacement component, and the additional feature; as well as The recurrent neural network model is trained using the training data.

13. The training method according to claim 12, wherein: The training method comprises the following steps: Creating a plurality of training data sets having different numbers of additional features or different data periods; creating a plurality of validation data having the same features as the plurality of training data; Training each of the multiple recurrent neural network models using each of the multiple training data; Calculating the prediction accuracy of each trained recurrent neural network model using each of the plurality of verification data; and Select the trained recurrent neural network model with relatively high prediction accuracy.

14. The training method according to claim 12, wherein: The training method comprises the following steps: Calculating the importance of each of the plurality of features; and The features with relatively high importance are extracted as the additional features.

15. The training method according to claim 12, wherein: The additional characteristics include at least one of pulse energy, spectral line width, central wavelength, pulse energy stability of the output laser light, and partial pressure of the halogen gas contained in the laser gas in the cavity.

16. The training method according to claim 12, wherein: Different training data are created when the first laser device outputs pulsed laser light of an ArF excimer laser wavelength and when the first laser device outputs pulsed laser light of a KrF excimer laser wavelength.

17. The training method according to claim 12, wherein: Different training data are created when the first laser device is a laser device composed of one cavity and when the first laser device is a laser device composed of two cavities.

18. The training method according to claim 12, wherein: The first laser device includes an oscillator and an amplifier. Different training data are created when the replacement component is a component constituting the oscillator and when the replacement component is a component constituting the amplifier.

19. A training method for a recurrent neural network model, wherein the recurrent neural network model predicts the performance of a first laser device, the first laser device having a first cavity into which laser gas is introduced and a pair of first electrodes disposed in the first cavity, wherein: The training method comprises the following steps: acquiring past data of a target characteristic, a replacement component, and a plurality of characteristics, wherein the target characteristic includes at least one of a gas pressure within a second cavity and a voltage applied between second electrodes of a second laser device different from the first laser device, the second laser device having the second cavity into which laser gas is introduced and a pair of second electrodes disposed within the second cavity; Extracting additional features for predicting the target feature from the plurality of features; Creating training data including data before and after replacement of the replacement component, wherein the training data includes the target feature, the number of pulses used for the replacement component, and the additional feature; as well as The recurrent neural network model is trained using the training data.

20. The training method according to claim 19, wherein: The training method comprises the following steps: Creating a plurality of training data sets having different numbers of additional features or different data periods; creating a plurality of validation data having the same features as the plurality of training data; Training each of the multiple recurrent neural network models using each of the multiple training data; Calculating the prediction accuracy of each trained recurrent neural network model using each of the plurality of verification data; and Select the trained recurrent neural network model with relatively high prediction accuracy.

Images