Control apparatus, lithography apparatus and article manufacturing method

Abstract

The invention relates to a control apparatus, a lithographic apparatus, and an article manufacturing method. Provided is a technique that contributes to shortening learning time and ensuring learning stability. A control device is provided with: a control unit that controls an object to be controlled in accordance with a learning model in which a parameter value has been determined by learning; and a learning unit that re-determines the parameter value of the learning model by learning using an operation history of the object to be controlled that is controlled by the control unit. The learning unit does not use a first operation history used in the first learning, and only uses a second operation history acquired after the first operation history to execute a second learning after the first learning.

Description

Technical Field

[0001] The present invention relates to a control device, a photolithography device and an article manufacturing method. Background Art

[0002] In recent years, demands for improved control accuracy have become increasingly stringent, and conventional feedback control alone sometimes fails to achieve the required accuracy. Therefore, a solution has been adopted, namely the use of neural network controllers (Patent Document 1). Neural network controllers adjust parameters through machine learning. For example, in fields where complex models require high control accuracy, the introduction of reinforcement learning, a type of machine learning, is increasing.

[0003] Reinforcement learning is a type of machine learning, alongside supervised and unsupervised learning. In reinforcement learning, learning is achieved by repeatedly acquiring learning data, calculating parameters, and updating those parameters. However, during learning, production by the equipment must be stopped. This means that the execution of learning leads to a decrease in productivity. To address this issue, a method has been proposed that reduces learning time by asynchronously performing parameter updates and learning data acquisition in parallel (Patent Document 2).

[0004] Prior art literature

[0005] Patent Literature

[0006] Patent Document 1: Japanese Patent Application Laid-Open No. 2022-046317

[0007] Patent Document 2: Japanese Patent Application Laid-Open No. 2022-121112 Summary of the Invention

[0008] Problems to be solved by the invention

[0009] Reinforcement learning of the neural network controller is performed by repeatedly acquiring learning data and calculating parameters using the acquired learning data by executing a predetermined learning sequence. Therefore, acquiring learning data and calculating parameters are time-consuming during learning.

[0010] In Patent Document 2, learning data acquisition and parameter calculation are performed asynchronously, thereby shortening learning time. Furthermore, if the learning data acquisition time is shorter than the parameter calculation time, the learning data is based on a consistent behavioral history with fixed parameter values, resulting in stable learning. However, if the learning data acquisition time is longer than the parameter calculation time, parameter values ​​are updated during the learning data acquisition process, resulting in a mixture of behavioral histories based on multiple parameter values ​​in the learning data. This results in inconsistent behavioral histories in the learning data, which may reduce the stability of learning.

[0011] The present invention provides a technology that is advantageous in shortening learning time and ensuring learning stability.

[0012] Technical solutions to problems

[0013] According to the first aspect of the present invention, a control device is provided, characterized in that it comprises a control unit for controlling a control object according to a learning model in which parameter values ​​are determined by learning; and a learning unit for re-determining the parameter values ​​of the learning model by learning the action history of the control object controlled by the control unit, wherein the learning unit does not utilize the action history used in the first learning, but only utilizes the action history obtained after the past action history to perform a second learning after the first learning.

[0014] According to a second aspect of the present invention, there is provided a photolithography apparatus for transferring a pattern of an original plate onto a substrate, the photolithography apparatus comprising the control device according to the first aspect configured to control the position of the substrate or the original plate.

[0015] According to the third aspect of the present invention, there is provided a method for manufacturing an article, which is characterized in that it includes a transfer process of transferring the pattern of the original plate to a substrate using the photolithography device of the above-mentioned second aspect and a processing process of processing the substrate after the transfer process, and an article is obtained from the substrate after the processing process.

[0016] Effects of the Invention

[0017] According to the present invention, it is possible to provide a technique that is advantageous in shortening the learning time and ensuring the learning stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a diagram showing the structure of the system.

[0019] Figure 2 Is to Figure 1 This is a hardware structure diagram of the system when it is applied to the stage device.

[0020] Figure 3 1 is a diagram showing the configuration of a system including a learning unit for determining parameter values.

[0021] Figure 4 Is to Figure 3 This is a control block diagram when the system is applied to the stage device.

[0022] Figure 5 It is a block diagram showing the detailed structure of the learning unit.

[0023] Figure 6 is a flowchart showing the procedure of the learning process.

[0024] Figure 7 It is a flowchart showing the details of the parameter calculation process.

[0025] Figure 8 This is a flowchart showing the details of the learning data acquisition process.

[0026] Figure 9 This is a diagram showing a configuration example of an exposure device.

[0027] (Explanation of symbols)

[0028] 101: Sequence unit; 102: Control unit; 103: Control object; 201: Learning unit; 202: Learning data storage unit; 203: Learning control unit; 204: Learning data copying unit; 205: Parameter calculation unit; 206: Parameter storage unit. DETAILED DESCRIPTION

[0029] The following embodiments are described in detail with reference to the accompanying drawings. The following embodiments do not limit the inventions described in the claims. Although various features are described in the embodiments, not all of these features are necessarily required for the inventions, and various features may be combined arbitrarily. Furthermore, in the accompanying drawings, identical or similar structures are denoted by the same reference numerals, and duplicate descriptions are omitted.

[0030] <First embodiment>

[0031] Figure 1 This figure shows the structure of a system SS according to the first embodiment. The system SS can be applied to, for example, a manufacturing apparatus for manufacturing an article. The manufacturing apparatus can include, for example, a processing device that processes the material or component of the article or a member that constitutes a portion of the article. The processing device can be, for example, a photolithography apparatus that transfers a pattern onto a material or component, a film-forming apparatus that forms a film on a material or component, an apparatus that etches a material or component, or a heating apparatus that heats a material or component.

[0032] System SS can include, for example, a sequencer 101, a control unit 102 (control device), and a controlled object 103. When the system is applied to a production system, a production sequence can be provided as sequencer 101. The production sequence can define the order for production. Based on the production sequence, sequencer 101 can generate target values ​​for controlling controlled object 103 and provide these target values ​​to controlr 102. Controlr 102 outputs manipulated variables to controlled object 103 based on the controlled variables detected from controlled object 103 and the target values ​​set by sequencer 101. By repeating this operation, controlr 102 controls controlled object 103.

[0033] Figure 2This diagram shows an example of the hardware configuration when the system SS is applied to a stage device. The stage device can include a controller 501, a current driver 502, a motor 503, a stage 504, and a sensor 505. The controller 501 outputs a current command to the current driver 502. The current driver 502 outputs a current to the motor 503, which is built into the stage 504. The sensor 505 observes the results of the stage 504's driving. The sensor 505 provides position information to the controller 501.

[0034] Figure 3 This diagram shows the structure of a system SS including a learning unit. The control unit 102 controls the controlled object 103 according to a learning model whose parameter values ​​have been determined through learning. The learning unit 201 is configured to redetermine the parameter values ​​of the learning model by learning the motion history of the controlled object 103 controlled by the control unit 102. The learning unit 201 sends a predetermined learning sequence to the sequence unit 101. The sequence unit 101 sets a target value according to the learning sequence. The control unit 102 sends the motion history of the controlled object 103 to the learning unit 201. The learning unit 201 calculates the parameter values ​​of the learning model using the motion history and sends the obtained parameter values ​​to the control unit 102. Reinforcement learning is used in the parameter adjustment method.

[0035] Figure 4 Is to Figure 3 This is a control block diagram for a system applied to a stage device. A controller 501, which includes a neural network as a learning model, receives as input the position deviation, which is the difference between the target position value and the position information of the stage 504 observed by the sensor 505. The controller 501 outputs a current command to the current driver 502. The current driver 502 supplies a current corresponding to the current command to the motor 503. The driving force of the motor 503 based on the supplied current is converted into thrust and acts on the stage 504. To learn the parameters used to determine the neural network, the learning data storage unit 202 of the learning unit 201 stores an operation history. The operation history stored in the learning data storage unit 202 can include the position deviation as the input of the neural network and the current command as the output, but may also include other information required for learning. The learning unit 201 can be configured, for example, by a CPU (Central Processing Unit) included in the stage device. Alternatively, the learning unit 201 can be configured by an external computer connected to the stage device via a network.

[0036] In this embodiment, the deterioration of learning stability caused by the mixing of action histories based on multiple new and old parameter values ​​in the learning data is suppressed. Specifically, in this embodiment, the learning unit 201 performs the second learning subsequent to the first learning, without utilizing the first action history already used in the first learning. Instead, it utilizes only the second action history acquired after the first action history to perform the second learning. The specific structure for achieving this is described below.

[0037] exist Figure 5 , a detailed configuration example of the learning unit 201 is shown. The learning unit 201 can include a storage unit 20, a learning control unit 203, a parameter calculation unit 205, and a parameter storage unit 206. The storage unit 20 includes a learning data storage unit 202 providing a first storage area and a learning data copying unit 204 providing a second storage area. The first storage area of ​​the learning data storage unit 202 and the second storage area of ​​the learning data copying unit 204 can be data storage areas obtained by logical division within a storage device shared by the learning data storage unit 202 and the learning data copying unit 204. Alternatively, the first storage area of ​​the learning data storage unit 202 and the second storage area of ​​the learning data copying unit 204 can be physically different data storage areas, that is, data storage areas in independent storage devices.

[0038] The learning control unit 203 can also function as a storage control unit that controls the learning data storage unit 202 and the learning data copying unit 204. The learning data storage unit 202 stores data containing the operation history of the control object 103, which is sent from the control unit 102 to the learning unit 201. In response to receiving a data copy instruction from the learning control unit 203, the learning data storage unit 202 transfers a copy of the learning data stored in the learning data storage unit 202 to the learning data copying unit 204. The learning data copying unit 204 stores the data transferred from the learning data storage unit 202. In this case, the learning data copying unit 204 overwrites the data transferred from the learning data storage unit 202 with the data stored therein and stores it. Alternatively, the learning data copying unit 204 may store the data transferred from the learning data storage unit 202 after deleting all the data stored therein. Furthermore, in response to receiving a data delete instruction from the learning control unit 203, the learning data storage unit 202 deletes the learning data stored therein. For example, after copying the learning data stored in the learning data storage unit 202 to the learning data copying unit 204, the learning control unit 203 outputs a data deletion command to the learning data storage unit 202. In response to the data deletion command, the learning data in the learning data storage unit 202 is deleted. This prevents the mixing of behavior histories based on multiple new and old parameter values ​​in the learning data.

[0039] The parameter calculation unit 205 calculates parameter values ​​by performing reinforcement learning using the learning data stored in the learning data copying unit 204. After completing the parameter calculation, the parameter calculation unit 205 transmits the parameter value to the parameter storage unit 206. The parameter storage unit 206 stores the parameter value calculated by the parameter calculation unit 205. The learning control unit 203 controls the transfer of the parameter value stored in the parameter storage unit 206 to the control unit 102. For example, the learning control unit 203 outputs a parameter update instruction. In response to the parameter update instruction output by the learning control unit 203, the parameter storage unit 206 outputs the updated parameter value to the control unit 102. The learning control unit 203 transmits a learning sequence to the sequence unit 101 and receives a learning sequence completion signal from the sequence unit 101. After receiving the learning sequence completion signal from the sequence unit 101 and before starting the learning sequence, the learning control unit 203 outputs a parameter update instruction to the parameter storage unit 206. Furthermore, after receiving the learning sequence completion signal, the learning control unit 203 outputs a data copy instruction to the learning data storage unit 202. In response to receiving the data copy instruction from the learning control unit 203, the learning data storage unit 202 transfers a copy of the learning data stored in the learning data storage unit 202 to the learning data copy unit 204. After outputting the data copy instruction, the learning control unit 203 outputs a data delete instruction to the learning data storage unit 202. In response to receiving the data delete instruction from the learning control unit 203, the learning data storage unit 202 deletes the learning data stored in the learning data storage unit 202. Alternatively, the learning data may be invalidated rather than actually deleted from the learning data storage unit 202, so that it is not used in subsequent learning.

[0040] Thus, the storage unit of the learning unit 201 includes a learning data storage unit 202 (first storage area) that stores learning data including the action history (motion history) of the controlled object 103, and a learning data copying unit 204 (second storage area). The learning unit 201 (learning control unit 203) copies the learning data stored in the learning data storage unit 202 to the learning data copying unit 204 and then deletes or invalidates the learning data in the learning data storage unit 202. The learning control unit 203 then concurrently performs learning using the learning data stored in the learning data copying unit 204 and writes new learning data provided by the control unit 102 to the learning data storage unit 202. The writing of new learning data provided by the control unit 102 to the learning data storage unit 202 occurs while the control unit 102 is controlling the controlled object 103. This allows the learning unit 201 to perform a second learning process subsequent to the first learning process, without using the first motion history already used in the first learning process, but only using the second motion history acquired after the first motion history. This limits the learning data used in learning to data acquired under fixed parameter values, preventing a decrease in learning stability. Furthermore, the storage of learning data including the operation history and the calculation of parameter values ​​during the learning phase can be performed in parallel, shortening the time required for learning without stopping production.

[0041] The structure of this embodiment is particularly useful in a situation where the time to acquire learning data is longer than the time to calculate parameters. For example, the time taken until the control object 103 is controlled by the control unit 102 and the amount of learning data required for the learning performed by the learning unit 201 is accumulated in the learning data storage unit 202 is set to Ta. This Ta is the learning data acquisition time. In addition, the time taken to execute the learning implemented by the learning unit 201 is set to Tb. This Tb is the parameter calculation time. Therefore, the situation where the time to acquire learning data is longer than the parameter calculation time refers to a situation where Ta>Tb. According to this embodiment, even in such a situation, it is possible to prevent the parameter value from being updated in the process of acquiring the learning data and the action history based on multiple parameter values ​​from being mixed in the learning data.

[0042] Figure 6 This is a flowchart illustrating the sequence of the learning process. In S400 , the learning control unit 203 sets parameters for the neural network (NN) of the control unit 102 . When learning parameters for the first time, the parameter values ​​can be initialized with random numbers. When learning again, previously used parameter values ​​can be used.

[0043] In S401 , the learning control unit 203 initializes the learning data storage unit 202 , the learning data copy unit 204 , and the parameter storage unit 206 . In S402 , the learning control unit 203 sends a learning sequence to the sequence unit 101 .

[0044] In S403, the learning data storage unit 202 stores the operation history data sent from the control unit 102. In S404, the learning control unit 203 waits for receipt of a learning sequence completion signal from the sequence unit 101. If the learning sequence completion signal is received, the process proceeds to S405.

[0045] In S405, the learning control unit 203 outputs a data copy command to the learning data storage unit 202, causing the learning data copy unit 204 to store a copy of the data stored in the learning data storage unit 202. Thereafter, in S406, the learning control unit 203 outputs a data delete command to the learning data storage unit 202, deleting (or invalidating) the data stored in the learning data storage unit 202.

[0046] In S407, the parameter calculation process is executed, and in S408, the learning data acquisition process is executed. The parameter calculation process in S407 and the learning data acquisition process in S408 can be executed in parallel. The details of the parameter calculation process in S407 and the learning data acquisition process in S408 are described later. After the parameter calculation process and the learning data acquisition process are completed, in S409, the learning control unit 203 outputs a data copy instruction to the learning data storage unit 202, causing the learning data copy unit 204 to store a copy of the data stored in the learning data storage unit 202.

[0047] In S410, the learning control unit 203 updates the parameters of the neural network of the control unit 102. Next, in S411, the learning control unit 203 outputs a data deletion instruction to the learning data storage unit 202, deleting (or invalidating) the data stored in the learning data storage unit 202. Thereafter, in S412, the learning control unit 203 determines whether learning is complete. This determination of learning completion can be made, for example, by determining whether the learning sequence has been executed a predetermined number of times. Alternatively, learning completion can be determined by determining whether the control accuracy has reached a predetermined accuracy value. If learning is not complete, the process returns to S407 and repeats the process.

[0048] Figure 7It is a flowchart showing the details of the parameter calculation process of S407. In S416, the parameter calculation unit 205 calculates the parameter value using the learning data stored in the learning data copying unit 204. In S426, the learning control unit 203 determines whether the calculation of the parameter value is completed. Specifically, for example, the learning control unit 203 determines whether the calculation of the parameter value using the learning data of the learning data copying unit 204 has been performed a prescribed number of times. In addition, the prescribed number of times is a number predetermined before the start of learning and can be any integer greater than 1. The execution time of the learning data acquisition process S408 is determined by the execution time of the learning sequence, so the prescribed number of times can also be determined in such a way that the execution time of the parameter calculation process S407 is less than the execution time of the learning data acquisition process S408. In the case where the calculation of the parameter value is not completed, the processing returns to S416 and the calculation of the parameter value is repeated. Thus, until the amount of learning data required for learning is accumulated in the learning data storage unit 202 through the learning data acquisition process in S408, the learning unit 201 can repeatedly perform learning using the learning data stored in the learning data copying unit 204. Once the calculation of the parameter values ​​is completed, in S436, the learning control unit 203 stores the calculated parameter values ​​in the parameter storage unit 206.

[0049] Figure 8 This is a flowchart showing the details of the learning data acquisition process in S408. In S417, the learning control unit 203 sends the learning sequence to the sequence unit 101. As a result, under the control of the control unit 102, the control object 103 is controlled according to the learning sequence. In S427, the action history is stored in the learning data storage unit 202. The action history can at least include the control amount output from the control unit 102, the operation amount, and the next control amount generated by the actual operation of the control object 103 on the control amount. In S437, the learning control unit 203 waits for a learning sequence completion signal to be received from the sequence unit 101. If the learning sequence completion signal is received, the process proceeds to S409.

[0050] Next, the description will Figures 6-8 The learning process sequence shown is applied to a learning method for a stage device. Learning is performed when the stage device is installed and at appropriate timings when the stage control accuracy needs to be improved due to environmental changes, equipment changes over time, etc.

[0051] First, the learning control unit 203 sends a learning sequence to the sequencer unit 101. The learning sequence can be a production sequence. Alternatively, it can be a learning sequence that causes the stage to vibrate, for example. The controller 501 outputs an operation variable based on the target value set by the sequencer unit 101, driving the stage 504 as the control target. The operation history of the stage 504 is stored in the learning data storage unit 202.

[0052] When the learning sequence for the stage 504 is completed, the sequencer 101 outputs a learning sequence completion signal to the learning control unit 203. Upon receiving the learning sequence completion signal, the learning control unit 203 outputs a data copy command to the learning data storage unit 202. In response to receiving the data copy command, the learning data storage unit 202 transfers a copy of the learning data stored in the learning data storage unit 202 to the learning data copy unit 204. After outputting the data copy command, the learning control unit 203 outputs a data delete command to the learning data storage unit 202. In response to receiving the data delete command from the learning control unit 203, the learning data storage unit 202 deletes the learning data stored in the learning data storage unit 202.

[0053] The parameter calculation unit 205 calculates parameter values ​​using the learning data stored in the learning data copying unit 204. In parallel with the parameter calculation, the learning control unit 203 sends a learning sequence to the sequencer 101, which begins acquiring the learning data, representing the movement history of the stage 504. This parallel execution of parameter calculation and the learning sequence shortens learning time. Furthermore, the area storing the learning data used for parameter calculation (the learning data copying unit 204) and the area storing the movement history during the learning sequence (the learning data storage unit 202) are separated into separate areas. This limits the data used in learning to data acquired with fixed parameter values, preventing a decrease in learning stability. Furthermore, if parameter calculation completes early relative to the learning sequence, the data stored in the learning data copying unit 204 can be learned multiple times. For example, if executing the learning sequence takes three times as long as the parameter calculation, parameter calculation can be performed three times. This improves learning stability.

[0054] According to this embodiment, when obtaining optimal parameter values ​​through learning according to the device installation environment and temporal changes, parameter value calculation and learning data acquisition can be performed in parallel while preventing a decrease in learning stability.

[0055] <Second embodiment>

[0056] Figure 9An example of the structure of the exposure device EXP according to the second embodiment is shown. The exposure device EXP can be configured as a scanning exposure device. The exposure device EXP can include, for example, an illumination light source 600, an illumination optical system 601, a mask mounting stage 603, a projection optical system 604, and a plate mounting stage 606. The illumination light source 600 can include, but is not limited to, a mercury lamp, an excimer laser light source, or an EUV light source. The exposure light 610 from the illumination light source 600 is shaped with uniform illumination by the illumination optical system 601 into the shape of the irradiation area of ​​the projection optical system 604. In one example, the exposure light 610 is shaped into a rectangle that is long in the X direction, which is an axis perpendicular to the plane formed by the Y-axis and the Z-axis. Depending on the type of projection optical system 604, the exposure light 610 can be shaped into an arc shape. The shaped exposure light 610 is irradiated onto the pattern of the mask 602 . The exposure light 610 that has passed through the pattern of the mask 602 forms an image of the pattern of the mask 602 on the surface of the plate 605 (substrate) via the projection optical system 604 .

[0057] The mask 602 is held on the mask stage 603 by vacuum suction or other means. The plate 605 is held by a chuck 607 on the plate stage 606 by vacuum suction or other means. The positions of the mask stage 603 and the plate stage 606 can be controlled by a multi-axis position control system comprising a position sensor 630 such as a laser interferometer or laser ruler, a drive system 631 such as a linear motor, and a controller 632. The position measurement value output by the position sensor 630 is supplied to the controller 632. The controller 632 generates a control signal (operation variable signal) based on the difference between the position target value and the position measurement value, i.e., the position control deviation, and supplies it to the drive system 631, thereby driving the mask stage 603 and the plate stage 606. By synchronously driving the mask stage 603 and the plate stage 606 in the Y direction while scanning and exposing the plate 605, the pattern on the mask 602 is transferred to the plate 605 (the photosensitive agent on it).

[0058] A case where the first embodiment is applied to the control of the plate mounting table 606 will be described. Figure 4The controller 501 corresponds to the controller 632, the current driver 502 and the motor 503 correspond to the drive system 631, the stage 504 corresponds to the plate stage 606, and the sensor 505 corresponds to the position sensor 630. By applying a controller having a neural network to the control of the plate stage 606, the position control deviation of the plate stage 606 can be reduced. As a result, the overlap accuracy can be improved. When the state of the control object changes or the interference environment changes, the parameter value of the neural network is learned. In addition, learning can also be performed periodically to maintain control performance. According to this embodiment, by performing parameter calculation and learning sequence in parallel, learning can be performed in a short time. In addition, during the period of acquiring learning data, by performing an amount of learning corresponding to an arbitrary number of times, the learning stability can be improved. As a result, the number of executions of the learning sequence can be reduced. According to the above, learning can be performed in a short time, so that the productivity reduction of the exposure device can be suppressed to a minimum and the control accuracy can be maintained.

[0059] This embodiment can be applied to the mask mounting stage 603 in the same manner as the case of being applied to the plate mounting stage 606. In this case as well, the same effects can be obtained.

[0060] The first embodiment has been described above as being applied to stage control in an exposure apparatus. However, the first embodiment is applicable not only to stage control in exposure apparatuses but also to stage control in other lithography apparatuses such as imprint apparatuses and electron beam lithography apparatuses. Furthermore, the first embodiment can also be applied to controlling movable parts of a transport mechanism that transports an object, such as a hand that holds the object.

[0061] The above-described photolithography apparatus can be used to implement a method for manufacturing various products (such as semiconductor IC components, liquid crystal display devices, and MEMS). This method includes a transfer step in which a pattern from a master plate is transferred to a substrate using the above-described photolithography apparatus, and a processing step in which the substrate undergoing the transfer step is treated, thereby obtaining an article from the substrate undergoing the processing step. If the photolithography apparatus is an exposure apparatus, the transfer step can include an exposure step in which the substrate is exposed through the master plate, and a development step in which the exposed substrate is developed.

[0062] The present invention is not limited to the above-described embodiments, and various changes and modifications can be made without departing from the spirit and scope of the invention. Therefore, in order to disclose the scope of the invention, the following claims are attached.

Claims

1. A control device, characterized in that: have: a control unit that controls a control object according to a learning model whose parameter values ​​are determined through learning; and a learning unit that re-determines parameter values ​​of the learning model by learning using the operation history of the control object controlled by the control unit, The learning unit performs second learning subsequent to the first learning by using only a second operation history acquired after the first operation history, without using the first operation history used in the first learning.

2. The control device according to claim 1, characterized in that The learning unit includes a first storage area for storing learning data including an operation history of the control object and a second storage area different from the first storage area. The learning unit copies the learning data stored in the first storage area to the second storage area, and then deletes or invalidates the learning data in the first storage area, and performs learning using the learning data stored in the second storage area and writing new learning data provided by the control unit to the first storage area in parallel.

3. The control device according to claim 2, characterized in that Writing of new learning data supplied from the control unit into the first storage area is performed while the control unit is controlling the controlled object.

4. The control device according to claim 3, characterized in that The time required for accumulating the learning data of an amount required for the learning by the learning unit in the first storage area is longer than the time required for executing the learning by the learning unit.

5. The control device according to claim 4, characterized in that The learning unit executes learning using the learning data stored in the second storage area a plurality of times until the amount of learning data required for learning by the learning unit is accumulated in the first storage area.

6. The control device according to claim 2, characterized in that The learning department also has: a parameter calculation unit that calculates parameter values ​​of the learning model by performing reinforcement learning using the learning data stored in the second storage area; a parameter storage unit for storing the parameter value calculated by the parameter calculation unit; as well as The learning control unit controls the transfer of the parameter value stored in the parameter storage unit to the control unit.

7. A photolithography apparatus for transferring a pattern of an original plate to a substrate, wherein: A control device according to any one of claims 1 to 6 is provided, configured to control the position of the substrate or the original plate.

8. A method for manufacturing an article, characterized in that: Include: a transfer step of transferring the pattern of the original plate to the substrate using the photolithography apparatus according to claim 7; and a processing step of processing the substrate after the transfer step, An article is obtained from the substrate that has undergone the treatment process.

Images