Control parameter adjustment device, control parameter adjustment method, and storage medium

Abstract

The present application provides a control parameter adjustment device, a control parameter adjustment method, and a storage medium, and provides a technology capable of improving the quality of substrate processing even when the causal relationship between the control parameter and the result of the substrate processing is not clear. A control unit (201) sets the connection relationship between nodes by a connection relationship setting process (S332) using an explanatory variable (x), an intermediate variable (f), and a target variable (y) as the nodes. In a first learning process (S333), the control unit learns the input-output relationship of the intermediate variable in a network (N) based on the connection relationship set by the connection relationship setting process using a data set of the explanatory variable and the intermediate variable. In a second learning process (S334), the control unit learns the input-output relationship of the target variable in the network using a data set of the explanatory variable, the intermediate variable, and the target variable (z). In the second learning process, a prior distribution that induces sparsity is used as the prior distribution of the weight parameter.

Description

Technical Field

[0001] The subject matter disclosed in this specification relates to control parameter adjustment devices, control parameter adjustment methods, and computer programs. Background Technology

[0002] The substrate processing apparatus operates based on control parameters set in the manufacturing process to process the substrate. In this apparatus, the set values ​​of the control parameters need to be appropriately adjusted to achieve the processing result within the target value or acceptable range. Typically, the control parameters are adjusted by repeatedly processing the substrate while changing the set values, confirming the results and making adjustments accordingly. This method requires numerous trials, thus the adjustment process is time-consuming and labor-intensive.

[0003] Bayesian optimization is an effective method for optimizing control parameters while minimizing costs. In Bayesian optimization, the control parameters are used as explanatory variables, the quality of substrate processing is used as the objective variable, and a learned model that outputs a predicted distribution showing the relationships between variables, the search range of the control parameters, and an acquisition function are used to obtain the control parameters within the search range that maximize the acquisition function.

[0004] In Bayesian optimization, with the aim of performing optimization more efficiently, there exists a gray-box optimization method that effectively utilizes information obtained during the optimization process. Specifically, this method considers using detection data (physical quantities) detected by sensors provided by the substrate processing device as intermediate variables.

[0005] For example, in Patent Document 1, an auxiliary device that flexibly utilizes machine learning is proposed to improve the quality of substrate processing. The purpose of this auxiliary device is to reduce the previous trial-and-error-based adjustment process and efficiently derive the optimal control parameters. Specifically, firstly, using detection data of physical quantities in substrate processing and quality data of the processing results, a first machine learning model is trained to construct a predictive model of the processing results. Then, a Bayesian optimization method is used to obtain recommended values ​​for the detection data in a manner close to the target quality. Next, a second machine learning model learns the relationship between the setpoints of the control parameters and the detection data, and based on the obtained recommended values ​​from the detection data, recommended values ​​for the control parameters are calculated.

[0006] However, in the method of Patent Document 1, the relationship between intermediate variables and the target variable, as well as the relationship between explanatory variables and intermediate variables, are learned, but the direct relationship between the explanatory variables and the target variable is not learned. Therefore, the accuracy of deriving the recommended values ​​of the optimal control parameters may be limited.

[0007] On the other hand, the causal relationship between explanatory variables (control parameters) and target variables (processing results) is often not clear. In actual substrate processing, the data generation process of the processing result (target variable) is rarely fully clear either. That is, the causal relationship between the setpoints of the control parameters (explanatory variables) and the measured data (intermediate variables) is mostly clear, but the causal relationship between the setpoints of the control parameters (explanatory variables) and the result of substrate processing (target variable) is mostly unclear. Therefore, it is difficult to properly design a network that correctly represents the data generation process of the target variable.

[0008] If the network representing the data generation process is improperly configured, the prediction accuracy and optimization accuracy of the target variable may decrease. In particular, due to the high dimensionality of the input to the target variable, a data dimensionality dilemma arises, which not only slows down learning convergence but also increases the risk of overlearning. As a result, the model's fitting performance may decrease, or the number of searches for optimal control parameters may increase.

[0009] Patent Document 1: Japanese Patent Application Publication No. 2024-046993 Summary of the Invention

[0010] The purpose of this invention is to provide a technique that can improve the quality of substrate processing even when the causal relationship between control parameters and the results of substrate processing is not clear.

[0011] To address the aforementioned problem, a first approach is a control parameter adjustment device that adjusts the set values ​​of the control parameters of a substrate processing apparatus that processes a substrate based on control parameters. The control parameter adjustment device includes a computational processing unit that calculates recommended values ​​for the control parameters. This computational processing unit is capable of performing: a connection relationship setting process, setting the control parameters as explanatory variables, setting measurement data related to physical quantities measured in the substrate processing apparatus as intermediate variables, setting quality data related to the quality of substrate processing based on the substrate processing apparatus as target variables, setting the explanatory variables, the intermediate variables, and the target variables as nodes, and setting connection relationships between the nodes; and a first learning process. Using the datasets of the explanatory variables and the intermediate variables, a second learning process is performed to learn the input-output relationship of the intermediate variables in the network based on the connection relationship set by the connection relationship setting process; a third learning process is performed to learn the input-output relationship of the target variable in the network using the datasets of the explanatory variables, the intermediate variables, and the target variable; and a search process is performed to perform Bayesian optimization using the prediction model, search range, and acquisition function obtained through the first and second learning processes to search for recommended values ​​of the control parameters that make the quality close to the target quality. In the second learning process, the computational processing unit uses a prior distribution that induces sparsity as the prior distribution of the weight parameters.

[0012] The second approach is that in the control parameter adjustment device of the first approach, the arithmetic processing unit sets all selectable explanatory variables, all selectable intermediate variables, or both as the parent node of the target variable during the connection relationship setting process.

[0013] The third approach involves the control parameter adjustment device of the first approach, in which the computational processing unit uses a semi-Cauchy distribution as the prior distribution in the second learning process.

[0014] The fourth method, in the control parameter adjustment device of the first method, further includes: an input unit that accepts input; and a display, wherein the arithmetic processing unit performs the following processing in the connection relationship setting process: displays a GUI for accepting input specifying the connection relationship between the node of the description variable and the node of the intermediate variable, input specifying the connection relationship between the nodes of the intermediate variable, or both, on the display, and sets the connection of each node according to the input accepted via the input unit.

[0015] The fifth method is in the control parameter adjustment device of the fourth method, in which the arithmetic processing unit performs the following processing in the connection relationship setting process: displaying a GUI for accepting inputs specifying the connection relationship between the node of the description variable and the node of the target variable, inputs specifying the connection relationship between the node of the intermediate variable and the node of the target variable, or both, on the display screen, and setting the connection of each node according to the inputs accepted by the input unit.

[0016] The sixth method is in the control parameter adjustment device of the fifth method, in which the arithmetic processing unit displays a GUI on the display screen. The GUI has options for selecting variables one by one from the selectable explanatory variables or the selectable intermediate variables as the parent node of the target variable, and options for selecting all the selectable explanatory variables or the selectable intermediate variables as the parent node of the target variable at once.

[0017] The seventh method is a control parameter adjustment method that adjusts the set values ​​of the control parameters of a substrate processing apparatus that processes a substrate based on control parameters. The method includes: a connection relationship setting process, where the control parameters are set as explanatory variables, measurement data related to physical quantities measured in the substrate processing apparatus are set as intermediate variables, quality data related to the quality of substrate processing based on the substrate processing apparatus are set as target variables, and the explanatory variables, intermediate variables, and target variables are set as nodes, and connection relationships are set between the nodes; a first learning process, using a dataset of the explanatory variables and intermediate variables, learning the input-output relationships of the intermediate variables in a network based on the connection relationships set by the connection relationship setting process; a second learning process, using a dataset of the explanatory variables, intermediate variables, and target variables, learning the input-output relationships of the target variables in the network; and a search process, performing Bayesian optimization using the prediction model, search range, and acquisition function obtained through the first and second learning processes to search for recommended values ​​of the control parameters that make the quality approach the target quality.

[0018] The eighth method is a computer program that the computer can execute, causing the computer to perform the various processes of the control parameter adjustment method of the seventh method.

[0019] According to methods one through eight, by using a prior distribution that induces sparsity, the values ​​of inputs important to the target variable are maintained at a large value, while the values ​​of unimportant inputs become smaller. This property allows for the appropriate learning of the important inputs to the target variable. Therefore, even when descriptive and intermediate variables with low importance are connected to the target variable in the network, the prediction accuracy of the network can be prevented from decreasing. This, in turn, prevents a decrease in the optimization accuracy of the control parameters, thereby improving the quality of substrate processing. Attached Figure Description

[0020] Figure 1 This is a diagram illustrating the general outline of the control parameter adjustment method in this embodiment.

[0021] Figure 2 It means Figure 1 A diagram illustrating an example of a quality inspection process.

[0022] Figure 3 This is a diagram showing the substrate processing system and inspection apparatus 300.

[0023] Figure 4 It is a top view schematically showing the structure of the substrate processing apparatus included in the substrate processing system.

[0024] Figure 5 It is a schematic representation Figure 4 A cross-sectional view of the structure of the substrate processing section shown.

[0025] Figure 6 This is a block diagram illustrating the structure of the control device in this embodiment.

[0026] Figure 7 It is a graph that represents a dataset containing setpoints for control parameters, measurement data, and quality data.

[0027] Figure 8 This is a flowchart illustrating the control parameter adjustment method of the implementation method.

[0028] Figure 9 It is a graph representing the preprocessed dataset DS2, which includes the setpoints of control parameters, preprocessed measurement data, and quality data.

[0029] Figure 10 yes Figure 8 The detailed flowchart of the learning process is shown.

[0030] Figure 11 It is a graph that conceptually represents network N.

[0031] Figure 12A It is a diagram representing the GUI used to specify the connection relationships of a network.

[0032] Figure 12B It is a diagram representing the GUI used to specify the connection relationships of a network.

[0033] Figure 12C It is a diagram representing the GUI used to specify the connection relationships of a network.

[0034] Figure 13 This is a diagram representing a variant of a GUI used to specify network connectivity. Detailed Implementation

[0035] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. Furthermore, in the drawings, for ease of understanding, the dimensions and quantities of various parts are sometimes exaggerated or simplified.

[0036] In this embodiment, "substrate" refers to various substrates, specifically including semiconductor wafers, photomask glass substrates, liquid crystal display glass substrates, plasma display glass substrates, FED (Field Emission Display) substrates, optical disc substrates, magnetic disk substrates, and optical disk substrates. Hereinafter, the control parameter adjustment method, control parameter adjustment device, and computer program for processing disk-shaped semiconductor wafers will be mainly described, but they can also be applied to the processing of the various substrates illustrated above. Furthermore, the shape of the substrate can be various.

[0037] <1. Implementation Method>

[0038] The purpose of the control parameter adjustment method in this embodiment is to adjust the setpoint X of the control parameters of the substrate processing apparatus 100. In this method, a machine learning model is used to obtain a recommended value RX (recommendation condition) for an appropriate control parameter, and the setpoint X of the control parameter is adjusted based on this recommended value RX. Here, the setpoint X of the control parameter refers to the value of various control parameters set in the substrate processing apparatus 100. Furthermore, the setpoint X of the control parameter is not limited to control parameters specified in the manufacturing process; control parameters not specified in the manufacturing process can also be considered.

[0039] Figure 1 This diagram illustrates the general outline of the control parameter adjustment method in this embodiment. The control parameter adjustment method includes a quality inspection step S1, a dataset generation step S2, and a recommended value acquisition step S3. The quality inspection step S1 to the recommended value acquisition step S3 are repeatedly executed until the quality of the substrate processing of the substrate processing apparatus 100 reaches the target value or the allowable value.

[0040] In the quality inspection step S1, a quality inspection (experiment) is performed. Quality indicators include, for example, the number of particles attached to the substrate W by substrate processing, the collapse rate, or the amount of etching. In the quality inspection step S1, the substrate processing apparatus 100 processes the unprocessed substrate W based on the values ​​of control parameters, and the processed substrate W is inspected to obtain quality data Y related to the quality of the substrate processing by the substrate processing apparatus 100. Furthermore, in the following description, the unprocessed substrate W is sometimes referred to as "unprocessed substrate Wa," and the processed substrate W is sometimes referred to as "processed substrate Wb."

[0041] In the dataset generation process S2, in the inspection apparatus 300, a dataset DS1 is generated based on quality data Y, the set value of control parameters X, and measurement data Z1. Measurement data Z1 represents a physical quantity measured by sensors or the like installed in the substrate processing apparatus 100.

[0042] In the recommended value acquisition step S3, the control device 200 uses a machine learning model to learn the dataset. Then, by using Bayesian optimization of the learned machine learning model (prediction model PM), a recommended value RX for the control parameter is obtained. After the setpoint X of the control parameter is adjusted to the recommended value RX, the quality check step S1 is performed again. If the result of the quality check (quality data Y) meets the acceptable value, the recommended value RX at this point becomes the optimized setpoint X of the control parameter. If the result of the quality check does not meet the acceptable value, the dataset generation step S2 is executed again.

[0043] Figure 2 It means Figure 1 The diagram shows an example of the quality inspection process S1. Here, the quality inspection process S1 is described as an inspection that measures the number of particles attached to the substrate W by substrate processing. The quality inspection process S1 includes a pre-processing inspection process S11, a substrate processing process S12, a post-processing inspection process S13, and a quality data generation process S14.

[0044] The pre-processing inspection step S11 is a step to inspect the untreated substrate Wa. When the quality inspection step S1 is related to particle count, the distribution of particles attached to the substrate Wa is detected in the pre-processing inspection step S11. More specifically, the coordinates of each particle attached to the untreated substrate Wa are obtained. The origin of the coordinates is the center of the substrate W. In the following description, the particles attached to the untreated substrate Wa are sometimes referred to as "pre-processing particles".

[0045] In the substrate processing step S12 following the pre-processing inspection step S11, the substrate processing apparatus 100 processes the unprocessed substrate Wa. If the quality inspection step S1 is related to particle count inspection, the substrate Wa is cleaned in the substrate processing step S12.

[0046] In the post-processing inspection step S13 following the substrate processing step S12, the processed substrate Wb is inspected. The same inspection as the pre-processing inspection step S11 is performed in the post-processing inspection step S13. If the quality inspection step S1 is related to particle count, the distribution of particles attached to the substrate Wb is detected in the post-processing inspection step S13. More specifically, the coordinates of each particle attached to the processed substrate Wb are obtained. In the following description, the particles attached to the processed substrate Wb are sometimes referred to as "processed particles".

[0047] In the quality data generation step S14 following the post-processing inspection step S13, quality data Y is generated. When the quality inspection step S1 is related to particle inspection, in the quality data generation step S14, based on the distribution of particles before and after processing, the number of particles attached to the substrate W by the substrate processing is calculated. Specifically, the following counting process is performed: It is determined whether any of the coordinates of the particles before processing are contained within a circle of radius Th centered on the coordinates of the processed particles of the inspection object. Then, if no coordinate of the particles before processing is contained within the circle of the inspection object, the processed particles of the inspection object are considered as particles attached to the substrate W by the substrate processing and counted. By performing this counting process on all processed particles, quality data Y representing the number of particles attached to the substrate W by the substrate processing is generated. When quality data Y is generated, the quality inspection step S1 ends.

[0048] Furthermore, the quality inspection step S1 is not limited to checking the particle count. For example, the quality inspection step S1 could also be checking the etching amount. In this case, the film thickness of substrate Wa is measured in the pre-processing inspection step S11, etching is performed in the substrate processing step S12, and the film thickness of substrate Wb is measured in the post-processing inspection step S13. Then, in the quality data generation step S14, quality data Y representing the difference between the film thickness of the untreated substrate Wa and the film thickness of the treated substrate Wb is generated.

[0049] Figure 3This diagram illustrates a substrate processing system 1000 and an inspection apparatus 300. The substrate processing system 1000 includes a substrate processing apparatus 100 and a control apparatus 200. The substrate processing apparatus 100 operates based on set control parameters to process an unprocessed substrate W. The control apparatus 200 controls the operation of the substrate processing apparatus 100 based on the control parameters. The inspection apparatus 300 detects the distribution of particles attached to the substrate W. The inspection apparatus 300 determines the distribution by acquiring the coordinates of each particle attached to the substrate W.

[0050] The inspection device 300 generates quality data Y. The inspection device 300 compares the distribution of particles attached to the treated substrate Wb (processed particles) with the distribution of particles attached to the untreated substrate Wa (unprocessed particles), and calculates the number of newly attached particles due to substrate processing. Furthermore, the inspection device 300 performs a predetermined counting process on each processed particle to determine the number of attached particles caused by substrate processing. The inspection device 300 generates quality data Y representing the calculated number of attached particles. In addition, the inspection device 300 stores the radius Th in a storage unit.

[0051] The generated quality data Y is input to the control device 200. The control device 200 is an example of a device that adjusts the control parameters of the board processing. The quality data Y may be transmitted from the inspection device 300 to the control device 200 via a cable or via wireless communication. Alternatively, it may be input to the control device 200 via a removable medium. The removable medium may include memory cards such as SD cards, storage media such as USB (Universal Serial Bus) memory, and optical discs such as DVDs (Digital Versatile Discs).

[0052] Alternatively, the control device 200 can generate quality data Y. In this case, the distribution of particles attached to the untreated substrate Wa (coordinates of each particle before processing) and the distribution of particles attached to the treated substrate Wb (coordinates of each particle after processing) are input from the inspection device 300 to the control device 200.

[0053] Figure 4 This is a top view schematically showing the structure of the substrate processing apparatus 100 included in the substrate processing system 1000. The substrate processing apparatus 100 processes the substrate W using a processing liquid. The substrate processing apparatus 100 of this embodiment is a monolithic apparatus that processes substrates W one by one.

[0054] like Figure 4 As shown, the substrate processing apparatus 100 includes multiple substrate processing units 1, fluid cabinets 100A, multiple fluid tanks 100B, multiple loading ports LP, indexing robot IR, and central robot CR.

[0055] Each loading port LP houses multiple substrates W in a stacked configuration. The indexing robot IR moves the substrates W between the loading ports LP and the central robot CR. The central robot CR moves the substrates W between the indexing robot IR and the substrate processing unit 1. Alternatively, a mounting stage (path) for temporarily placing the substrates W can be configured between the indexing robot IR and the central robot CR. In this case, the indexing robot IR and the central robot CR can indirectly transfer the substrates W via the mounting stage.

[0056] Multiple substrate processing sections 1 are formed with multiple towers TW. Figure 4 In the example shown, the substrate processing apparatus 100 has four towers TW. The towers TW are arranged around a central robot CR when viewed from above. Each tower TW includes multiple substrate processing units 1 stacked in the vertical direction. Figure 4 In the example, three substrate processing units 1 are stacked in each tower TW layer.

[0057] Fluid tank 100A stores the processing liquid used for substrate processing. Multiple fluid tanks 100B correspond to different towers TW. Fluid tank 100A supplies the stored processing liquid to each fluid tank 100B. Each fluid tank 100B supplies processing liquid to all substrate processing sections 1 within its corresponding tower TW.

[0058] Each substrate processing unit 1 supplies a processing solution to the substrate W to process the substrate W. The processing solution may contain a chemical solution and a rinsing solution. Examples of chemical solutions include dilute hydrofluoric acid (DHF), hydrofluoric acid (HF), fluoronitric acid (a mixture of hydrofluoric acid and nitric acid (HNO3)), buffered hydrofluoric acid (BHF), ammonium fluoride, HFEG (a mixture of hydrofluoric acid and ethylene glycol), phosphoric acid (H3PO4), sulfuric acid, acetic acid, nitric acid, hydrochloric acid, ammonia, hydrogen peroxide, organic acids (e.g., citric acid, oxalic acid), organic bases (e.g., TMAH: tetramethylammonium hydroxide), sulfuric acid-hydrogen peroxide mixture (SPM), ammonia-hydrogen peroxide mixture (SC1), hydrochloric acid-hydrogen peroxide mixture (SC2), isopropanol (IPA), surfactants, or corrosion inhibitors. Examples of rinsing solutions include pure water (deionized water, etc.), carbonated water, electrolyzed ionized water, hydrogen water, ozone water, or hydrochloric acid water diluted to a concentration of approximately 10 PPM to 100 PPM.

[0059] The control device 200 controls various parts of the board processing device 100. For example, the control device 200 controls the board processing unit 1, the loading port LP, the indexing robot IR, and the central robot CR. The control device 200 includes a control unit 201 and a storage unit 202.

[0060] The control unit 201 includes a processor 201a. For example, the control unit 201 includes a CPU (Central Processing Unit) or an MPU (Micro Processing Unit). The control unit 201 may include a general-purpose arithmetic unit, a dedicated arithmetic unit, a GPU (Graphics Processing Unit), an NPU (Neural Network Processing Unit), or a quantum computer. Dedicated arithmetic units include, for example, ASICs (Application Specific Integrated Circuits).

[0061] The control unit 201 controls each part of the board processing apparatus 100 based on the data stored in the storage unit 202. The control unit 201 executes... Figure 1 The dataset generation process S2 and the recommendation value acquisition process S3 are shown. The control unit 201 is an example of a "computation processing unit".

[0062] Storage unit 202 includes a main storage device 202a. The main storage device is, for example, a semiconductor memory. Storage unit 202 may also include an auxiliary storage device 202b. The auxiliary storage device 202b is, for example, a semiconductor memory or a hard disk drive. Storage unit 202 may also include removable media.

[0063] The auxiliary storage device 202b stores data used to control the operation of various parts of the substrate processing apparatus 100. Additionally, the auxiliary storage device 202b stores data used for performing... Figure 1 The dataset shown represents the data for process S2 (generating the dataset) and process S3 (obtaining the recommended values).

[0064] The auxiliary storage device 202b stores a computer program. The computer program includes a control program for controlling the operation of the substrate processing apparatus 100. Additionally, the computer program includes an adjustment program PR. The adjustment program PR includes a computer program for performing the dataset generation step S2 and the recommended value acquisition step S3.

[0065] The auxiliary storage device 202b stores process data. The process data represents the processing content, processing conditions, and processing sequence of the substrate W. Various control parameters are specified as processing conditions in the process. Additionally, the data includes various control parameters not specified in the process. Hereinafter, the set values ​​of control parameters specified in the process will sometimes be recorded as "process set values." Furthermore, the set values ​​of control parameters not specified in the process will sometimes be recorded as "external set values."

[0066] Process settings include flow rate settings for various processing fluids, substrate rotation speed settings, pressure settings, and time settings. External settings include temperature settings for specific processing fluids.

[0067] The processor 201a reads the adjustment program PR stored in the auxiliary storage device 202b to the main storage device 201a and executes it to realize the dataset generation process S2 and the recommendation value acquisition process S3.

[0068] Figure 5 It is a schematic representation Figure 4 The diagram shows a cross-sectional view of the structure of the substrate processing unit 1. The substrate processing unit 1 includes a chamber 2, a rotary chuck 3, a rotary motor 4, a protective unit 5, a lifting unit 6, a nozzle 11, a first nozzle moving unit 110, and a second nozzle moving unit 120. The substrate processing apparatus 100 includes a first supply unit 130, a second supply unit 140, and a third supply unit 150.

[0069] Chamber 2 houses a rotary chuck 3, a rotary motor 4, a protective unit 5, a lifting unit 6, a camera 7, a nozzle 11, a first nozzle moving part 110, and a second nozzle moving part 120. Substrate W is housed one by one in chamber 2.

[0070] The rotary chuck 3 is a substrate holding part that holds the substrate W horizontally. The rotary chuck 3 has multiple chuck components 31 and a rotating base 33. The multiple chuck components 31 are arranged along the periphery of the substrate W on the rotating base 33. The multiple chuck components 31 hold the substrate W in a horizontal position. The rotating base 33 is generally circular and supports the multiple chuck components 31 in a horizontal position. The rotary chuck 3 is controlled by a control device 200.

[0071] The rotary motor 4 causes the substrate W and the rotary chuck 3 to rotate integrally about the rotation axis AX. The rotation axis AX extends in a generally vertical direction. The rotation axis AX is an example of a central axis, and the rotary motor 4 is an example of a substrate rotation part. The rotary motor 4 is connected to the rotating base 33 about the rotation axis AX, causing the rotating base 33 to rotate. Therefore, the rotating base 33 rotates about the rotation axis AX. As a result, the substrate W held by the plurality of chuck components 31 provided on the rotating base 33 rotates about the rotation axis AX.

[0072] The rotary motor 4 has a motor body 41 and a shaft 43. The shaft 43 is coupled to a rotating base 33. The motor body 41 rotates the shaft 43. As a result, the rotating base 33 rotates. The motor body 41 is controlled by a control device 200.

[0073] Nozzle 11 supplies processing liquid to substrate W. In this embodiment, nozzle 11 supplies a first processing liquid, a second processing liquid, and a third processing liquid to substrate W. More specifically, nozzle 11 has a first nozzle 11a, a second nozzle 11b, and a third nozzle 11c. First nozzle 11a supplies the first processing liquid to substrate W. First nozzle 11a sprays the first processing liquid toward the rotating substrate W.

[0074] The first processing fluid is a pharmaceutical solution. The first supply unit 130 supplies the first processing fluid to the first nozzle 11a. The first supply unit 130 includes a first piping PA1, a first on / off valve V1, a first control valve V2, a first flow sensor F1, a temperature sensor T, and a heating element H.

[0075] The first piping PA1 allows the first treatment fluid to flow to the first nozzle 11a. Furthermore, a portion of the first piping PA1 is housed within the chamber 2. A first on / off valve V1 is disposed on the first piping PA1. The first on / off valve V1 switches the supply of the first treatment fluid to the first nozzle 11a and stops the supply. The actuator of the first on / off valve V1 is, for example, a pneumatic actuator or an electric actuator. The first on / off valve V1 is controlled by a control device 200.

[0076] The first control valve V2 controls the flow rate of the first treatment fluid flowing in the first piping PA1. The opening degree of the first control valve V2 is adjustable. The actuator of the first control valve V2 is, for example, an electric actuator. The first control valve V2 is controlled by the control device 200.

[0077] The first flow sensor F1 detects the flow rate of the first processing fluid flowing in the first piping PA1. The flow rate of the first processing fluid is an example of a "physical quantity". The setpoint X of the control parameter includes the setpoint for the flow rate of the first processing fluid. The setpoint for the flow rate of the first processing fluid is an example of a process setpoint. The detection result SE of the first flow sensor F1 is an example of measurement data Z1 and is input to the control device 200.

[0078] Heating element H heats the first processing liquid flowing in the first piping PA1. Heating element H is controlled by control device 200. Temperature sensor T detects the temperature of the first processing liquid flowing in the first piping PA1. The temperature of the first processing liquid is an example of a "physical quantity". The setpoint X of the control parameter includes the setpoint for the temperature of the first processing liquid. The setpoint for the temperature of the first processing liquid is an example of an external setpoint. The detection result SE of temperature sensor T is an example of measurement data Z1 and is input to control device 200.

[0079] The first nozzle moving part 110 moves the first nozzle 11a in a generally vertical direction and a generally horizontal direction. Specifically, the first nozzle moving part 110 has a first arm 111, a first rotating shaft 113, and a first nozzle moving mechanism 115.

[0080] The first arm 111 extends in a generally horizontal direction. A first nozzle 11a is disposed at the front end of the first arm 111. The first arm 111 is coupled to a first rotating shaft 113. The first rotating shaft 113 extends in a generally vertical direction.

[0081] The first nozzle moving mechanism 115 rotates the first rotating shaft 113 about a rotation axis in a generally vertical direction, causing the first arm 111 to rotate in a generally horizontal plane. As a result, the first nozzle 11a moves in a generally horizontal plane. The first nozzle moving mechanism 115 also raises and lowers the first rotating shaft 113 in a generally vertical direction, thereby raising and lowering the first arm 111. As a result, the first nozzle 11a moves in a generally vertical direction. The first nozzle moving mechanism 115 may, for example, include a ball screw mechanism and an electric motor that applies driving force to the ball screw mechanism. The first nozzle moving mechanism 115 is controlled by the control device 200.

[0082] The second nozzle 11b supplies the second processing liquid to the substrate W. Specifically, the second nozzle 11b sprays the second processing liquid toward the rotating substrate W. The second processing liquid is a pharmaceutical solution.

[0083] The second supply unit 140 supplies the second processing fluid to the second nozzle 11b. Specifically, the second supply unit 140 includes a second piping PA2, a second on / off valve V11, a second control valve V12, and a second flow sensor F2.

[0084] The second supply unit 140 supplies the second processing fluid to the second nozzle 11b. Specifically, the second supply unit 140 includes a second piping PA2, a second on / off valve V11, a second control valve V12, and a second flow sensor F2.

[0085] The second piping PA2 allows the second processing fluid to flow to the second nozzle 11b. Chamber 2 houses a portion of the second piping PA2. A second on / off valve V11 is disposed on the second piping PA2. The second on / off valve V11 switches the supply of the second processing fluid to the second nozzle 11b and stops the supply. The actuator of the second on / off valve V11 is, for example, a pneumatic actuator or an electric actuator. The second on / off valve V11 is controlled by a control device 200.

[0086] The second control valve V12 controls the flow rate of the second processing fluid flowing in the second piping PA2. The second control valve V12 can adjust the opening degree of the second piping PA2. The actuator of the second control valve V12 is, for example, an electric actuator. The second control valve V12 is controlled by the control device 200.

[0087] The second flow sensor F2 detects the flow rate of the second processing fluid flowing in the second piping PA2. The flow rate of the second processing fluid is an example of a "physical quantity". The setpoint X of the control parameter includes the setpoint for the flow rate of the second processing fluid. The setpoint for the flow rate of the second processing fluid is an example of a process setpoint. The detection result SE of the second flow sensor F2 is an example of measurement data Z1 and is input to the control device 200.

[0088] The second nozzle moving part 120 moves the second nozzle 11b in a generally vertical and generally horizontal direction. Specifically, the second nozzle moving part 120 has a second arm 121, a second rotating shaft 123, and a second nozzle moving mechanism 125. The second nozzle moving mechanism 125 is controlled by the control device 200. The structure of the second nozzle moving part 120 is substantially the same as that of the first nozzle moving part 110, so its description is omitted.

[0089] The third nozzle 11c supplies a third processing liquid to the substrate W. The third nozzle 11c sprays the third processing liquid toward the rotating substrate W. The third processing liquid is a rinsing liquid. The third nozzle 11c is a stationary nozzle.

[0090] The third supply unit 150 supplies the third processing fluid to the third nozzle 11c. Specifically, the third supply unit 150 includes a third piping PA3, a third on / off valve V21, a third control valve V22, and a third flow sensor F3.

[0091] The third piping PA3 allows the third processing fluid to flow to the third nozzle 11c. Chamber 2 houses a portion of the third piping PA3. A third on / off valve V21 is disposed on the third piping PA3. The third on / off valve V21 switches the supply of the third processing fluid to the third nozzle 11c and stops the supply. The actuator of the third on / off valve V21 is, for example, a pneumatic actuator or an electric actuator. The third on / off valve V21 is controlled by a control device 200.

[0092] The third control valve V22 controls the flow rate of the third treatment fluid flowing in the third piping PA3. The third control valve V22 can adjust the opening degree of the third piping PA3. The actuator of the third control valve V22 is, for example, an electric actuator. The third control valve V22 is controlled by the control device 200.

[0093] The third flow sensor F3 detects the flow rate of the third processing fluid flowing in the third piping PA3. The flow rate of the third processing fluid is an example of a "physical quantity". The setpoint X of the control parameter includes the setpoint for the flow rate of the third processing fluid. The setpoint for the flow rate of the third processing fluid is an example of a process setpoint. The detection result SE of the third flow sensor F3 is input to the control device 200 as measurement data Z1.

[0094] A protective unit 5 is disposed outside the rotary chuck 3 and the rotary motor 4. The protective unit 5 is generally cylindrical. In other words, the protective unit 5 surrounds the rotary chuck 3 and the rotary motor 4. The protective unit 5 catches the processing liquid (first processing liquid to third processing liquid) that splashes from the rotating substrate W. Figure 5 In the example shown, the protective unit 5 includes three protective components.

[0095] The lifting unit 6 causes the three protective components to rise and fall individually. The lifting unit 6 is controlled by the control device 200. Specifically, the lifting unit 6 causes the protective unit 5 (here, three protective components) to rise and fall between the liquid receiving position and the retracted position. The liquid receiving position is a position higher than the retracted position.

[0096] via the central robot CR ( Figure 4 When the substrate W is moved into the chamber 2, the protective unit 5 is positioned in the retracted position. Additionally, when the substrate W is removed from the chamber 2 by the central robot CR, the protective unit 5 is positioned in the retracted position. Each protective component is positioned in the liquid receiving position when receiving the processing liquid.

[0097] Chamber 2 houses camera 7. Camera 7, for example, captures images of the upper surface of substrate W held on rotating base 33. Camera 7 includes an optical system, an image sensor, signal processing circuitry, and a control system (microcomputer). The image sensor is a semiconductor device such as CCD (Charge-Coupled Device) or CMOS (Complementary Metal Oxide Semiconductor). The image data acquired by the image sensor represents physical quantities such as the brightness of the photographed object. Camera 7 inputs the acquired image data as measurement data Z1 to control device 200.

[0098] Furthermore, although the illustration is omitted, several airflow sensors for measuring airflow can be installed at specific locations within chamber 2. The detection results of these airflow sensors can also be input to the control device 200 as measurement data Z1.

[0099] Figure 6 This is a block diagram showing the structure of the control device 200 according to this embodiment. The control device 200 is an example of a "control parameter adjustment device". In addition to the control unit 201 and the storage unit 202, the control device 200 also includes an input unit 203 and a display 204.

[0100] The input unit 203 is a user interface device operated by the operator. The input unit 203 inputs instructions (input signals) corresponding to the operator's operation to the control unit 201. Additionally, the input unit 203 inputs data corresponding to the operator's operation to the control unit 201. The input unit 203 includes a keyboard 203a and a mouse 203b. The input unit 203 may also include a touch sensor that overlaps with the display surface of the display 204. A graphical user interface can also be constructed by overlapping a touch sensor with the display surface of the display 204. For example, the operator can operate the input unit 203 to initiate the processing of the substrate W by the substrate processing apparatus 100.

[0101] In addition, the input unit 203 may also include an interface capable of accessing removable media. The interface may include a memory card slot, a USB terminal, and a reading device for reading data from optical discs. A USB cable with a memory card reader may also be connected to the USB terminal.

[0102] The display 204 displays various screens or images. The display 204 may include, for example, a liquid crystal display device. The display 204 displays screens for the operator to use in operating the control device 200 or the substrate processing device 100. Additionally, the display 204 may also display recommended values ​​RX for control parameters.

[0103] The substrate processing apparatus 100 inputs measurement data Z1 to the control device 200. The inspection apparatus 300 inputs quality data Y to the control device 200. The control unit 201 stores the input measurement data Z1 and quality data Y in the auxiliary storage device 202b.

[0104] The auxiliary storage device 202b stores the setpoint X of the control parameters, measurement data Z1, mass data Y, acquisition function AF, search range SW, and adjustment program PR. The control unit 201 uses the setpoint X of the control parameters, measurement data Z1, mass data Y, acquisition function AF, search range SW, and adjustment program PR to generate a recommended value RX for the control parameters. Specifically, the control unit 201 executes... Figure 1 The dataset generation process S2 and the recommendation value acquisition process S3 are shown.

[0105] The tuning procedure (PR) contains a machine learning model. Additionally, the tuning procedure PR includes a procedure for performing Bayesian optimization. The function AF and search range SW are obtained for Bayesian optimization. Alternatively, the tuning procedure PR may also include the function AF and search range SW.

[0106] Figure 7 This is a graph representing the dataset DS1, which includes the setpoint X of the control parameter, the measurement data Z1, and the quality data Y. For example... Figure 7As shown, the setpoint X of the control parameter represents the setpoint P (setpoint P1, P2, ...) of various control parameters. The measurement data Z1 represents the set of detection results SE (SE1, SE2, ...) of various sensors installed on the substrate processing device 100. The auxiliary storage device 202b is for... Figure 1 Each quality inspection process S1 shown stores the set value of the control parameter X, the measurement data Z1, and the quality data Y in an interrelated manner.

[0107] Figure 8 This is a flowchart illustrating the control parameter adjustment method of the implementation method. Figure 8 This indicates the processing flow executed by the control unit 201. Figure 8 The control parameter adjustment methods shown include Figure 1 The dataset generation process S2 and the recommendation value acquisition process S3 are shown. The dataset generation process S2 includes a preprocessing process S31 and a preprocessed dataset generation process S32. The recommendation value acquisition process S3 includes a learning process S33, an optimization process S34, and a saving process S35. Figure 8 The process shown is initiated by instructing the operator to obtain the recommended value RX via the operator operation input unit 203.

[0108] like Figure 8 As shown, if the control unit 201 is instructed to acquire the recommended value RX of the control parameter, it executes the preprocessing step S31. Specifically, the control unit 201 preprocesses each detection result SE contained in the measurement data Z1 to generate preprocessed measurement data Z2. Preprocessing reduces the number of data points for each detection result SE. Hereinafter, the preprocessed detection result SE will sometimes be referred to as "preprocessed data ASE". The preprocessed measurement data Z2 contains preprocessed data ASE (ASE1, ASE2, ASE3, etc.). The preprocessed measurement data Z2 is an example of "measurement data related to physical quantities".

[0109] Preprocessing can be, for example, the process of extracting features from the raw data (detection result SE). Preprocessing may include processes such as reducing the dimensionality of the raw data (dimensionality reduction), calculating summary statistics of the raw data, or extracting one or more arbitrary data points from the raw data.

[0110] Dimensionality reduction can be linear or nonlinear. Linear dimensionality reduction includes, for example, principal component analysis (PCA), independent component analysis (ICA), or tensor decomposition (TFD). Nonlinear dimensionality reduction includes, for example, UMAP (Uniform Manifold Approximation and Projection) or autoencoders. For instance, PCA yields low-dimensional features of the detection result SE (original data). Therefore, when preprocessing is PCA, the preprocessed data ASE represents the low-dimensional features of the detection result SE (original data).

[0111] Summary statistics, such as mean, median, or mode, are used to calculate the summary statistics of the test results SE (raw data). The mean, median, or mode of the test results SE is obtained by calculating these summary statistics. Therefore, if the preprocessing involves calculating summary statistics, the preprocessed data ASE represents the mean, median, or mode of the test results SE.

[0112] In the preprocessed dataset generation step S32, the control unit 201 generates a preprocessed dataset DS2 based on the setpoint X of the control parameter, the preprocessed measurement data Z2, and the quality data Y. The preprocessed dataset DS2 is an example of training data used for machine learning. Figure 9 This is a graph representing the preprocessed dataset DS2, which includes the setpoint X of the control parameter, the preprocessed measurement data Z2, and the quality data Y. For example... Figure 9 As shown, in the preprocessed dataset DS2, the set values ​​of the control parameters X (P1, P2, P3, etc.) in each quality inspection, the preprocessed measurement data Z2 (ASE1, ASE2, ASE3, etc.), and the quality data Y are recorded in a corresponding manner.

[0113] In the learning process S33, the control unit 201 performs machine learning using a machine learning model and the preprocessed dataset DS2. As a result, the control unit 201 obtains a predictive model PM that predicts the quality of the substrate processing. The predictive model PM is, for example, a model that predicts the number of particles that adhere to the substrate W through substrate processing. Details of the learning process S33 will be described later.

[0114] In optimization step S34, control unit 201 performs Bayesian optimization based on prediction model PM, acquisition function AF, and search range SW to search for recommended values ​​RX of control parameters that bring the substrate processing quality close to the target quality. Specifically, control unit 201 obtains recommended values ​​RX of control parameters for maximizing the acquisition function AF within the search range SW through Bayesian optimization. Furthermore, based on the substrate processing quality indicators, it is also conceivable that control unit 201 is designed to obtain recommended values ​​RX of control parameters that minimize the acquisition function. Optimization step S34 is an example of a "search process".

[0115] The function AF can be, for example, EI (Expected Improvement), MI (Mutual Information), PI (Probability of Improvement), or UCB (Upper Confidence Bound). The search range SW can be any range. The search range SW can be set, for example, based on the value input by the operator's operation input unit 203.

[0116] After the optimization process S34, the control unit 201 saves the recommended value RX obtained in the saving process S35 to the auxiliary storage device 202b. Then, the control unit 201 ends the processing of the recommended value acquisition process S3.

[0117] Figure 10 yes Figure 8 The detailed flowchart of learning step S33 is shown. As described above, in this embodiment, a recommended value RX of the control parameter that makes the quality of substrate processing close to the target quality is obtained through Bayesian optimization, and therefore a network N representing the data generation process of quality data Y (target variable) is constructed.

[0118] In the learning process S33, the control unit 201 acquires the preprocessed dataset DS2 for learning in the dataset acquisition process S331. The control unit 201 acquires the preprocessed dataset DS2 from the auxiliary storage device 202b based on the operator's input to the input unit 203. Alternatively, the control unit 201 may automatically acquire the dataset DS2 stored in the auxiliary storage device 202b based on predetermined rules. After the dataset acquisition process S331, the control unit 201 executes the connection relationship setting process S332.

[0119] Figure 11 It is a conceptual representation of a network N. For example... Figure 11As shown, network N uses the explanatory variable x, intermediate variable f, and target variable y as nodes, with each node connected by directed edges representing causal relationships. Network N is a graph representing the data generation process of the target variable y. Here, the explanatory variable x corresponds to the setpoint X of the control parameter, the intermediate variable f corresponds to the preprocessed measurement data Z2, and the target variable y corresponds to the quality data Y. To implement Bayesian optimization, in network N, the nodes for each intermediate variable f and the node for the target variable y are designed to output the predicted distribution of each variable (mean (predicted value) and variance (indicator of uncertainty)).

[0120] return Figure 10 In the connection relationship setting process S332, the control unit 201 sets the connection relationships between each node. Specifically, the control unit 201 receives the operator's specification regarding the parent nodes of each intermediate variable f and the target variable y via a GUI (Graphical User Interface) displayed on the display 204. Then, based on this specification, the control unit 201 sets the parent nodes of each intermediate variable f and the target variable y. The GUI is an input interface displayed on the display 204, allowing the operator to visually select or specify the parent nodes of the intermediate variable f and the target variable y. By setting the connection relationships between nodes, the control unit 201 constructs a connection relationship such as... Figure 10 The network N is shown. Furthermore, it is possible to specify an intermediate variable as an input to other intermediate variables. However, it is not possible to specify variables in a loop.

[0121] Figure 12A , Figure 12B as well as Figure 12C This is a diagram representing the GUI used to specify the connection relationships of network N. In this example, a list of intermediate variables f (f0~f3) and target variable y is arranged horizontally on display 204. Furthermore, a drop-down menu for selecting the parent node is positioned directly below each variable.

[0122] The operator can view a list of selectable parent nodes by opening a drop-down menu. For example, when the operator opens the drop-down menu directly below the intermediate variable f3, a list of options for selectable descriptive variables x (x0~x3) and intermediate variables f (f0~f2) is displayed, allowing the operator to select the appropriate parent node one by one. When a parent node is selected, the control unit 201 can display its network structure on the display 204. Figure 11 This allows the operator to visually confirm the network structure.

[0123] like Figure 12B As shown, the drop-down menu for the target variable y includes an option for "All Variables". Figure 12CAs shown, by selecting this option, the operator can select all optional explanatory variables x (x0~x3) and intermediate variables f (f0~f3) as the parent node of the target variable y at once.

[0124] Furthermore, in the connection relationship setting process S332, the control unit 201 can automatically set all optional descriptive variables x and all optional intermediate variables f as the parent node of the target variable y. In this case, the GUI (drop-down menu) for selecting the parent node of the target variable y can be omitted.

[0125] The causal relationships between each explanatory variable x (the setpoint of the control parameter X) and each intermediate variable f (the pre-processed measurement data Z2) are mostly clear. For example, the causal relationship between the flow rate setpoint of a certain treatment fluid and the flow rate is clear. Therefore, the connection relationship between the explanatory variable x and the intermediate variable f can be specified with high precision. On the other hand, the causal relationship between the explanatory variable x and the target variable y (mass data Y) or the causal relationship between the intermediate variable f and the target variable y is usually unclear. For example, it is difficult to properly determine whether the information of the explanatory variable x is aggregated into the intermediate variable f, or whether it should be connected to the target variable y as a variable that also affects the target variable y. In this case, it is advisable to consider connecting all optional explanatory variables x and intermediate variables f to the target variable y.

[0126] Furthermore, in this embodiment, the configuration is such that a parent node can be specified for both the multiple declarative variables x and the multiple intermediate variables f. However, it is possible to specify only one of the multiple declarative variables x and the multiple intermediate variables f as the parent node.

[0127] Figure 13 This is a diagram illustrating a variant of the GUI used to specify the connectivity relationships of network N. Figure 13 The GUI shown is configured to accept the designation of child nodes for the explanatory variable x and the target variable y. Specifically, a list of explanatory variables x (x0~x3) and a list of intermediate variables f (f0~f3) are arranged vertically. Furthermore, a drop-down menu for selecting child nodes is provided next to each variable. The operator can view the list of variables that can be selected as child nodes by opening the drop-down menu. Additionally, as... Figure 13 As shown, checkboxes are positioned next to each variable in the drop-down menu. When the operator selects a checkbox, the control unit 201 sets the corresponding variable as a child node.

[0128] Furthermore, the GUI used to specify the connection relationships between nodes is not limited to Figures 12A-12C as well as Figure 13 The GUI shown is illustrated. For example, it can be constructed as follows: Figure 10A diagram illustrating the network structure is used to specify the connections. Specifically, the control unit 201 displays boxes representing nodes and icons on the display 204 for selecting the edges (arrows) used to connect these nodes. Each node is labeled with a variable name. Furthermore, to identify the descriptive variable x, intermediate variable f, and target variable y, each node can be displayed with different shapes or colors.

[0129] The operator configures nodes by dragging and dropping, and uses edge icons to select arrows that specify parent (or child) nodes. Furthermore, error messages and warning icons can be displayed in case of incorrect connections. This allows the operator to construct the appropriate network N.

[0130] return Figure 10 After the connection relationship setting process S332, the control unit 201 executes the first learning process S333. The first learning process S333 uses the preprocessed dataset DS2 to learn the input-output relationships of each intermediate variable f in the network N. That is, in the first learning process S333, for each intermediate variable f, the control unit 201 uses the setpoint X of the control parameters of the preprocessed dataset DS2, the preprocessed measurement data Z2, and a machine learning model to learn the relationship between the explanatory variable x connected to each intermediate variable f and other intermediate variables f. The machine learning model in the first learning process S333 uses a model capable of generating a predicted distribution (mean and variance) of the intermediate variable f. Specifically, Gaussian process regression, Bayesian linear regression, hierarchical Bayesian models, Bayesian deep learning, or NGBoost (Natural Gradient Boosting) can be used. Through the first learning process S333, the control unit 201 acquires the same number of learned machine learning models as the number of intermediate variables f.

[0131] After the first learning process S333, the control unit 201 executes the second learning process S334. In the second learning process S334, the control unit 201 uses the preprocessed dataset DS2 to learn the input-output relationship of the target variable y in the network N. That is, in the second learning process S334, the control unit 201 uses the preprocessed dataset DS2 and the machine learning model to learn the relationship between the explanatory variable x and the intermediate variable f connected to the target variable y and the target variable y. The machine learning model in the second learning process S334 uses a model capable of generating a predicted distribution (mean and variance) of the target variable y. Specifically, Gaussian process regression, Bayesian linear regression, hierarchical Bayesian models, Bayesian deep learning, or NGBoost (Natural Gradient Boosting) can be used.

[0132] Furthermore, in the second learning process S334, the control unit 201 applies an induced sparsity prior distribution as the prior distribution of the weight parameters for the explanatory variable x and the intermediate variable f. Induced sparsity refers to making unwanted inputs in the inputs to the target variable y close to zero.

[0133] Specifically, based on the high-dimensional Bayesian optimization method proposed by David Eriksson et al. ("High-dimensional Bayesian optimization with sparse axis-aligned subspaces", Proceedings of Machine Learning Research (2021): 493-503), this method uses a prior distribution that induces sparsity. In this method, a half-Cauchy distribution is used as the prior distribution of the hyperparameters in the kernel function of the Gaussian process, controlling the importance of each explanatory variable. The half-Cauchy distribution is a positively skewed distribution limited to a positive range. According to this property, the weights for low-importance inputs (explanatory variable x or intermediate variable f) or the weights for intermediate variable f tend to concentrate near 0, while large weights are assigned to important inputs with sufficient information. Thus, unwanted dimensions can be automatically suppressed, and the essential structure of the data can be captured.

[0134] By applying a prior distribution that induces sparsity, even when the data generation process of the target variable y is unclear or the input to the target variable y is high-dimensional, important explanatory variables x and intermediate variables f can be automatically identified from the input for learning. Especially within the scope of this implementation, it can reduce the computational cost of optimization and, compared to conventional Bayesian optimization methods, achieve high-precision predictions and optimal solutions with fewer trials.

[0135] Furthermore, in this embodiment, a semi-Cauchy distribution is used as the prior distribution for inducing sparsity, but it is not limited to this, and other appropriate prior distributions, including the Laplace distribution and sparse Gaussian processes, can also be applied.

[0136] Additionally, as a second learning process for inference in S334, NUTS (No-U-Turn Sampling) can be used. NUTS is an adaptive method of the Hamiltonian Monte Carlo method and is a standard method for inference of complex models. Alternatively, Maximum A posteriori probability estimation (MAP) can be used instead of NUTS.

[0137] The control unit 201 constructs a prediction model PM by combining several learned machine learning models obtained in the first learning process S333 and the second learning process S334. Then, the control unit 201... Figure 8 In the Bayesian optimization process S34 (search process) shown, the constructed prediction model PM is used to perform Bayesian optimization.

[0138] As described above, according to the control parameter adjustment method and control parameter adjustment apparatus of this embodiment, even when the causal relationship between the control parameters (as explanatory variables) or the measurement data (as intermediate variables) and the target variable (as the quality of substrate processing) is unclear, the important inputs of the target variable can be appropriately learned by using a prior distribution with induced sparsity as the prior distribution of the weight parameters. Therefore, even if the explanatory variable x and the intermediate variable f, which have low importance, are connected to the target variable in network N, the reduction in the prediction accuracy of network N can be avoided. In addition, the reduction in the optimization accuracy of the control parameters can be avoided, thereby improving the quality of substrate processing.

[0139] Especially in network N, when multiple explanatory variables x or intermediate variables f are connected to the target variable y, the input of the target variable y becomes high-dimensional, leading to a decrease in the prediction accuracy of the learned model and a decrease in the sampling efficiency in Bayesian optimization. Therefore, in this embodiment, by using a prior distribution that induces sparsity, even when the input of the target variable y is high-dimensional, particularly important inputs can be learned from that input. Therefore, in the connection relationship setting process S332, even if explanatory variables x or intermediate variables f are arbitrarily specified as parent nodes of the target variable y, or if all explanatory variables x or intermediate variables f are set, the decrease in the prediction accuracy of the learned network N can be avoided. In addition, the decrease in optimization accuracy caused by Bayesian optimization can be avoided. Therefore, the quality of substrate processing can be improved.

[0140] <2. Variations>

[0141] The embodiments have been described above, but the present invention is not limited to the above-described manner and various modifications are possible.

[0142] For example, the above implementation uses the preprocessed dataset DS2 for machine learning, thus reducing the likelihood of overlearning. Therefore, the prediction accuracy of the machine learning model is higher, improving the quality of substrate processing. However, machine learning can also be performed using the unprocessed raw data dataset DS1.

[0143] In the above embodiments, the substrate processing apparatus 100 performs etching or cleaning processes, but the substrate processing is not limited to etching and cleaning processes. For example, the substrate processing may also be brushing, photosensitive film coating, development, annealing, or painting.

[0144] The present invention has been described in detail, but the foregoing description is illustrative in all respects and the invention is not limited thereto. Numerous variations not illustrated are to be understood as examples that can be conceived without departing from the scope of the invention. The structures described in the above embodiments and variations can be appropriately combined or omitted as long as they do not contradict each other.

[0145] <3. Summary of Implementation Methods>

[0146] The first approach is a control parameter adjustment device that adjusts the set values ​​of the control parameters of a substrate processing apparatus that processes a substrate based on control parameters. The control parameter adjustment device includes a computational processing unit that calculates recommended values ​​for the control parameters. This computational processing unit is capable of performing: a connection relationship setting process, setting the control parameters as explanatory variables, setting measurement data related to physical quantities measured in the substrate processing apparatus as intermediate variables, setting quality data related to the quality of substrate processing based on the substrate processing apparatus as target variables, setting the explanatory variables, the intermediate variables, and the target variables as nodes, and setting connection relationships between the nodes; and a first learning process, using the explanatory parameters... The dataset of the specified variable and the intermediate variable is used to learn the input-output relationship of the intermediate variable in the network based on the connection relationship set by the connection relationship setting process; the second learning process uses the dataset of the specified variable, the intermediate variable and the target variable to learn the input-output relationship of the target variable in the network; and the search process performs Bayesian optimization using the prediction model, search range and acquisition function obtained by the first learning process and the second learning process to search for recommended values ​​of the control parameters that make the quality close to the target quality. In the second learning process, the computational processing unit uses the prior distribution of induced sparsity as the prior distribution of the weight parameters.

[0147] The second approach is that in the control parameter adjustment device of the first approach, the arithmetic processing unit sets all selectable explanatory variables, all selectable intermediate variables, or both as the parent node of the target variable during the connection relationship setting process.

[0148] The third approach involves the control parameter adjustment device of the first approach, in which the computational processing unit uses a semi-Cauchy distribution as the prior distribution in the second learning process.

[0149] The fourth method, in the control parameter adjustment device of the first method, further includes: an input unit that accepts input; and a display, wherein the arithmetic processing unit performs the following processing in the connection relationship setting process: displays a GUI for accepting input specifying the connection relationship between the node of the description variable and the node of the intermediate variable, input specifying the connection relationship between the nodes of the intermediate variable, or both, on the display, and sets the connection of each node according to the input accepted via the input unit.

[0150] The fifth method is in the control parameter adjustment device of the fourth method, in which the arithmetic processing unit performs the following processing in the connection relationship setting process: displaying a GUI for accepting inputs specifying the connection relationship between the node of the description variable and the node of the target variable, inputs specifying the connection relationship between the node of the intermediate variable and the node of the target variable, or both, on the display screen, and setting the connection of each node according to the inputs accepted by the input unit.

[0151] The sixth method is in the control parameter adjustment device of the fifth method, in which the arithmetic processing unit displays a GUI on the display screen. The GUI has options for selecting variables one by one from the selectable explanatory variables or the selectable intermediate variables as the parent node of the target variable, and options for selecting all the selectable explanatory variables or the selectable intermediate variables as the parent node of the target variable at once.

[0152] The seventh method is a control parameter adjustment method that adjusts the set values ​​of the control parameters of a substrate processing apparatus that processes a substrate based on control parameters. The method includes: a connection relationship setting process, where the control parameters are set as explanatory variables, measurement data related to physical quantities measured in the substrate processing apparatus are set as intermediate variables, quality data related to the quality of substrate processing based on the substrate processing apparatus are set as target variables, and the explanatory variables, intermediate variables, and target variables are set as nodes, and connection relationships are set between the nodes; a first learning process, using a dataset of the explanatory variables and intermediate variables, learning the input-output relationships of the intermediate variables in a network based on the connection relationships set by the connection relationship setting process; a second learning process, using a dataset of the explanatory variables, intermediate variables, and target variables, learning the input-output relationships of the target variables in the network; and a search process, performing Bayesian optimization using the prediction model, search range, and acquisition function obtained through the first and second learning processes to search for recommended values ​​of the control parameters that make the quality approach the target quality.

[0153] The eighth method is a computer program that the computer can execute, causing the computer to perform the various processes of the control parameter adjustment method of the seventh method.

[0154] According to methods one through eight, by using a prior distribution that induces sparsity, the values ​​of inputs important to the target variable are maintained at a large value, while the values ​​of unimportant inputs are reduced. This property allows for the appropriate learning of the important inputs to the target variable. Therefore, even when descriptive and intermediate variables with low importance are connected to the target variable in the network, a decrease in the network's prediction accuracy can be avoided. This, in turn, prevents a decrease in the optimization accuracy of the control parameters, thereby improving the quality of substrate processing.

[0155] According to the control parameter adjustment device of the third method, since the semi-Cauchy distribution is a positively skewed distribution limited to a positive range, the weights of variables with low importance tend to concentrate near 0, while the weights of important variables become larger. Therefore, the fitting efficiency can be improved.

[0156] According to the control parameter adjustment device of the fourth method, the parent node of the intermediate variable can be set via GUI based on the operator's specified input.

[0157] According to the control parameter adjustment device of the fifth method, the parent node of the target variable can be set based on the operator's specified input.

[0158] According to the control parameter adjustment device of the sixth method, the operator can specify one by one from the explanatory variables or intermediate variables, or can specify all the explanatory variables or intermediate variables at once.

[0159] Explanation of reference numerals in the attached figures

[0160] 100: Substrate processing apparatus

[0161] 203: Input Department

[0162] 204: Monitor

[0163] AF: Get Function

[0164] DS2: Preprocessed dataset

[0165] N: Network

[0166] PM: Predictive Model

[0167] PR: Adjusting procedures

[0168] RX: Recommended value

[0169] SW: Search range

[0170] W: substrate.

Claims

1. A control parameter adjustment device for adjusting the set value of the control parameters of a substrate processing apparatus that processes a substrate based on control parameters, characterized in that, The control parameter adjustment device includes a calculation processing unit for calculating recommended values ​​for the control parameters. The processing unit is capable of performing: The connection relationship setting process involves setting the control parameters as explanatory variables, setting measurement data related to physical quantities measured in the substrate processing apparatus as intermediate variables, setting quality data related to the quality of substrate processing based on the substrate processing apparatus as target variables, setting the explanatory variables, the intermediate variables, and the target variables as nodes, and setting the connection relationship between the nodes. The first learning process uses the dataset of the descriptive variables and the intermediate variables to learn the input-output relationship of the intermediate variables in the network based on the connection relationship set by the connection relationship setting process; The second learning process involves using the dataset of the explanatory variables, the intermediate variables, and the target variable to learn the input-output relationship of the target variable in the network. as well as The search process performs Bayesian optimization using the prediction model, search range, and acquisition function obtained through the first and second learning processes to search for recommended values ​​of the control parameters that make the quality approach the target quality. In the second learning process, the computational processing unit uses the prior distribution of induced sparsity as the prior distribution of the weight parameters.

2. The control parameter adjustment device according to claim 1, characterized in that, In the connection relationship setting process, the computation processing unit sets all selectable explanatory variables, all selectable intermediate variables, or both of these as the parent node of the target variable.

3. The control parameter adjustment device according to claim 1 or 2, characterized in that, The computational processing unit uses a semi-Cauchy distribution as the prior distribution in the second learning process.

4. The control parameter adjustment device according to claim 1, characterized in that, The control parameter adjustment device also includes: The input section, which accepts input; and monitor, The arithmetic processing unit performs the following processing in the connection relationship setting process: displays a GUI for accepting the connection relationship between the node specifying the descriptive variable and the node of the intermediate variable, the input specifying the connection relationship between the nodes of the intermediate variable, or both, on the display, and sets the connection of each node according to the input accepted via the input unit.

5. The control parameter adjustment device according to claim 4, characterized in that, The arithmetic processing unit performs the following processing in the connection relationship setting process: displays a GUI for accepting inputs specifying the connection relationship between the node of the description variable and the node of the target variable, inputs specifying the connection relationship between the node of the intermediate variable and the node of the target variable, or both, on the display screen, and sets the connection of each node according to the inputs accepted by the input unit.

6. The control parameter adjustment device according to claim 5, characterized in that, The processing unit displays a GUI on the display screen. The GUI has options for selecting variables one by one from the selectable explanatory variables or the selectable intermediate variables as the parent node of the target variable, and for selecting all the selectable explanatory variables or the selectable intermediate variables as the parent node of the target variable at once.

7. A method for adjusting control parameters, wherein the set value of the control parameters of a substrate processing apparatus that processes a substrate based on control parameters is adjusted, characterized in that, The control parameter adjustment method includes: The connection relationship setting process involves setting the control parameters as explanatory variables, setting measurement data related to physical quantities measured in the substrate processing apparatus as intermediate variables, setting quality data related to the quality of substrate processing based on the substrate processing apparatus as target variables, setting the explanatory variables, the intermediate variables, and the target variables as nodes, and setting the connection relationship between the nodes. The first learning process uses the dataset of the descriptive variables and the intermediate variables to learn the input-output relationship of the intermediate variables in the network based on the connection relationship set by the connection relationship setting process; The second learning process uses the dataset containing the explanatory variables, the intermediate variables, and the target variable to learn the input-output relationship of the target variable in the network; and The search process performs Bayesian optimization using the prediction model, search range, and acquisition function obtained through the first and second learning processes to search for recommended values ​​of the control parameters that make the quality approach the target quality.

8. A storage medium storing a computer program executable by a computer, characterized in that, The computer program causes the computer to perform each process of the control parameter adjustment method according to claim 7.

Images