Substrate processing apparatus and substrate processing method

Abstract

The present invention provides a substrate processing apparatus and a substrate processing method that enable maintenance of resin components at the appropriate time. [Solution] The processor, by executing a program, predicts the weight change of a resin component using a learning model 600 based on the processing conditions when substrate processing is performed using a processing liquid in a chamber containing a resin component that comes into contact with the processing liquid. The learning model 600 is an inference model that has learned an input-output relationship in which processing conditions are input and main ground truth data and sub-ground truth data are output. The main ground truth data shows the weight change of the resin component after substrate processing is performed in the chamber under the processing conditions. The sub-ground truth data shows physical properties related to the substrate processing.

Description

[Technical Field]

[0001] The present invention relates to a substrate processing apparatus and a substrate processing method. [Background Art]

[0002] A substrate processing apparatus is used for processing substrates such as semiconductor wafers, glass substrates for liquid crystal display devices, glass substrates for photomasks, and glass substrates for optical disks with a processing liquid such as a developer, a cleaning liquid, a rinsing liquid, or a photoresist liquid. For example, in the cleaning apparatus described in Patent Document 1, a substrate is horizontally held by a spin chuck. A cup is disposed so as to surround the lateral side and lower side of the substrate held by the spin chuck. A cleaning liquid supply nozzle is disposed above the substrate held by the spin chuck.

[0003] While the substrate is held and rotated by the spin chuck, the cleaning liquid is supplied from the cleaning liquid supply nozzle to the central portion of the surface of the substrate. In this case, the resist liquid at the central portion of the surface is spread over the entire surface by the centrifugal force generated along with the rotation of the substrate, thereby cleaning the surface of the substrate. Further, the cleaning liquid scattered from the surface of the substrate to the periphery is caught by the cup and then recovered into the drainage part. [Prior Art Literature] [Patent Literature]

[0004] [Patent Document 1] Japanese Patent No. 4347785 [Summary of the Invention] [Problem to be Solved by the Invention]

[0005] In substrate processing equipment, various components such as cups, nozzles, and piping are made of resin. These resin components gradually corrode or deteriorate upon contact with the processing liquid. Therefore, maintenance, such as replacing the resin components or replacing the substrate processing equipment, is necessary before the corrosion or deterioration of the resin components becomes severe. However, the rate at which corrosion or deterioration of resin components progresses varies depending on the operating period of the substrate processing equipment and the content of the substrate processing, making it difficult to maintain the substrate processing equipment at the appropriate time.

[0006] The object of the present invention is to provide a substrate processing apparatus and a substrate processing method that can maintain resin components at an appropriate timing. [Means for solving the problem]

[0007] A substrate processing apparatus according to one aspect of the present invention includes a processor, which, by executing a program, predicts the weight change of a resin member in a chamber containing a processing liquid, based on processing conditions when substrate processing is performed using the processing liquid, using a learning model. The learning model is an inference model that has learned an input-output relationship, with the processing conditions as input and main correct data and sub-correct data as output. The main correct data shows the weight change of the resin member after substrate processing is performed in the chamber under the processing conditions, and the sub-correct data shows physical properties related to the substrate processing.

[0008] A substrate processing method according to another aspect of the present invention is a substrate processing method performed by a processor in a substrate processing apparatus, wherein the processor, by executing a program, predicts the weight change of a resin member using a learning model based on processing conditions when substrate processing using the processing liquid is performed in a chamber containing a resin member in contact with the processing liquid, the learning model is an inference model that has learned an input-output relationship with the processing conditions as input and main correct data and sub-correct data as output, the main correct data shows the weight change of the resin member after substrate processing is performed in the chamber under the processing conditions, and the sub-correct data shows physical property values ​​related to the substrate processing. [Effects of the Invention]

[0009] According to the present invention, resin components can be maintained at the appropriate time. [Brief explanation of the drawing]

[0010] [Figure 1] This figure shows an example of the configuration of a substrate processing apparatus according to one embodiment of the present invention. [Figure 2] This figure shows an example of the configuration of an information processing device. [Figure 3] This is a schematic diagram showing the inside of the processing unit viewed horizontally. [Figure 4] This figure shows an example of the structure of a learning model. [Figure 5] Figure 1 is a block diagram showing the functional configuration of the learning device. [Figure 6] This figure shows an example of a dataset acquired by the data acquisition unit. [Figure 7] This figure illustrates the main ground truth data in the dataset shown in Figure 6. [Figure 8] Figure 2 is a block diagram showing the functional configuration of the information processing device. [Figure 9] This is a flowchart illustrating an example of the learning process flow. [Figure 10] This is a flowchart showing an example of the lifespan evaluation process. [Figure 11] It is a diagram showing an example structure of a learning model in a first modification. [Figure 12] It is a diagram showing an example structure of a learning model in a second modification. [Figure 13] It is a diagram showing an example structure of a learning model in another embodiment. MODE FOR CARRYING OUT THE INVENTION

[0011] 1. Substrate Processing Apparatus Hereinafter, a substrate processing apparatus and a substrate processing method according to embodiments of the present invention will be described with reference to the drawings. In the following description, a substrate refers to a semiconductor substrate (wafer), a substrate for FPD (Flat Panel Display) such as a liquid crystal display device or an organic EL (Electro Luminescence) display device, an optical disc substrate, a magnetic disc substrate, a magneto-optical disc substrate, a photomask substrate, a ceramic substrate, a solar cell substrate, or the like. FIG. 1 is a diagram showing an example configuration of a substrate processing apparatus according to an embodiment of the present invention. As shown in FIG. 1, the substrate processing apparatus 500 includes a processing unit 100, a database storage device 200, an information processing device 300, and a learning device 400.

[0012] The processing unit 100, the database storage device 200, the information processing device 300, and the learning device 400 are connected to a network 501 and can transmit and receive data to and from each other. The network 501 may be, for example, a local area network (LAN) or a wide area network (WAN). Alternatively, the network 501 may be the Internet. The connection form of the network 501 may be wired or wireless. Further, the processing unit 100 and the information processing device 300 may be connected via a dedicated network instead of the network 501.

[0013] The processing unit 100 includes a plurality of processing units 110 and a transfer mechanism 120. The transfer mechanism 120 sequentially transfers a substrate to be processed to the plurality of processing units 110 in accordance with predetermined transfer conditions. Each processing unit 110 is a chamber including various resin members that come into contact with a processing liquid. The structure of the processing unit 110 will be described later. Each processing unit 110 sequentially performs a series of processes using the processing liquid on a plurality of substrates in accordance with a processing recipe that describes processing details for the substrate. In this example, each processing unit 110 forms a solid film or a liquid film on the substrate by supplying the processing liquid to the substrate under predetermined processing conditions, and then removes the solid film or the liquid film from the substrate and dries the substrate.

[0014] Specifically, each processing unit 110 applies a processing liquid, which is a mixed solution of a predetermined concentration obtained by dissolving a sublimation agent in a solvent, onto the surface of a substrate on which a pattern is formed, and then evaporates the solvent while rotating the substrate to thereby precipitate the sublimation agent on the substrate, performing the series of processes. Thereafter, each processing unit 110 sprays an inert gas onto the substrate to perform sublimation drying. However, the substrate processing in each processing unit 110 is not limited to sublimation drying of the substrate, and may be, for example, development, cleaning, or coating film formation.

[0015] The database storage device 200 includes a large-capacity storage device such as a server. Molecular descriptors of various resin materials and various solutions are stored in the database storage device 200. In this example, the molecular descriptors of the resin material and the solution are molecular descriptors included in force field descriptors or the mass of the resin material. A force field descriptor is a force field parameter used in MD (molecular dynamics) calculations that describe the behavior of molecules in a resin material.

[0016] The information processing device 300 is comprised of, for example, a general-purpose computer. Figure 2 shows an example of the configuration of the information processing device 300. As shown in Figure 2, the information processing device 300 includes a CPU (Central Processing Unit) 310, RAM (Random Access Memory) 320, ROM (Read-Only Memory) 330, storage device 340, operation unit 350, display device 360, input / output interface 370, and bus 380. The CPU 310, RAM 320, ROM 330, storage device 340, operation unit 350, display device 360, and input / output interface 370 are connected to the bus 380.

[0017] RAM320 is used as a workspace for CPU310. ROM330 stores the system program. The storage device 340 includes a storage medium such as a hard disk or semiconductor memory and stores a life evaluation program for executing the life evaluation process described later. The life evaluation program may be stored in ROM330 or other external storage devices. The storage device 340 may also store processing recipes or transport conditions. Furthermore, the storage device 340 may store history information indicating the history of substrate processing in the processing unit 100.

[0018] The operation unit 350 is an input device such as a keyboard, mouse, or touch panel. The user can give predetermined instructions to the information processing device 300 by operating the operation unit 350. The display device 360 ​​is a display device such as a liquid crystal display device, and displays a GUI (Graphical User Interface) for receiving instructions from the user or processing results from life evaluation processing, etc. The input / output I / F 370 is connected to the network 501 in Figure 1.

[0019] The learning device 400 is configured, for example, by a general-purpose computer. Therefore, the learning device 400 has basically the same configuration as the information processing device 300, except for the following points, so a detailed explanation of the configuration of the learning device 400 will be omitted. The storage device of the learning device 400 stores a learning program for executing the learning process described later. The learning program may be stored in the ROM of the learning device 400 or in other external storage devices. Note that the information processing device 300 and the learning device 400 may be configured with the same hardware. In this case, the programs installed in the information processing device 300 and the learning device 400 will be different. Therefore, the processes that the information processing device 300 and the learning device 400 execute are different from each other.

[0020] 2. Processing Unit The configuration of the processing unit 110 of the processing unit 100 in Figure 1 will be described below. Figure 3 is a schematic diagram of the inside of the processing unit 110 viewed horizontally. Figure 3 shows the configuration of one processing unit 110, but the configuration of other processing units 110 is the same as in Figure 3. As shown in Figure 3, the processing unit 110 includes a processing mechanism 1 and a control device 2. The control device 2 can communicate with the information processing device 300 in Figure 2 and controls the processing mechanism 1 based on commands from the information processing device 300. The processing mechanism 1 includes a box-shaped chamber 3, a spin chuck 10 that holds a single substrate W horizontally within the chamber 3 and rotates it around a vertical rotation axis A1 passing through the center of the substrate W, and a cylindrical processing cup 21 that surrounds the spin chuck 10 around the rotation axis A1.

[0021] The spin chuck 10 includes a disc-shaped spin base 12 held in a horizontal position, a plurality of chuck pins 11 that hold the substrate W in a horizontal position above the spin base 12, a spin shaft 13 extending downward from the center of the spin base 12, and a spin motor 14 that rotates the spin base 12 and the plurality of chuck pins 11 by rotating the spin shaft 13.

[0022] The processing cup 21 includes a plurality of guards 24 that receive the processing liquid discharged outward from the substrate W, a plurality of cups 23 that receive the processing liquid guided downward by the plurality of guards 24, and a cylindrical outer wall member 22 that surrounds the plurality of guards 24 and the plurality of cups 23. The processing cup 21 is an example of a resin material. The plurality of guards 24 can be individually raised and lowered by a guard lifting unit 27.

[0023] The processing mechanism 1 includes a chemical nozzle 31 for discharging a chemical solution, a rinse nozzle 35 for discharging a rinse solution, a processing nozzle 39 for discharging a processing solution, and a replacement liquid nozzle 43 for discharging a replacement liquid. The chemical nozzle 31, rinse nozzle 35, processing nozzle 39, and replacement liquid nozzle 43 are each independently movable horizontally within the chamber 3 by nozzle moving units 34, 38, 42, and 46 provided in conjunction with each other.

[0024] The processing liquid nozzle 39 is connected to a processing liquid piping 40 that guides the processing liquid to the processing liquid nozzle 39. The processing liquid nozzle 39 and the processing liquid piping 40 are other examples of resin components. When the processing liquid valve 41 interposed in the processing liquid piping 40 is opened, the processing liquid is continuously discharged downward from the discharge port of the processing liquid nozzle 39. The processing liquid is a mixture containing a sublimation agent and a solvent that dissolves in the sublimation agent.

[0025] The processing liquid nozzle 39 is connected to a nozzle moving unit 42. The nozzle moving unit 42 moves the processing liquid nozzle 39 in at least one of the vertical and horizontal directions. The nozzle moving unit 42 moves the processing liquid nozzle 39 horizontally between a processing position where the processing liquid discharged from the processing liquid nozzle 39 is supplied to the upper surface of the substrate W and a standby position where the processing liquid nozzle 39 is located around the processing cup 21 in a plan view.

[0026] The processing mechanism 1 includes a disc-shaped blocking member 51 positioned above the spin chuck 10. The blocking member 51 includes a disc portion 52 positioned horizontally above the spin chuck 10. The blocking member 51 is horizontally supported by a cylindrical support shaft 53 extending upward from the center of the disc portion 52. The center line of the disc portion 52 is positioned on the rotation axis A1 of the substrate W. The lower surface of the disc portion 52 corresponds to the lower surface 51L of the blocking member 51. The lower surface 51L of the blocking member 51 is parallel to the upper surface of the substrate W and has an outer diameter greater than or equal to the diameter of the substrate W.

[0027] The blocking member 51 is connected to a blocking member lifting unit 54 that moves the blocking member 51 vertically up and down. The blocking member lifting unit 54 moves the blocking member 51 to any position from the upper position (the position shown in Figure 3) to the lower position.

[0028] A central nozzle 55 is positioned within a through-hole that penetrates the central part of the shut-off member 51 vertically. The central nozzle 55 moves up and down together with the shut-off member 51. The central nozzle 55 is connected to an upper gas pipe 56 that guides inert gas to the central nozzle 55. The processing unit 110 includes an upper temperature controller 59 that heats or cools the inert gas discharged from the central nozzle 55. When the upper gas valve 57 interposed in the upper gas pipe 56 is opened, the inert gas is continuously discharged downward from the outlet of the central nozzle 55 at a flow rate corresponding to the opening of a flow rate adjustment valve 58 that changes the flow rate of the inert gas. The inert gas discharged from the central nozzle 55 is nitrogen gas.

[0029] 3. Learning Model The information processing device 300 in Figure 1 uses a learning model to predict the lifespan of the resin component of the processing unit 100 based on the processing conditions of the substrate. The learning model performs a main task and one or more subtasks. Specifically, the main task outputs the predicted result of the weight change of the resin component. Each subtask outputs the predicted result of predetermined physical properties related to the substrate processing.

[0030] The predetermined physical properties related to substrate processing include physical properties related to the resin component, physical properties related to the processing solution, and physical properties related to the interaction between the resin component and the processing solution. The physical properties related to the resin component include, for example, electron affinity, tensile strength at fracture, band gap, or solubility parameter. The physical properties related to the processing solution include, for example, solubility parameter, boiling point, or density. The physical properties related to the interaction between the resin component and the processing solution include calculated values ​​of the χ parameter, measured values ​​of the χ parameter, penetration depth, contact angle, adsorption energy, or solvation free energy.

[0031] In this embodiment, the learning model includes multiple neural networks (NNs). Of these NNs, some operate as lower-level NNs, and others operate as higher-level NNs. The lower-level NNs are input with the processing conditions of the substrate. The lower-level NNs output intermediate prediction results. The physical properties of the intermediate prediction results may be unknown. The higher-level NNs are input with the intermediate prediction results output from the lower-level NNs. Some of the substrate processing conditions may be input to the higher-level NNs. The higher-level NNs output prediction results corresponding to the main task or one or more subtasks.

[0032] Figure 4 shows an example of the structure of a learning model. As shown in Figure 4, the learning model 600 includes two lower NNs 601: a resin NN 610 and a processing liquid NN 620. The learning model 600 also includes three upper NNs 602: a resin prediction NN 630, a processing liquid prediction NN 640, and an interaction prediction NN 650. In other words, the learning model 600 includes five NNs. Each NN contains multiple layers. In this example, the number of layers in the lower NNs 601 is greater than the number of layers in the upper NNs 602.

[0033] Specifically, the resin NN610 includes four layers 611, 612, 613, and 614. Layer 611 is the input layer, layers 612 and 613 are hidden layers (intermediate layers), and layer 614 is the output layer. Layers 611 and 612, 612 and 613, and 613 and 614 are connected by multiple paths. Layer 611 receives resin descriptors as input information, which are part of the substrate processing conditions and represent the molecular descriptors of the resin component in the processing unit 110 shown in Figure 3. The information input to layer 611 is passed to layer 614 through layers 612 and 613. Based on the given information, layer 614 outputs intermediate prediction results related to the resin component.

[0034] The processing solution NN620 has the same configuration as the resin NN610 and includes four layers 621, 622, 623, and 624. Layer 621 receives information as processing solution descriptors, which indicate the molecular descriptors of the processing solution used for substrate processing among the substrate processing conditions. The information input to layer 621 is passed to layer 624 through layers 622 and 623. Based on the given information, layer 624 outputs intermediate prediction results related to the processing solution.

[0035] The resin prediction NN630 includes three layers 631, 632, and 633. Layer 631 is the input layer, layer 632 is the hidden layer, and layer 633 is the output layer. Layers 631 and 632, and layers 632 and 633 are connected by multiple paths. Intermediate prediction results related to resin members output by the resin NN610 are input as information to layer 631. The information input to layer 631 is passed to layer 633 through layer 632. Based on the given information, layer 633 outputs material property values ​​corresponding to one or more subtasks related to the resin member as prediction results.

[0036] The processing liquid prediction NN640 has the same configuration as the resin prediction NN630 and includes three layers 641, 642, and 643. Layer 641 receives intermediate prediction results related to the processing liquid output by the processing liquid NN620 as input information. The information input to layer 641 is passed to layer 643 through layer 642. Based on the given information, layer 643 outputs physical property values ​​corresponding to one or more subtasks related to the processing liquid as prediction results.

[0037] The interaction prediction NN650 has the same configuration as the resin prediction NN630 and includes three layers 651, 652, and 653. Layer 651 receives information such as intermediate prediction results related to the resin component output by the resin NN610 and intermediate prediction results related to the processing liquid output by the processing liquid NN620. Layer 651 also receives information such as predetermined setting conditions among the substrate processing conditions. The setting conditions include the rotation speed of the substrate during substrate processing, or the time the resin component is in contact with the processing liquid (contact time).

[0038] Information input to layer 651 is passed to layer 653 through layer 652. Based on the given information, layer 653 outputs a prediction result corresponding to the main task. Layer 653 also outputs prediction results for zero or more subtasks related to the interaction between the resin component and the processing liquid, based on the given information. In Figure 4, the prediction result corresponding to the main task is indicated by "○", and the prediction result corresponding to the subtask is indicated by "□". Based on the prediction result corresponding to the main task, the information processing device 300 in Figure 1 predicts the lifespan of the resin component. Details will be described later.

[0039] 4. Learning device Figure 5 is a block diagram showing the functional configuration of the learning device 400 in Figure 1. As shown in Figure 5, the learning device 400 includes a data acquisition unit 401, a descriptor acquisition unit 402, and a model generation unit 403 as functional units. The functional units of the learning device 400 are realized when the CPU of the learning device 400 executes a learning program. Some or all of the functional units of the learning device 400 may be realized by hardware such as electronic circuits.

[0040] The data acquisition unit 401 acquires multiple pre-generated datasets. Each dataset is generated by a user of the processing unit 100 through preliminary experiments. The user may input the generated datasets into the data acquisition unit 401 using an operation unit (not shown). Alternatively, the user may store the multiple generated datasets in a storage device (not shown) of the learning device 400. In this case, the multiple datasets are acquired from the storage device.

[0041] Figure 6 shows an example of a dataset acquired by the data acquisition unit 401. As shown in Figure 6, each dataset includes a set of substrate processing conditions and corresponding ground truth data. The substrate processing conditions include, for example, the resin component of the processing unit 110 in Figure 3, the processing liquid, the temperature of the processing liquid, the concentration of the processing liquid, and the setting conditions described above. The ground truth data is the prediction result that should be output when the corresponding processing conditions are input to the learning model 600 in Figure 4. The ground truth data includes main ground truth data, which is the prediction result that should be output as the main task, and sub-ground truth data, which is the prediction result that should be output as a subtask.

[0042] Each dataset does not need to contain all the ground truth data corresponding to the main task and one or more subtasks output by the learning model 600. Each dataset only needs to contain at least one ground truth data corresponding to the main task and one or more subtasks output by the learning model 600. In the example in Figure 6, the cells marked with "-" do not contain ground truth data. Therefore, "Dataset 1," "Dataset 4," and "Dataset 5" contain the main ground truth data, while the other datasets do not.

[0043] Figure 7 is a diagram illustrating the main ground truth data in the dataset shown in Figure 6. As shown in Figure 7, the main ground truth data represents the weight change of the resin component. Specifically, the weight change of the resin component corresponding to each processing condition is the initial weight of the resin component minus the weight of the resin component at the time the processing liquid under that processing condition has been in contact with the resin component for a predetermined period of time. In this example, the weight of the resin component is normalized so that the initial weight is 1 (100%). Therefore, the weight of the resin component in this example is a relative value based on the initial weight and is treated as a dimensionless quantity without units.

[0044] In the example shown in Figure 7, the dataset includes the weight change of the resin component when the processing liquid is in contact with the resin component for 1 day, 1 week, and 2 weeks (hereinafter referred to as the liquid contact time). The weight of the resin component decreases due to corrosion or degradation as the liquid contact time increases. Therefore, the weight change of the resin component increases as the liquid contact time increases.

[0045] In preliminary experiments for dataset generation, a resin component with a known initial weight and made of the same material as the resin component of the processing unit 110 is immersed in the same processing solution as the processing conditions. Subsequently, the weight of the resin component is measured when the immersion time (contact time) of the resin component reaches a predetermined time. This determines the weight change of the resin component for each contact time as the main ground truth data corresponding to that processing condition. A dataset is generated consisting of each processing condition and the weight change of the resin component for each contact time determined under that processing condition.

[0046] On the other hand, sub-ground data represents physical properties related to substrate processing and is obtained from literature or databases. Alternatively, sub-ground data can be obtained by performing calculations using values ​​published in literature or databases. For example, if the sub-ground data is the calculated value of the χ parameter, it can be obtained by performing a simulation based on the COSMO (Conductor-like Screening Model) method using values ​​published in literature or databases. Therefore, sub-ground data can be obtained more easily and in larger quantities than main ground data.

[0047] The descriptor acquisition unit 402 acquires molecular descriptors of resin materials used as resin components in the dataset from the molecular descriptors of various resin materials stored in the database storage device 200, and assigns them as resin descriptors. In this case, the data acquisition unit 401 associates the resin descriptors acquired by the descriptor acquisition unit 402 with the resin components in the dataset. Similarly, the descriptor acquisition unit 402 acquires molecular descriptors of solutions used as processing solutions in the dataset from the molecular descriptors of various solutions stored in the database storage device 200, and assigns them as processing solution descriptors. In this case, the data acquisition unit 401 associates the processing solution descriptors acquired by the descriptor acquisition unit 402 with the processing solutions in the dataset.

[0048] The model generation unit 403 uses the substrate processing conditions and liquid contact time in each dataset acquired by the data acquisition unit 401 as explanatory variables, and the main ground truth data and one or more sub-ground truth data as objective variables to perform multi-task learning on a predetermined inference model. In this example, the inference model has the same structure as the learning model 600 in Figure 4. That is, the inference model includes a resin NN, a processing liquid NN, a resin prediction NN, a processing liquid prediction NN, and an interaction prediction NN.

[0049] In each dataset, the resin descriptor is input to the input layer of the resin NN, the processing solution descriptor is input to the input layer of the processing solution NN, and the setting conditions are input to the input layer of the interaction prediction NN. As a result, intermediate prediction results related to the resin component are output from the output layer of the resin NN. Intermediate prediction results related to the processing solution are also output from the output layer of the processing solution NN.

[0050] Intermediate prediction results output from the resin NN are input to the respective input layers of the resin prediction NN and the interaction prediction NN. Intermediate prediction results output from the processing solution NN are input to the respective input layers of the processing solution prediction NN and the interaction prediction NN. As a result, prediction results corresponding to one or more subtasks are output from the respective output layers of the processing solution prediction NN and the interaction prediction NN. In addition, prediction results corresponding to the main task and prediction results corresponding to one or more subtasks are output from the output layer of the interaction prediction NN.

[0051] Here, the weights of the paths between layers through which information has passed in each neural network are adjusted so that the prediction results corresponding to the main task approach the main ground truth data of each dataset, and the prediction results corresponding to the subtask approach the sub-ground truth data of each dataset. For example, when generating a learning model 600 to predict electron affinity as a subtask, the weights of the paths between layers are adjusted so that the loss function L given by equation (1) below is minimized.

[0052]

number

[0053] In equation (1), a M This is the learning weight coefficient for the main task. M w0 is the number of datasets containing the main ground truth data. i This is the value of the weight change of the resin component in the i-th dataset (main ground truth data). S This is the learning weight coefficient for the subtask. Se is the number of datasets containing sub-ground truth data for electron affinity. e0 is the predicted value of electron affinity. j This represents the electron affinity value (sub-ground truth data) in the j-th dataset.

[0054] The user performs a predetermined operation using the control unit of the learning device 400 shown in Figure 5, thereby obtaining equation (1) a M ,a S This can be set to any value. This adjusts the balance between learning for the main task and learning for subtasks in multitask learning. By performing the above multitask learning using multiple datasets, the learning model 600 shown in Figure 4 is generated, which predicts the weight change of the resin component as the main task and various physical properties for substrate processing as a subtask, based on the substrate processing conditions and wetting time. The generated learning model 600 is stored in the storage device 340 of the information processing device 300 shown in Figure 2 and is then mounted on the information processing device 300.

[0055] 5. Information Processing Device Figure 8 is a block diagram showing the functional configuration of the information processing device 300 shown in Figure 2. As shown in Figure 8, the information processing device 300 includes a condition determination unit 301, a weight prediction unit 302, a calculation unit 303, a life evaluation unit 304, a presentation unit 305, and a processing control unit 306 as functional units. The functional units of the information processing device 300 are realized when the CPU 310 in Figure 2 executes a life evaluation program. Some or all of the functional units of the information processing device 300 may be realized by hardware such as electronic circuits.

[0056] The condition determination unit 301 determines the processing conditions and wetting time for the substrate. The processing conditions for the substrate include a processing solution descriptor, a resin descriptor, the temperature of the processing solution, and the concentration of the processing solution. The temperature and concentration of the processing solution are described in the processing recipe. Therefore, the temperature or concentration of the processing solution may be determined based on the processing recipe.

[0057] The processing solution descriptor is obtained from molecular descriptors of various processing solutions stored in the database storage device 200, based on information about the processing solution (type of processing solution, temperature, or concentration). Alternatively, the processing solution descriptor is determined by calculation by the information processing device 300, based on information about the processing solution (type of processing solution, temperature, or concentration). Similarly, the resin descriptor is obtained from molecular descriptors of various resin components stored in the database storage device 200, based on information about the resin component. Alternatively, the resin descriptor is determined by calculation by the information processing device 300, based on information about the resin component.

[0058] The wetting time is proportional to the time the processing solution is supplied to the substrate. The supply time of the processing solution to the substrate is described in the processing recipe. Therefore, the wetting time can also be determined based on the processing recipe. For example, the wetting time may be determined by multiplying the supply time of the processing solution by a predetermined proportionality coefficient. Alternatively, the user can instruct the condition determination unit 301 on the processing conditions and wetting time of the substrate by operating the operation unit 350. Therefore, some or all of the processing conditions and wetting time of the substrate may be determined based on the user's instructions.

[0059] The weight prediction unit 302 uses the learning model 600 shown in Figure 4, mounted on the information processing device 300, to predict the weight change of the resin member after a series of substrate processing steps, based on the processing conditions and wetting time determined by the condition determination unit 301, as the main task. Furthermore, if the processing recipe to be executed in the processing unit 110 shown in Figure 2 is changed, the weight prediction unit 302 updates the predicted weight change based on the changed processing conditions or wetting time. In addition, the weight prediction unit 302 may calculate the remaining weight of the resin member after a series of substrate processing steps. In this example, similar to the weight change, the remaining weight is also treated as a relative value based on the initial weight of the resin member. In this case, the remaining weight is calculated by subtracting the predicted weight change from 1 (100%).

[0060] The calculation unit 303 calculates reference information regarding the weight change prediction of the resin member by the weight prediction unit 302. The reference information includes the contribution rate and prediction accuracy of each processing condition. As a method for calculating the contribution rate, SHAP (SHapley Additive exPlanations) values ​​or the importance of the explanatory variables in the extra tree may be calculated.

[0061] The life evaluation unit 304 evaluates the life of the resin component in relation to the processing liquid based on the weight change of the resin component predicted by the weight prediction unit 302. In this example, the life is evaluated as the point at which the cumulative weight change predicted by the weight prediction unit 302 reaches a preset threshold. For example, the threshold is set to 0.1. That is, the life of the resin component is defined as the point at which the remaining weight of the resin component reaches 0.9 (90%). In this example, if the weight change is predicted to be 0.01 for a liquid contact time of 100 hours, the life is evaluated as 100 hours × 0.1 / 0.01 = 1000 hours. The evaluated life serves as an indicator of when the resin component should be replaced.

[0062] The display unit 305 presents the user with various information, such as the weight change of the resin component predicted by the weight prediction unit 302, reference information calculated by the calculation unit 303, or the lifespan of the resin component evaluated by the lifespan evaluation unit 304. In this example, various information is presented to the user by displaying a predetermined screen on the display device 360.

[0063] The processing control unit 306 controls the operation of the transport mechanism 120 in Figure 1 to transport the substrate to be processed to one of the multiple processing units 110 based on the transport conditions. The processing control unit 306 also controls the operation of the processing mechanism 1 of each processing unit 110 via the control device 2 in Figure 3 of each processing unit 110 based on the processing recipe.

[0064] The processing control unit 306 may change the transport conditions of the transport mechanism 120 so as to change the transport order of the substrates based on the lifespan evaluated by the lifespan evaluation unit 304. When the transport conditions are changed, the processing unit into which the substrate is loaded is changed. Since the processing recipe is tied to the substrate, if the processing recipes tied to each substrate are different, changing the processing unit into which the substrate is loaded will change the processing content performed in the processing unit 110. In addition, the processing control unit 306 may change the processing conditions of the substrate based on the lifespan evaluated by the lifespan evaluation unit 304. When the processing conditions are changed, the processing recipe related to those processing conditions is changed.

[0065] By changing the substrate transport conditions or substrate processing conditions, the lifespan of multiple processing units 110 can be made uniform. In this case, the timing for maintaining multiple resin components in multiple processing units 110 can be aligned. Therefore, it is not necessary to maintain multiple resin components at different times. This improves maintenance efficiency. In addition, the reduction in the operating time of the substrate processing device 500 due to maintenance can be minimized.

[0066] 6. Learning Process Figure 9 is a flowchart illustrating an example of the learning process flow. The learning process is performed by the CPU of the learning device 400 when the CPU executes a learning program. The learning process in Figure 9 will be explained below with reference to the learning device 400 in Figure 5. First, the data acquisition unit 401 acquires a pre-generated dataset (step S1).

[0067] Next, the descriptor acquisition unit 402 acquires molecular descriptors of the resin material used as the resin component in the dataset acquired in step S1 from the database storage device 200 as resin descriptors (step S2). Subsequently, the data acquisition unit 401 associates the resin descriptors acquired in step S2 with the resin components in the dataset acquired in step S1 (step S3).

[0068] Similarly, the descriptor acquisition unit 402 acquires the molecular descriptor of the solution used as the processing solution in the dataset acquired in step S1 from the database storage device 200 as a processing solution descriptor (step S4). Subsequently, the data acquisition unit 401 associates the processing solution descriptor acquired in step S4 with the processing solution in the dataset acquired in step S1 (step S5). Steps S2, S3 and steps S4, S5 may be executed in any order or simultaneously.

[0069] Subsequently, the data acquisition unit 401 determines whether or not to terminate the acquisition of the dataset (step S6). If a sufficient number of datasets have been acquired, the data acquisition unit 401 may determine to terminate the acquisition of the dataset. Alternatively, the data acquisition unit 401 may determine to terminate the acquisition of the dataset based on instructions from the user. The user can instruct the data acquisition unit 401 to terminate the acquisition of the dataset by operating an operation unit (not shown).

[0070] If the acquisition of the dataset is not completed, the process returns to step S1. Steps S1 to S5 are repeated until the acquisition of the dataset is completed. If the acquisition of the dataset is completed, the model generation unit 403 generates the learned model 600 shown in Figure 4 by training the inference model using the dataset to which the resin descriptors and processing liquid descriptors were associated in steps S3 and S5 (step S7). This completes the learning process.

[0071] 7. Life evaluation process Figure 10 is a flowchart showing an example of the life evaluation process flow. The life evaluation process is performed by the CPU 310 in Figure 2 of the information processing device 300 when the CPU 310 executes a life evaluation program. The life evaluation process in Figure 10 will be described below with reference to the information processing device 300 in Figure 8. First, the condition determination unit 301 determines the processing conditions and wetting time of the substrate in each processing unit 110 in Figure 1 (step S11).

[0072] Next, the weight prediction unit 302 uses the learning model 600 generated in step S7 of the learning process to predict the weight change of the resin member after a series of substrate processing in each processing unit 110 as the main task, based on the processing conditions and wetting time determined in step S11 (step S12). The calculation unit 303 also calculates reference information regarding the prediction of the weight change of the resin member in step S12 (step S13).

[0073] Next, the life evaluation unit 304 evaluates the lifespan of the resin components of each processing unit 110 based on the weight change of the resin components predicted in step S12 (step S14). The presentation unit 305 then presents various information to the user, such as the weight change of the resin components predicted in step S12, the reference information calculated in step S13, and the lifespan of the resin components evaluated in step S14 (step S15).

[0074] Subsequently, the processing control unit 306 determines whether the lifespan of the evaluated resin members is uniform for all processing units 110 (step S16). If the difference between the longest and shortest lifespans among the evaluated lifespans is less than or equal to a predetermined threshold, it is determined that the lifespan of the resin members is uniform. If the lifespan of the resin members is uniform, the processing control unit 306 determines whether the processing recipe has been changed (step S17).

[0075] If the processing recipe has not been changed, the process returns to step S16. In this case, steps S16 and S17 are repeated until the lifespan of the resin component becomes uneven or the processing recipe is changed. On the other hand, if the processing recipe is changed, the process returns to step S11. In this case, in step S11, the processing conditions and wetting time of the substrate in each processing unit 110 are determined by the condition determination unit 301 to correspond to the changed processing recipe. After that, the process from step S12 onward is executed again.

[0076] In step S16, if the lifespan of the resin components is not uniform, the processing control unit 306 changes the substrate transport conditions or the substrate processing conditions in any of the processing units 110 based on the lifespan evaluated in step S16 (step S18). Specifically, the substrate transport conditions are changed so that the lifespan of the resin components evaluated by the lifespan evaluation unit 304 approaches uniformity. Since the processing recipe is tied to the substrate, if the processing recipes tied to each substrate are different, the processing unit 110 containing a resin component evaluated to have a longer lifespan than other processing units 110 may be modified to transport a substrate tied to a processing recipe that results in a longer wetting time. Also, the processing unit 110 containing a resin component evaluated to have a shorter lifespan than other processing units 110 may be modified to transport a substrate tied to a processing recipe that results in a shorter wetting time.

[0077] Even if the processing recipe associated with each substrate is the same, the substrate transport conditions may be changed so that the substrate transport order is moved up for processing units 110 that include resin components evaluated to have a longer lifespan than other processing units 110. Alternatively, the substrate transport conditions may be changed so that the substrate transport order is moved down for processing units 110 that include resin components evaluated to have a shorter lifespan than other processing units 110.

[0078] Furthermore, for a processing unit 110 that includes a resin component evaluated to have a longer lifespan than other processing units 110, the processing conditions of the substrate may be changed so that the wetting time is longer than before the change. Alternatively, for a processing unit 110 that includes a resin component evaluated to have a shorter lifespan than other processing units 110, the processing conditions of the substrate may be changed so that the wetting time is shorter than before the change.

[0079] After step S18 is executed, the process returns to step S12. In this case, in step S12, the weight prediction unit 302 predicts the weight change of the resin component after a series of substrate processing in each processing unit 110, based on the modified substrate processing conditions. Subsequently, the processing from step S13 onwards is executed again.

[0080] 8. Variations In this embodiment, the information output from the resin prediction NN630 and the processing liquid prediction NN640 is not input to the interaction prediction NN650, but the embodiment is not limited thereto. The information output from at least one of the resin prediction NN630 and the processing liquid prediction NN640 may be input to the interaction prediction NN650. In this case, the interaction prediction NN650 can further use the information output by the resin prediction NN630 or the processing liquid prediction NN640 to perform a main task and zero or more subtasks.

[0081] Figure 11 shows an example of the structure of the learning model 600 in the first modified example. As shown in Figure 11, in the first modified example, the prediction result corresponding to at least one subtask related to the resin member, output by layer 633 of the resin prediction NN630, is input to layer 651 of the interaction prediction NN650. In addition, the prediction result corresponding to at least one subtask related to the processing liquid, output by layer 643 of the processing liquid prediction NN640, is input to layer 651 of the interaction prediction NN650.

[0082] Figure 12 shows an example of the structure of the learning model 600 in the second modified example. As shown in Figure 12, in the second modified example, the intermediate prediction results related to the resin member output by layer 632 of the resin prediction NN630 are input to layer 651 of the interaction prediction NN650. In addition, the intermediate prediction results related to the processing liquid output by layer 642 of the processing liquid prediction NN640 are input to layer 651 of the interaction prediction NN650.

[0083] 9. Effects In this substrate processing apparatus 500, the processing unit 110 includes a resin component that comes into contact with the processing liquid. When substrate processing using the processing liquid is performed in the processing unit 110, the weight change of the resin component is predicted by the weight prediction unit 302 using a learning model 600 based on the processing conditions. Therefore, the user can uniformly determine the timing for maintenance of the resin component by recognizing the predicted weight change of the resin component. This allows the resin component to be maintained at the appropriate time. In addition, the lifespan of the resin component is evaluated by the lifespan evaluation unit 304 based on the weight change of the resin component predicted by the weight prediction unit 302. In this case, the resin component can be maintained at the appropriate time more easily.

[0084] Here, the learning model 600 is an inference model that has learned an input-output relationship, taking processing conditions in the dataset as input and outputting main ground truth data and sub-ground truth data. The main ground truth data shows the weight change of the resin component after substrate processing is performed by the processing unit 110 under the processing conditions. The sub-ground truth data shows the physical properties related to the substrate processing. Therefore, by acquiring a large number of main ground truth data, a learning model 600 capable of predicting the weight change of the resin component with high accuracy is generated.

[0085] On the other hand, obtaining a large amount of main ground truth data can not always be easy. Even in such cases, as described above, sub-ground truth data can be obtained more easily and in larger quantities than main ground truth data. Therefore, by using sub-ground truth data, the learning model 600 can learn the input-output relationships necessary to predict the weight change of the resin component after substrate processing with high accuracy. As a result, even when the amount of main ground truth data is relatively small, it is possible to generate a learning model 600 that can predict the weight change of the resin component with sufficiently high accuracy.

[0086] The learning model 600 includes one or more lower-level NN601 and one or more upper-level NN602. Each lower-level NN601 outputs an intermediate prediction result based on the processing conditions. Each upper-level NN602 outputs a prediction result corresponding to the main ground truth data and a prediction result corresponding to the sub-ground truth data, based on the intermediate prediction result output by any of the lower-level NN601.

[0087] In this case, each lower NN601 does not need to be trained to output prediction results corresponding to the main and sub-ground data. Similarly, the upper NN602 does not need to be trained to output intermediate prediction results. Therefore, both the lower NN601 and upper NN602 are simplified. This makes it possible to generate a learning model 600 that can predict the weight change of resin components with sufficiently high accuracy, even when the amount of main and sub-ground data is relatively small.

[0088] In this embodiment, the learning model 600 includes two lower-level NN601: a resin NN610 and a processing liquid NN620. The resin NN610 outputs intermediate prediction results related to the resin member based on the processing conditions. The processing liquid NN620 outputs intermediate prediction results related to the processing liquid based on the processing conditions. In this case, the resin NN610 does not need to be trained to output intermediate prediction results related to the processing liquid. Similarly, the processing liquid NN620 does not need to be trained to output intermediate prediction results related to the resin member.

[0089] This simplifies each lower-level NN601. Furthermore, the interpretability of the intermediate prediction results output by each lower-level NN601 is improved. Therefore, each upper-level NN602 can be easily trained to output prediction results corresponding to the main and sub-ground data. In addition, the intermediate prediction results output by each lower-level NN601 can be easily reused for other tasks.

[0090] Furthermore, in this embodiment, the learning model 600 includes three higher-level NN602: a resin prediction NN630, a processing liquid prediction NN640, and an interaction prediction NN650. The resin prediction NN630 outputs physical properties related to the resin member as prediction results corresponding to the sub-ground truth data. The processing liquid prediction NN640 outputs physical properties related to the processing liquid as prediction results corresponding to the sub-ground truth data. The interaction prediction NN650 outputs prediction results corresponding to the main ground truth data, and also outputs physical properties related to the interaction between the resin member and the processing liquid as prediction results corresponding to the sub-ground truth data.

[0091] In this case, the resin prediction NN630 does not need to be trained to output prediction results corresponding to the main ground truth data and physical properties related to the processing liquid. The processing liquid prediction NN640 does not need to be trained to output prediction results corresponding to the main ground truth data and physical properties related to the resin component. The interaction prediction NN650 does not need to be trained to output physical properties related to the resin component and the processing liquid. This simplifies each of the upper-level NN602s.

[0092] Furthermore, since the interaction prediction NN650 learns to output physical property values ​​related to the interaction between the resin component and the processing liquid, it is possible to output prediction results corresponding to the main ground truth data with sufficiently high accuracy even when the main ground truth data is relatively small. In particular, the sub-ground truth data includes the χ parameter, which indicates the interaction between the polymer compound constituting the resin component and the processing liquid. With this configuration, even when the main ground truth data is relatively small, it is possible to more reliably generate a learning model 600 that can predict the weight change of the resin component with sufficiently high accuracy.

[0093] Furthermore, the number of layers in each lower-level NN601 is greater than the number of layers in each upper-level NN602. In this case, the interpretability of the intermediate prediction results output by each lower-level NN601 is improved. This makes it easier for each upper-level NN602 to train to output prediction results corresponding to the main and sub-ground data. In addition, the intermediate prediction results output by each lower-level NN601 can be easily reused for other tasks.

[0094] 10. Other Embodiments (1) In the above embodiment, the learning model 600 is realized by five NNs, but the embodiment is not limited thereto. By integrating some or all of the five NNs, the learning model 600 may be realized by one or more and four or fewer NNs. Figure 13 is a diagram showing an example of the structure of the learning model 600 in another embodiment. As shown in Figure 13, in the other embodiment, the learning model 600 includes a resin / processing liquid NN660 as a single lower NN601, instead of the resin NN610 and processing liquid NN620 in Figure 4. In this case, the learning model 600 is realized by four NNs.

[0095] The resin / processing solution NN660 has a configuration in which the resin NN610 and the processing solution NN620 are integrated, and includes four layers 661, 662, 663, and 664. Layer 661 receives information about the substrate processing conditions (resin descriptor, processing solution descriptor, and setting conditions). The information input to layer 661 is passed to layer 664 through layers 662 and 663. Layer 664 outputs an intermediate prediction result based on the given information.

[0096] The intermediate prediction results output by the resin / processing liquid NN660 are a superposition of the intermediate prediction results output by the resin NN610 and the intermediate prediction results output by the processing liquid NN620. Therefore, the interpretability of the intermediate prediction results output by the resin / processing liquid NN660 is lower than that of the intermediate prediction results output by the resin NN610 or the processing liquid NN620. Even in this case, the intermediate prediction results can be used as information to be input to the higher-level NN602. Accordingly, the intermediate prediction results output by the resin / processing liquid NN660 are input as information to each of the layers: layer 631 of the resin prediction NN630, layer 641 of the processing liquid prediction NN640, and layer 651 of the interaction prediction NN650.

[0097] As another example of the learning model 600, the learning model 600 may include a single higher-level NN602 that integrates the resin prediction NN630, the treatment liquid prediction NN640, and the interaction prediction NN650, instead of the resin prediction NN630, the treatment liquid prediction NN640, and the interaction prediction NN650. In this case, the learning model 600 is implemented by three NNs.

[0098] Alternatively, the learning model 600 may not include either or both of the resin NN610 and the processing liquid NN620. Furthermore, if the interaction prediction NN650 performs a subtask, the learning model 600 may not include either or both of the resin prediction NN630 and the processing liquid prediction NN640. In these cases as well, the learning model 600 is implemented using one or more but no more than four neural networks (NNs).

[0099] (2) In the above embodiment, the interaction prediction NN650 performs subtasks, but the embodiment is not limited thereto. The interaction prediction NN650 does not have to perform subtasks.

[0100] (3) In the above embodiment, the number of layers in the lower NN601 is greater than the number of layers in the upper NN602, but the embodiment is not limited thereto. The number of layers in the lower NN601 may be equal to the number of layers in the upper NN602, or it may be less than the number of layers in the upper NN602.

[0101] (4) In the above embodiment, the weight of the resin member and the remaining weight are treated as relative values ​​with respect to the initial weight, but the embodiment is not limited thereto. The weight of the resin member and the remaining weight may be treated as absolute values ​​with units.

[0102] (5) In the above embodiment, the substrate processing apparatus 500 includes a plurality of processing units 110, but the embodiment is not limited thereto. The substrate processing apparatus 500 may include one processing unit 110. Even in this case, the processing control unit 306 can adjust the timing at which the resin member should be maintained in the single processing unit 110 by changing the processing conditions of the substrate by the processing unit 110 based on the lifespan of the resin member evaluated by the lifespan evaluation unit 304.

[0103] (6) In the above embodiment, the substrate processing apparatus 500 includes the processing unit 110, but the embodiment is not limited thereto. The substrate processing apparatus 500 does not have to include the processing unit 110, as long as the information processing apparatus 300 is configured to be connectable to the processing unit 110 which includes a resin member.

[0104] (7) In the above embodiment, the substrate processing apparatus 500 includes a life evaluation unit 304, but the embodiment is not limited thereto. The user can uniformly determine the timing for maintenance of the resin member by recognizing the weight change of the resin member predicted by the weight prediction unit 302. For this reason, the substrate processing apparatus 500 does not have to include a life evaluation unit 304.

[0105] (8) In the above embodiment, the substrate processing apparatus 500 includes a processing control unit 306, but the embodiment is not limited thereto. If control based on the weight change of the resin member predicted by the weight prediction unit 302 is not performed on the processing unit 110 or the transport mechanism 120, the substrate processing apparatus 500 does not need to include a processing control unit 306.

[0106] (9) In the above embodiment, the treatment solution is a mixture containing a sublimation agent and a solvent that dissolves with the sublimation agent, but the embodiment is not limited thereto. The treatment solution may be sulfuric acid, nitric acid, hydrochloric acid, hydrofluoric acid, phosphoric acid, acetic acid, aqueous ammonia, aqueous hydrogen peroxide, organic acid (e.g., citric acid or oxalic acid), organic alkali (e.g., TMAH: tetramethylammonium hydroxide), or organic solvent (e.g., IPA: isopropyl alcohol). Alternatively, the treatment solution may be a mixture of any of these liquids.

[0107] (10) In the above embodiment, the rotation speed of the substrate is given as an example of the setting conditions, but the embodiment is not limited to this. The setting conditions may include, for example, other parameters described in the processing recipe. For example, the setting conditions may include the supply time of the processing liquid, the discharge flow rate of the processing liquid, the processing time, the behavior of the supply nozzle, the behavior of the shut-off plate, the supply time of the inert gas, or the discharge flow rate of the inert gas.

[0108] 11. Correspondence between each component of the claim and each part of the embodiment The following describes examples of the correspondence between each component of the claims and each element of the embodiments, but the present invention is not limited to the following examples. Various other elements having the configuration or function described in the claims can also be used as each component of the claims.

[0109] In the above embodiment, the processing cup 21, processing liquid nozzle 39, or processing liquid piping 40 are examples of resin members, the processing unit 110 is an example of a processing unit, the learning model 600 is an example of a learning model, and the weight prediction unit 302 is an example of a weight prediction unit. The substrate processing device 500 is an example of a substrate processing device, the lower NN601 is an example of a lower neural network, and the upper NN602 is an example of a higher neural network.

[0110] Resin NN610 is an example of a resin neural network. Processing liquid NN620 is an example of a processing liquid neural network, resin prediction NN630 is an example of a resin prediction neural network, processing liquid prediction NN640 is an example of a processing liquid prediction neural network, interaction prediction NN650 is an example of an interaction prediction neural network, and lifetime evaluation unit 304 is an example of a lifetime evaluation unit.

[0111] 12. Summary of Embodiments (Paragraph 1) The substrate processing apparatus relating to Paragraph 1 is: Equipped with a processor, The processor, by executing a program, predicts the weight change of the resin member using a learning model based on the processing conditions when substrate processing using the processing liquid is performed in a chamber containing a resin member that comes into contact with the processing liquid. The aforementioned learning model is an inference model that learns an input-output relationship, taking the processing conditions as input and outputting the main correct data and sub-correct data. The aforementioned main correct data shows the weight change of the resin member after substrate processing is performed in the chamber under the aforementioned processing conditions. The aforementioned sub-correct data represents physical properties related to substrate processing.

[0112] In this substrate processing apparatus, the weight change of the resin component when substrate processing is performed using a processing solution in the chamber is predicted using a learning model. In this case, the user can uniformly determine the timing for maintenance of the resin component by recognizing the weight change. This allows for maintenance of the resin component at the appropriate time.

[0113] Here, the learning model learns the input-output relationship using the main ground truth data and sub-ground truth data. Therefore, even when it is not easy to obtain a large amount of main ground truth data, the learning model can learn the input-output relationship necessary to predict the weight change of the resin component after substrate processing with high accuracy by using sub-ground truth data. As a result, even when the amount of main ground truth data is relatively small, it is possible to generate a learning model that can predict the weight change of the resin component with sufficiently high accuracy.

[0114] (Paragraph 2) In the substrate processing apparatus described in Paragraph 1, The aforementioned learning model, One or more lower-level neural networks that output intermediate prediction results from the processing conditions described above, The system may also include one or more higher-level neural networks that output prediction results corresponding to the main ground truth data and prediction results corresponding to the sub-ground truth data based on the intermediate prediction results output by the one or more lower-level neural networks.

[0115] In this case, each lower-level neural network does not need to be trained to output prediction results corresponding to the main and sub-ground data. Similarly, each higher-level neural network does not need to be trained to output intermediate prediction results. Therefore, both the lower-level and higher-level neural networks are simplified. This makes it possible to generate a learning model that can predict the weight change of resin components with sufficiently high accuracy, even when the amount of main and sub-ground data is relatively small.

[0116] (3) In the substrate processing apparatus described in paragraph 2, The one or more sub-neural networks mentioned above are: A resin neural network that outputs the intermediate prediction results related to the resin member from the processing conditions, The system may include at least one of the following: a processing solution neural network that outputs the intermediate prediction results related to the processing solution from the processing conditions.

[0117] In this case, the resin neural network does not need to learn to output intermediate prediction results related to the processing liquid. Similarly, the processing liquid neural network does not need to learn to output intermediate prediction results related to the resin component. Therefore, the lower-level neural networks are simplified. This makes it possible to generate a learning model that can predict the weight change of the resin component with sufficiently high accuracy, even when the number of main and sub-ground truth data is relatively small.

[0118] (Article 4) In the substrate processing apparatus described in Article 2 or Article 3, The above one or more higher-level neural networks are, It includes an interaction prediction neural network that outputs prediction results corresponding to the main ground truth data, A resin prediction neural network that outputs physical property values ​​related to the resin member as prediction results corresponding to the sub-correction data, The system may include at least one of the following: a neural network that predicts the processing liquid and outputs physical properties related to the processing liquid as prediction results corresponding to the sub-correction data.

[0119] In this case, the interaction prediction neural network does not need to be trained to output physical properties related to the resin component and the processing liquid. The resin prediction neural network does not need to be trained to output prediction results corresponding to the main ground truth data and physical properties related to the processing liquid. Similarly, the processing liquid prediction neural network does not need to be trained to output prediction results corresponding to the main ground truth data and physical properties related to the resin component. Therefore, the higher-level neural networks are simplified. This makes it possible to generate a learning model that can predict the weight change of the resin component with sufficiently high accuracy, even when the number of main and sub-ground truth data is relatively small.

[0120] (Item 5) In the substrate processing apparatus described in Item 4, The interaction prediction neural network may further output physical property values ​​related to the interaction between the resin member and the processing liquid as prediction results corresponding to the sub-correction data.

[0121] With this configuration, the interaction prediction neural network learns to output physical property values ​​related to the interaction between the resin component and the processing liquid, enabling it to output prediction results corresponding to the main ground truth data with sufficiently high accuracy, even when the main ground truth data is relatively small.

[0122] (Paragraph 6) In the substrate processing apparatus described in paragraph 2 or 3, The one or more higher-level neural networks mentioned above may include an interaction prediction neural network that outputs prediction results corresponding to the main ground truth data, and also outputs physical property values ​​related to the interaction between the resin member and the processing liquid as prediction results corresponding to the sub-ground truth data.

[0123] With this configuration, the interaction prediction neural network learns to output physical property values ​​related to the interaction between the resin component and the processing liquid, enabling it to output prediction results corresponding to the main ground truth data with sufficiently high accuracy, even when the main ground truth data is relatively small.

[0124] (Paragraph 7) In the substrate processing apparatus described in any one of paragraphs 2 to 6, The number of layers in the lower neural network may be greater than the number of layers in the upper neural network.

[0125] In this case, the interpretability of the intermediate prediction results output by the lower-level neural network is improved. This makes it easier for the higher-level neural network to learn to output prediction results corresponding to the main and sub-ground data.

[0126] (Paragraph 8) In the substrate processing apparatus described in any one of paragraphs 1 to 7, The aforementioned sub-correct data may include a χ parameter indicating the interaction between the polymer compound constituting the resin member and the processing solution.

[0127] This configuration makes it possible to more reliably generate a learning model that can predict the weight change of resin components with sufficiently high accuracy, even when the main ground truth data is relatively small.

[0128] (Paragraph 9) The substrate processing apparatus described in any one of paragraphs 1 to 8 is: The processor may, by executing the program, further evaluate the lifespan of the resin member based on the predicted weight change of the resin member.

[0129] In this case, the lifespan of the resin component is evaluated, making it easier to maintain the resin component at the appropriate time.

[0130] (Paragraph 10) The substrate processing method relating to Paragraph 10 is: A substrate processing method performed by a processor in a substrate processing apparatus, The processor, by executing a program, predicts the weight change of the resin member using a learning model based on the processing conditions when substrate processing using the processing liquid is performed in a chamber containing a resin member that comes into contact with the processing liquid. The aforementioned learning model is an inference model that learns an input-output relationship, taking the processing conditions as input and outputting the main correct data and sub-correct data. The aforementioned main correct data shows the weight change of the resin member after substrate processing is performed in the chamber under the aforementioned processing conditions. The aforementioned sub-correct data represents physical properties related to substrate processing.

[0131] According to this substrate processing method, the weight change of the resin component when substrate processing is performed using a processing solution in the chamber is predicted using a learning model. In this case, the user can uniformly determine the timing for maintaining the resin component by recognizing the weight change. This allows for maintenance of the resin component at the appropriate time. [Explanation of symbols]

[0132] 1…Processing mechanism, 2…Control device, 3…Chamber, 10…Spin chuck, 11…Chuck pin, 12…Spin base, 13…Spin shaft, 14…Spin motor, 21…Processing cup, 22…Outer wall member, 23…Cup, 24…Guard, 27…Guard lifting unit, 31…Chemical solution nozzle, 34, 38, 42, 46…Nozzle movement unit, 35…Rinse solution nozzle, 39…Processing solution nozzle, 40…Processing solution piping, 41…Processing solution valve, 43…Displacement solution nozzle, 51…Shut-off member, 51L…Bottom surface, 52…Disc section, 53…Support shaft, 54…Shut-off member lifting unit, 55…Center nozzle, 56…Upper gas piping, 57…Upper gas valve, 58…Flow rate adjustment valve, 59…Upper temperature controller, 100…Processing unit, 110…Processing unit, 120…Transportation mechanism, 200…Database storage device, 300…Information processing 301…Condition Determination Unit, 302…Weight Prediction Unit, 303…Calculation Unit, 304…Lifespan Evaluation Unit, 305…Presentation Unit, 306…Processing Control Unit, 310…CPU, 320…RAM, 330…ROM, 340…Storage Device, 350…Operation Unit, 360…Display Device, 370…Input / Output I / F, 380…Bus, 400…Learning Device, 401…Data Acquisition Unit, 402…Descriptor Acquisition Unit, 403…Model Generation Part, 500…Substrate processing apparatus, 501…Network, Learning model 600…, 601…Lower NN, 602…Upper NN, 610…Resin NN, 611~614, 621~624, 631~633, 641~643, 651~653, 661~664…Layer, 620…Processing liquid NN, 630…Resin prediction NN, 640…Processing liquid prediction NN, 650…Interaction prediction NN, A1…Rotation axis, W…Substrate

Claims

1. Equipped with a processor, The processor, by executing a program, predicts the weight change of the resin member using a learning model based on the processing conditions when substrate processing using the processing liquid is performed in a chamber containing a resin member that comes into contact with the processing liquid. The aforementioned learning model is an inference model that learns an input-output relationship, taking the processing conditions as input and outputting the main correct data and sub-correct data. The aforementioned main correct data shows the weight change of the resin member after substrate processing is performed in the chamber under the aforementioned processing conditions. The aforementioned sub-correct data represents a substrate processing apparatus that exhibits physical properties related to substrate processing.

2. The aforementioned learning model, One or more lower-level neural networks that output intermediate prediction results from the processing conditions, The substrate processing apparatus according to claim 1, comprising one or more higher-level neural networks that output prediction results corresponding to the main ground truth data and prediction results corresponding to the sub-ground truth data based on the intermediate prediction results output by the one or more lower-level neural networks.

3. The one or more sub-neural networks mentioned above are: A resin neural network that outputs the intermediate prediction results related to the resin member from the processing conditions, The substrate processing apparatus according to claim 2, comprising at least one of the following: a processing liquid neural network that outputs the intermediate prediction results related to the processing liquid from the processing conditions.

4. The above one or more higher-level neural networks are, It includes an interaction prediction neural network that outputs prediction results corresponding to the main ground truth data, A resin prediction neural network that outputs physical property values ​​related to the resin member as prediction results corresponding to the sub-correction data, The substrate processing apparatus according to claim 2 or 3, comprising at least one of the following: a processing liquid prediction neural network that outputs physical property values ​​related to the processing liquid as prediction results corresponding to the sub-correct data.

5. The substrate processing apparatus according to claim 4, wherein the interaction prediction neural network further outputs physical property values ​​related to the interaction between the resin member and the processing liquid as prediction results corresponding to the sub-correct data.

6. The substrate processing apparatus according to claim 2 or 3, wherein the one or more higher-level neural networks include an interaction prediction neural network that outputs prediction results corresponding to the main ground truth data and, as prediction results corresponding to the sub-ground truth data, outputs physical property values ​​related to the interaction between the resin member and the processing liquid.

7. The substrate processing apparatus according to claim 2 or 3, wherein the number of layers in the lower neural network is greater than the number of layers in the upper neural network.

8. The substrate processing apparatus according to any one of claims 1 to 3, wherein the subcorrect data includes a χ parameter indicating the interaction between the polymer compound constituting the resin member and the processing liquid.

9. The substrate processing apparatus according to any one of claims 1 to 3, wherein the processor further evaluates the lifespan of the resin member based on the predicted weight change of the resin member by executing the program.

10. A substrate processing method performed by a processor in a substrate processing apparatus, The processor, by executing a program, predicts the weight change of the resin member using a learning model based on the processing conditions when substrate processing using the processing liquid is performed in a chamber containing a resin member that comes into contact with the processing liquid. The aforementioned learning model is an inference model that learns an input-output relationship, taking the processing conditions as input and outputting the main correct data and sub-correct data. The aforementioned main correct data shows the weight change of the resin member after substrate processing is performed in the chamber under the aforementioned processing conditions. The aforementioned sub-correct data represents a substrate processing method, showing physical properties related to substrate processing.

Images