Learning apparatus, information processing apparatus, substrate processing apparatus, learning model generation method, and processing condition determination method

Abstract

The learning apparatus includes: a processing condition acquisition unit that acquires a processing condition for driving a substrate processing apparatus that performs processing of a coating film by supplying a processing liquid to a substrate on which the coating film is formed; a first processing result acquisition unit that acquires a first processing result including a difference in film thickness before and after processing of a coating film at each of a plurality of radially different positions of the substrate after driving the substrate processing apparatus under a processing condition to perform processing of the coating film; a statistical value calculation unit for statistically processing the first processing result and calculating a statistical value; a conversion unit that converts the first processing result into a second processing result using the statistical value; a first learning model generation unit that generates a first learning model by machine learning the processing conditions and the second processing results; and a second learning model generation unit that generates a second learning model by machine learning the processing conditions and the statistical values.

Description

Technical Field

[0001] This invention relates to a learning device, an information processing device, a substrate processing device, a learning model generation method, and a processing condition determination method. Background Technology

[0002] One of the semiconductor manufacturing processes is the cleaning process. In the cleaning process, the film thickness of the coating formed on the substrate is adjusted by etching through the supply of a chemical solution to the substrate. In this film thickness adjustment, it is important to perform etching in a way that ensures uniform etching across the entire surface of the substrate, or to make the substrate surface flat through etching. When the etching solution is sprayed from a nozzle onto a portion of the substrate, it is necessary to move the nozzle radially relative to the substrate.

[0003] Patent Document 1 describes a liquid processing apparatus capable of etching a substrate by spraying etchant from an etching nozzle onto the substrate. An example is described where, in order to uniformly distribute the in-plane temperature of a wafer while etching the central region of the substrate, the etching nozzle is repeatedly moved between a first position and a second position while spraying etchant. The first position is the location where the sprayed etchant passes through the center of the wafer, and the second position is a location closer to the periphery of the wafer than the central position.

[0004] Etching is a complex process where the amount of coating processed varies depending on the movement of the nozzle, as well as differences in the concentration and temperature of the etching solution and the rotation speed of the substrate. Therefore, machine learning is considered, utilizing artificial intelligence learning models to predict the processing volume.

[0005] Patent Document 1: Japanese Patent Application Publication No. 2015-103656 Summary of the Invention

[0006] The problem that the invention aims to solve

[0007] The amount of material processed by etching for coating varies depending on processing conditions such as nozzle movement, etchant concentration, etchant temperature, and substrate rotation speed. Because the amount of material processed is affected by numerous processing conditions, the magnitude (scale) of the multiple amounts obtained for multiple processing conditions can vary significantly. Therefore, when using multiple amounts of material processed by etching for coating as training data for machine learning models, the scale difference among these multiple amounts becomes a problem. Consequently, the learned model generated from these multiple amounts of material with large scale differences in machine learning may suffer from issues.

[0008] One of the objectives of this invention is to provide a learning device and a learning model generation method that can improve the accuracy of the learning model that models the coating process performed by the substrate processing device.

[0009] Another object of the present invention is to provide an information processing apparatus capable of indicating a plurality of processing conditions for the processing results of a complex process used to process a substrate, a substrate processing apparatus having the information processing apparatus, and a method for determining processing conditions.

[0010] Methods for solving problems

[0011] A learning apparatus according to one embodiment of the present invention includes: a processing condition acquisition unit for acquiring processing conditions for driving a substrate processing apparatus, the substrate processing apparatus performing a coating process by supplying a processing liquid to a substrate having a coating formed thereon; a first processing result acquisition unit for acquiring a first processing result after the substrate processing apparatus is driven to perform the coating process under the processing conditions, the first processing result including the difference in film thickness before and after the coating process at each of a plurality of radially different locations on the substrate; a statistical value calculation unit for statistically processing the first processing result and calculating a statistical value; a conversion unit for converting the first processing result into a second processing result using the statistical value; a first learning model generation unit for generating a first learning model by machine learning processing conditions and the second processing result; and a second learning model generation unit for generating a second learning model by machine learning processing conditions and the statistical value.

[0012] Another embodiment of the learning apparatus of the present invention includes: a processing condition acquisition unit for acquiring processing conditions for driving a substrate processing apparatus, the substrate processing apparatus performing a coating process by supplying a processing liquid to a substrate having a coating formed; a first processing result acquisition unit for acquiring a first processing result after the substrate processing apparatus is driven to perform the coating process under the processing conditions, the first processing result including the difference in film thickness before and after the coating process at each of a plurality of radially different positions on the substrate; a statistical value calculation unit for statistically processing the first processing result and calculating a statistical value; a conversion unit for converting the first processing result into a second processing result using the statistical value; and a multi-task learning model generation unit for machine learning learning data and generating a multi-task learning model, the learning data including the processing conditions, the second processing result, and the statistical value.

[0013] An information processing apparatus according to another embodiment of the present invention is used to manage a substrate processing apparatus, wherein the substrate processing apparatus supplies processing liquid to a substrate having a coating formed on it to perform a coating process under processing conditions including varying conditions that change over time; the information processing apparatus includes a processing condition determination unit that uses a predictor to determine the processing conditions for driving the substrate processing apparatus, the predictor predicting a processing result representing the difference in film thickness before and after the coating process on the substrate before the coating process is performed by the substrate processing apparatus; the predictor includes: a first learning model, which is obtained by machine learning using statistical processing of the first processing result after the coating process is performed by driving the substrate processing apparatus with the processing conditions. The processing condition is generated by converting the statistical value of the first processing result into a second processing result and the processing conditions. The first processing result includes the difference in film thickness before and after processing at each of a plurality of radially different locations on the substrate. The processing condition determination unit includes a recovery unit that recovers the first predicted processing result based on a second predicted processing result inferred by assigning temporary processing conditions to the first learning model and a predicted statistical value inferred by assigning temporary processing conditions to the second learning model. If the first predicted processing result recovered by the recovery unit meets the allowable conditions, the temporary processing conditions are determined as the processing conditions for driving the substrate processing apparatus.

[0014] An information processing apparatus according to another embodiment of the present invention is used to manage a substrate processing apparatus, wherein the substrate processing apparatus supplies processing liquid to a substrate having a coating formed on it to perform a coating process under processing conditions including changing conditions that vary over time; the information processing apparatus includes a processing condition determination unit that uses a predictor to determine processing conditions for driving the substrate processing apparatus, the predictor predicting a processing result representing the difference in film thickness before and after the coating process on the substrate prior to the coating process by the substrate processing apparatus; the predictor has a multi-task learning model generated by machine learning learning data, the learning data including data used in processing conditions After the substrate processing device performs a coating process, the statistical value obtained by statistically processing the first processing result is converted into a second processing result. The statistical value is used as the target variable and the processing conditions are used as the explanatory variable. The first processing result includes the difference in film thickness before and after coating at each of a plurality of different radial positions of the substrate. The processing condition determination unit includes a restoration unit, which restores the first predicted processing result based on the second predicted processing result and the predicted statistical value inferred by the multi-task learning model with temporary processing conditions. If the first predicted processing result restored by the restoration unit meets the allowable conditions, the temporary processing conditions are determined as the processing conditions for driving the substrate processing device.

[0015] A substrate processing apparatus according to another embodiment of the present invention includes the information processing apparatus described above.

[0016] Another embodiment of the present invention provides a learning model generation method that causes a learning device to perform: a processing condition acquisition step, acquiring processing conditions for driving a substrate processing apparatus, the substrate processing apparatus performing a coating process by supplying a processing liquid to a substrate having a coating; a first processing result acquisition step, after driving the substrate processing apparatus to perform the coating process under the processing conditions, acquiring a first processing result, the first processing result including the difference in film thickness before and after coating processing at each of a plurality of radially different locations on the substrate; a statistical value calculation step, statistically processing the first processing result and calculating a statistical value; a conversion step, using the statistical value to convert the first processing result into a second processing result; a first learning model generation step, using the machine learning processing conditions and the second processing result to generate a first learning model; and a second learning model generation step, using the machine learning processing conditions and the statistical value to generate a second learning model.

[0017] Another embodiment of the present invention provides a learning model generation method that causes a learning device to perform: a processing condition acquisition step, acquiring processing conditions for driving a substrate processing apparatus, the substrate processing apparatus performing coating processing by supplying a processing liquid to a substrate having a coating formed; a first processing result acquisition step, acquiring a first processing result after driving the substrate processing apparatus to perform coating processing under the processing conditions, the first processing result including the difference in film thickness before and after coating processing at each of a plurality of radially different locations on the substrate; a statistical value calculation step, statistically processing the first processing result and calculating a statistical value; a conversion step, converting the first processing result into a second processing result using the statistical value; and a multi-task learning model generation step, machine learning using data and generating a multi-task learning model, the learning data including the processing conditions, the second processing result, and the statistical value.

[0018] A processing condition determination method according to another embodiment of the present invention is used by an information processing device for managing a substrate processing apparatus, wherein the substrate processing apparatus supplies processing liquid to a substrate having a coating formed with a coating to perform a coating process under processing conditions including varying conditions that change over time; the processing condition determination method includes a processing condition determination step using a predictor to determine processing conditions for driving the substrate processing apparatus, the predictor speculating a predicted processing result representing the difference in film thickness before and after the coating process on the substrate formed before the coating process by the substrate processing apparatus; the predictor has: a first learning model, which uses machine learning to perform statistical processing of the first processing after the coating process is performed by driving the substrate processing apparatus with processing conditions. The statistical value obtained is converted into a second processing result from the first processing result and the processing conditions. The first processing result includes the difference in film thickness before and after processing at each of a plurality of radially different locations on the substrate. The second learning model is generated by machine learning the processing conditions and statistical values. The processing condition determination step includes: a restoration step, restoring the first predicted processing result based on the second predicted processing result inferred by assigning temporary processing conditions to the first learning model and the predicted statistical value inferred by assigning temporary processing conditions to the second learning model; and a step of determining the temporary processing conditions as the processing conditions for driving the substrate processing apparatus if the restored first predicted processing result in the restoration step meets the allowable conditions.

[0019] A processing condition determination method according to another embodiment of the present invention is used by an information processing device for managing a substrate processing apparatus. The substrate processing apparatus supplies a processing liquid to a substrate having a coating formed on it to perform a coating process under processing conditions that vary over time. The processing condition determination method includes a processing condition determination step using a predictor to determine the processing conditions for driving the substrate processing apparatus. The predictor, for the coating formed on the substrate before the coating process by the substrate processing apparatus, predicts a processing result representing the difference in film thickness before and after the coating process. The predictor has a multi-task learning model generated by machine learning data. The method includes using statistical values ​​obtained by statistically processing a first processing result after performing a coating process with a substrate processing apparatus driven by processing conditions to transform the first processing result into a second processing result, the statistical values, and processing conditions. The first processing result includes the difference in film thickness before and after coating at each of a plurality of radially different locations on the substrate. The processing condition determination step includes: a restoration step, restoring the first predicted processing result based on a second predicted processing result inferred by a multi-task learning model and the predicted statistical values; and a step of determining the temporary processing conditions as processing conditions for driving the substrate processing apparatus if the restored first predicted processing result in the restoration step meets the allowable conditions.

[0020] The effects of the invention

[0021] A learning apparatus and a learning model generation method are provided that can improve the accuracy of a learning model that models a coating process performed by a substrate processing apparatus. Additionally, an information processing apparatus that can provide multiple processing conditions for the processing results of complex processes on a substrate, a substrate processing apparatus having the information processing apparatus, and a method for determining processing conditions are provided. Attached Figure Description

[0022] Figure 1 This is a diagram illustrating the structure of a substrate processing system according to an embodiment of the present invention.

[0023] Figure 2 This diagram illustrates an example of the structure of an information processing device.

[0024] Figure 3 A diagram illustrating an example of the structure of a learning device.

[0025] Figure 4 This figure illustrates an example of the structure of a substrate processing system to demonstrate its functionality.

[0026] Figure 5 A diagram illustrating one example of the functions of the predictor generation unit.

[0027] Figure 6 This is a figure showing an example of film thickness characteristics.

[0028] Figure 7 A diagram showing an example of the first processing result obtained by coating process through a substrate processing apparatus.

[0029] Figure 8 A diagram illustrating an example of the second processing result.

[0030] Figure 9 This is a diagram used to illustrate the first learning model.

[0031] Figure 10 This is a diagram used to illustrate the second learning model.

[0032] Figure 11 A diagram illustrating one example of the functions of the prediction unit.

[0033] Figure 12 A flowchart illustrating an example of the learning process.

[0034] Figure 13 This is a flowchart illustrating an example of the predictor generation process.

[0035] Figure 14 This is a flowchart illustrating an example of a process where processing conditions determine the course of action.

[0036] Figure 15 This is a diagram illustrating one example of the functionality of the predictor generation unit in the modified example.

[0037] Figure 16 This is a diagram used to illustrate an example of a multi-task learning model. Detailed Implementation

[0038] Hereinafter, a substrate processing system according to an embodiment of the present invention will be described in detail with reference to the accompanying drawings. In the following description, substrate refers to a semiconductor substrate (semiconductor wafer), a substrate for a flat panel display (FPD) such as a liquid crystal display device or an organic EL (Electroluminescence) display device, a substrate for an optical disc, a substrate for a magnetic disk, a substrate for an optical disk, a substrate for a photomask, a ceramic substrate, or a substrate for a solar cell, etc.

[0039] (1) Overall structure of the substrate processing system

[0040] Figure 1 This is a diagram illustrating the structure of a substrate processing system according to an embodiment of the present invention. Figure 1The substrate processing system 1 includes an information processing device 100, a learning device 200, and a substrate processing device 300. The learning device 200 is, for example, a server, and the information processing device 100 is, for example, a personal computer.

[0041] The learning device 200 and the information processing device 100 are used to manage the substrate processing device 300. Furthermore, the substrate processing device 300 managed by the learning device 200 and the information processing device 100 is not limited to one unit, but can also manage multiple substrate processing devices 300.

[0042] In the substrate processing system 1 of this embodiment, the information processing device 100, the learning device 200, and the substrate processing device 300 are connected to each other via wired or wireless communication lines or communication network. The information processing device 100, the learning device 200, and the substrate processing device 300 are each connected to a network and can send and receive data from each other. The network may be, for example, a local area network (LAN) or a wide area network (WAN). Alternatively, the network may be the Internet. Furthermore, the information processing device 100 and the substrate processing device 300 may also be connected via a dedicated communication network. The network connection method can be wired or wireless.

[0043] Furthermore, the learning device 200 does not necessarily need to be connected to the board processing device 300 and the information processing device 100 via a communication cable or communication line network. In this case, data generated in the board processing device 300 can also be transferred to the learning device 200 via a recording medium. Similarly, data generated in the learning device 200 can also be transferred to the information processing device 100 via a recording medium.

[0044] The substrate processing apparatus 300 is equipped with a display device (not shown), a sound output device, and an operation unit. The substrate processing apparatus 300 operates according to preset processing conditions (processing procedures).

[0045] (2) Overview of the substrate processing apparatus

[0046] The substrate processing apparatus 300 includes a control device 10 and a plurality of substrate processing units WU. The control device 10 controls the plurality of substrate processing units WU. The plurality of substrate processing units WU perform coating processing on the substrate W by supplying a processing solution to the substrate W on which a coating has been formed. In this embodiment, although the substrate W to be processed has a diameter of 300 mm, the present invention is not limited thereto. The processing solution includes an etching solution, and the substrate processing units WU perform etching processing. The etching solution is a chemical solution. Examples of etching solutions include hydrofluoric acid (a mixture of hydrofluoric acid (HF) and nitric acid (HNO3)), hydrofluoric acid, buffered hydrofluoric acid (BHF), ammonium fluoride, HFEG (hydrofluoric acid ethylene glycol) (a mixture of hydrofluoric acid and ethylene glycol), or phosphoric acid (H3PO4).

[0047] The substrate processing unit WU includes a rotary chuck SC, a rotary motor SM, a nozzle 311, and a nozzle moving mechanism 301. The rotary chuck SC holds the substrate W horizontally. The substrate W is held by the rotary chuck SC with the center of the substrate W aligned with the first rotation axis AX1 of the rotary motor SM. The rotary motor SM has a first rotation axis AX1. The first rotation axis AX1 extends in the vertical direction. The rotary chuck SC is mounted on the upper end of the first rotation axis AX1 of the rotary motor SM. When the rotary motor SM rotates, the rotary chuck SC rotates about the first rotation axis AX1 as its center. The rotary motor SM is a stepper motor. The substrate W held by the rotary chuck SC rotates about the first rotation axis AX1 as its center. Therefore, the rotational speed of the substrate W is the same as the rotational speed of the stepper motor. Furthermore, if an encoder is provided to generate a speed signal representing the rotational speed of the rotary motor SM, the rotational speed of the substrate W can also be obtained from the speed signal generated by the encoder. In this case, a motor other than a stepper motor can be used for the rotary motor SM.

[0048] The nozzle 311 supplies etching solution to the surface (upper surface) of the substrate W held by the rotating chuck SC. An etching solution supply unit (not shown) supplies etching solution to the nozzle 311. The nozzle 311 sprays etching solution toward the surface of the rotating substrate W.

[0049] The nozzle moving mechanism 301 moves the nozzle 311 generally horizontally. Specifically, the nozzle moving mechanism 301 includes a nozzle motor 303 and a nozzle arm 305, the nozzle motor 303 having a second rotation axis AX2. The nozzle motor 303 is arranged generally vertically with respect to the second rotation axis AX2. The nozzle arm 305 has a linearly extending elongated shape. One end of the nozzle arm 305 is mounted to the upper end of the second rotation axis AX2 in a direction different from the second rotation axis AX2 along its length. The nozzle 311 is mounted at the other end of the nozzle arm 305 with the etchant outlet facing downwards.

[0050] When the nozzle motor 303 is activated, the nozzle arm 305 rotates in the horizontal plane around the second rotation axis AX2. Consequently, the nozzle 311, mounted at the other end of the nozzle arm 305, moves (rotates) horizontally around the second rotation axis AX2. The nozzle 311 sprays etching solution toward the substrate W while moving horizontally. The nozzle motor 303 is, for example, a stepper motor.

[0051] The control device 10 includes a CPU (central processing unit) and a memory. The CPU executes the program stored in the memory, thereby controlling the entire substrate processing device 300. The control device 10 controls the rotary motor SM and the nozzle motor 303.

[0052] Experimental data is input from the substrate processing apparatus 300 to the learning apparatus 200. The learning apparatus 200 uses the experimental data to generate a predictor including a first predictor and a second predictor, and outputs the predictor to the information processing apparatus 100.

[0053] The information processing device 100 uses a predictor to determine the processing conditions for processing the predetermined substrate W to be processed by the substrate processing device 300 next. The information processing device 100 outputs the determined processing conditions to the substrate processing device 300.

[0054] Figure 2 This figure illustrates an example of the structure of the information processing apparatus 100. (Refer to...) Figure 2 The information processing device 100 is composed of a CPU 101, RAM (Random Access Memory) 102, ROM (Read Only Memory) 103, storage device 104, operation unit 105, display device 106, and input / output interface (I / F) 107. The CPU 101, RAM 102, ROM 103, storage device 104, operation unit 105, display device 106, and input / output interface 107 are connected to a bus 108.

[0055] RAM 102 is used as the operating area of ​​CPU 101. The system program is stored in ROM 103. Storage device 104 includes a recording medium such as a hard disk or semiconductor memory, storing the program. The program can also be stored in ROM 103 or other external storage devices.

[0056] The CD-ROM (Compact Disc Read Only Memory) 109 can be installed in and removed from the storage device 104. The recording medium for storing the program executed by the CPU 101 is not limited to the CD-ROM 109; it can also be a semiconductor memory medium such as an optical disc (MO, Mini Disc, DVD), an integrated circuit card, an optical card, a mask ROM, or an EPROM. Furthermore, the program stored in the storage device 104 can be loaded into the RAM 102 and executed by the CPU 101 by either downloading the program from a computer connected to a network and storing it in the storage device 104, or by writing the program to the storage device 104 from a computer connected to a network. The programs mentioned here include not only those that can be directly executed by the CPU101, but also original programs, compressed programs, and encrypted programs.

[0057] The operation unit 105 is an input device such as a keyboard, mouse, or touch panel. Users can issue predetermined instructions to the information processing device 100 by operating the operation unit 105. The display device 106 is a display device such as a liquid crystal display (LCD) that displays a GUI (Graphical User Interface) for receiving user instructions. The input / output interface 107 is connected to a network.

[0058] Figure 3 This is a diagram illustrating an example of the structure of the learning device 200. (Refer to...) Figure 3 The learning device 200 is composed of a CPU 201, RAM 202, ROM 203, storage device 204, operation unit 205, display device 206, and input / output interface 207. The CPU 201, RAM 202, ROM 203, storage device 204, operation unit 205, display device 206, and input / output interface 207 are connected to a bus 208.

[0059] RAM 202 is used as the operating area of ​​CPU 201. The system program is stored in ROM 203. Storage device 204 includes a recording medium such as a hard disk or semiconductor memory, storing the program. The program can also be stored in ROM 203 or other external storage devices. CD-ROM 209 can be mounted on and removed from storage device 204.

[0060] The operation unit 205 is an input device such as a keyboard, mouse, or touch panel. The input / output interface 207 is connected to a network.

[0061] (3) Functional structure of substrate processing system 1

[0062] Figure 4 This figure illustrates an example of the functional structure of a substrate processing system according to one embodiment. (Refer to...) Figure 4 The control device 10 of the substrate processing apparatus 300 controls the substrate processing unit WU and performs a coating process on the substrate W according to the processing conditions during the processing time. The processing time is the time determined for the coating process on the substrate W. In this embodiment, the processing time is the time during which the nozzle 311 sprays etching solution onto the substrate W. The processing conditions are the conditions used by the substrate processing unit WU when performing the coating process.

[0063] The processing conditions include the temperature of the etching solution, the concentration of the etching solution, the flow rate of the etching solution, the rotation speed of the substrate W, and the position of the nozzle 311. The concentration of the etching solution is expressed as the mixing ratio of a plurality of solutions. The position of the nozzle 311 is expressed as the relative position between the nozzle 311 and the substrate W at each of a plurality of moments during the coating process. The relative position between the nozzle 311 and the substrate W is expressed as the rotation angle of the nozzle motor 303.

[0064] The processing conditions include: fixed conditions that do not change over time; and variable conditions that change over time. In this embodiment, the fixed conditions are the temperature of the etching solution, the concentration of the etching solution, the flow rate of the etching solution, and the rotation speed of the substrate W, while the variable condition is the relative position between the nozzle 311 and the substrate W.

[0065] The learning device 200 includes an experimental data acquisition unit 210, a predictor generation unit 220, and a predictor transmission unit 230. The functions of the learning device 200 are implemented by the CPU 201, which executes a learning model generation program stored in RAM 202.

[0066] The experimental data acquisition unit 210 acquires experimental data from the substrate processing apparatus 300. The experimental data includes: processing conditions; and a first processing result, indicating the result of the substrate processing apparatus 300 performing a coating process on the substrate W under the processing conditions.

[0067] Experimental data is input from the experimental data acquisition unit 210 to the predictor generation unit 220. The predictor generation unit 220 uses the experimental data to generate a predictor and outputs the generated predictor to the predictor transmission unit 230. A detailed description of the predictor generation unit 220 will be provided later. The predictor transmission unit 230 transmits the predictor generated by the predictor generation unit 220 to the information processing device 100.

[0068] The information processing apparatus 100 includes a predictor receiving unit 110, a processing condition determining unit 120, a prediction unit 130, an evaluation unit 140, and a processing condition transmitting unit 150. The functions of the information processing apparatus 100 are implemented by the CPU 101, which executes a processing condition determining program stored in RAM 102.

[0069] The predictor receiving unit 110 receives the predictor sent from the learning device 200 and outputs the received predictor to the prediction unit 130. The processing condition determining unit 120 determines the processing conditions for the substrate W, which will be processed by the substrate processing device 300 next, and outputs the variable conditions and fixed conditions included in the processing conditions to the prediction unit 130.

[0070] The prediction unit 130 uses a predictor to infer a first prediction processing result based on variable conditions and fixed conditions. The prediction unit 130 inputs the fixed conditions and the variable conditions input from the processing condition determination unit 120 to the predictor, and outputs the first prediction processing result output by the predictor to the evaluation unit 140.

[0071] Evaluation unit 140 evaluates the first prediction processing result input from prediction unit 130 and outputs the evaluation result to processing condition determination unit 120. Specifically, evaluation unit 140 acquires the pre-processing film thickness characteristics of a predetermined substrate W, which is the target of processing by substrate processing apparatus 300. Evaluation unit 140 calculates the predicted film thickness characteristics after etching processing based on the first prediction processing result input from prediction unit 130 and the pre-processing film thickness characteristics of substrate W, and compares it with the target film thickness characteristics. If the comparison result meets the evaluation criteria, the processing conditions determined by processing condition determination unit 120 are output to processing condition sending unit 150. For example, evaluation unit 140 calculates a deviation characteristic and determines whether the deviation characteristic meets the evaluation criteria. The deviation characteristic is the difference between the film thickness characteristics of the substrate W after etching processing and the target film thickness characteristics. The evaluation criteria can be arbitrarily determined. For example, the evaluation criteria can be that the maximum value of the difference in the deviation characteristic is below a threshold, or that the average value of the difference is below a threshold.

[0072] The processing condition sending unit 150 sends the processing conditions determined by the processing condition determining unit 120 to the control device 10 of the substrate processing apparatus 300. The substrate processing apparatus 300 processes the substrate W according to the processing conditions.

[0073] If the evaluation result does not meet the evaluation criteria, the evaluation unit 140 outputs the evaluation result to the processing condition determination unit 120. The evaluation result includes the film thickness characteristics predicted after etching or the difference between the film thickness characteristics predicted after etching and the target film thickness characteristics.

[0074] The processing condition determination unit 120 determines new processing conditions for the prediction unit 130 to make predictions based on the evaluation results input from the evaluation unit 140. The processing condition determination unit 120 selects one of a plurality of pre-prepared variable conditions using experimental planning, pairwise methods, or Bayesian inference, and determines the processing conditions, including the fixed conditions and the selected variable conditions, as the new processing conditions for the prediction unit 130 to make predictions.

[0075] The processing condition determination unit 120 can also use Bayesian inference to explore the processing conditions. When the evaluation unit 140 outputs a plurality of evaluation results, there are a plurality of groups of processing conditions and evaluation results. The process is explored based on the tendency of the first predicted processing result in each of the plurality of groups: processing conditions that make the film thickness uniform; or processing conditions that minimize the difference between the predicted film thickness characteristics after etching and the target film thickness characteristics.

[0076] Specifically, the processing condition determination unit 120 explores processing conditions in a manner that minimizes the objective function. The objective function is a function representing the uniformity of the coating thickness or a function representing the consistency between the coating thickness characteristics and the target coating thickness characteristics. For example, the objective function is a function that, as a parameter, represents the difference between the predicted coating thickness characteristics and the target coating thickness characteristics. Here, the parameter is the corresponding variation condition. The corresponding variation condition is the variation condition used by the predictor to infer the first prediction processing result. The processing condition determination unit 120 selects the parameter, i.e., the variation condition, determined through exploration from a plurality of variation conditions, and determines a new processing condition that includes the fixed conditions and the selected variation condition.

[0077] Next, the detailed structure of the predictor generation unit 220 will be explained. Figure 5 The diagram illustrates an example of the functions of the predictor generation unit 220. The predictor generation unit 220 includes a processing condition acquisition unit 221, a first processing result acquisition unit 222, a statistical value calculation unit 223, a conversion unit 224, a first learning model generation unit 225, and a second learning model generation unit 226.

[0078] The processing condition acquisition unit 221 acquires processing conditions from the experimental data acquisition unit 210. The processing condition acquisition unit 221 outputs the processing conditions to the first learning model generation unit 225 and the second learning model generation unit 226. The processing conditions include fixed conditions and variable conditions. Fixed conditions include the temperature of the etching solution, the concentration of the etching solution, the flow rate of the etching solution, and the rotational speed of the substrate W. Variable conditions are represented by the relative position between the nozzle 311 and the substrate W at each of a plurality of moments during the coating process. The first processing result acquisition unit 222 acquires a first processing result from the experimental data acquisition unit 210. The first processing result acquired by the first processing result acquisition unit 222 is the difference in film thickness before and after coating at each of a plurality of different radial positions of the substrate W obtained by performing coating processing in the substrate processing apparatus 300 under the processing conditions acquired by the processing condition acquisition unit 221.

[0079] The statistical value calculation unit 223 performs statistical processing on the first processing result obtained by the first processing result acquisition unit 222, thereby calculating a statistical value. In this embodiment, the statistical value is the average value of the difference in film thickness before and after the coating process at each of a plurality of different positions in the radial direction of the substrate W.

[0080] The first processing result and statistical values ​​are input from the statistical value calculation unit 223 to the conversion unit 224. The conversion unit 224 uses the statistical values ​​to convert the first processing result into a second processing result. In this embodiment, the conversion unit 224 divides the first processing result by the average of the differences in film thickness before and after processing at each of a plurality of different locations in the radial direction of the substrate W, thereby calculating the second processing result. As a result, the difference in scale between the plurality of second processing results becomes smaller than the difference in scale between the plurality of first processing results. The scale is the average value of the first processing results. Furthermore, the scale is not limited to the average value; it can also be expressed as the difference between the maximum and minimum values, i.e., the range, or it can be expressed as the average value and the dispersion.

[0081] Figure 6 This is a graph used to illustrate the results of the first processing step. Figure 6 In the diagram, the vertical axis represents the film thickness, and the horizontal axis represents the radial position of the substrate W. Furthermore, the origin of the horizontal axis represents the center of the substrate W. Solid lines represent the film thickness formed on the substrate W before processing by the substrate processing apparatus 300. The substrate processing apparatus 300 performs a coating process for applying an etching solution according to processing conditions, thereby adjusting the film thickness of the coating formed on the substrate W. Dashed lines represent the film thickness formed on the substrate W after the coating process performed by the substrate processing apparatus 300.

[0082] The difference between the film thickness of the film formed on the substrate W before processing by the substrate processing apparatus 300 and the film thickness of the film formed on the substrate W after processing by the substrate processing apparatus 300 is the first processing result (etching amount). In other words, the first processing result represents the reduction in film thickness at each of a plurality of locations in the radial direction of the substrate W due to the coating process of the substrate processing apparatus 300.

[0083] It is desirable that the film thickness formed by the substrate processing apparatus 300 is uniform across the entire surface of the substrate W. Therefore, a target film thickness is determined for the processing performed by the substrate processing apparatus 300. The target film thickness is represented by a dashed line. The deviation characteristic is the difference between the film thickness formed on the substrate W after processing by the substrate processing apparatus 300 and the target film thickness. The deviation characteristic includes the difference at each of a plurality of locations in the radial direction of the substrate W.

[0084] Figure 7 This is a diagram showing an example of the first processing result obtained by the coating process of the substrate processing apparatus 300. Figure 7 The vertical axis represents the difference in film thickness between the film formed on substrate W before coating and the film formed on substrate W after coating. The horizontal axis represents the radial position of substrate W. (Refer to...) Figure 7The diagram shows the first processing results ERa to ERc after the substrate processing apparatus 300 performed a coating process under three different processing conditions. The first processing results ERa to ERc each have different dimensions. The difference in dimensions between the first processing results ERa to ERc is due to the differences in processing conditions.

[0085] Figure 8 A diagram illustrating an example of the second processing result. Figure 8 It is shown that the conversion unit 224 has been used to convert the data. Figure 7 The first processing result, ERa, is transformed into ERc, resulting in a second processing result, ERA, which is then converted to ERC. In the second processing result, the scale of the vertical axis values ​​is normalized for the ERA to ERC results. Figure 8 In the process, since the scale is standardized among the second processing results ERA, ERB, and ERC, it can be known that the deviation of the vertical axis values ​​among the second processing results ERA, ERB, and ERC is greater than that of the first processing results ERA. Figure 7 The deviation of the vertical axis values ​​among ERa, ERb, and ERc in the first processing result shown is small.

[0086] When the first processing result is used as teacher data, conditions with larger vertical axis values ​​(larger membrane thickness differences) are relatively emphasized in machine learning, while conditions with smaller vertical axis values ​​(smaller membrane thickness differences) are relatively disregarded. On the other hand, when the second processing result is used as teacher data, because the scale of the vertical axis values ​​is standardized, the weights in machine learning are reflected and applied equally regardless of the magnitude of the membrane thickness difference. Therefore, even when considering the processing condition of smaller membrane thickness differences and when the membrane thickness difference is small, learning results capable of high-precision prediction can be obtained.

[0087] In this embodiment, since the statistical value is the average of the first processing results, the first processing results ERa, ERb, and ERc are divided by the average of the first processing results ERa, ERb, and ERc, thereby unifying the scale of the second processing results ERA, ERB, and ERC to 1.

[0088] Return to Figure 5The first learning model generation unit 225 uses learning data to enable the first learning model to learn machine. Processing conditions are input to the first learning model generation unit 225 from the processing condition acquisition unit 221, and a second processing result is input to the first learning model generation unit 225 from the conversion unit 224. The learning data includes input data as explanatory variables and positive solution data as target variables. The input data includes fixed conditions and variable conditions included in the processing conditions. The positive solution data includes the second processing result. The first learning model generation unit 225 inputs the input data into the first learning model and determines the parameters of the first learning model by minimizing the difference between the second predicted processing result inferred by the first learning model and the positive solution data, i.e., the second processing result.

[0089] The first learning model used by the first learning model generation unit 225 to perform machine learning includes a neural network. The first learning model generation unit 225 determines the values ​​of a plurality of parameters used to construct the neural network. The first learning model defines a loss function and uses gradient descent to update the weights in the direction where the loss decreases, thereby performing learning. During learning, the value of the loss function is calculated for each of the positive data, i.e., a plurality of second processing results, and the average of the calculated values ​​of the plurality of loss functions is defined as the loss of the learning model. In the loss function, if the scale of the second processing result assigned as positive data is larger, the output value of the loss function for that second processing result is larger; conversely, if the scale of the second processing result assigned as positive data is smaller, the value of the loss function for that second processing result is smaller. The difference in scale between the values ​​of the plurality of second processing results is smaller than the difference in scale between the values ​​of the plurality of first processing results. Therefore, since the first learning model generation unit 225 enables the first learning model to learn a plurality of second processing results, the accuracy of the first learning model is improved compared to the case of learning a plurality of first processing results.

[0090] The second learning model generation unit 226 uses learning data to enable the second learning model to learn machine. Processing conditions are input to the second learning model generation unit 226 from the processing condition acquisition unit 221, and statistical values ​​are input to the second learning model generation unit 226 from the statistical value calculation unit 223. The learning data includes input data as explanatory variables and positive solution data as target variables. The input data includes fixed conditions and variable conditions included in the processing conditions. The positive solution data includes statistical values ​​calculated by the statistical value calculation unit 223.

[0091] The second learning model generation unit 226 inputs input data into the second learning model and determines the parameters of the second learning model by reducing the difference between the predicted statistical value inferred by the second learning model and the correct answer data, i.e., the statistical value. Specifically, the second learning model generation unit 226 inputs input data into the second learning model and causes the second learning model to infer predicted statistical values. The second learning model used by the second learning model generation unit 226 to learn includes a neural network. The second learning model generation unit 226 determines the values ​​of a plurality of parameters used to constitute the neural network. The second learning model generation unit 226 calculates the difference between the predicted statistical value inferred by the second learning model and the statistical value calculated by the statistical value calculation unit 223 as the error, and causes the second learning model to learn by reducing the calculated error.

[0092] The predictor generation unit 220 generates: a first predictor, which incorporates parameters set on a first learned model after learning; and a second predictor, which incorporates parameters set on a second learned model after learning. The first predictor is an inference program that incorporates the parameters set on the first learned model. The second predictor is an inference program that incorporates the parameters set on the second learned model.

[0093] Figure 9 This is a diagram used to illustrate the first learning model. (See reference...) Figure 9 The first learning model has layers A1 to C1 arranged sequentially from the input side to the output side (from the top layer to the bottom layer). A first convolutional neural network CNN1 is set in layer A1, a first fully connected neural network NN1 is set in layer B1, and a second convolutional neural network CNN2 is set in layer C1.

[0094] Input the variable conditions into the first convolutional neural network CNN1. Input the output of the first convolutional neural network CNN1 and the fixed conditions into the first fully connected neural network NN1. Input the output of the first fully connected neural network NN1 into the second convolutional neural network CNN2.

[0095] The first convolutional neural network CNN1 includes a plurality of layers. In this embodiment, the first convolutional neural network CNN1 includes three layers. Within the first convolutional neural network CNN1, a first layer L1a, a second layer L1b, and a third layer L1c are arranged sequentially from the input side (upper layer side) to the output side (lower layer side). Furthermore, this embodiment describes the case of including three layers as a plurality of layers, but it may also include more than three layers.

[0096] The first layer L1a, the second layer L1b, and the third layer L1c each include convolutional layers and pooling layers. Multiple filters are applied in the convolutional layers. The pooling layers compress the output of the convolutional layers. The number of filters in the convolutional layer of the second layer L1b is set to twice the number of filters in the convolutional layer of the first layer L1a. The number of filters in the convolutional layer of the third layer L1c is set to twice the number of filters in the convolutional layer of the second layer L1b. Therefore, as many features as possible can be extracted from the changing conditions. Here, the changing conditions include the relative position of the nozzle relative to the substrate W, which changes over time. Since the first convolutional neural network CNN1 uses multiple filters to extract features, multiple features including time-related factors are extracted for the change in the relative position of the nozzle relative to the substrate W. Furthermore, although an example is shown here where the number of filters in the convolutional layer of the second layer L1b is set to twice the number of filters in the convolutional layer of the first layer L1a, it is not necessary to set it to twice the number of filters. The number of filters in the convolutional layer of the second layer L1b can be any number greater than the number of filters in the convolutional layer of the first layer L1a. Furthermore, the number of filters in the third convolutional layer L1c does not necessarily have to be twice the number of filters in the second convolutional layer L1b. The number of filters in the third convolutional layer L1c can be more than the number of filters in the second convolutional layer L1b.

[0097] The first fully connected neural network NN1 has a plurality of layers. Figure 9 In the example, the first fully connected neural network NN1 has two layers: a b1a layer on the input side and a b1b layer on the output side. Figure 9 In the example, each layer includes a plurality of nodes. Although in Figure 9 In the example, five nodes are shown in layer b1a and four nodes in layer b1b, but the number of nodes is not limited to this. The number of nodes in layer b1a is set such that it equals the sum of the number of nodes on the output side of the first convolutional neural network CNN1 and the number of fixed conditions. The number of nodes in layer b1b is set such that it equals the number of nodes on the input side of the second convolutional neural network CNN2. The outputs of the nodes in layer b1a are connected to the inputs of the nodes in layer b1b. Parameters include weighting coefficients for the outputs of the nodes in layer b1a. One or more intermediate layers can also be set between layers b1a and b1b.

[0098] The second convolutional neural network CNN2 includes a plurality of layers. In this embodiment, the second convolutional neural network CNN2 includes three layers. In the second convolutional neural network CNN2, a fourth layer L1d, a fifth layer L1e, and a sixth layer L1f are arranged sequentially from the input layer (upper layer side) to the output layer (lower layer side). Furthermore, although this embodiment describes the case of including three layers as a plurality of layers, it may also include more than three layers.

[0099] The fourth layer L1d, the fifth layer L1e, and the sixth layer L1f each include convolutional layers and pooling layers. Multiple filters are applied in the convolutional layers. The pooling layers compress the output of the convolutional layers. The number of filters in the convolutional layer of the fifth layer L1e is set to half the number of filters in the convolutional layer of the fourth layer L1d. Furthermore, the number of filters in the convolutional layer of the sixth layer L1f is set to half the number of filters in the convolutional layer of the fifth layer L1e. Therefore, as many features as possible can be extracted from the second processing result. The multiple processing values ​​included in the second processing result are values ​​with different positions in the radial direction of the substrate W. Since the second convolutional neural network CNN2 uses multiple filters, multiple features including the radial position of the substrate W are extracted from the second processing result. Furthermore, although an example is shown here where the number of filters in the convolutional layer of the fifth layer L1e is set to half the number of filters in the convolutional layer of the fourth layer L1d, it is not necessary to set it to half. The number of filters in the fifth convolutional layer L1e should be less than the number of filters in the fourth convolutional layer L1d. Furthermore, the number of filters in the sixth convolutional layer L1f does not necessarily have to be half the number of filters in the fifth convolutional layer L1e. The number of filters in the sixth convolutional layer L1f should be less than the number of filters in the fifth convolutional layer L1e.

[0100] Figure 10 This is a diagram used to illustrate the second learning model. (See reference...) Figure 10 The second learning model consists of layers A2 and B2 arranged sequentially from the input side to the output side (from the top layer to the bottom layer). A third convolutional neural network CNN3 is set in layer A2, and a second fully connected neural network NN2 is set in layer B2.

[0101] Input the variable conditions into the third convolutional neural network CNN3. Input the output of the third convolutional neural network CNN3 and the fixed conditions into the second fully connected neural network NN2.

[0102] The third convolutional neural network (CNN3) includes a plurality of layers. In this embodiment, the third convolutional neural network (CNN3) includes three layers. Within the third convolutional neural network (CNN3), a first layer L2a, a second layer L2b, and a third layer L2c are arranged sequentially from the input side (upper layer side) to the output side (lower layer side). Furthermore, although this embodiment describes the case of including three layers as a plurality of layers, it may also include more than three layers.

[0103] The first layer L2a, the second layer L2b, and the third layer L2c each include convolutional layers and pooling layers. Multiple filters are applied in the convolutional layers. The pooling layers compress the output of the convolutional layers. The number of filters in the convolutional layer of the second layer L2b is set to twice the number of filters in the convolutional layer of the first layer L2a. The number of filters in the convolutional layer of the third layer L2c is set to twice the number of filters in the convolutional layer of the second layer L2b. Therefore, as many features as possible can be extracted from the changing conditions. Here, the changing conditions include the relative position of the nozzle relative to the substrate W, which changes over time. Since the third convolutional neural network CNN3 uses multiple filters to extract features, more features including time-related elements are extracted for changes in the relative position of the nozzle relative to the substrate W. Furthermore, although an example is shown here where the number of filters in the convolutional layer of the second layer L2b is set to twice the number of filters in the convolutional layer of the first layer L2a, it is not necessary to set it to twice the number of filters. The number of filters in the convolutional layer of the second layer L2b can simply be more than the number of filters in the convolutional layer of the first layer L2a. Furthermore, the number of filters in the third L2c convolutional layer does not necessarily have to be twice the number of filters in the second L2b convolutional layer. The number of filters in the third L2c convolutional layer can simply be more than the number of filters in the second L2b convolutional layer.

[0104] The second fully connected neural network NN2 has a plurality of layers. Figure 9 In the example, the second fully connected neural network NN2 has five layers arranged sequentially from the input side to the output side: b2a, b2b, b2c, b2d, and b2e. Figure 9 In the example, there are multiple nodes at each level. Although in Figure 9 In the example, five nodes are shown in layer b2a, four nodes in layer b2b, three nodes in layer b2c, two nodes in layer b2d, and one node in layer b2e, but the number of nodes is not limited to these. The number of nodes in layer b2a is set to be equal to the sum of the number of nodes on the output side of the third convolutional neural network CNN3 and the number of fixed conditions. The number of nodes in layer b2e is set to be equal to the amount of data being predicted. Since the data predicted by the second learning model is statistical value, the number of nodes in layer b2e is set to 1 here. Between layers b2c and b2e, in addition to layers b2b to b2c, there can also be many intermediate layers.

[0105] In this embodiment, the filters set in the first convolutional neural network CNN1 of the first learning model and the filters set in the third convolutional neural network CNN3 of the second learning model are set to be the same. Since the first learning model and the second learning model learn the same changing conditions respectively, the learning of changing conditions can be correlated between the first learning model and the second learning model.

[0106] Next, the detailed functions of the prediction unit 130 will be explained. Figure 11 This diagram illustrates an example of the function of the prediction unit 130. The prediction unit 130 includes a processing condition acquisition unit 131, a first prediction unit 133, a second prediction unit 135, and a restoration unit 137. The processing condition acquisition unit 131 acquires the processing conditions output by the processing condition determination unit 120. Furthermore, the processing condition acquisition unit 131 outputs the acquired processing conditions to the first prediction unit 133 and the second prediction unit 135.

[0107] The first predictor is input from the predictor receiving unit 110 to the first predictor unit 133, and processing conditions are input from the processing condition acquisition unit 131 to the first predictor unit 133. The first predictor unit 133 inputs the processing conditions to the first predictor and causes the first predictor to predict a second prediction processing result. The first predictor unit 133 outputs the second prediction processing result predicted by the first predictor to the restoration unit 137.

[0108] The second predictor is input from the predictor receiving unit 110 to the second predictor unit 135, and processing conditions are input from the processing condition acquisition unit 131 to the second predictor unit 135. The second predictor unit 135 inputs the processing conditions to the second predictor and causes the second predictor to predict statistical values. The second predictor unit 135 outputs the predicted statistical values ​​predicted by the second predictor to the recovery unit 137.

[0109] The restoration unit 137 restores the first prediction processing result based on the second prediction processing result output from the first prediction unit 133 and the prediction statistics output from the second prediction unit 135. In this embodiment, the restoration unit 137 restores the first prediction processing result by multiplying the second prediction processing result and the prediction statistics. The restoration unit 137 outputs the restored second prediction processing result to the evaluation unit 140.

[0110] Figure 12 The flowchart illustrates an example of the learning process. The learning process is a process executed by the CPU 201 of the learning device 200, which executes a learning model generation program stored in RAM 202.

[0111] Reference Figure 12The CPU 201 of the learning device 200 acquires experimental data. The CPU 201 controls the input / output interface 107 to acquire experimental data from the substrate processing device 300 (step S01). Experimental data can also be acquired by reading experimental data recorded on a recording medium such as a CD-ROM 209 using the storage device 104. The acquired experimental data is multiple. The experimental data includes processing conditions and a first processing result of coating processing performed in the substrate processing device 300 under those processing conditions.

[0112] In the next step S02, a predictor including a first predictor and a second predictor is generated based on the acquired experimental data, and the process proceeds to step S03. In step S03, the CPU 201 sends the predictor to the information processing device 100 through the control input / output interface 107.

[0113] Figure 13 The flowchart illustrates an example of the predictor generation process. Predictor generation is a process performed by CPU 201 in step S02 of the learning process. Experimental data is acquired in the stage prior to the execution of the predictor generation process.

[0114] Reference Figure 13 The CPU 201 of the learning device 200 acquires a dataset of multiple groups, including processing conditions and a first processing result, from the experimental data (step S11). In step S12, a first processing condition is selected, and the process proceeds to step S13. A dataset is selected from the acquired dataset of multiple groups, and the processing conditions included in the selected dataset are selected.

[0115] In step S13, a first processing result corresponding to the selected first processing condition is selected. In step S14, a statistical value is calculated based on the selected first processing result. The statistical value is the average of the difference in film thickness before and after processing at each of a plurality of different locations in the radial direction of the substrate W. In step S15, the selected first processing result is converted into a second processing result using the calculated average value. The second processing result, obtained by dividing the first processing result by the average value, is calculated.

[0116] In step S16, CPU 201 uses the selected processing conditions as input data and the transformed second processing result as the forward solution data to enable the first learning model to learn, and then proceeds to step S17. Here, CPU 201 inputs the input data into the first learning model and determines the filter and parameters in a way that the error between the output of the first learning model and the forward solution data decreases. Thus, the filter and parameters of the first learning model are adjusted.

[0117] In step S17, CPU 201 uses the selected processing conditions as input data and the statistical values ​​calculated from the first processing result as positive solution data to enable the second learning model to learn, thus proceeding to step S18. Here, the input data is fed into the second learning model, and the filter and parameters are determined in such a way that the error between the output of the second learning model and the positive solution data decreases. Therefore, the filter and parameters of the second learning model are adjusted.

[0118] In step S18, it is determined whether there are any processing conditions that have not been selected as processing objects. If there are unselected processing conditions, the processing returns to step S12; if there are no unselected processing conditions, the processing ends and returns to the learning process. By repeatedly performing the processing steps S12 to S17, a first predictor and a second predictor are generated.

[0119] Figure 14 This is a flowchart illustrating an example of a processing condition determination process. The processing condition determination process is executed by the CPU 101 of the information processing device 100, which executes a processing condition determination program stored in RAM 102. The CPU 101 of the information processing device 100 selects a processing condition from a plurality of pre-prepared processing conditions (step S21) and proceeds to step S22. Here, a processing condition is selected from the plurality of pre-prepared processing conditions using methods such as experimental planning, pairwise methods, or Bayesian inference.

[0120] In step S22, the first predictor is used to infer the second prediction processing result based on the processing conditions, and the process proceeds to step S23. Here, processing conditions including variable conditions and fixed conditions are input to the first predictor, and the second prediction processing result output by the first predictor is obtained. In step S23, the second predictor is used to infer the prediction statistics based on the processing conditions, and the process proceeds to step S24. Here, processing conditions including variable conditions and fixed conditions are input to the second predictor, and the prediction statistics output by the second predictor are obtained.

[0121] In step S24, the second prediction result predicted in step S22 is restored to the first prediction result using the predicted statistical value predicted in step S23. The first prediction result is restored by multiplying the second prediction result predicted in step S22 with the predicted statistical value predicted in step S23.

[0122] In step S25, the restored first prediction processing result is compared with the target film thickness characteristic. The film thickness characteristic of the substrate W to be processed by the substrate processing apparatus 300 before processing is calculated based on the first prediction processing result restored in step S24. Then, the processed film thickness characteristic is compared with the target film thickness characteristic. Here, the difference between the processed film thickness characteristic and the target film thickness characteristic is calculated.

[0123] In step S26, it is determined whether the comparison result meets the evaluation criteria. If the comparison result meets the evaluation criteria (yes in step S26), the process proceeds to step S27; otherwise, the process proceeds to step S21. Here, for example, if the maximum value of the difference is below a threshold, it is determined that the evaluation criteria are met. Additionally, if the average value of the difference is below a threshold, it is determined that the evaluation criteria are met.

[0124] In step S27, processing conditions including the variable conditions selected in step S21 are set in the candidate processing conditions for the drive substrate processing device 300, and the process proceeds to step S28. In step S28, it is determined whether an end instruction for exploration has been accepted. If the end instruction has been accepted by the user of the operation information processing device 100, the process proceeds to step S29; otherwise, the process returns to step S21. Alternatively, the end instruction may be changed to one input by the user, and it is determined whether a pre-defined number of processing conditions have been set as candidates.

[0125] In step S29, a processing condition is selected from one or more candidate processing conditions, and the process proceeds to step S30. Alternatively, the user of the operation information processing device 100 can select a processing condition from one or more candidate processing conditions. Therefore, the user has a wider range of choices. Furthermore, the simplest nozzle operation variation condition can be automatically selected from the variation conditions included in the plurality of processing conditions. For example, the variation condition with the fewest speed change points can be selected. Thus, the processing results for the complex nozzle operation used to process the substrate W can be displayed under multiple variation conditions. If a variation condition that facilitates nozzle control is selected from the plurality of variation conditions, the control of the substrate processing device 300 becomes easier.

[0126] In step S30, the processing conditions determined in step S29 are sent to the substrate processing apparatus 300, and the processing ends. The CPU 101 controls the input / output interface 107 to send the processing conditions to the substrate processing apparatus 300. Upon receiving the processing conditions from the information processing apparatus 100, the substrate processing apparatus 300 processes the substrate W according to the processing conditions.

[0127] (4) Specific examples

[0128] In this embodiment, the variation condition is time-series data obtained by sampling at a nozzle action processing time of 60 seconds and a sampling interval of 0.01 seconds. The variation condition consists of 6001 values. Therefore, the variation condition can represent complex nozzle actions. In particular, it can accurately represent nozzle actions with a large number of speed change points that alter the nozzle's movement speed. Conversely, because the variation condition has a high dimensionality, overfitting may sometimes occur when a fully connected neural network model learns from the time-series data of the variation condition.

[0129] In this embodiment, the predictor generation unit 220 includes... Figure 9 as well as Figure 10 The convolutional neural network learning model shown has both variable and fixed conditions for machine learning. The inventors discovered through experiments that: Figure 9 as well as Figure 10 The first and second learning models shown learn from the varying and fixed conditions consisting of 6001 values ​​representing complex nozzle actions to generate a predictor, so that the processing results predicted by the predictor can become appropriate results.

[0130] Furthermore, in this embodiment, since the processing condition determination unit 120 explores processing conditions corresponding to situations different from the first processing result when exploring appropriate processing conditions, processing conditions corresponding to different plurality of first processing results are selected. Therefore, the processing condition determination unit 120 can efficiently explore processing conditions that predict the first processing result as the target from a plurality of processing conditions.

[0131] Furthermore, although the example of setting the sampling interval to 0.01 seconds was given, the sampling interval is not limited to this. It can also be set to a sampling interval longer than 0.01 seconds, or shorter than 0.01 seconds. For example, the sampling interval can also be set to 0.1 seconds, or it can be set to 0.005 seconds.

[0132] (5) Technical effects of the implementation method

[0133] According to the learning apparatus 200 of the above embodiment, statistical values ​​are calculated through statistical processing of the first processing result, and the first processing result is converted into a second processing result using these statistical values. Therefore, the scale of the plurality of second processing results learned by the first learning model is unified before the first learning model learns. Therefore, the accuracy of the first learning model can be improved. Furthermore, the second learning model learns the unified statistical values ​​used for the scale. Therefore, the first processing result before scale unification can be inferred based on the value inferred from a processing condition using the learned first learning model and the value inferred from a processing condition using the learned second learning model.

[0134] Furthermore, since the fixed conditions and the changing conditions that vary over time are mechanically learned in order to process the substrate W, it is possible to generate a learning model that can predict the processing results based on complex processing conditions.

[0135] Furthermore, since the variation condition is a value that changes over time, features considering time can be extracted using the first convolutional neural network CNN1 and the third convolutional neural network CNN3. Additionally, since the number of learning parameters input to the first fully connected neural network NN1 and the second fully connected neural network NN2 can be reduced by using the first convolutional neural network CNN1 and the third convolutional neural network CNN3, the generalization performance of the first and second learning models can be improved. Moreover, since the second processing result is determined at several different positions in the radial direction of the substrate W, features considering the radial position of the substrate W can be extracted using the second convolutional neural network CNN2. Therefore, the generalization performance of the first and second learning models can be improved.

[0136] Furthermore, according to the information processing apparatus 100 of the above embodiment, temporary processing conditions are assigned to a first learning model, and a second predicted processing result is inferred using the first learning model. Similarly, temporary processing conditions are assigned to a second learning model, and a predicted statistical value is inferred using the second learning model. Moreover, if the first predicted processing result, reconstructed from the second predicted processing result and the predicted statistical value, satisfies an allowable condition, the temporary processing conditions are determined as processing conditions for driving the substrate processing apparatus. Therefore, a plurality of temporary processing conditions can be determined for a first predicted processing result that satisfies an allowable condition. As a result, a plurality of processing conditions can be indicated for the processing result of a complex process used to process the substrate W.

[0137] (6) Other implementation methods

[0138] (6-1) Figure 15 This diagram illustrates an example of the functionality of the predictor generation unit in other embodiments. Figure 15 The predictor generation unit 220A in the other embodiments shown has the same functions as Figure 5 The differences in functionality shown are: the first learning model generation unit 225 is changed to a multi-task learning model generation unit 227; and the second learning model generation unit 226 is deleted. Due to other functions... Figure 5 The functions shown are the same, so they will not be explained again here.

[0139] Processing conditions are input from the processing condition acquisition unit 221 to the multi-task learning model generation unit 227, and statistical values ​​and second processing results are input from the conversion unit 224 to the multi-task learning model generation unit 227. The multi-task learning model generation unit 227 inputs the input data into the multi-task learning model and enables the multi-task learning model to predict the processing results and predict statistical values.

[0140] Figure 16 This is a diagram used to illustrate an example of a multi-task learning model. (See reference...) Figure 16 The multi-task learning model has layers A through C arranged sequentially from the input side to the output side (from the top layer to the bottom layer). Layer A contains a first convolutional neural network (CNN1), layer B contains a first fully connected neural network (NN1), and layer C contains a second convolutional neural network (CNN2) and a second fully connected neural network (NN2).

[0141] The variable conditions are input into the first convolutional neural network CNN1. The output of the first convolutional neural network CNN1 and the fixed conditions are input into the first fully connected neural network NN1. A portion of the output of the first fully connected neural network NN1 is input into the second convolutional neural network CNN2, and the remaining portion of the output of the first fully connected neural network NN1 is input into the second fully connected neural network NN2. The second convolutional neural network CNN2 outputs the second processing result, and the second fully connected neural network NN2 outputs a statistical value.

[0142] Return to Figure 15 The multi-task learning model generation unit 227 calculates the error by taking the difference between the second prediction processing result and the predicted statistical value inferred by the multi-task learning model and the correct solution data, i.e., the second processing result and the statistical value, as the error, and then trains the multi-task learning model to reduce the calculated error. Specifically, the multi-task learning model generation unit 227 updates the values ​​of the weight parameters determined by the plurality of filters of the first convolutional neural network CNN1 and the plurality of nodes of the first fully connected neural network NN1, the values ​​of the weight parameters determined by the plurality of nodes of the second fully connected neural network NN2, and the values ​​of the weight parameters determined by the plurality of filters of the second convolutional neural network CNN2 and the plurality of nodes of the third fully connected neural network NN3.

[0143] (6-2) In the above embodiment, although the average value of the first processing result is used as the statistical value obtained by the statistical value calculation unit 223 statistically processing the first processing result, it is not limited to this. The statistical value is a value in which the difference between the scales of a plurality of second processing results is smaller than the scales of a plurality of first processing results and the value of the first processing result can be restored from the second processing results.

[0144] As an example of a statistical value, the maximum value, minimum value, and median value can be used instead of the average value. Similarly to the case of the average value, in this case, the conversion unit 224 calculates the second processing result by dividing the first processing result by the statistical value.

[0145] Furthermore, the conversion process performed by the conversion unit 224 may also include a process for normalizing the first processing result. Specifically, the conversion unit 224 performs a process of dividing the deviation from the minimum value of the first processing result by the range of the first processing result (= maximum value - minimum value). In this case, the statistical values ​​are the maximum and minimum values, and there are multiple such values.

[0146] Furthermore, the conversion unit 224 may also include a process for standardizing the first processing result. Specifically, the conversion unit 224 performs a process of dividing the deviation from the average value of the first processing result by the standard deviation. In this case, the statistical values ​​are the average value and the standard deviation, and there are multiple such values.

[0147] (7) Summary of implementation methods

[0148] (Item 1) A learning device according to an embodiment of the present invention, wherein, have: The processing condition acquisition unit acquires processing conditions for driving the substrate processing apparatus, which performs coating processing by supplying a processing liquid to a substrate on which a coating is formed. The first processing result acquisition unit acquires a first processing result after driving the substrate processing apparatus to perform the coating process under the processing conditions. The first processing result includes the difference in film thickness before and after the coating process at each of a plurality of radially different positions of the substrate. The statistical value calculation unit statistically processes the first processing result and calculates the statistical value; The conversion unit uses the statistical values ​​to convert the first processing result into a second processing result. A first learning model generation unit generates a first learning model by machine learning the processing conditions and the second processing result; and The second learning model generation unit generates a second learning model by machine learning the processing conditions and the statistical values.

[0149] According to the learning apparatus described in claim 1, a statistical value is calculated through statistical processing of the first processing result, and the first processing result is transformed into a second processing result using this statistical value. Therefore, the scale of the plurality of second processing results learned by the first learning model is unified before the first learning model learns. Therefore, the accuracy of the first learning model can be improved. Furthermore, the second learning model learns the unified statistical value used for the scale. Therefore, the first processing result before the scale is unified can be inferred based on the value inferred from a processing condition by the learned first learning model and the value inferred from a processing condition by the learned second learning model. As a result, a learning apparatus can be provided that can generate a learning model with improved accuracy in inferring processing results based on processing conditions.

[0150] (Item 2) In the learning device described in Item 1, The processing conditions may also include: The conditions change over time; and Fixed conditions other than the aforementioned variable conditions; The first learning model may also include: The first convolutional neural network is input with the changing conditions in the processing conditions; A first fully connected neural network is input to the output of the first convolutional neural network and the fixed condition in the processing conditions; and The second convolutional neural network is input into the output of the first fully connected neural network and outputs the second processing result. The second learning model may also include: A third convolutional neural network is input to the changing conditions in the processing conditions; and The second fully connected neural network is fed into the output of the third convolutional neural network and the fixed condition in the processing conditions, and outputs the statistical value.

[0151] According to the learning apparatus described in item 2, since the variation condition is a value that changes over time, features considering time factors can be extracted by using the first and third convolutional neural networks. Furthermore, by using the first and third convolutional neural networks, the number of learning parameters input to the first and second fully connected neural networks, respectively, can be reduced, thus improving the generalization performance of the first and second learning models. Additionally, since the second processing result is determined at each of a plurality of radially different positions on the substrate, features considering radial position factors of the substrate are extracted by using the second convolutional neural network. Therefore, the generalization performance of the first and second learning models can be improved.

[0152] (Item 3) In the learning device described in Item 2, The first convolutional neural network and the third convolutional neural network have the same filters.

[0153] According to the learning apparatus described in item 3, since the number of filters in the first convolutional neural network of the first learning model and the number of filters in the third convolutional neural network of the second learning model are the same, the conditions for extracting features based on the same processing conditions are set to be the same in both the first and second learning models. Therefore, it is possible to establish a correlation between the machine learning performed by the first learning model and the machine learning performed by the second learning model.

[0154] (Item 4) A learning device according to another embodiment of the present invention, wherein, have: The processing condition acquisition unit acquires processing conditions for driving the substrate processing apparatus, which performs coating processing by supplying a processing liquid to a substrate on which a coating is formed. The first processing result acquisition unit acquires a first processing result after driving the substrate processing apparatus to perform the coating process under the processing conditions. The first processing result includes the difference in film thickness before and after the coating process at each of a plurality of different radial positions of the substrate. The statistical value calculation unit statistically processes the first processing result and calculates the statistical value; The conversion unit uses the statistical value to convert the first processing result into a second processing result; and The multi-task learning model generation unit generates a multi-task learning model from machine learning data, wherein the learning data includes the processing conditions, the second processing result, and the statistical values.

[0155] According to the learning apparatus described in item 4, a statistical value is calculated through statistical processing of the first processing result, and the first processing result is transformed into a second processing result using this statistical value. Therefore, the scale of the second processing result learned by the multi-task learning model is unified before the multi-task learning model learns it. Therefore, the accuracy of the multi-task learning model can be improved. Furthermore, the multi-task learning model learns the unified statistical value for the scale. Therefore, the first processing result before the scale is unified can be inferred based on the second processing result inferred by the learned multi-task learning model and the statistical value. As a result, a learning apparatus can be provided that can generate a multi-task learning model with improved accuracy in inferring processing results based on processing conditions.

[0156] (Item 5) In the learning device described in Item 4, The processing conditions may also include: The conditions change over time; and Fixed conditions other than the aforementioned variable conditions; The multi-task learning model may also include: The first convolutional neural network is input with the changing conditions in the processing conditions; The first fully connected neural network is input to the output of the first convolutional neural network and the fixed condition in the processing conditions; The second convolutional neural network, which receives a portion of the output of the first fully connected neural network, outputs the second processing result; and The second fully connected neural network, which is input into the other part of the output of the first fully connected neural network, outputs the statistical value.

[0157] According to the learning apparatus described in item 5, since the variation condition is a value that changes over time, features considering time factors can be extracted using a first convolutional neural network. Furthermore, since the number of learning parameters also input to the first fully connected neural network can be reduced by using the first convolutional neural network, the generalization performance of the first fully connected neural network can be improved. Moreover, since a second processing result is determined for each of a plurality of radially different positions of the substrate, features considering radial position factors of the substrate can be extracted using a second convolutional neural network. Therefore, the generalization performance of the multi-task learning model can be improved.

[0158] (Item 6) In the learning device described in any one of Items 2, 3, or 5, The substrate processing apparatus can also supply the processing liquid to the substrate by moving a nozzle that supplies the processing liquid to the substrate. The variation conditions may also include nozzle movement conditions, which represent the relative position of the nozzle with respect to the substrate as it changes over time.

[0159] According to the learning apparatus described in item 6, the nozzle movement conditions are input into the first convolutional neural network. Therefore, a learning model capable of predicting processing results can be generated based on the nozzle movement conditions.

[0160] (Item 7) An information processing apparatus according to another embodiment of the present invention, used for managing a substrate processing apparatus, wherein, The substrate processing apparatus supplies processing liquid to the substrate with the coating to perform the coating process under processing conditions that vary over time. The information processing apparatus has a processing condition determination unit that uses a predictor to determine the processing conditions for driving the substrate processing apparatus. The predictor predicts a processing result that represents the difference in film thickness before and after the coating process is performed on the substrate before the coating process is performed by the substrate processing apparatus. The predictor has: The first learning model is generated through machine learning using statistical values ​​obtained by statistically processing a first processing result after performing the coating process with the substrate processing apparatus driven by the processing conditions, transforming the first processing result into a second processing result, and the processing conditions. The first processing result includes the difference in film thickness before and after the coating process at each of a plurality of radially different locations on the substrate; and The second learning model is generated through machine learning of the processing conditions and the statistical values. The processing condition determination unit includes a recovery unit, which recovers the first prediction processing result based on a second predicted processing result inferred by assigning temporary processing conditions to the first learning model and a predicted statistical value inferred by assigning the temporary processing conditions to the second learning model. If the first predicted processing result restored by the restoration section meets the allowable conditions, the temporary processing conditions are determined as the processing conditions for driving the substrate processing apparatus.

[0161] According to the information processing apparatus described in item 7, temporary processing conditions are assigned to a first learning model, and a second predicted processing result is inferred using the first learning model. Furthermore, temporary processing conditions are assigned to a second learning model, and a predicted statistical value is inferred using the second learning model. Moreover, if the first predicted processing result, reconstructed from the second predicted processing result and the predicted statistical value, satisfies an allowable condition, the temporary processing conditions are determined as the processing conditions for driving the substrate processing apparatus. Therefore, a plurality of temporary processing conditions can be determined for a first predicted processing result that satisfies the allowable condition. As a result, a plurality of processing conditions can be indicated for the processing results of a complex process used to process a substrate.

[0162] (Item 8) An information processing apparatus according to another embodiment of the present invention, used for managing a substrate processing apparatus, wherein, The substrate processing apparatus supplies processing liquid to the substrate with the coating to perform the coating process under processing conditions that vary over time. The information processing apparatus has a processing condition determination unit that uses a predictor to determine the processing conditions for driving the substrate processing apparatus. The predictor predicts a processing result that represents the difference in film thickness before and after the coating process is performed on the substrate before the coating process is performed by the substrate processing apparatus. The predictor has a multi-task learning model, which is generated by machine learning learning data. The learning data includes statistical values ​​obtained by statistically processing a first processing result after driving the substrate processing device to perform the coating process under the processing conditions, transforming the first processing result to obtain a second processing result, and the statistical values ​​as target variables and including the processing conditions as explanatory variables. The first processing result includes the difference in film thickness before and after the coating process at each of a plurality of different radial positions of the substrate. The processing condition determination unit includes a restoration unit, which restores the first predicted processing result based on a second predicted processing result inferred by assigning temporary processing conditions to the multi-task learning model and a predicted statistical value. If the first predicted processing result restored by the restoration section meets the allowable conditions, the temporary processing conditions are determined as the processing conditions for driving the substrate processing apparatus.

[0163] According to the information processing apparatus described in item 8, temporary processing conditions are assigned to a multi-task learning model, and a second predicted processing result and a predicted statistical value are inferred through the multi-task learning model. Furthermore, if the first predicted processing result reconstructed from the second predicted processing result and the predicted statistical value meets an allowable condition, the temporary processing conditions are determined as the processing conditions for driving the substrate processing apparatus. Therefore, a plurality of temporary processing conditions can be determined for a first predicted processing result that meets an allowable condition. As a result, a plurality of processing conditions can be indicated for the processing results of a complex process used to process the substrate.

[0164] (Item 9) A substrate processing apparatus according to another embodiment of the present invention, having the information processing apparatus described in item 7 or 8.

[0165] According to the substrate processing apparatus described in claim 9, a substrate processing apparatus is provided that can perform a coating process under appropriate processing conditions based on the results of a complex process for processing the substrate.

[0166] (Item 10) Another embodiment of the present invention provides a learning model generation method, which causes a learning device to perform: The processing condition acquisition step acquires processing conditions for driving the substrate processing apparatus, wherein the substrate processing apparatus performs coating processing by supplying a processing liquid to a substrate on which a coating has been formed. The first processing result acquisition step involves acquiring a first processing result after driving the substrate processing device to perform the coating process under the processing conditions. The first processing result includes the difference in film thickness before and after the coating process at each of a plurality of different radial positions of the substrate. The statistical value calculation steps involve statistically processing the first processing result and calculating the statistical value. The conversion step involves using the statistical values ​​to convert the first processing result into a second processing result. The first learning model generation step involves generating a first learning model using the processing conditions described in the machine learning process and the second processing result; and The second learning model generation step involves using the processing conditions and statistical values ​​obtained from machine learning to generate a second learning model.

[0167] According to the learning model generation method described in item 10, a statistical value is calculated by statistical processing of the first processing result, and the first processing result is transformed into a second processing result using this statistical value. Therefore, the scale of the plurality of second processing results learned by the first learning model is unified before the first learning model learns. Thus, the accuracy of the first learning model can be improved. Furthermore, the second learning model learns the unified statistical value used for the scale. Therefore, the first processing result before the scale is unified can be inferred based on the value inferred from a processing condition by the learned first learning model and the value inferred from a processing condition by the learned second learning model.

[0168] (Item 11) Another embodiment of the present invention provides a learning model generation method, which causes a learning device to perform: The processing condition acquisition step acquires processing conditions for driving the substrate processing apparatus, wherein the substrate processing apparatus performs coating processing by supplying a processing liquid to a substrate on which a coating has been formed. The first processing result acquisition step involves acquiring a first processing result after driving the substrate processing device to perform the coating process under the processing conditions. The first processing result includes the difference in film thickness before and after the coating process at each of a plurality of different radial positions of the substrate. The statistical value calculation steps involve statistically processing the first processing result and calculating the statistical value. The conversion step involves using the statistical value to convert the first processing result into a second processing result; and The multi-task learning model generation steps involve using machine learning data to generate a multi-task learning model, wherein the learning data includes the processing conditions, the second processing result, and the statistical values.

[0169] According to the learning model generation method described in item 11, a statistical value is calculated by statistical processing of the first processing result, and the first processing result is transformed into a second processing result using this statistical value. Therefore, the scale of the second processing result learned by the multi-task learning model is unified before the multi-task learning model learns it. Therefore, the accuracy of the multi-task learning model can be improved. Furthermore, the multi-task learning model learns the unified statistical value used for the scale. Therefore, the first processing result before the scale is unified can be inferred based on the second processing result inferred by the learned multi-task learning model and the statistical value. As a result, a learning apparatus can be provided that can generate a multi-task learning model with improved accuracy in inferring processing results based on processing conditions.

[0170] (Item 12) Another embodiment of the present invention provides a method for determining processing conditions, which is executed by an information processing apparatus for managing a substrate processing apparatus, wherein... The substrate processing apparatus supplies processing liquid to the substrate with the coating to perform the coating process under processing conditions that vary over time. The processing condition determination method includes a processing condition determination step using a predictor to determine the processing conditions for driving the substrate processing apparatus, wherein the predictor, for the coating formed on the substrate before the coating process is performed by the substrate processing apparatus, infers a predicted processing result representing the difference in film thickness before and after the coating process. The predictor has: The first learning model is generated through machine learning using statistical values ​​obtained by statistically processing a first processing result after performing the coating process with the substrate processing apparatus driven by the processing conditions, transforming the first processing result into a second processing result, and the processing conditions. The first processing result includes the difference in film thickness before and after the coating process at each of a plurality of radially different locations on the substrate; and The second learning model is generated through machine learning of the processing conditions and the statistical values. The steps for determining the processing conditions include: The recovery step involves restoring the first prediction result based on the second predicted processing result inferred by assigning temporary processing conditions to the first learning model and the predicted statistical value inferred by assigning the temporary processing conditions to the second learning model; and If the first predicted processing result restored in the restoration step meets the allowable conditions, the temporary processing conditions are determined as the processing conditions for driving the substrate processing apparatus.

[0171] According to the processing condition determination method described in item 12, temporary processing conditions are assigned to a first learning model, and a second predicted processing result is inferred using the first learning model. Similarly, temporary processing conditions are assigned to a second learning model, and a predicted statistical value is inferred using the second learning model. Furthermore, if the first predicted processing result, reconstructed from the second predicted processing result and the predicted statistical value, satisfies an allowable condition, the temporary processing conditions are determined as the processing conditions for driving the substrate processing apparatus. Therefore, it is possible to determine a plurality of temporary processing conditions for a first predicted processing result that satisfies an allowable condition. As a result, it is possible to suggest a plurality of processing conditions for the processing results of a complex process used to process a substrate.

[0172] (Item 13) Another embodiment of the present invention provides a method for determining processing conditions, which is executed by an information processing apparatus for managing a substrate processing apparatus, wherein... The substrate processing apparatus supplies processing liquid to the substrate with the coating to perform the coating process under processing conditions that vary over time. The processing condition determination method includes a processing condition determination step using a predictor to determine the processing conditions for driving the substrate processing apparatus, wherein the predictor, for the coating formed on the substrate before the coating process is performed by the substrate processing apparatus, infers a predicted processing result representing the difference in film thickness before and after the coating process. The predictor has a multi-task learning model, which is generated by machine learning learning data. The learning data includes statistical values ​​obtained by statistically processing a first processing result after driving the substrate processing device to perform the coating process under the processing conditions, transforming the first processing result to obtain a second processing result, the statistical values, and the processing conditions. The first processing result includes the difference in film thickness before and after the coating process at each of a plurality of radially different locations on the substrate. The steps for determining the processing conditions include: The recovery step involves recovering the first prediction result based on the second prediction result inferred from the temporary processing conditions applied to the multi-task learning model, and the prediction statistics; and If the first predicted processing result restored in the restoration step meets the allowable conditions, the temporary processing conditions are determined as the processing conditions for driving the substrate processing apparatus.

[0173] According to the processing condition determination method described in item 13, temporary processing conditions are assigned to a multi-task learning model, and the second predicted processing result and predicted statistics are inferred through the multi-task learning model. Furthermore, if the first predicted processing result reconstructed from the second predicted processing result and predicted statistics meets the allowable conditions, the temporary processing conditions are determined as the processing conditions for driving the substrate processing apparatus. Therefore, a plurality of temporary processing conditions can be determined for a first predicted processing result that meets the allowable conditions. As a result, a plurality of processing conditions can be indicated for the processing results of complex processes used to process substrates.

Claims

1. A learning device, wherein, have: The processing condition acquisition unit acquires processing conditions for driving the substrate processing apparatus, which performs coating processing by supplying a processing liquid to a substrate on which a coating is formed. The first processing result acquisition unit acquires a first processing result after driving the substrate processing apparatus to perform the coating process under the processing conditions. The first processing result includes the difference in film thickness before and after the coating process at each of a plurality of radially different positions of the substrate. The statistical value calculation unit statistically processes the first processing result and calculates the statistical value; The conversion unit uses the statistical values ​​to convert the first processing result into a second processing result. The first learning model generation unit generates a first learning model by machine learning the processing conditions and the second processing result; as well as The second learning model generation unit generates a second learning model by machine learning the processing conditions and the statistical values.

2. The learning device as claimed in claim 1, wherein, The processing conditions include: The conditions change over time; and Fixed conditions other than the aforementioned variable conditions; The first learning model includes: The first convolutional neural network is input with the changing conditions in the processing conditions; A first fully connected neural network is input to the output of the first convolutional neural network and the fixed condition in the processing conditions; and The second convolutional neural network is input into the output of the first fully connected neural network and outputs the second processing result. The second learning model includes: A third convolutional neural network is input to the changing conditions in the processing conditions; and The second fully connected neural network is fed into the output of the third convolutional neural network and the fixed condition in the processing conditions, and outputs the statistical value.

3. The learning device as described in claim 2, wherein, The first convolutional neural network and the third convolutional neural network have the same filters.

4. A learning device, wherein, have: The processing condition acquisition unit acquires processing conditions for driving the substrate processing apparatus, which performs coating processing by supplying a processing liquid to a substrate on which a coating is formed. The first processing result acquisition unit acquires a first processing result after driving the substrate processing apparatus to perform the coating process under the processing conditions. The first processing result includes the difference in film thickness before and after the coating process at each of a plurality of different radial positions of the substrate. The statistical value calculation unit statistically processes the first processing result and calculates the statistical value; The conversion unit uses the statistical values ​​to convert the first processing result into a second processing result. as well as The multi-task learning model generation unit generates a multi-task learning model from machine learning data, wherein the learning data includes the processing conditions as explanatory variables, the second processing result as the objective variable, and the statistical values.

5. The learning device as described in claim 4, wherein, The processing conditions include: The conditions change over time; and Fixed conditions other than the aforementioned variable conditions; The multi-task learning model includes: The first convolutional neural network is input with the changing conditions in the processing conditions; The first fully connected neural network is input to the output of the first convolutional neural network and the fixed condition in the processing conditions; The second convolutional neural network, which receives a portion of the output of the first fully connected neural network, outputs the second processing result; and The second fully connected neural network, which is input into the other part of the output of the first fully connected neural network, outputs the statistical value.

6. The learning device as described in any one of claims 2, 3, and 5, wherein, The substrate processing apparatus supplies the processing liquid to the substrate by moving a nozzle that supplies the processing liquid to the substrate. The variation conditions include nozzle movement conditions, which represent the relative position of the nozzle with respect to the substrate as time progresses.

7. An information processing apparatus for managing a substrate processing apparatus, wherein, The substrate processing apparatus supplies processing liquid to the substrate with the coating to perform the coating process under processing conditions that vary over time. The information processing apparatus has a processing condition determination unit that uses a predictor to determine the processing conditions for driving the substrate processing apparatus. The predictor predicts a processing result that represents the difference in film thickness before and after the coating process is performed on the substrate before the coating process is performed by the substrate processing apparatus. The predictor has: The first learning model is generated by machine learning using statistical values ​​obtained by statistically processing the first processing result after performing the coating process with the substrate processing device driven by the processing conditions, transforming the first processing result into a second processing result, and the processing conditions. The first processing result includes the difference in film thickness before and after the coating process at each of a plurality of radially different locations on the substrate. as well as The second learning model is generated through machine learning of the processing conditions and the statistical values. The processing condition determination unit includes a recovery unit, which recovers the first prediction processing result based on a second predicted processing result inferred by assigning temporary processing conditions to the first learning model and a predicted statistical value inferred by assigning the temporary processing conditions to the second learning model. If the first predicted processing result restored by the restoration section meets the allowable conditions, the temporary processing conditions are determined as the processing conditions for driving the substrate processing apparatus.

8. An information processing apparatus for managing a substrate processing apparatus, wherein, The substrate processing apparatus supplies processing liquid to the substrate with the coating to perform the coating process under processing conditions that vary over time. The information processing apparatus has a processing condition determination unit that uses a predictor to determine the processing conditions for driving the substrate processing apparatus. The predictor predicts a processing result that represents the difference in film thickness before and after the coating process is performed on the substrate before the coating process is performed by the substrate processing apparatus. The predictor has a multi-task learning model, which is generated by machine learning learning data. The learning data includes statistical values ​​obtained by statistically processing a first processing result after driving the substrate processing device to perform the coating process under the processing conditions, transforming the first processing result to obtain a second processing result, and the statistical values ​​as target variables and including the processing conditions as explanatory variables. The first processing result includes the difference in film thickness before and after the coating process at each of a plurality of different radial positions of the substrate. The processing condition determination unit includes a restoration unit, which restores the first predicted processing result based on a second predicted processing result inferred by assigning temporary processing conditions to the multi-task learning model and a predicted statistical value. If the first predicted processing result restored by the restoration section meets the allowable conditions, the temporary processing conditions are determined as the processing conditions for driving the substrate processing apparatus.

9. A substrate processing apparatus comprising the information processing apparatus of claim 7 or 8.

10. A method for generating a learning model, which enables a learning device to perform: The processing condition acquisition step acquires processing conditions for driving the substrate processing apparatus, wherein the substrate processing apparatus performs coating processing by supplying a processing liquid to a substrate on which a coating has been formed. The first processing result acquisition step involves acquiring a first processing result after driving the substrate processing device to perform the coating process under the processing conditions. The first processing result includes the difference in film thickness before and after the coating process at each of a plurality of different radial positions of the substrate. The statistical value calculation steps involve statistically processing the first processing result and calculating the statistical value. The conversion step involves using the statistical values ​​to convert the first processing result into a second processing result. The first learning model generation step involves using the processing conditions described in the machine learning process and the second processing result to generate the first learning model. as well as The second learning model generation step involves using the processing conditions and statistical values ​​obtained from machine learning to generate a second learning model.

11. A method for generating a learning model, which enables a learning device to perform: The processing condition acquisition step acquires processing conditions for driving the substrate processing apparatus, wherein the substrate processing apparatus performs coating processing by supplying a processing liquid to a substrate on which a coating has been formed. The first processing result acquisition step involves acquiring a first processing result after driving the substrate processing device to perform the coating process under the processing conditions. The first processing result includes the difference in film thickness before and after the coating process at each of a plurality of different radial positions of the substrate. The statistical value calculation steps involve statistically processing the first processing result and calculating the statistical value. The conversion step involves using the statistical values ​​to convert the first processing result into a second processing result. as well as The multi-task learning model generation steps involve using machine learning data to generate a multi-task learning model, wherein the learning data includes the processing conditions, the second processing result, and the statistical values.

12. A method for determining processing conditions, used by an information processing apparatus to manage a substrate processing device, wherein, The substrate processing apparatus supplies processing liquid to the substrate with the coating to perform the coating process under processing conditions that vary over time. The processing condition determination method includes a processing condition determination step using a predictor to determine the processing conditions for driving the substrate processing apparatus, wherein the predictor, for the coating formed on the substrate before the coating process is performed by the substrate processing apparatus, infers a predicted processing result representing the difference in film thickness before and after the coating process. The predictor has: The first learning model is generated by machine learning using statistical values ​​obtained by statistically processing the first processing result after performing the coating process with the substrate processing device driven by the processing conditions, transforming the first processing result into a second processing result, and the processing conditions. The first processing result includes the difference in film thickness before and after the coating process at each of a plurality of radially different locations on the substrate. as well as The second learning model is generated through machine learning of the processing conditions and the statistical values. The steps for determining the processing conditions include: The restoration step involves restoring the first prediction result based on the second predicted processing result inferred by assigning temporary processing conditions to the first learning model and the predicted statistical value inferred by assigning the temporary processing conditions to the second learning model. as well as If the first predicted processing result restored in the restoration step meets the allowable conditions, the temporary processing conditions are determined as the processing conditions for driving the substrate processing apparatus.

13. A method for determining processing conditions, used by an information processing apparatus to manage a substrate processing device, wherein, The substrate processing apparatus supplies processing liquid to the substrate with the coating to perform the coating process under processing conditions that vary over time. The processing condition determination method includes a processing condition determination step using a predictor to determine the processing conditions for driving the substrate processing apparatus, wherein the predictor, for the coating formed on the substrate before the coating process is performed by the substrate processing apparatus, infers a predicted processing result representing the difference in film thickness before and after the coating process. The predictor has a multi-task learning model, which is generated by machine learning learning data. The learning data includes statistical values ​​obtained by statistically processing a first processing result after driving the substrate processing device to perform the coating process under the processing conditions, transforming the first processing result to obtain a second processing result, the statistical values, and the processing conditions. The first processing result includes the difference in film thickness before and after the coating process at each of a plurality of radially different locations on the substrate. The steps for determining the processing conditions include: The restoration step involves restoring the first prediction result based on the second prediction result inferred by assigning temporary processing conditions to the multi-task learning model and the prediction statistics. as well as If the first predicted processing result restored in the restoration step meets the allowable conditions, the temporary processing conditions are determined as the processing conditions for driving the substrate processing apparatus.

Images