JP7219190B2 - Teaching data generation method and discharge state determination method - Google Patents

Description

TECHNICAL FIELD The present invention relates to a teacher data generation method and an ejection state determination method.

Conventionally, there has been proposed a substrate processing apparatus that supplies a processing liquid to a substrate. This substrate processing apparatus includes a substrate holder that holds a substrate in a horizontal position, a rotation mechanism that rotates the substrate holder to rotate the substrate in a horizontal plane, and a discharge nozzle that discharges a processing liquid from above the substrate. ing.

When the processing liquid is discharged from the discharge nozzle toward the substrate, the processing liquid lands near the center of the substrate, spreads on the substrate due to the centrifugal force accompanying the rotation of the substrate, and scatters from the periphery of the substrate. Thereby, the processing liquid acts on the entire surface of the substrate, and the substrate can be processed according to the processing liquid. As the treatment liquid, SC1 liquid (mixture of ammonia water, hydrogen peroxide water and water), SC2 liquid (hydrochloric acid, hydrogen peroxide water and water mixture) and DHF liquid (dilute hydrofluoric acid), or , pure water or the like is used.

A technique has also been proposed in which a camera is used to monitor the discharge state of the processing liquid from the discharge nozzles (for example, Japanese Unexamined Patent Application Publication No. 2002-100003). In Japanese Patent Application Laid-Open No. 2002-201000, an imaging region including the tip of an ejection nozzle is imaged by a camera, and based on the image data acquired by the camera, it is determined whether or not the treatment liquid is ejected from the ejection nozzle.

JP 2017-29883 A

It is conceivable to perform classification processing using a machine-learned classifier on image data acquired by a camera. As the classification category, a category corresponding to the discharge state of the treatment liquid from the discharge nozzle can be adopted. Specifically, the categories include a category indicating that the ejection state of the treatment liquid is normal, a category indicating that the treatment liquid has not yet been ejected, and a category indicating that the ejection state of the treatment liquid is unsatisfactory. The category shown can be adopted. A classifier classifies the image data acquired by the camera into the appropriate category.

In order to generate such a classifier by machine learning, it is necessary to generate teacher data.

SUMMARY OF THE INVENTION Accordingly, it is an object of the present application to provide a teaching data generating method capable of generating teaching data corresponding to the discharge state of a discharge nozzle.

A first aspect of a teaching data generating method is a method of generating teaching data about the discharge state of a discharge nozzle in a substrate processing apparatus, comprising: a holding step of horizontally holding a substrate in a substrate holding mechanism; an image pickup adjusted to include at least part of the treatment liquid ejected in the treatment liquid ejection step of ejecting the treatment liquid toward the main surface of the substrate, the ejection nozzle, and the treatment liquid ejection step. an imaging step of capturing an image of the area with a camera over a period including at least part of the treatment liquid discharging step to acquire a plurality of image data; Image data is misclassified into a certain category by a teacher data generating step of generating teacher data by attaching a label corresponding to the ejection state of the treatment liquid to the image data, and a classifier generated based on the teacher data. adopting the image data as training data, and assigning the training data a second label different from the first label corresponding to the category, wherein the category corresponds to the processing A category indicating a dripping state in which the treatment liquid drops as droplets when liquid ejection is stopped, and the second label is a pattern formed on the surface of the substrate immediately below the ejection nozzle in the image data. This is a label indicating that the ejection stop state has been misclassified as the dripping state due to the similarity to the dripping pattern .
A second aspect of the teaching data generating method is a method of generating teaching data about the discharge state of a discharge nozzle in a substrate processing apparatus, comprising: a holding step of horizontally holding a substrate on a substrate holding mechanism; a treatment liquid ejection step of ejecting the treatment liquid toward the main surface of the substrate; an imaging region adjusted to include the ejection nozzle and at least part of the treatment liquid ejected in the treatment liquid ejection step; is imaged by a camera over a period including at least a part of the treatment liquid ejection step to obtain a plurality of image data; The image data indicates the discharge state of the treatment liquid by a teacher data generating step of generating teacher data by adding a label corresponding to the discharge state of the liquid to the image data, and a classifier generated based on the teacher data. When the image data is classified into one of a plurality of categories and the image data is misclassified into an incorrect category different from the correct ejection state, the image data is adopted as teacher data, and the teacher data is: a step of applying a second label indicating classification into an erroneous category different from a first label corresponding to the category classified according to the discharge state of the treatment liquid, wherein the second label is: A label indicating misclassification.

A third aspect of the teacher data generation method is the teacher data generation method according to the first or second aspect, wherein the teacher data generation step includes timing at which the ejection state of the treatment liquid changes, and a storing step of storing image data acquired by the camera during a period shorter than the period of the liquid ejection step in a storage medium as candidates for teaching data; and a selecting step of selecting image data to be employed as teaching data from the candidates. and a label assigning step of assigning a label to the teacher data according to the discharge state of the treatment liquid reflected in the teacher data.

A mode of the discharge state determination method includes a step of performing machine learning using teacher data generated by the teacher data generation method according to any one of the first to third modes to generate a classifier; a step of horizontally holding a substrate on a substrate holding mechanism; a step of ejecting a processing liquid from the ejection nozzle toward a main surface of the substrate; acquiring a plurality of image data by capturing an image of an imaging region updated to include at least part of the treatment liquid by a camera for a period including at least part of the treatment liquid discharging step; provisionally determining whether the discharge state of the treatment liquid is good or bad based on the statistic of pixel values in an area extending in the discharge direction of the treatment liquid from the tip of the discharge nozzle in the image data; and classifying the image data into a category according to the ejection state by the classifier when it is tentatively determined.

According to the first aspect of the teaching data generation method, it is possible to generate teaching data labeled according to the ejection state of the treatment liquid. Moreover, it is possible to generate teacher data for re-learning, and to generate teacher data that contributes to suppression of erroneous determination of dropping.
According to the second aspect of the teaching data generation method, it is possible to generate teaching data labeled according to the ejection state of the treatment liquid. Moreover, teacher data for re-learning can be generated, and classification accuracy can be improved.

According to the third aspect of the training data generation method, image data acquired in a shorter candidate period are stored as training data candidates. The number of training data candidates can be reduced compared to the case of storing data candidates. Therefore, it becomes easy to select training data.

According to the aspect of the ejection state determination method, when the ejection state is determined to be defective by the provisional determination based on the statistics, the classifier classifies the image data into categories. In comparison, processing can be simplified.

It is a figure which shows a schematic example of a structure of a substrate processing apparatus. FIG. 2 is a plan view showing a schematic example of the configuration of a processing unit; FIG. 2 is a side view showing a schematic example of the configuration of a processing unit; 3 is a functional block diagram schematically showing an example of the configuration of a control section; FIG. It is a figure which shows roughly an example of an ejection state in tabular form. FIG. 11 is a flowchart showing an example of processing for generating a classifier; FIG. 6 is a flow chart showing an example of a process of generating teacher data; It is a figure which shows an example of an input screen roughly. It is a figure which shows roughly an example of a part of image data. FIG. 5 is a functional block diagram schematically showing another example of the configuration of the control section; It is a flow chart which shows an example of update processing of a classifier. It is a figure which shows an example of a browsing screen roughly. FIG. 10 is a diagram schematically showing another example of a browsing screen; FIG. 5 is a functional block diagram schematically showing another example of the configuration of the control section; It is a figure which shows an example of image data roughly.

Embodiments will be described below with reference to the attached drawings. It should be noted that the drawings are shown schematically, and for the convenience of explanation, the configuration may be omitted or simplified as appropriate. Also, the mutual relationship of sizes and positions of the configurations shown in the drawings are not necessarily described accurately, and may be changed as appropriate.

In addition, in the description given below, the same components are denoted by the same reference numerals, and their names and functions are also the same. Therefore, a detailed description thereof may be omitted to avoid duplication.

<Overview of substrate processing equipment>
FIG. 1 is a diagram showing the overall configuration of a substrate processing apparatus 100. As shown in FIG. The substrate processing apparatus 100 is an apparatus that processes the substrate W by supplying a processing liquid to the substrate W. As shown in FIG. The substrate W is, for example, a semiconductor substrate. This substrate W has a substantially disk shape.

This substrate processing apparatus 100 can supply the main surface of the substrate W with the processing liquid. For example, the substrate processing apparatus 100 can perform a cleaning process by supplying a rinse liquid to the substrate W after supplying the substrate W with a chemical solution for cleaning. The chemical solution is typically SC1 solution (a mixture of ammonia water, hydrogen peroxide solution and water), SC2 solution (a mixture of hydrochloric acid, hydrogen peroxide solution and water), or DHF solution (dilute fluoride). acid) and the like are used. Pure water, for example, is used as the rinsing liquid. In this specification, the chemical liquid and the rinse liquid are collectively referred to as "treatment liquid". In addition to the cleaning process, the "processing liquid" includes a coating liquid such as a photoresist liquid for film formation, a chemical liquid for removing unnecessary films, a chemical liquid for etching, and the like.

A substrate processing apparatus 100 includes an indexer 102 , a plurality of processing units 1 and a main transfer robot 103 . The indexer 102 has a function of loading unprocessed substrates W received from outside the apparatus into the apparatus, and unloading processed substrates W that have undergone cleaning processing from the apparatus. The indexer 102 mounts a plurality of carriers and has a transfer robot (not shown). As a carrier, adopt a FOUP (front opening unified pod) or SMIF (Standard Mechanical Interface) pod that stores the substrate W in a sealed space, or an OC (open cassette) that exposes the substrate W to the outside while stored. can be done. The transfer robot transfers the substrate W between the carrier and the main transfer robot 103 .

Twelve processing units 1 are arranged in the substrate processing apparatus 100 . The detailed arrangement configuration is such that four towers in which three processing units 1 are stacked are arranged so as to surround the main transfer robot 103 . In other words, the four processing units 1 arranged around the main transfer robot 103 are stacked in three layers, and FIG. 1 shows one of them. The number of processing units 1 mounted on the substrate processing apparatus 100 is not limited to twelve, and may be, for example, eight or four.

A main transport robot 103 is installed in the center of four towers containing stacked processing units 1 . The main transport robot 103 loads unprocessed substrates W received from the indexer 102 into the respective processing units 1 and unloads the processed substrates W from the respective processing units 1 to deliver them to the indexer 102 .

Next, the processing unit 1 will be explained. One of the 12 processing units 1 mounted on the substrate processing apparatus 100 will be described below, but the other processing units 1 are also the same. FIG. 2 is a plan view of the processing unit 1. FIG. 3 is a longitudinal sectional view of the processing unit 1. FIG. 2 shows a state in which the substrate W is not held by the substrate holding portion 20, and FIG. 3 shows a state in which the substrate W is held by the substrate holding portion 20. As shown in FIG.

In the chamber 10, the processing unit 1 includes, as main elements, a substrate holding part 20 that holds the substrate W in a horizontal posture (a posture in which the normal line of the substrate W is along the vertical direction), and a substrate held by the substrate holding part 20. Three processing liquid supply units 30, 60, 65 for supplying the processing liquid to the upper surface of W, a processing cup 40 surrounding the substrate holding unit 20, and a camera 70 for imaging the space above the substrate holding unit 20. Prepare. A partition plate 15 is provided around the processing cup 40 in the chamber 10 to partition the inner space of the chamber 10 into upper and lower parts.

The chamber 10 includes side walls 11 extending in the vertical direction, a ceiling wall 12 that closes the upper side of the space surrounded by the side walls 11, and a floor wall 13 that closes the lower side. A space surrounded by the side walls 11, the ceiling wall 12, and the floor wall 13 serves as a processing space for the substrates W. As shown in FIG. A loading/unloading port for loading/unloading the substrate W by the main transfer robot 103 into/from the chamber 10 and a shutter for opening/closing the loading/unloading port are provided on a part of the side wall 11 of the chamber 10 (both are shown in the figure). omit).

A fan filter unit (FFU) 14 is attached to the ceiling wall 12 of the chamber 10 for further purifying the air in the clean room in which the substrate processing apparatus 100 is installed and supplying it to the processing space in the chamber 10 . . The fan filter unit 14 is equipped with a fan and a filter (for example, a HEPA filter) for taking in air in the clean room and sending it out into the chamber 10 , and forms clean air downflow in the processing space inside the chamber 10 . In order to uniformly disperse the clean air supplied from the fan filter unit 14 , a punching plate having a large number of blowout holes may be provided immediately below the ceiling wall 12 .

The substrate holder 20 is, for example, a spin chuck. The substrate holding unit 20 includes a disk-shaped spin base 21 fixed in a horizontal posture to the upper end of a rotating shaft 24 extending along the vertical direction. A spin motor 22 that rotates a rotating shaft 24 is provided below the spin base 21 . The spin motor 22 rotates the spin base 21 in the horizontal plane via the rotation shaft 24 . A tubular cover member 23 is provided to surround the spin motor 22 and the rotating shaft 24 .

The outer diameter of the disk-shaped spin base 21 is slightly larger than the diameter of the circular substrate W held by the substrate holder 20 . Therefore, the spin base 21 has a holding surface 21a facing the entire lower surface of the substrate W to be held.

A plurality of (four in this embodiment) chuck pins 26 are erected on the periphery of the holding surface 21 a of the spin base 21 . The plurality of chuck pins 26 are spaced evenly along the circumference corresponding to the outer circumference of the circular substrate W (in the case of four chuck pins 26 as in this embodiment, they are spaced at 90° intervals). ) are placed. A plurality of chuck pins 26 are interlocked and driven by a link mechanism (not shown) housed in the spin base 21 . The substrate holding unit 20 holds the substrate W in a horizontal position above the spin base 21 and close to the holding surface 21a by holding the substrate W with each of the plurality of chuck pins 26 in contact with the outer peripheral edge of the substrate W. (see FIG. 3), and each of the plurality of chuck pins 26 can be separated from the outer peripheral edge of the substrate W to release the grip.

In a state where the substrate holding unit 20 holds the substrate W by gripping with a plurality of chuck pins 26, the spin motor 22 rotates the rotation shaft 24, thereby rotating the rotation axis CX along the vertical direction passing through the center of the substrate W. substrate W can be rotated.

The processing liquid supply unit 30 is configured by attaching a discharge nozzle 31 to the tip of a nozzle arm 32 (see FIG. 2). The base end side of the nozzle arm 32 is fixedly connected to the nozzle base 33 . The nozzle base 33 is rotatable about a vertical axis by a motor (not shown). As the nozzle base 33 rotates, the discharge nozzle 31 moves between the processing position above the substrate holder 20 and the waiting position outside the processing cup 40 as indicated by the arrow AR34 in FIG. Moves in an arc along the horizontal direction.

In the examples of FIGS. 2 and 3, the processing liquid supply unit 30 is configured to supply a plurality of types of processing liquids. Specifically, the treatment liquid supply unit 30 has a plurality of ejection nozzles 31 . Two ejection nozzles 31 a and 31 b are shown as the ejection nozzle 31 in the examples of FIGS. 2 and 3 . The ejection nozzles 31 a and 31 b are fixed to a nozzle base 33 via a nozzle arm 32 . Therefore, the ejection nozzles 31a and 31b move in synchronization with each other. The ejection nozzles 31a and 31b are provided adjacent to each other in the horizontal plane.

As illustrated in FIG. 3, the discharge nozzle 31a is connected to a processing liquid supply source 37a through a pipe 34a, and the discharge nozzle 31b is connected to a processing liquid supply source 37b through a pipe 34b. On-off valves 35a and 35b are provided in the middle of the pipes 34a and 34b, respectively. By opening the on-off valve 35a, the processing liquid from the processing liquid supply source 37a flows inside the pipe 34a and is discharged from the discharge nozzle 31a, and by opening the on-off valve 35b, the processing liquid from the processing liquid supply source 37b is discharged. It flows inside the pipe 34b and is discharged from the discharge nozzle 31b. The SC1 liquid, for example, is discharged from the discharge nozzle 31a, and pure water, for example, is discharged from the discharge nozzle 31b. The processing liquid ejected while the ejection nozzles 31 a and 31 b are stopped at the processing position lands on the upper surface of the substrate W held by the substrate holder 20 .

Suck-back valves 36a and 36b may be provided in the middle of the pipes 34a and 34b, respectively. The suck back valve 36a draws in the processing liquid from the tip of the discharge nozzle 31a by sucking the processing liquid in the pipe 34a when the discharge of the processing liquid is stopped. This makes it difficult for the treatment liquid to drop as relatively large chunks (droplets) from the tip of the ejection nozzle 31a when the ejection is stopped. The same applies to the suck back valve 36b.

The processing unit 1 of the present embodiment is further provided with two processing liquid supply units 60 and 65 in addition to the processing liquid supply unit 30 described above. The processing liquid supply units 60 and 65 of this embodiment have the same configuration as the processing liquid supply unit 30 described above. That is, the treatment liquid supply unit 60 is configured by attaching a discharge nozzle 61 to the tip of a nozzle arm 62, and the discharge nozzle 61 is connected to an arrow AR64 by a nozzle base 63 connected to the base end of the nozzle arm 62. , moves in an arc between the processing position above the substrate holding part 20 and the standby position outside the processing cup 40. As shown in FIG. Similarly, the treatment liquid supply unit 65 is configured by attaching a discharge nozzle 66 to the tip of a nozzle arm 67 , and the discharge nozzle 66 is moved by a nozzle base 68 connected to the base end of the nozzle arm 67 . , move in an arc between the processing position above the substrate holder 20 and the standby position outside the processing cup 40 . The processing liquid supply units 60 and 65 may also be configured to supply a plurality of types of processing liquid, or may be configured to supply a single processing liquid.

The processing liquid supply units 60 and 65 discharge the processing liquid onto the upper surface of the substrate W held by the substrate holding unit 20 with the discharge nozzles 61 and 66 positioned at the processing positions. At least one of the discharge nozzles 61 and 66 is a two-fluid system that mixes a cleaning liquid such as pure water with a pressurized gas to generate droplets, and jets the mixed fluid of the droplets and the gas onto the substrate W. It may be a nozzle. Further, the number of processing liquid supply units provided in the processing unit 1 is not limited to three, and may be one or more. However, in the present embodiment, since it is assumed that the two treatment liquids are sequentially switched and ejected, two or more ejection nozzles are provided as a whole. Each of the discharge nozzles of the processing liquid supply units 60 and 65 is also connected to the processing liquid supply source through a pipe in the same manner as the processing liquid supply unit 30, and an on-off valve is provided in the middle of the pipe, and a suckback valve is provided in the middle of the pipe. may be provided. Processing using the processing liquid supply unit 30 will be representatively described below.

The processing cup 40 is provided so as to surround the substrate holder 20 . The processing cup 40 has an inner cup 41 , a middle cup 42 and an outer cup 43 . The inner cup 41, the middle cup 42 and the outer cup 43 are provided so as to be able to move up and down. When the inner cup 41 , the middle cup 42 and the outer cup 43 are raised, the processing liquid scattered from the peripheral edge of the substrate W hits the inner peripheral surface of the inner cup 41 and drops. The dropped treatment liquid is appropriately recovered by the first recovery mechanism. In a state where the inner cup 41 is lowered and the middle cup 42 and the outer cup 43 are raised, the processing liquid scattered from the peripheral edge of the substrate W hits the inner peripheral surface of the middle cup 42 and falls. The dropped treatment liquid is appropriately recovered by the second recovery mechanism. In a state where the inner cup 41 and the middle cup 42 are lowered and the outer cup 43 is raised, the processing liquid scattered from the peripheral edge of the substrate W hits the inner peripheral surface of the outer cup 43 and falls. The dropped treatment liquid is appropriately recovered by the third recovery mechanism. According to this, different processing liquids can be recovered appropriately.

The partition plate 15 is provided around the processing cup 40 so as to vertically partition the inner space of the chamber 10 . The partition plate 15 may be a single plate-like member surrounding the processing cup 40, or may be a plurality of plate-like members joined together. Further, the partition plate 15 may be formed with a through hole or a notch that penetrates in the thickness direction. A through hole (not shown) is formed for passing a support shaft for support.

The outer peripheral edge of the partition plate 15 is connected to the side wall 11 of the chamber 10 . Further, the edge portion of the partition plate 15 surrounding the processing cup 40 is formed in a circular shape with a diameter larger than the outer diameter of the outer cup 43 . Therefore, the partition plate 15 does not hinder the lifting of the outer cup 43 .

An exhaust duct 18 is provided near the floor wall 13 of the side wall 11 of the chamber 10 . The exhaust duct 18 is communicatively connected to an exhaust mechanism (not shown). Of the clean air supplied from the fan filter unit 14 and flowing down inside the chamber 10 , the air that has passed between the processing cup 40 and the partition plate 15 is discharged from the exhaust duct 18 to the outside of the apparatus.

The camera 70 is installed inside the chamber 10 above the partition plate 15 . The camera 70 includes, for example, an imaging device (such as a CCD (Charge Coupled Device)) and an optical system such as an electronic shutter and a lens. The discharge nozzle 31 of the processing liquid supply unit 30 is moved by the nozzle base 33 to a processing position above the substrate W held by the substrate holding unit 20 (a solid line position in FIG. 3) and a standby position outside the processing cup 40 ( dotted line position in FIG. 3). The processing position is a position where the processing liquid is discharged from the processing liquid supply unit 30 onto the upper surface of the substrate W held by the substrate holding unit 20 to perform the cleaning process. The standby position is a position where the treatment liquid supply unit 30 stops discharging the treatment liquid and waits when the treatment liquid supply unit 30 does not perform the cleaning process. A standby pod that accommodates the ejection nozzles 31 of the processing liquid supply unit 30 may be provided at the standby position.

The camera 70 is installed so that its imaging area includes at least the tip of the ejection nozzle 31 at the processing position. More specifically, the camera 70 is installed such that the tip of the ejection nozzle 31 and the treatment liquid ejected from the tip are included in the imaging area. In the present embodiment, as shown in FIG. 3, a camera 70 is installed at a position that captures an image of the discharge nozzle 31 at the processing position from above the front. Therefore, the camera 70 can capture an imaging area including the tip of the ejection nozzle 31 at the processing position. Similarly, the camera 70 can also image the tips of the ejection nozzles 61 and 66 of the treatment liquid supply units 60 and 65 at the treatment position and the imaging region including the treatment liquid ejected from the tips. In other words, the position of the camera 70 is adjusted so that the tips of these ejection nozzles and the ejected treatment liquid are included in the imaging area.

When the camera 70 is installed at the position shown in FIG. 2, the discharge nozzles 31 and 66 of the processing liquid supply units 30 and 65 move laterally within the imaging field of the camera 70. Although it is possible to appropriately image the movement in the vicinity, the ejection nozzle 61 of the processing liquid supply unit 60 moves in the depth direction within the imaging field of the camera 70. It may not be possible to take an image immediately. In such a case, a camera dedicated to the treatment liquid supply section 60 may be provided separately from the camera 70 .

This camera 70 acquires a plurality of image data by capturing an image of an imaging region over a period including at least part of the period during which the treatment liquid is discharged onto the substrate W. FIG. The camera 70 outputs the acquired image data to the control section 9 .

Further, as shown in FIG. 3, an illumination unit 71 is provided inside the chamber 10 and above the partition plate 15 . Since the chamber 10 is normally a dark room, when the camera 70 takes an image, the illumination unit 71 irradiates the discharge nozzles 31, 61, 66 of the processing liquid supply units 30, 60, 65 near the processing position with light. Image data acquired by the camera 70 is output to the control section 9 .

The user interface 80 has a display section 81 and an input section 82 . The display unit 81 is, for example, a liquid crystal display or an organic EL (Electro Luminescence) display. The input unit 82 is, for example, a touch panel, mouse or keyboard. This user interface 80 is connected to the control unit 9 . A display unit 81 displays a display image based on a display signal from the control unit 9 . This display image includes image data from the camera 70, for example. The input unit 82 outputs input information input by the operator to the control unit 9 . The control unit 9 can control various configurations according to input information.

The control unit 9 controls various components of the substrate processing apparatus 100 to proceed with the processing of the substrate W. FIG. Also, the control unit 9 performs image processing on the image data acquired by the camera 70 . The control unit 9 determines the ejection state of the treatment liquid from each ejection nozzle by this image processing. This image processing will be described in detail later.

The hardware configuration of the control unit 9 is similar to that of a general computer. That is, the control unit 9 stores a CPU that performs various arithmetic processes, a ROM that is a read-only memory that stores basic programs, a RAM that is a readable and writable memory that stores various information, control software, data, and the like. It is configured with a magnetic disk or the like to store. Each operation mechanism of the substrate processing apparatus 100 is controlled by the control unit 9 by the CPU of the control unit 9 executing a predetermined processing program, and processing in the substrate processing apparatus 100 proceeds. Further, image processing is performed by the CPU of the control unit 9 executing a predetermined processing program. Part or all of the functions of the control unit 9 may be realized by dedicated hardware.

FIG. 4 is a functional block diagram schematically showing an example of the internal configuration of the control section 9. As shown in FIG. The control unit 9 includes a classifier 91 , a machine learning unit 92 , a candidate data extraction unit 93 , a display control unit 94 and a teacher data generation unit 95 .

The image data from the camera 70 are sequentially input to the classifier 91 . Here, for the sake of simplicity, the image data will be described assuming that the ejection nozzles 61 of the treatment liquid supply unit 60 are shown. Classifier 91 classifies the input image data into one of a plurality of categories. Categories may also be called classes. Here, the category is a category according to the ejection state of the treatment liquid from the ejection nozzle 61 . FIG. 5 is a diagram schematically showing an example of ejection states. The discharge state includes a state related to the flow shape of the processing liquid discharged from the tip of the discharge nozzle 61 . Specifically, as illustrated in FIG. 5, the ejection state is a normal ejection state in which the treatment liquid flows down as a continuous flow from the ejection nozzle 61, and the treatment liquid drops as droplets when the ejection of the treatment liquid is stopped. Examples include a dripping state and an ejection stop state in which the treatment liquid is not ejected. Dripping is not desirable, and the dripping condition is also a type of defective ejection condition. Another example of the defective ejection state is a liquid splash state in which the processing liquid splashes at the liquid landing point on the substrate W, for example. Here, the dripping state is adopted as the defective ejection state.

Each category is set according to each discharge state. For example, category C1 is a category indicating a normal ejection state, category C2 is a category indicating a certain defective ejection state (here, a dripping state), and category C3 is a category indicating an ejection stop state.

By classifying the image data into categories by the classifier 91, the control section 9 substantially determines the ejection state. For example, when the classifier 91 classifies the image data into the category C2, it can be determined that the spillage has occurred. may be notified.

This classifier 91 is generated by the machine learning section 92 using a plurality of teacher data. In other words, it can be said that this classifier 91 is a machine-learned classifier. The machine learning unit 92 uses, for example, a neighborhood method, a support vector machine, a random forest, or a neural network (including deep learning) as a machine learning algorithm.

The training data includes image data and labels indicating to which category the image data should be classified. The image data may be acquired by camera 70 . Labeling can be performed, for example, by an operator's operation on the user interface 80 (input unit 82). This operation will be detailed later. The machine learning unit 92 performs machine learning based on these teacher data to generate the classifier 91 .

As an example, the classifier 91 that classifies image data by the neighborhood method will be described. The classifier 91 includes a feature vector extraction unit 911, a determination unit 912, and a storage medium in which a determination database 913 is stored. Image data from the camera 70 are sequentially input to the feature vector extraction unit 911 . A feature vector extraction unit 911 extracts a feature vector of image data according to a predetermined algorithm. This feature vector is a vector that can indicate a feature corresponding to the ejection state of the ejection nozzle. A known algorithm can be adopted as the algorithm. Feature vector extraction section 911 outputs the feature vector to determination section 912 .

A determination database 913 stores a plurality of feature vectors (hereinafter referred to as reference vectors) generated from a plurality of teacher data by the machine learning unit 92, and the reference vectors are classified into categories C1 to C3. there is Specifically, the machine learning unit 92 applies the same algorithm as the feature vector extraction unit 911 to multiple teacher data to generate multiple reference vectors. Then, the machine learning unit 92 assigns a teacher data label (correct category) to the reference vector.

A determination unit 912 classifies image data (frames) based on the feature vector input from the feature vector extraction unit 911 and a plurality of reference vectors stored in the determination database 913 . For example, the determining unit 912 may identify a reference vector whose feature vectors are closest, and classify frames into the category of the identified reference vector (nearest neighbor method). As a result, the determination unit 912 can classify the frame input to the classifier 91 (feature vector extraction unit 911) into one of the categories C1 to C3.

The candidate data extraction unit 93, the display control unit 94, and the teacher data generation unit 95 are functional units related to generation of teacher data. Image data acquired by the camera 70 is input to the candidate data extraction unit 93 . Of the image data, the candidate data extracting unit 93 stores image data acquired within a candidate period described below in a storage medium as candidates for teacher data (hereinafter referred to as candidate data).

The candidate period is a period that includes at least part of the period during which the treatment liquid is ejected from the ejection nozzles 61 . Further, the candidate period includes the timing for changing the ejection state of the treatment liquid. More specifically, it is a period including, for example, a discharge stop timing for stopping the discharge of the processing liquid from the discharge nozzles 61 . This candidate period can be set shorter than the ejection period during which the treatment liquid is ejected from the ejection nozzle 61 onto the substrate W. FIG. In this candidate period, the treatment liquid is initially ejected from the ejection nozzles 61, and the ejection of the treatment liquid is stopped in the middle of the candidate period. and image data showing a state in which the discharge of the treatment liquid from the discharge nozzles 61 is stopped. In addition, when ink drops occur when ejection is stopped, image data showing a state in which ink drops have occurred is also included. In other words, by storing a plurality of data within the candidate period as candidate data, image data corresponding to each category C1 to C3 can be stored as candidate data.

The storage medium in which the candidate data is stored may be the same as the storage medium in which the determination database 913 is stored, or may be a different storage medium. A plurality of candidate data stored in the storage medium is hereinafter also referred to as a candidate database 931 .

The display control unit 94 reads the candidate data from the storage medium in response to the operator's input on the input unit 82 and causes the display unit 81 to display the candidate data. Thereby, the worker can visually recognize the candidate data on the display unit 81 .

The operator visually recognizes the candidate data displayed on the display unit 81 and selects the candidate data to be adopted as teacher data. The operator performs input to the input unit 82 to specify candidate data to be adopted as teacher data, and inputs a label to be given to the candidate data to the input unit 82 . For example, a label of "normal ejection state" is assigned to image data representing a normal ejection state.

The teacher data generation unit 95 receives the candidate data being displayed from the display control unit 94 in response to the input regarding the label to the input unit 82, and transmits the label information input to the input unit 82 by the operator to the input unit. Received from 82. The teacher data generation unit 95 associates the candidate data and the label with each other and stores them in the storage medium as teacher data.

The operator sequentially visually recognizes other candidate data and determines whether or not to adopt them as teacher data. Then, the operator inputs a label for candidate data to be adopted as teacher data. As a result, a plurality of pieces of teacher data are stored in the storage medium.

The storage medium storing the teacher data may be the same as the storage medium storing the determination database 913 or the candidate database 931, or may be a storage medium different from these. A plurality of teacher data stored in the storage medium is hereinafter also referred to as a teacher database 921 .

FIG. 6 is a flow chart showing an example of processing for generating the classifier 91. As shown in FIG. Here, as an example, a case where the classifier 91 is generated before installation of the substrate processing apparatus 100 (that is, before shipment) will be described. As illustrated in FIG. 6, first, in step S1, teacher data is generated.

FIG. 7 is a flow chart showing an example of the teaching data generation process in step S1 of FIG. First, in step S<b>11 , the substrate W is held by the substrate holding part 20 . Specifically, the substrate W is transported onto the substrate holder 20 by the main transport robot 103 . The substrate holding part 20 holds the transported substrate W substantially horizontally.

Next, in step S12, the control unit 9 controls, for example, the rotation mechanism of the processing liquid supply unit 60 to move the discharge nozzle 61 to the processing position. Here, as an example, the processing liquid supply unit 60 is used for explanation, but the processing liquid supply units 30 and 65 may be used.

Next, in step S<b>13 , the control unit 9 causes the camera 70 to take an image when the discharge nozzle 61 is moved, for example. The camera 70 captures an image of the imaging area at a predetermined frame rate (for example, 60 frames/second) and sequentially outputs the acquired image data to the control section 9 .

Next, in step S14, the control unit 9 controls the spin motor 22 to rotate the substrate W in the horizontal plane, and in step S15, the treatment liquid is discharged. Specifically, the control unit 9 outputs an open signal to the on-off valve connected to the discharge nozzle 61 to discharge the treatment liquid onto the upper surface of the substrate W from the discharge nozzle 61 . Then, for example, when a predetermined processing period elapses, the control unit 9 outputs a close signal to the on-off valve to stop the ejection of the processing liquid from the ejection nozzle 61 . If a suck-back valve is provided, a suction signal is output to the suck-back valve. Next, in step S16, the control unit 9 causes the camera 70 to finish imaging, and in step S17, the control unit 9 controls the spin motor 22 to stop the rotation of the substrate W. FIG.

Next, in step S18, the candidate data extraction unit 93 stores the image data acquired by the camera 70 during the candidate period in the storage medium as candidate data. The candidate period is, for example, several seconds.

The candidate data extraction unit 93 can determine this candidate period based on the discharge stop timing of the discharge nozzles 61 . Specifically, the candidate data extraction unit 93 may determine a period from a timing earlier than the discharge stop timing by a first predetermined amount to a timing later than the discharge stop timing by a second predetermined amount as the candidate period. .

By the way, the opening/closing valve that governs the discharge/stop of the processing liquid from the discharge nozzle 61 is controlled by the controller 9 as described above. After the controller 9 outputs a close signal to the open/close valve, the ejection of the treatment liquid from the ejection nozzle 61 is stopped after a delay time corresponding to the length of the pipe. Since this delay time is usually not so long, the candidate data extractor 93 may determine the candidate period based on the output timing of outputting the close signal to the on-off valve. That is, this output timing may be regarded as the discharge stop timing. Alternatively, if a suck-back valve is provided, the candidate period may be determined based on the output timing of the suction signal to the suck-back valve. Alternatively, if another discharge stop signal is used, that discharge stop signal may be employed.

Alternatively, the candidate data extraction unit 93 may specify the discharge stop timing based on the image data from the camera 70 . In the image data from the camera 70, the treatment liquid from the ejection nozzle 61 is captured in the ejection region R1 extending from the tip of the ejection nozzle 61 in the ejection direction of the treatment liquid (see also FIG. 15). Specifically, when the treatment liquid is being ejected, the liquid columnar treatment liquid is reflected in the ejection region R1, and when the treatment liquid is not being ejected, the treatment liquid is not reflected in the ejection region R1. When the treatment liquid is reflected in the ejection region R1, the luminance value (or pixel value in gray scale) of the pixels in the ejection region R1 becomes relatively high, and the distribution thereof varies. Conversely, when the treatment liquid is not reflected in the ejection region R1, the upper surface of the substrate W is reflected, so the luminance values of the pixels in the ejection region R1 are relatively low and the distribution is more uniform.

Therefore, the statistic (for example, the sum or variance) of the pixel values in the ejection region R1 when the treatment liquid is ejected is the statistic (for example, the sum) of the pixel values in the ejection region R1 when the treatment liquid is not ejected. or variance). Therefore, the candidate data extracting unit 93 determines that the processing liquid is being discharged when the statistic is higher than the reference value, and the discharge of the processing liquid is stopped when the statistic is smaller than the reference value. I judge. The reference value can be set by simulation or experiment, for example. The candidate data extraction unit 93 may specify the discharge stop timing based on this determination result. According to this, the ejection stop timing can be specified with high accuracy. Therefore, even if the candidate period is set short, the ejection stop timing can be included in the candidate period. As a result, it is possible to obtain image data corresponding to the categories C1 to C3 in the candidate period.

The candidate data extraction unit 93 stores the plurality of candidate data acquired in this candidate period as a group of candidate data in a storage medium. Further, the candidate data extraction unit 93 preferably stores information accompanying the candidate data (hereinafter also referred to as accompanying information) in the storage medium in association with the candidate data group. The accompanying information includes, for example, information identifying the processing unit 1 to which the camera 70 that acquired the candidate data belongs, date and time information that the candidate data was acquired, and information that identifies the processing liquid supply unit shown in the candidate data. can be done.

By performing the series of operations described above while changing the processing unit 1 , the candidate data group can be stored in the storage medium for each processing unit 1 . Further, by changing the processing liquid supply section in each processing unit 1, the candidate data group can be stored in the storage medium for each processing liquid supply section.

Next, in step S19, the operator assigns a label to the candidate data by inputting to the input unit 82 to generate teacher data. Specifically, the operator first causes the display unit 81 to display the candidate data. FIG. 8 is a diagram schematically showing an example of an input screen IS1 for selecting candidate data. The input screen IS1 is provided with an image display area IR1, a list display area TR1, and a button area BR1. The image display area IR1 is a rectangular area in which candidate data is displayed. In the example of FIG. 8, the image display area IR1 is positioned at the upper left within the input screen IS1.

The button area BR1 includes a plurality of buttons for designating candidate data to be displayed in the image display area IR1. In the example of FIG. 8, the button area BR1 is positioned to the right of the image display area IR1. Operations on the buttons in the button region BR1 are realized by inputting to the input unit 82. FIG. The button area BR1 includes a unit selection area U1, a date selection area D1, an image selection area Ca1 and a label selection area L1.

The unit selection area U1 includes buttons for selecting one of the twelve processing units 1. FIG. For example, the unit selection area U1 includes buttons “1” to “12”, and these numbers indicate the identification numbers of the processing units 1. FIG.

The date selection area D1 includes buttons for designating the date when the candidate data was obtained. For example, the date selection area D1 includes a "previous day" button and a "next day" button, and a date can be selected by operating these buttons. Also, the date selection area D1 may include a frame in which a date such as "2016/3/25" is displayed. By inputting to the input unit 82, the date may be directly input by operating this frame.

The image selection area Ca1 includes buttons for designating candidate data from the candidate data group. For example, the image selection area Ca1 includes a "previous image" button and a "next image" button, and by operating these buttons, candidate data can be selected. Also, the image selection area Ca1 may include a frame for displaying the candidate data number relative to the total number of candidate data in the candidate data group, such as "1/64". By inputting to the input unit 82, the candidate data may be specified directly by performing an operation on this frame.

The label selection area L1 includes buttons for designating labels. For example, the label selection area L1 includes buttons for "normal discharge state", "dripping state", and "discharge stop state".

The operator operates buttons in the unit selection area U1 to select one of the processing units 1 by inputting to the input unit 82, and operates buttons in the date selection area D1 to designate a date. In response to the input to the input unit 82, the display control unit 94 reads a candidate data group that satisfies the conditions (processing unit 1 and date) (for example, reads from the candidate database 931, or reads the candidate data group (read from the history database 961 (see FIG. 10)), and the list is displayed in the list display area TR1 in tabular form. In the example of FIG. 8, the list display region TR1 displays the acquisition time of the candidate data group and the treatment liquid supply unit appearing in the candidate data group.

The operator can select one from the list of this table. This selection is also realized by input to the input section 82 . When the operator selects one of the candidate data groups from the table of the list display region TR1 by inputting to the input unit 82, the display control unit 94 displays one candidate data included in the selected candidate data group as an image. It is displayed in the display area IR1. For example, the first candidate data included in the candidate data group is displayed in the image display area IR1.

The operator operates buttons or the like in the image selection area Ca1 by inputting to the input unit 82 to sequentially display the candidate data included in the candidate data group in the image display area IR1. For example, when the operator operates a button indicating "next image", the display control unit 94 terminates the display of the candidate data displayed in the image display area IR1, and displays the next image in chronological order. Positioned candidate data is displayed in the image display area IR1.

The operator visually recognizes the candidate data displayed in the image display area IR1 and determines whether or not to employ the candidate data as teacher data. That is, the operator selects image data to be employed as teacher data from candidate data. When the operator determines that the candidate data is to be used as teacher data, the correct category is given as a label to the image data. Specifically, the operator operates buttons in the label selection area L1 by inputting to the input unit 82 to input the correct label for the candidate data. In response to the input to the input unit 82, the teacher data generation unit 95 receives the candidate data displayed in the image display area IR1 at that time from the display control unit 94, and generates the candidate data according to the input. The associated labels are associated with each other, and the associated candidate data and labels are stored in a storage medium as teacher data.

The operator can generate a plurality of teacher data (teacher database 921) by repeating the above operation. A plurality of teacher data are stored in a storage medium.

As described above, the teacher data generation unit 95 assigns labels to the image data according to the discharge state of the treatment liquid appearing in at least one of the image data based on the operator's input to the input unit 82, Generate teacher data.

Next, in step S2, the machine learning unit 92 reads teacher data from the teacher database 921 and generates the classifier 91 by learning processing based on the teacher data. Classifier 91 may also be referred to as a trained model. Note that the machine learning section 92 may generate the classifier 91 for each processing unit 1 or for each processing liquid supply section. Alternatively, the machine learning section 92 may generate the classifier 91 according to the combination of the processing unit 1 and the processing liquid supply section.

Next, in step S3, the accuracy of the classifier 91 is evaluated. For example, image data is input to the classifier 91 and the classifier 91 classifies the image data. The classifier 91 calculates the degree to which the input image data corresponds to each category (degree of relevance), and classifies the image data into the category with the highest degree of relevance. For example, the "normal discharge state" is classified into category C1, the "dripping state" into category C2, and the "discharge stopped state" into category C3. In this classification process, if the degree of relevance for the correct category is sufficiently higher than the degree of relevance for other categories, it can be said that the classifier 91 has correctly classified. Conversely, when the degree of correspondence is low, it can be said that the classifier 91 still has room for improvement.

Therefore, the control unit 9 determines whether the degree of relevance to the correct category is higher or lower than the reference value. Update the classifier 91. The machine learning unit 92 may perform learning processing to update the classifier 91 after newly adding teacher data.

Classifier 91 can be judged to be correct when the degree of relevance to the correct category is higher than the reference value. However, if the accuracy evaluation is performed only for the classification result of one image data, it can be said that the evaluation is insufficient. Therefore, it is desirable to perform the accuracy evaluation for the classification results of a plurality of image data. More specifically, the classifier 91 is judged to be appropriate when the degree of relevance to the correct category is higher than the reference value in the classification results of a predetermined number of times. If the classifier 91 determines that it is appropriate, the process ends.

As described above, according to this operation, the image data obtained by the camera 70 and showing various ejection states are adopted as the teacher data. Therefore, it is possible to generate a classifier 91 that can classify the discharge state.

Moreover, in the above example, the control unit 9 stores image data within a short candidate period including the timing at which the ejection state of the treatment liquid changes (e.g., ejection stop timing) as candidates for the teacher data (candidate data) in the storage medium. there is Therefore, the number of candidate data can be reduced compared to the case where all image data within the processing period are stored as candidate data. Therefore, the operator can reduce the time and effort required for checking candidate data, and the selection of image data is facilitated. Also, the required capacity of the storage medium can be reduced.

It may be desired to update the classifier 91 even after installation of the substrate processing apparatus 100 (that is, after shipment). For example, when the classifier 91 misclassifies image data, it is desirable to update the classifier 91 so as not to misclassify image data similar to that image data.

FIG. 9 is a diagram schematically showing an example of a portion of misclassified image data. FIG. 9 shows an enlarged portion of the image data. In this image data, the treatment liquid is not ejected from the ejection nozzles 61 . Further, the upper surface of the substrate W of this image data includes a substantially circular pattern P1 discretely. This pattern P1 is a pattern formed on the upper surface of the substrate W, for example. In the example of FIG. 9, the patterns P1 are arranged substantially vertically at intervals directly below the tip of the ejection nozzle 61 . The pixel values of the pixels in this circular pattern P1 may be greater than the pixel values of the pixels on the upper surface of the other substrate W. FIG. In this case, among the image data, the distribution of pixel values in the ejection region R1 immediately below the tip of the ejection nozzle 61 can be similar to the distribution of pixel values in the ejection region when ink drops occur. Therefore, the classifier 91 can classify the image data into category C2 (dripping state) even though the dripping does not actually occur.

Also, in the example of FIG. 9, the pattern P1 is formed in a relatively wide area of the substrate W, but it may be formed in a predetermined area on the upper surface of the substrate W in some cases. The pattern P1 can be formed by reflecting the light from the illumination section 71 on the upper surface of the substrate W. As shown in FIG. For example, while the substrate W is rotating, the light diffusely reflected by the metal pattern of the predetermined region on the upper surface of the substrate W forms the pattern P1. Then, when the predetermined area enters directly under the nozzle 61 as the substrate W rotates, the image data captured by the camera 70 also includes the pattern P1 in the ejection area R1 directly under the nozzle 61 . Even in this case, the distribution of pixel values in the ejection region R1 can be similar to the distribution of pixel values in the ejection region when ink drops occur. Therefore, the classifier 91 can classify the image data into category C2 (dripping state) even though the dripping does not actually occur.

After that, in order to reduce the possibility of such misclassification, it is desirable to adopt the misclassified image data as teacher data and update the classifier 91 by re-learning based on the teacher data.

FIG. 10 is a diagram schematically showing an example of the internal configuration of the control section 9A. The control section 9A has the same configuration as the control section 9 except for the presence or absence of the history data storage control section 96. FIG. Image data acquired by the camera 70 and classification result information by the classifier 91 are input to the history data storage control unit 96 . The history data storage control unit 96 associates each image data with its classification result and stores them in a storage medium as history data. The history data storage controller 96 may store a history data group for each process performed on the substrate W in a storage medium. In other words, each time a substrate W is processed, a plurality of pieces of history data acquired by the processing may be stored in the storage medium as a collection of history data groups. Further, the history data storage control unit 96 associates accompanying information similar to the candidate data (for example, identification information of the processing unit 1, date and time information, and identification information of the treatment liquid supply unit) with the history data and stores them in the storage medium. You may

The storage medium in which history data is stored may be the same as or different from the storage medium in which at least one of judgment database 913, candidate database 931, and teacher database 921 is stored. A plurality of pieces of history data stored in the storage medium are hereinafter also referred to as a history database 961 .

The input unit 82 can receive input for instructing display of history data. When the operator makes the input, the display control unit 94 causes the display unit 81 to display history data (image data and classification results) in response to the input.

The operator visually confirms whether the classification result of the image data displayed on the display unit 81 is correct. When this classification is incorrect, that is, when the image data is misclassified, the operator inputs a correct category other than the misclassified category as a label by inputting to the input unit 82 . The teacher data generation unit 95 associates the image data displayed on the display unit 81 and the label input to the input unit 82 with each other, and stores them in the storage medium as new teacher data. That is, the teacher database 921 is updated.

FIG. 11 is a flowchart showing an example of update processing of the classifier 91 by re-learning. First, in step S21, the operator selects the misclassified image data as teacher data. Specifically, first, the operator causes the display unit 81 to display the browsing screen PS1 by inputting to the input unit 82 .

FIG. 12 is a diagram schematically showing an example of the browsing screen PS1. The viewing screen PS1 is similar to the input screen IS1 except for the following points. In other words, not the candidate data but the history data is displayed in the image display area IR1. In the list display area TR1, a list of historical data groups for the designated processing unit 1 and date and time is displayed. Further, the table of the list display area TR1 is provided with an item of "pass/fail". The "good or bad" item indicates the good or bad of the ejection state of the treatment liquid. As a more specific example, if the classification result of the history data included in the history data group includes category C2 (spotted state), the ejection state of the treatment liquid is defective, so it is defective. A black dot indicating that is shown for that historical data group. If nothing is displayed in the item of "good or bad", it means that the treatment liquid was discharged normally and stopped normally.

When the operator selects a history data group in which the discharge state is determined to be defective by inputting to the input unit 82, the display control unit 94 reads out the selected history data group from the history database 961 in response to the input. , to display the image data of the history data group in the image display area IR1. Then, the operator operates the buttons in the image selection area Ca1 by inputting to the input unit 82 to sequentially display the image data included in the history data group in the image display area IR1. In particular, the operator visually determines whether or not the image data is smeared when the ejection of the treatment liquid is stopped.

If dripping occurs, the classification result by the classifier 91 is appropriate, so the operator ends the process.

On the other hand, if no dripping occurs, the classification result by the classifier 91 is inappropriate, so that image data is adopted as teacher data. That is, a correct label is assigned to the image data. Specifically, while the image data is displayed in the image display area IR1, the operator operates the buttons in the label selection area L1 by inputting to the input unit 82 to input the correct label for the image data. do. For example, for the image data shown in FIG. 9, the label "ejection stop state" is input.

In response to the input to the input unit 82, the teacher data generation unit 95 receives from the display control unit 94 the image data displayed in the image display area IR1 at that time, and assigns a label corresponding to the input to the image data. are stored in a storage medium as teacher data. Thereby, teacher data for re-learning can be generated.

Next, in step S22, the machine learning unit 92 updates the classifier 91 by re-learning based on teacher data including new teacher data. Next, steps S23 and S24 are executed. Steps S23 and S24 are the same as steps S3 and S4, respectively.

By the above operation, the correct category for misclassified image data can be reflected in the classifier 91 . Therefore, it is possible to reduce the possibility of misclassification by the updated classifier 91 .

Note that in the above example, the label "ejection stop state" is given to the erroneously classified image data. However, the feature vector of the erroneously classified image data was erroneously classified because it was close to the category C2 of "dripping state". The latter classifier 91 may easily misclassify the image data that should be classified into category C2 of "dripping state" into category C3 of "ejection stopped state".

Therefore, a new label may be assigned to the erroneously classified image data. In other words, another category C4 may be provided in addition to the categories C1 to C3 as the type of classification. This category C4 may be a category indicating misclassification. As a more specific example, in the image data, a category (label) indicating that a pattern (for example, pattern) P1 similar to dripping is formed on the upper surface of the substrate W may be employed.

FIG. 13 is a diagram schematically showing an example of the viewing screen PS1A. The browsing screen PS1A of FIG. 13 is the same as that of FIG. 12 except for the types of buttons in the label selection area L1. In FIG. 13, the label selection area L1 includes a button of "pattern similar to dripping" as a label corresponding to category C4. When the operator operates this button by inputting to the input unit 82, the teacher data generation unit 95 responds to the input and adds the input label to the history data displayed in the image display area IR1 at that time. are associated with each other and stored in the storage medium as new teacher data. As described above, a dedicated label (label corresponding to category C4) can be assigned to erroneously classified image data.

Next, conditions for executing classification processing by the classifier 91 generated (or updated) as described above will be described. The classifier 91 classifies all of the image data sequentially input from the camera 70 during the processing of the substrate W using the processing liquid into each category, thereby determining whether the discharge state of the processing liquid is good or bad. good. Specifically, the control section 9 (or the control section 9A) may determine that the ejection state is defective when the image data is classified into category C2. However, it is not necessarily limited to this. Before the classification processing by the classifier 91, the control unit 9 (or the control unit 9A) may temporarily determine the defective ejection state by performing simpler processing on the image data.

FIG. 14 is a functional block diagram schematically showing an example of the internal configuration of the control section 9B. The control section 9B is the same as the control section 9 except for the presence or absence of the provisional determination section 97. FIG. Image data acquired by the camera 70 is input to the provisional determination unit 97 . A provisional determination unit 97 performs image processing on the image data to provisionally determine whether or not the ejection state is defective. Specifically, the provisional determination unit 97 determines whether or not the ejection state is the dripping state.

This provisional determination is performed by focusing on the ejection area described below in the image data. FIG. 15 is a diagram schematically showing an example of image data. As shown in FIG. 15, for example, the ejection region R1 is a rectangular region extending in the ejection direction of the treatment liquid from the tip of the ejection nozzle 61 in the image data, and has a width wider than the width of the treatment liquid. The ejection region R1 is set, for example, to a length that does not include the liquid contact point between the processing liquid and the substrate W. As shown in FIG.

The provisional determination unit 97 calculates the statistic of the pixel values within the ejection region R1. This statistic is a value corresponding to the area of the image of the treatment liquid within the ejection region R1, and is, for example, the variance (eg, standard deviation) of the pixel values within the ejection region.

Now, when the processing liquid is normally discharged from the discharge nozzle 61, the processing liquid extends from end to end in the vertical direction within the discharge region R1, and the upper surface of the substrate W is reflected on both sides in the horizontal direction. Since the luminance value of the pixels in which the treatment liquid is reflected increases due to the reflection of light on the treatment liquid, etc., the variation in the luminance value within the ejection region R1 becomes relatively high.

Note that the camera 70 may be of a type that acquires grayscale image data, or may be a type of camera that acquires color image data. In the former case, it can be said that the pixel value of the image data indicates the luminance value. A camera that acquires grayscale image data will be described below as an example, but in the case of color, the luminance value is calculated from the pixel value and used.

Now, when dripping occurs, the treatment liquid is arranged discretely in the vertical direction in the ejection region R1 (see also FIG. 5). Therefore, the variation in pixel values within the ejection region R1 is even higher than in the normal ejection state. When the treatment liquid is not ejected from the ejection nozzles 61, the upper surface of the substrate W is reflected in the ejection region R1, so the pixel values in the ejection region R1 are relatively small, and the variation thereof is also relatively small. Therefore, the dripping state (defective ejection state) can be detected according to the dispersion (for example, standard deviation) of the pixel values in the ejection region R1.

For example, the provisional determination unit 97 calculates the variance within the ejection region R1 and determines whether or not the statistic (variance) is greater than a reference value. When the variance is greater than the reference value, the provisional determination unit 97 provisionally determines that the falling state has occurred, and notifies the classifier 91 to that effect.

The classifier 91 performs classification processing on the image data with the notification from the temporary determination unit 97 as a trigger. In other words, if the classifier 91 does not receive a notification from the provisional determination section 97, it does not perform the classification process.

According to such a control unit 9B, the classification process by the classifier 91 is performed when a provisional determination is made by simpler image processing and it is provisionally determined that a defective ejection state has occurred. Therefore, compared to the case where the classification processing by the classifier 91 is always performed, the processing in the control section 9B can be lightened.

Although the embodiments of the teaching data generation method and the ejection determination method of the ejection state have been described above, this embodiment can be modified in various ways other than those described above without departing from the gist thereof. . The various embodiments and modifications described above can be implemented in appropriate combinations.

For example, as the substrate W, a semiconductor substrate was used in the description, but the substrate W is not limited to this. For example, substrates such as photomask glass substrates, liquid crystal display glass substrates, plasma display glass substrates, FED (Field Emission Display) substrates, optical disk substrates, magnetic disk substrates, or magneto-optical disk substrates may be used. good.

Furthermore, the present embodiment can be applied to any apparatus that performs a predetermined process by discharging a processing liquid from a movable nozzle onto a substrate. For example, in addition to the single-wafer cleaning processing apparatus of the above embodiment, a spin coating apparatus (spin coater) that applies resist by discharging a photoresist solution from a nozzle onto a rotating substrate, and a substrate with a film formed on its surface. The technique according to the present embodiment may be applied to a device that discharges a film removing liquid from a nozzle onto the edge of a substrate, or a device that discharges an etchant from a nozzle onto the surface of a substrate.

Further, in the present embodiment, the treatment liquid is ejected as a continuous stream from the ejection nozzles, but this is not necessarily the case. The treatment liquid may be ejected in the form of mist from the ejection nozzle.

Further, the history data storage control unit 96 may store all of the image data during the processing period and the classification result for the image data as history data in the recording medium. This makes it possible to check all the image data during the processing period. Alternatively, the history data storage control unit 96 may store, for example, image data only in the candidate period and classification results for the image data as history data in the recording medium.

Also, the teacher data may be stored in an external server or the like. The teacher data stored in the external server may be transmitted to the control unit of another substrate processing apparatus, and the classifier may be generated in the control unit.

20 Substrate holding mechanism (substrate holding part)
31, 61, 66 discharge nozzle 70 camera 90 user interface 91 classifier 100 substrate processing apparatus W substrate

Claims (4)

A method for generating teacher data about the discharge state of a discharge nozzle in a substrate processing apparatus, comprising:
a holding step of horizontally holding the substrate in the substrate holding mechanism;
a treatment liquid ejection step of ejecting the treatment liquid from the ejection nozzle toward the main surface of the substrate;
An imaging region adjusted to include the ejection nozzle and at least a portion of the treatment liquid ejected in the treatment liquid ejection step is imaged by a camera for a period including at least a portion of the treatment liquid ejection step. and an imaging step of acquiring a plurality of image data;
a teacher data generation step of generating teacher data by adding a label to at least one of the plurality of image data in accordance with the ejection state of the treatment liquid shown in the image data;
When image data is misclassified into a certain category by a classifier generated based on the training data, the image data is adopted as training data, and a first class corresponding to the category is applied to the training data. a step of applying a second label different from the label;
with
The category includes a category indicating a dripping state in which the treatment liquid drops as droplets when the discharge of the treatment liquid is stopped,
In the second label, the pattern formed on the surface of the substrate immediately below the ejection nozzle in the image data is similar to the dripping pattern, and thus the discharge stop state is misclassified as the dripping state. The teacher data generation method, which is a label indicating that . A method for generating teacher data about the discharge state of a discharge nozzle in a substrate processing apparatus, comprising:
a holding step of horizontally holding the substrate in the substrate holding mechanism;
a treatment liquid ejection step of ejecting the treatment liquid from the ejection nozzle toward the main surface of the substrate;
An imaging region adjusted to include the ejection nozzle and at least a portion of the treatment liquid ejected in the treatment liquid ejection step is imaged by a camera for a period including at least a portion of the treatment liquid ejection step. and an imaging step of acquiring a plurality of image data;
a teacher data generation step of generating teacher data by adding a label to at least one of the plurality of image data in accordance with the ejection state of the treatment liquid shown in the image data;
The classifier generated based on the training data classifies the image data into one of a plurality of categories indicating the ejection state of the treatment liquid, and the image data is erroneously classified into an erroneous category different from the correct ejection state. When classified, the image data is adopted as teacher data, and the teacher data is put into an incorrect category different from the first label corresponding to the category classified according to the ejection state of the treatment liquid. a step of assigning a second label indicating that it has been classified;
with
The teacher data generation method, wherein the second label is a label indicating misclassification. The teacher data generation method according to claim 1 or claim 2 ,
The training data generation step includes:
a storage step of storing in a storage medium image data obtained by the camera during a period shorter than the period of the treatment liquid ejection step, including the timing at which the ejection state of the treatment liquid changes, as a candidate for teacher data;
a selection step of selecting image data to be adopted as training data from the candidates;
A teaching data generating method, comprising a labeling step of adding a label to teaching data in accordance with the discharge state of the treatment liquid reflected in the teaching data. A step of performing machine learning using the teacher data generated by the teacher data generation method according to any one of claims 1 to 3 to generate a classifier;
horizontally holding the substrate on the substrate holding mechanism;
a step of ejecting the treatment liquid from the ejection nozzle toward the main surface of the substrate;
An imaging region updated to include the ejection nozzle and at least a portion of the treatment liquid ejected in the treatment liquid ejection step is imaged by a camera for a period including at least a portion of the treatment liquid ejection step. and acquiring a plurality of image data;
provisional determination of whether the ejection state of the treatment liquid from the ejection nozzle is good or bad based on a statistic of pixel values in an area extending from the tip of the ejection nozzle in the ejection direction of the treatment liquid in the image data; classifying the image data into a category according to the discharge state by the classifier when it is provisionally determined that the state is defective ;
A method for determining an ejection state, comprising :

Images