Method and device for estimating likelihood of fine powder generation in polysilicon, method for manufacturing polysilicon, device and method for generating trained model

Abstract

The ease of occurrence of fine powder in polysilicon is estimated. Provided is a method for estimating the ease of fine powder occurrence in polysilicon that includes an estimation step that estimates the ease of occurrence of fine powder when vibration is applied to a second polysilicon from second imaging data that include images of the second polysilicon to be estimated, using a first trained model generated using as teaching data determination data on whether the amount of fine powder occurring due to vibration is at or below a first threshold in a first polysilicon to which vibration has been applied under predetermined conditions and first imaging data that include images of the first polysilicon.

Description

[Technical field]

[0001] The present invention relates to a method and an apparatus for estimating the likelihood of fine powder generation in polysilicon, a method for producing polysilicon, and an apparatus and method for generating a trained model. [Background technology]

[0002] In polysilicon, fine powder may be generated due to vibration during transportation, for example. Such fine powder may not only make the polysilicon difficult to handle, but may also cause contamination of the polysilicon.

[0003] One of the factors that causes the generation of such fine powder is thought to be unevenness that occurs on the surface depending on the precipitation state of polysilicon. Conventionally, methods for classifying polysilicon based on its surface state have been proposed (for example, Patent Documents 1 to 3). [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japan Special Publication No. 2022-537014 [Patent Document 2] Japan Special Publication No. 2016-516666 [Patent Document 3] Japan Special Publication No. 2009-532319 Summary of the Invention [Problem to be solved by the invention]

[0005] However, the techniques disclosed in Patent Documents 1 to 3 do not focus on the fine powder generated from polysilicon, and therefore cannot be said to be suitable for separating polysilicon, which is prone to generating fine powder.

[0006] An object of one aspect of the present invention is to provide a method capable of estimating the likelihood of generation of fine powder in polysilicon. [Means for solving the problem]

[0007] In order to solve the above-mentioned problems, a method for estimating the likelihood of fine powder generation in polysilicon according to one embodiment of the present invention includes an estimation step of estimating the likelihood of fine powder generation in second polysilicon when vibration is applied from second imaging data including an image of second polysilicon that is the estimation target, using a first trained model generated using judgment data indicating whether or not an amount of fine powder generation due to vibration of a first polysilicon applied with predetermined conditions is equal to or less than a predetermined first threshold value, and first imaging data including an image of the first polysilicon as training data.

[0008] In order to solve the above-mentioned problems, a method for producing polysilicon according to one embodiment of the present invention includes an estimation step of estimating the likelihood of fine powder generation in a second polysilicon when vibration is applied from second imaging data including an image of a second polysilicon to be estimated, using a first trained model generated using as training data determination data on whether an amount of fine powder generation due to vibration of a first polysilicon applied with vibration under predetermined conditions is equal to or less than a predetermined first threshold value and first imaging data including an image of the first polysilicon, and a first sorting step of sorting the second polysilicon based on the likelihood of fine powder generation estimated in the estimation step.

[0009] In order to solve the above-mentioned problems, an estimation device for estimating the likelihood of fine powder generation in polysilicon according to one embodiment of the present invention includes an estimation unit that uses a first trained model generated using, as training data, determination data on whether an amount of fine powder generated by vibration of a first polysilicon to which vibration under predetermined conditions is applied is equal to or less than a predetermined first threshold value, and first imaging data including an image of the first polysilicon, to estimate the likelihood of fine powder generation in the second polysilicon when vibration is applied from second imaging data including an image of the second polysilicon that is the estimation target.

[0010] In order to solve the above problem, a trained model generation device according to one embodiment of the present invention includes an acquisition unit that acquires, as teacher data, determination data on whether an amount of fine powder generated due to vibration of a first polysilicon applied with a predetermined condition is equal to or less than a predetermined first threshold value, and first imaging data including an image of the first polysilicon, and a generation unit that generates a first trained model from input second imaging data including an image of a second polysilicon that is the subject of estimation, and outputs an estimation result of the likelihood of fine powder generation in the second polysilicon when vibration is applied.

[0011] In order to solve the above-mentioned problem, a method for generating a trained model according to one embodiment of the present invention includes an acquisition step of acquiring, as teacher data, judgment data on whether or not an amount of fine powder generated due to vibration of a first polysilicon applied with a predetermined condition is equal to or less than a predetermined first threshold value, and first imaging data including an image of the first polysilicon, and a generation step of generating, from input second imaging data including an image of a second polysilicon that is the subject of estimation, a first trained model that outputs an estimation result of the likelihood of fine powder generation in the second polysilicon when vibration is applied. Effect of the Invention

[0012] According to one aspect of the present invention, it is possible to realize a method capable of estimating the likelihood of fine powder generation in polysilicon. [Brief description of the drawings]

[0013] [Figure 1] 1 is a schematic diagram outlining a part of a method for producing polysilicon; FIG. [Diagram 2] 2 is a block diagram showing functional configurations of a generation device, an estimation device, and a separation control device according to an embodiment of the present invention. FIG. [Diagram 3] 1 is a flowchart illustrating a method for manufacturing polysilicon in accordance with an embodiment of the present invention. [Figure 4] FIG. 1 is a diagram showing an example of classification of polysilicon according to its surface state. [Diagram 5]FIG. 1 is a diagram showing the size distribution of crushed polysilicon particles packed in a PE bag. [Figure 6] FIG. 13 is a diagram showing the direction and intensity of vibration applied to fragments of polysilicon. [Figure 7] FIG. 13 is a diagram showing the amount of fine powder in a PE bag packed with crushed polysilicon before and after test transportation. [Figure 8] FIG. 4 is a schematic diagram showing an overview of the first acquisition step to the estimation step shown in FIG. 3. [Figure 9] 4 is a schematic diagram showing an overview from the third acquisition step to the step of calculating the area ratio of the concave-convex region shown in FIG. 3. [Figure 10] 4 is a schematic diagram showing an overview of a third acquisition step to a dent number calculation step shown in FIG. 3. [Figure 11] FIG. 13 is a diagram showing an example of threshold setting in classification based on the total area of ​​polysilicon and the area ratio of uneven regions. [Figure 12] FIG. 13 is a diagram showing an example of threshold setting in classification based on the total area number of depressions in polysilicon. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0014] Summary of the Invention One embodiment of the present invention relates to a method for producing polysilicon. Polysilicon can be produced by depositing trichlorosilane as a raw material using, for example, a bell jar. The deposited rod-shaped polysilicon may have a surface condition that varies depending on the deposition position, etc.

[0015] Specifically, the surface of the deposited polysilicon may be uneven, with the degree of the unevenness varying depending on the position on the polysilicon surface. Such unevenness is sometimes called popcorn or cone because of its shape similar to popcorn.

[0016] The polysilicon surface is inferior in quality in a portion where the surface is highly uneven. For example, the polysilicon surface is often fragile and is likely to break during transportation. In addition, the portion where the surface is highly uneven is prone to generating fine powder due to friction between the polysilicon particles or between the polysilicon and the packaging. Such fine powder not only makes the polysilicon difficult to handle, but can also cause contamination of the polysilicon.

[0017] The fine powder may be polysilicon fragments or powder smaller than a predetermined size. For example, when the size of the polysilicon to be the product is more than 5 mm and not more than 120 mm, the fine powder may be polysilicon having a size of 5 mm or less.

[0018] In this way, the size of the fine powder may be appropriately determined according to the size of the polysilicon. The size of the polysilicon and the fine powder respectively indicates the major axis. The major axis indicates the distance between two points where the distance is maximum in the polysilicon or the fine powder. In addition, since it may be difficult to accurately measure the size of the fine powder with a major axis of 5 mm or less, for example, when the fine powder is defined as a size of 5 mm or less, it may be considered that the fine powder passes through a 5 mm sieve.

[0019] The rod-shaped polysilicon may be divided into divided bodies by a dividing process such as crushing or cutting for storage or transportation. The surface condition of such divided bodies differs depending on the position in the rod-shaped polysilicon before division. Therefore, the physical properties such as the tendency to generate fine powder also differ for each divided body. In this specification, when the term "polysilicon" is simply referred to, the shape is not particularly limited, and the polysilicon may be in the form of either a rod or a divided body.

[0020] FIG. 1 shows an example of a method for crushing polysilicon to which the method for producing polysilicon according to one embodiment of the present invention can be applied. Rod-shaped polysilicon P has an uneven area P1 on a part of its surface, where unevenness of a predetermined roughness or more exists. In order to obtain crushed polysilicon that is less likely to generate fine powder, it is preferable to remove the crushed material including the uneven area P1 from the crushed material group. In FIG. 1, the crushed material including the uneven area P1 is shown as "with cone", the crushed material not including the uneven area P1 is shown as "without cone", and the crushed material group in which the crushed material including the uneven area P1 and the crushed material not including the uneven area P1 are mixed is shown as "cone MIX".

[0021] In order to remove the crushed object including the uneven region P1, the crushing method described above includes a rough crushing step S1, a crushing step S2, and a chipping step S3.

[0022] The rough crushing step S1 is a step for separating the polysilicon P into a portion having the rough region P1 on its surface and a portion not having the rough region P1 in order to remove the rough region P1. The crushing step S2 is a step for crushing the portion having the rough region P1 on its surface and the portion not having the rough region P1 obtained in the rough crushing step S1, respectively, to obtain crushed objects. If the portion of the polysilicon P not having the rough region P1 is further crushed, a group of crushed objects not having the rough region P1 is obtained.

[0023] On the other hand, when the portion of the polysilicon P that includes the uneven region P1 is crushed, a group of crushed materials is obtained that is a mixture of crushed materials that include the uneven region P1 and crushed materials that do not include the uneven region P1. The crushed materials that include the uneven region P1 are further crushed and only the uneven region P1 is scraped off, that is, a chipping process S3 is performed, whereby the crushed materials that do not include the uneven region P1 and the crushed materials that are mostly made up of the uneven region P1 can be separated.

[0024] The classification of the crushed materials in the chipping process S3 can be done by a human being by visually inspecting the surface condition of the crushed materials. However, it was not clear to what extent, among the various surface conditions shown in Fig. 4, the contribution of these to the generation of fine powder by vibration. In addition, there was a problem that it was difficult to standardize the judgment criteria by visual inspection, etc.

[0025] According to one embodiment of the present invention, it is possible to classify polysilicon based on a criterion for the correlation between the uneven area P1 and the likelihood of fine powder generation, which is learned based on the likelihood of fine powder generation under any vibration condition.

[0026] In order to realize such a crushing method, the manufacturing method of polysilicon exemplified in the embodiment 1 of the present invention described later uses a trained model generated by using images of polysilicon that are likely to generate fine powder as training data to estimate the likelihood of fine powder generation in polysilicon. Then, the polysilicon can be sorted based on the result of the estimation. With this configuration, polysilicon that does not include an uneven area P1 on the surface and is unlikely to generate fine powder can be selected with high accuracy.

[0027] In addition, the method for producing polysilicon exemplified in the second embodiment of the present invention described later estimates the area ratio of the uneven region P1 on the surface of the polysilicon using a trained model generated using the surface state of polysilicon that is prone to generating fine powder as training data. By separating the polysilicon based on the area ratio, etc., it is possible to select polysilicon suitable for the chipping process S3. With this configuration, it is possible to efficiently remove the uneven region P1, and the production efficiency of polysilicon can be greatly improved.

[0028] Here, extremely high purity is required for polysilicon products. Therefore, polysilicon is etched for purification in the final stage of commercialization. In order to maintain the cleanliness of the surface, it is desirable to package the etched polysilicon so that it does not come into contact with anything other than polysilicon.

[0029] Inspecting the surface condition before etching and before crushing in order to remove the uneven region P1 is a labor-intensive task. Therefore, all of the conventional surface condition inspection methods target polysilicon after etching, and serve as final product inspections, with the aim of not only inspecting the surface condition but also preventing the inclusion of foreign matter.

[0030] Therefore, there are many opportunities for contact with the gloves of inspection workers during post-etching inspection work, and there is a risk of impairing the cleanliness of the surface during the post-etching inspection process.

[0031] As an example, the weight percentage of polysilicon P before the rough crushing step S1 is 100%. In this case, after passing through the rough crushing step S1, the crushing step S2, and the chipping step S3, the weight percentage of the crushed material containing the uneven regions P1 unsuitable for use as a product is, for example, about 25%, depending on the deposition conditions of the polysilicon P. If the crushed material containing the uneven regions P1 can be detected before etching, inspection of the surface condition after etching is unnecessary and the surface cleanliness is not impaired, so the surface cleanliness of the polysilicon is improved compared to the conventional method. In addition, the burden of inspecting the surface condition of the polysilicon can be reduced.

[0032] According to the method for producing polysilicon according to an embodiment of the present invention, the likelihood of generating fine powder can be estimated before etching of polysilicon, and the polysilicon can be classified based on the result of the estimation. Since the shape of the unevenness on the surface of polysilicon does not change significantly before and after etching, the likelihood of generating fine powder in the polysilicon product after etching can be accurately estimated by the inspection before etching.

[0033] This makes it possible to exclude polysilicon that is prone to generating fine powder from being etched. In order to obtain more valuable polysilicon for semiconductors, it is only necessary to etch polysilicon that is less likely to generate fine powder. This makes it possible to greatly improve the efficiency of polysilicon production compared to conventional methods. This effect will also contribute to the achievement of Goal 12 of the Sustainable Development Goals (SDGs) advocated by the United Nations, "Ensure sustainable consumption and production patterns."

[0034] [Embodiment 1] 3, the method for producing polysilicon according to this embodiment includes at least an estimation step S14 and a first classification step S15, and may further include a first acquisition step S11, a first generation step S12, and a second acquisition step S13. In the estimation step S14, for example, estimation is performed using the first trained model obtained in the first generation step S12. First, a trained model generation method according to this embodiment for acquiring the first trained model will be described.

[0035] <Generation of the first trained model> The trained model generation method according to the present embodiment may be executed, for example, by a trained model generation device 10 illustrated in Fig. 2. Hereinafter, the trained model generation method will be described using an example in which the method is executed by the generation device 10.

[0036] 2, the generating device 10 has a first acquisition unit (acquisition unit) 11 and a first generation unit (generation unit) 12. The generating device 10 may be a control device capable of executing the processes of the first acquisition unit 11 and the first generation unit 12, and may be, for example, a CPU (Central Processing Unit). Moreover, the generating device 10 may be a device that includes such a control device, as well as a storage device, a communication device, and the like.

[0037] (First acquisition process and first teacher data) As shown in FIG. 3, the first acquisition unit 11 executes a first acquisition step (acquisition step) S11 for acquiring first teacher data used to generate a first trained model. The first teacher data is data including judgment data and first imaging data. The judgment data is data indicating whether or not the amount of fine powder generated by vibration in the first polysilicon to which vibration under a predetermined condition is applied is equal to or less than a predetermined first threshold. The first imaging data is data including an image of the first polysilicon.

[0038] The first polysilicon is polysilicon for which the tendency of fine powder generation due to vibration is known. For example, it is preferable that the first polysilicon has information on the amount of fine powder generation due to vibration under a predetermined condition.

[0039] The first polysilicon may be, for example, a final polysilicon product, and the amount of fine powder generated before and after transportation is known. The first polysilicon may be a polysilicon sample or the like, and vibration is applied under conditions similar to those in the transportation of the product, and the amount of fine powder generated before and after the vibration is known. The vibration applied to the first polysilicon does not have to be under conditions similar to those in the transportation of the product, and may be vibration under conditions different from those in the transportation of the product, such as an accelerated test.

[0040] The first threshold value in the judgment data may be, for example, a tolerance value for the amount of fine powder generated for vibration under a predetermined condition. In this case, if the amount of fine powder generated is equal to or less than the first threshold value, the first polysilicon can be said to be polysilicon that is unlikely to generate fine powder. The method of setting the first threshold value is not particularly limited, and a value suitable for separating polysilicon that is unlikely to generate fine powder may be appropriately set.

[0041] The determination of whether the amount of fine powder generated in the first polysilicon is equal to or less than the first threshold value may be performed for each first polysilicon based on a measurement value of the actual amount of fine powder generated. The determination may also be performed based on the surface condition of the first polysilicon. This is based on the finding experimentally revealed by the present inventors that the rougher the surface of polysilicon is, the more likely fine powder is generated.

[0042] As shown in FIG. 4, the inventors classified the polysilicon fragments into groups according to their surface conditions. Specifically, the group with the smoothest surface was classified into level 1, and the group with the roughest surface was classified into level 8, and the polysilicon fragments were relatively classified into eight groups. FIG. 4 shows an example of the surface of each group of polysilicon fragments photographed at the same magnification. In this case, the inventors found that the polysilicon fragments classified into levels 5 to 8 have a higher amount of fine powder than the polysilicon fragments classified into levels 1 to 4. In other words, the inventors found that if the polysilicon has a surface condition classified into level 4 or lower, the amount of fine powder generated is likely to be within the allowable range for commercialization.

[0043] Specifically, 5 kg of nuggets were prepared so as to have the weight distribution (weight%, left vertical axis) and cumulative weight distribution (weight%, right vertical axis) for the size (long diameter, horizontal axis) of the crushed polysilicon (nuggets) as shown in FIG. 5, and were packed into a PE (polyethylene) bag. The size of the nuggets was obtained by measuring the long diameter, which is the maximum side length of the nuggets, with a vernier caliper. The size of the nuggets was 10 to 110 mm, and the median was about 70 mm. The PE bags were packed into cardboard boxes, and a transportation test was performed for a round trip of about 400 km (about 200 km one way). These conditions are equivalent to the transportation conditions for polysilicon products. The cardboard boxes were loaded into the rear of the loading space of a transportation truck, and a vibration meter with a three-axis acceleration sensor was installed on the cardboard boxes.

[0044] Figure 6 shows the results of measuring the acceleration (G) over time in each direction measured by a vibrometer, with the X-axis representing the direction of travel of the transport truck, the Y-axis representing the left-right direction relative to the direction of travel, and the Z-axis representing the up-down direction. For example, when a transport truck stops while traveling, the X-axis takes a positive value because acceleration occurs in the direction of travel, and when the truck starts moving, the X-axis takes a negative value. When a transport truck curves left while traveling, acceleration acts on the right side, so the Y-axis takes a positive value, and when the truck curves right, the Y-axis takes a negative value. Due to the influence of gravity, the Z-axis is approximately -1G when stationary. When a cardboard box bounces upwards, the Z-axis takes a negative value more than -1G, and when the cardboard box is pushed from above, the Z-axis takes a positive value more than -1G.

[0045] As shown in Figure 6, it was found that the cardboard boxes containing the nuggets moved significantly up and down, and tilted more left and right than front and back. It is believed that the generation of fine powder from the nuggets occurs when the nuggets are subjected to such vibrations during transportation, causing them to collide or rub against each other. In this case, it is believed that the greater the unevenness of the nugget's surface, the more likely it is that fine powder will be generated by collisions or rubbing against each other.

[0046] Furthermore, if there is a depression such as a hole or a crack on the surface of the nugget, fine powder that cannot be removed by etching may remain in the depression. The fine powder in such a depression is also discharged from the depression by vibration, which causes the generation of fine powder.

[0047] Figure 7 shows the results of two tests conducted under conditions similar to those for product transportation. In test 1, a comparison was made between conditions in which nuggets with surface conditions classified as levels 1 to 4 were packed in PE bags and conditions in which nuggets with surface conditions classified as levels 1 to 8 were packed in PE bags. In test 1, 25 PE bags (N=25) were randomly sampled before transportation (before shipping), and 14 PE bags (N=14) were randomly sampled after the test, and the amount of fine powder was measured.

[0048] Test 2 was conducted on a different day and under the same transportation conditions as Test 1, and a comparison was made between conditions in which nuggets with surface conditions classified as levels 1 to 6 were packed in PE bags and conditions in which nuggets with surface conditions classified as levels 5 or 6 were packed in PE bags. In Test 2, 10 PE bags (N=10) were randomly sampled both before transportation (before shipping) and after the test, and the amount of fine powder was measured.

[0049] The nuggets in the PE bag were separated based on size. First, the nuggets in the PE bag were separated using a 5 mm sieve. The fine powder that passed through the 5 mm sieve was further separated using a 1 mm sieve. This resulted in three groups: fine powder with sizes of 1 mm or less, fine powder with sizes of 1 to 5 mm, and nuggets with sizes over 5 mm. The weight (g) of the fine powder group was measured.

[0050] As shown in Fig. 7, in Test 1, the amount of fine powder that increased after the test was greater in the levels 1 to 8 groups compared to the levels 1 to 4 groups. This is thought to be because the levels 1 to 8 groups contained many nuggets with rough surfaces, which increased the amount of fine powder generated compared to the levels 1 to 4 groups.

[0051] Moreover, in Test 2, it was shown that the amount of fine powder that increased after the test was greater in the levels 5 and 6 groups compared to the levels 1 to 6 groups. This is thought to be because the proportion of nuggets with rough surfaces was greater in the levels 5 and 6 groups than in the levels 1 to 6 groups, which increased the amount of fine powder generated.

[0052] As described above, it has been experimentally shown that the rougher the surface condition of polysilicon, the more likely it is that fine powder will be generated by vibration. Based on this finding, whether the amount of fine powder generated by vibration in the first polysilicon is equal to or less than a first threshold value may be determined in a pseudo manner based on whether the surface condition of the first polysilicon is equal to or less than a predetermined roughness.

[0053] The first imaging data may include at least one image of the first polysilicon. The first imaging data is preferably a plurality of images of the first polysilicon taken from different directions. More specifically, the first imaging data preferably includes an image of a first surface, which is one surface of the first polysilicon, and an image of a second surface, which is a surface including the back side of the first surface.

[0054] According to this configuration, the first imaging data includes images of the front and back surfaces of the first polysilicon, and is therefore suitable data as an index showing the correlation between the surface state of the first polysilicon and the tendency of the first polysilicon to generate fine powder. Therefore, according to this first imaging data, it is possible to generate a first trained model that can estimate with high accuracy the tendency of the second polysilicon to generate fine powder, which is the estimation target.

[0055] The first imaging data preferably includes a plurality of images of the first surface of the first polysilicon taken from a plurality of directions, and a plurality of images of the second surface taken from a plurality of directions. The image of the first polysilicon included in the first imaging data is preferably as small as possible in shadow. For example, the imaging of the first polysilicon is preferably performed under a condition where light is irradiated on the first polysilicon from a plurality of directions, for example, by providing a plurality of light sources or by diffusely reflecting light in a dome.

[0056] When the first imaging data includes a plurality of images, the first imaging data may be a composite image obtained by combining the plurality of images. If the first imaging data is such a composite image, the first generating unit 12 only needs to process a relatively small number of images as the first teacher data, and the processing load can be reduced. The first imaging data may be archive-format data that individually includes a plurality of images.

[0057] (1st generation process) Next, the first generating unit 12 executes a first generating step (generating step) S12 of generating a first trained model using the judgment data and the first imaging data acquired by the first acquiring unit 11 as teacher data (first teacher data). The first trained model functions as an algorithm that outputs an estimation result of the tendency of the second polysilicon to generate fine powder when vibration is applied, from the input second imaging data including an image of the second polysilicon that is the estimation target.

[0058] The second polysilicon is a polysilicon for which the tendency of fine powder generation is estimated using the first trained model. The second polysilicon is preferably a polysilicon for which the tendency of fine powder generation is unknown, but is not limited thereto, and may be, for example, a polysilicon for which the tendency of fine powder generation has been estimated by some method.

[0059] The second imaging data is data including an image of the second polysilicon. The second imaging data may include at least one image of the second polysilicon. The second imaging data is preferably a plurality of images of the second polysilicon taken from different directions. More specifically, the second imaging data preferably includes an image of a third surface, which is one surface of the second polysilicon, and an image of a fourth surface, which is a surface including the back side of the third surface.

[0060] According to this configuration, the second imaging data includes images of the front and back surfaces of the second polysilicon, and is therefore suitable data as an index showing the correlation between the surface state of the second polysilicon and the tendency of the second polysilicon to generate fine powder. Therefore, such second imaging data is suitable as input data for estimating the tendency of the second polysilicon to generate fine powder with high accuracy.

[0061] The second imaging data preferably includes a plurality of images of the third surface of the second polysilicon taken from a plurality of directions and a plurality of images of the fourth surface taken from a plurality of directions. The image of the second polysilicon included in the second imaging data is preferably as small as possible in shadow, as is the image of the first polysilicon included in the first imaging data. For example, the imaging of the second polysilicon is preferably performed under a condition where light is irradiated onto the second polysilicon from a plurality of directions, such as by providing a plurality of light sources or by diffusely reflecting light inside a dome.

[0062] The image format of the second imaging data is preferably the same as the image format of the first imaging data. For example, when the first imaging data is a composite image of a plurality of images, the second imaging data is also preferably a composite image of a plurality of images. In this case, the number of images included in the second imaging data is preferably the same as the number of images included in the first imaging data. With this configuration, it is possible to easily perform estimation from the second imaging data using the first trained model.

[0063] The first generation step S12 executed by the first generation unit 12 can be said to be a step of optimizing the machine learning algorithm by inputting the first teacher data into the machine learning algorithm. The first trained model can be said to be a machine learning algorithm optimized for estimating the likelihood of generation of fine powder of the second polysilicon by inputting the first teacher data. Such a machine learning algorithm is not particularly limited, and may be, for example, a neural network. The machine learning algorithm may be, for example, a CNN (convolutional neural network).

[0064] <Estimation using the first trained model> The method for estimating the likelihood of fine powder generation in polysilicon according to this embodiment is performed using a first trained model. The estimation method may be performed by an estimation device 20 for the likelihood of fine powder generation in polysilicon shown in Fig. 2. Hereinafter, the method for estimating the likelihood of fine powder generation in polysilicon will be described using an example in which the method is performed by the estimation device 20.

[0065] 2, the estimation device 20 has a second acquisition unit 21 and an estimation unit 22. The estimation device 20 may be a control device capable of executing the processes of the second acquisition unit 21 and the estimation unit 22, and may be, for example, a CPU (Central Processing Unit). The estimation device 20 may be a device that includes such a control device and further includes a storage device, a communication device, and the like.

[0066] 3, the second acquisition unit 21 executes a second acquisition step S13 of acquiring second imaging data. Since the second imaging data has already been described, a description thereof will be omitted here.

[0067] Next, the estimation unit 22 executes an estimation step S14 of estimating the likelihood of generation of fine powder in the second polysilicon when vibration is applied, from the second imaging data acquired by the second acquisition unit 21, using the first learned model. The first learned model used by the estimation unit 22 may be stored in advance in the estimation device 20, or may be acquired by the second acquisition unit 21 from the generation device 10 as needed.

[0068] The estimation result of the tendency of the second polysilicon to generate fine powder may be, for example, an estimation result of whether or not the amount of fine powder generated when vibration of a predetermined condition is applied to the second polysilicon is equal to or less than a first threshold value corresponding to the vibration of the predetermined condition. The amount of fine powder generated from the second polysilicon is considered to be positively correlated with the strength of the applied vibration. Therefore, such an estimation result is useful as an index of the tendency of the second polysilicon to generate fine powder, regardless of the actual strength of the vibration.

[0069] In other words, the first trained model is generated by vibration under a specific condition and a first threshold value of the amount of fine powder generation corresponding to the condition. However, the estimation unit 22 can use such a first trained model to accurately estimate the relative likelihood of fine powder generation in the second polysilicon regardless of the condition of the vibration applied to the second polysilicon.

[0070] Fig. 8 shows an overview of the steps from the first acquisition step S11 to the estimation step S14. In the example shown in Fig. 8, the first imaging data is a composite image including a plurality of images of the first polysilicon. The determination data is data showing the result of determining that the amount of fine powder generated is equal to or less than a first threshold value when the surface state of the first polysilicon is at level 1 to 4, and that the amount of fine powder generated is greater than the first threshold value when the surface state is at level 5 to 8.

[0071] The first generating unit 12 generates a first trained model from the first imaging data and the judgment data having such a correspondence relationship. The estimating unit 22 inputs a composite image including a plurality of images of the second polysilicon as the second imaging data to the first trained model, and estimates the tendency of the second polysilicon to generate fine powder. Specifically, the estimating unit 22 may display a result of whether or not the surface state of the second polysilicon is at level 4 or lower as an estimated result of the tendency of the second polysilicon to generate fine powder.

[0072] <Separation of second polysilicon> The first sorting step S15 is a step of sorting the second polysilicon based on the likelihood of generating fine powder estimated in the estimation step S14. The first sorting step S15 may be executed by a sorting control device 30 shown in Fig. 2. The first sorting step S15 may be a step performed on the crushed polysilicon obtained in the crushing step S2 shown in Fig. 1.

[0073] 2, the sorting control device 30 has a first sorting unit 31. The sorting control device 30 may be a control device capable of executing the processing of the first sorting unit 31, and may be, for example, a CPU (Central Processing Unit). The sorting control device 30 may be a device that includes such a control device, as well as a storage device, a communication device, and the like.

[0074] The first sorting unit 31 may perform control for sorting the second polysilicon into the third polysilicon, which is the second polysilicon estimated in the estimation step S14 to have a tendency to generate fine powder exceeding a predetermined second threshold value, and the other second polysilicon. The second threshold value may be, for example, a value that indicates that the second polysilicon is unsuitable for commercialization when the tendency to generate fine powder exceeds the second threshold value.

[0075] For example, when the estimation result in the estimation step S14 indicates that the surface condition of the second polysilicon is equal to or less than a predetermined roughness, the first sorting unit 31 may determine that the likelihood of generating fine powder of the second polysilicon is equal to or less than the second threshold value.

[0076] The first sorting unit 31 may transmit a control signal to a physical mechanism for sorting the second polysilicon for control of sorting the second polysilicon. Such a physical mechanism may be included in the sorting control device 30, or may be included in a sorting device (not shown) for the second polysilicon that is different from the sorting control device 30.

[0077] According to the method for producing polysilicon according to the present embodiment having the above-mentioned configuration, it is possible to accurately estimate the likelihood of generating fine powder for the second polysilicon. Also, by estimating in advance the second polysilicon that is unlikely to generate fine powder, etching can be performed only on the second polysilicon, and therefore the efficiency of producing polysilicon can be greatly improved.

[0078] 2 may be realized by a single control device. Also, each control block of the generation device 10 and the estimation device 20 may be realized by a single control device, or each control block of the estimation device 20 and the separation control device 30 may be realized by a single control device.

[0079] [Embodiment 2] A second embodiment of the present invention will be described below. For ease of explanation, the same reference numerals will be given to members having the same functions as those described in the first embodiment, and the description thereof will not be repeated.

[0080] The method for producing polysilicon according to this embodiment includes a method for efficiently carrying out the chipping step S3 on the third polysilicon which is estimated to be prone to generating fine powder in the method for producing polysilicon according to the first embodiment.

[0081] 3, the method for producing polysilicon according to this embodiment includes a third acquisition step S16, a second generation step S17, an area ratio calculation step S18, a depression number calculation step S19, and a second sorting step S20 in addition to the steps of the method for producing polysilicon according to the first embodiment. The area ratio calculation step S18 and the depression number calculation step S19 each perform a calculation process using, for example, a second trained model generated by the second generation step S17. First, a method for generating the second trained model will be described.

[0082] <Generation of the second trained model> (Second acquisition process and second teacher data) The second trained model functions as an algorithm for estimating whether the third polysilicon is suitable for the chipping process S3. The generation of the second trained model may be executed by the generation device 10. As shown in FIG. 2, the generation device 10 may include a second generation unit 13 for generating the second trained model.

[0083] As shown in Fig. 3, in this embodiment, the first acquisition unit 11 may execute a third acquisition step S16 of acquiring second teacher data used to generate a second trained model. The second teacher data includes surface condition data indicating a surface condition of the first polysilicon in which the amount of fine powder generation exceeds a first threshold. Such a surface condition of the first polysilicon may be, for example, a surface condition having irregularities or depressions, as shown as levels 5 to 8 in Fig. 4.

[0084] The surface condition data may be data including information that identifies (annotates) as an uneven region a region where the surface unevenness corresponds to levels 5 to 8 in the first imaging data of the first polysilicon in which the amount of fine powder generation exceeds a first threshold value, as shown in Fig. 9. The annotation may be performed by a user inputting the visual identification result, or may be performed mechanically by image processing.

[0085] Furthermore, as shown in FIG. 10, the surface condition data may include information specifying depressions such as holes or cracks formed on the surface of the first polysilicon in addition to information specifying the uneven region. It is preferable that the image included in the first imaging data does not include shadow areas. However, depending on the way the light hits the image or the surface condition during imaging, it is difficult to completely eliminate shadow areas from the image included in the first imaging data. Therefore, it may be difficult to determine whether a dark area in the image included in the first imaging data is a depression or a shadow.

[0086] Therefore, the surface condition data may be data including information on identifying (annotating) recessed portions in the first imaging data of the first polysilicon in which the amount of fine powder generation exceeds the first threshold value. The annotation may be performed by a user inputting the identification result by visual inspection, or may be performed mechanically by image processing.

[0087] Such surface condition data is suitable as second training data for creating a second trained model capable of identifying portions of the third polysilicon having a surface condition that may cause the generation of fine powder, as shown in Figures 9 and 10.

[0088] The surface condition data may be acquired by the first acquisition unit 11 at the same time as acquiring the determination data and the first imaging data in the first acquisition step S11. In this case, the third acquisition step S16 may be omitted.

[0089] (Second generation process) Next, the second generating unit 13 executes a second generating step S17 of generating a second trained model using the surface state data acquired by the first acquiring unit 11 as teacher data (second teacher data). Specifically, the second trained model functions as an algorithm for calculating the area ratio of at least a part of the surface of the third polysilicon that is estimated to be similar to the surface state of the first polysilicon in which the amount of fine powder generation exceeds the first threshold value, relative to the total area of ​​the surface of the third polysilicon. The calculation of the area ratio will be described later as an explanation of the area ratio calculation step S18.

[0090] The second generation step S17 executed by the second generation unit 13 can be said to be a step of optimizing the machine learning algorithm by inputting the second teacher data to the machine learning algorithm. The second trained model can be said to be a machine learning algorithm optimized for estimating the portion causing the generation of fine powder in the third polysilicon by inputting the second teacher data.

[0091] Such a machine learning algorithm is not particularly limited, and may be, for example, a neural network or a convolutional neural network (CNN). The machine learning algorithm used to generate the second trained model may be the same as or different from the machine learning algorithm used to generate the first trained model.

[0092] <Estimation using the second trained model> The second trained model is used in estimation processing using the second imaging data of the third polysilicon as an input in an area ratio calculation step S18 and a depression number calculation step S19. The area ratio calculation step S18 and the depression number calculation step S19 may be executed by an estimation device 20. As shown in Fig. 2, the estimation device 20 may have an area ratio calculation unit 23 that executes the area ratio calculation step S18 and a depression number calculation unit 24 that executes the depression number calculation step S19.

[0093] (Area ratio calculation process) The area ratio calculation unit 23 calculates, from the second imaging data of the third polysilicon, the area ratio of at least a portion of the surface of the third polysilicon that is estimated to be similar to the surface state of the first polysilicon in which the amount of fine powder generated exceeds a first threshold value, to the total area of ​​the surface of the third polysilicon.

[0094] The second imaging data of the third polysilicon may be data acquired by the second acquisition unit 21 in the second acquisition step S13. In this case, the area ratio calculation unit 23 may select only the second imaging data of the third polysilicon estimated to have a tendency to generate fine powder exceeding a predetermined second threshold value from the second imaging data of the second polysilicon in which the estimation result estimated in the estimation step S14 exists. In addition, the second acquisition unit 21 may acquire the second imaging data of the third polysilicon again before the area ratio calculation step S18.

[0095] The area ratio calculation unit 23 calculates the total area of ​​the projected portion of the third polysilicon based on the second imaging data of the third polysilicon captured at a magnification of 1, preferably 1 to 3. The total area of ​​the projected portion of the third polysilicon can be calculated from the image included in the second imaging data. In this case, the area ratio calculation unit 23 may calculate the approximate total area in two dimensions from the image included in the second imaging data, or may estimate the three-dimensional shape of the third polysilicon to calculate the total area more accurately.

[0096] For example, the area ratio calculation unit 23 calculates the sum of the areas of the projected portions of the third polysilicon captured from a plurality of directions as an algebra of the number of pixels of the projected portions of the third polysilicon. Then, the value obtained by dividing the sum by the number of imaging directions may be used as the index of the total area on the surface of the third polysilicon. However, the method of calculating the total area of ​​the surface of the third polysilicon by the area ratio calculation unit 23 is not limited to this.

[0097] In addition, when the image included in the second imaging data does not include a portion of the surface of the third polysilicon, the area ratio calculation unit 23 may regard the total area of ​​the surface appearing in the image included in the second imaging data as the total area of ​​the surface of the third polysilicon.

[0098] Next, the area ratio calculation unit 23 inputs the second imaging data of the third polysilicon to the second trained model, and estimates the position of the uneven region, which is a surface region similar to the surface state of the first polysilicon in which the amount of fine powder generation exceeds the first threshold. The uneven region in the third polysilicon can be said to be a surface region of at least a part of the third polysilicon, which is a portion estimated to have a large amount of fine powder generation.

[0099] The area ratio calculation unit 23 calculates the area of ​​the estimated uneven region. The area ratio calculation unit 23 calculates the area of ​​the uneven region using the same method as that for calculating the total area. When the area ratio calculation unit 23 estimates that a plurality of independent uneven regions exist in the third polysilicon, the area ratio calculation unit 23 may individually calculate the area of ​​each uneven region and its total area, or may calculate only the total area of ​​all the uneven regions.

[0100] For example, the area ratio calculation unit 23 may use the number of pixels in the portion regarded as the uneven region as an algebra for the sum of the areas of the portions regarded as the uneven region, as in the case of the surface area, and may divide the sum by the number of imaging directions to obtain the exponent of the area of ​​the uneven region. However, the method of calculating the area of ​​the uneven region on the surface of the third polysilicon by the area ratio calculation unit 23 is not limited to this.

[0101] The second trained model learns the surface state data of the first polysilicon in which the amount of fine powder generation exceeds the first threshold value as teacher data. Therefore, by using the second trained model, the area ratio calculation unit 23 can accurately estimate which parts of the surface of the third polysilicon are uneven areas.

[0102] Then, the area ratio calculation unit 23 calculates the area ratio of the concave-convex region to the total area of ​​the third polysilicon. The area ratio of the concave-convex region is used for separating the third polysilicon in the second separation step S20 described later.

[0103] (Counting the number of recesses) The dent number calculation unit 24 executes a dent number calculation step S19 of calculating the number of dents, which is the number of portions estimated to be dents formed on the surface of the third polysilicon, from the second imaging data of the third polysilicon using the second learned model. Note that, in the method for producing polysilicon according to this embodiment, the dent number calculation step S19 may be omitted.

[0104] The second trained model is a model for detecting depressions such as holes or cracks formed on the surface of the first polysilicon (for example, a depression of 1 mm in an image captured at 1x magnification). 2 Surface state data including information identifying the pits formed on the surface of the third polysilicon (holes or cracks having a total area of ​​100 mm or more) may be generated as teacher data. In this case, the pit number calculation unit 24 can distinguish between pits and shadows formed on the surface of the third polysilicon by using the second trained model, and can estimate the pits with high accuracy. Since the pits formed on the surface of the third polysilicon cause the generation of fine powder after etching as described above, it is considered that the more pits there are, the more likely the third polysilicon is to generate fine powder.

[0105] The number of recesses calculated by the recess number calculation unit 24 may be used to separate the third polysilicon in a second separation step S20 described later.

[0106] <Separation of the third polysilicon> The second sorting step S20 is a step of sorting the third polysilicon based on the total area and the area ratio calculated in the area ratio calculation step S18. The second sorting step S20 may be a step of sorting the third polysilicon based on the number of recesses calculated in the recess number calculation step S19 in addition to the total area and the area ratio. The sorting by the second sorting step S20 may be aimed at sorting the third polysilicon to be subjected to the chipping step S3 shown in FIG.

[0107] The second sorting step S20 may be performed by a sorting control device 30 shown in Fig. 2. As shown in Fig. 2, the sorting control device 30 may have a second sorting section 32 that performs the second sorting step S20.

[0108] The third polysilicon to be subjected to the chipping process S3 preferably has a total area equal to or greater than a predetermined third threshold value. This is because when a portion including an uneven region is scraped off from the third polysilicon by crushing or the like, if the third polysilicon is too small, the work becomes difficult. Therefore, it is preferable that the second sorting unit 32 uses the total area of ​​the third polysilicon as a sorting criterion for the third polysilicon.

[0109] The third threshold value may be set appropriately depending on the method of the chipping step S3. For example, the third threshold value may be set to a different value when the chipping step S3 is performed manually with a hammer or the like and when the chipping step S3 is performed mechanically with a crushing device.

[0110] Moreover, the third polysilicon to be subjected to the chipping step S3 preferably has an area ratio of the uneven region equal to or less than a predetermined fourth threshold. If the area ratio of the uneven region is too large, most of the surface of the third polysilicon will be the uneven region, and therefore, even if the chipping step S3 is performed, almost no third polysilicon containing the uneven region can be obtained. From the viewpoint of manufacturing efficiency, it is preferable that the second sorting unit 32 uses the area ratio of the uneven region in the third polysilicon as a sorting criterion for the third polysilicon.

[0111] The fourth threshold value may be appropriately set according to the target product size of the third polysilicon. The fourth threshold value may also be appropriately adjusted according to the total area of ​​the third polysilicon. When the total area of ​​the third polysilicon is sufficiently large, a predetermined amount or more of the third polysilicon not including the uneven region can be obtained after the chipping step S3 even if the fourth threshold value is set to a large value.

[0112] Moreover, the third polysilicon to be subjected to the chipping step S3 preferably has a number of dents equal to or less than a predetermined fifth threshold. Moreover, the third polysilicon to be subjected to the chipping step S3 preferably has a number of dents equal to or more than a predetermined sixth threshold. In other words, the third polysilicon to be subjected to the chipping step S3 preferably has a number of dents equal to or more than the sixth threshold and equal to or less than the fifth threshold.

[0113] In the third polysilicon having a large number of recesses, fine powder that cannot be removed by etching may exist in the recesses, so that the amount of fine powder generated is likely to be large regardless of the area ratio of the uneven region. In addition, when the number of recesses is extremely small, the estimation result of the estimation step S14 for the third polysilicon, that the tendency of fine powder generation exceeds the second threshold, is considered to be due to a false detection of a shadow or the like included in the second imaging data. For these reasons, it is preferable that the second sorting unit 32 uses the number of recesses in the third polysilicon as the sorting criterion for the third polysilicon.

[0114] The fifth threshold and the sixth threshold may be appropriately set, for example, empirically. The fifth threshold and the sixth threshold may also be appropriately adjusted according to the total area of ​​the third polysilicon. If the number of depressions is the same, the larger the total area of ​​the third polysilicon, the smaller the number of depressions per unit area. It is preferable to consider the tendency of the third polysilicon to generate fine powder per unit area.

[0115] 11 and 12 show the relationship between the results of visually separating polysilicon and the total area, area ratio, and number of dents calculated by area ratio calculation unit 23 and number of dents calculation unit 24. Here, the results of separating the "cone mix" obtained in the crushing step S2 shown in FIG. 1, which is a polysilicon crushed material containing third polysilicon, are shown. Each graph shown in FIG. 11 shows the correlation between the total area and the area ratio of the uneven area for each separated group, and each graph shown in FIG. 12 shows the correlation between the total area and the number of dents for each separated group. The total area of ​​polysilicon is expressed as the number of pixels of the polysilicon projection portion calculated from the image included in the second imaging data, as an index of the total area.

[0116] In Fig. 11, graph 111 shows the distribution of all crushed objects, graph 112 shows the distribution of groups determined to not require the chipping step S3, graph 113 shows the distribution of groups determined to be suitable for the chipping step S3, and graph 114 shows the distribution of groups determined to be unsuitable for the chipping step S3, even though they contain parts that are likely to generate fine powder.

[0117] 12, graph 121 shows the distribution of groups obtained by excluding third polysilicon having a total area index of 10 or less from all crushed objects. Graphs 122 to 124 are similar in that third polysilicon having a total area index of 10 or less is excluded. Graph 122 shows the distribution of groups determined to not require the chipping process S3, graph 123 shows the distribution of groups determined to be suitable for the chipping process S3, and graph 124 shows the distribution of groups determined to be unsuitable for the chipping process S3, although they contain portions that are likely to generate fine powder.

[0118] 11, the region shown as "threshold range" is a region in which the index of the total area is greater than 10 and the area ratio of the uneven regions is 65% or less. Polysilicon classified into such a region has a total area that is not too small and an area ratio of the uneven regions that is not excessive, and therefore is highly likely to be suitable for the chipping step S3. Such a threshold range may be an example of setting the third and fourth thresholds.

[0119] In addition, in FIG. 12, the region where the index of the total area is greater than 10 and is sandwiched between the "threshold range (upper limit)" and the "threshold range (lower limit)" is an empirically set region in which crushed material that is unlikely to generate fine powder is likely to be plotted. It was shown that polysilicon classified in such a region has a high possibility of being suitable as a target for the chipping process S3, since the number of depressions relative to the total area is neither too small nor too large. Such a region may be an example of setting the third threshold, the fifth threshold, and the sixth threshold.

[0120] Each of the third to sixth thresholds may be set based on the actual measurement results as described above, may be set empirically, or may be set by other methods.

[0121] The second sorting unit 32 may sort the third polysilicon based on the total area and the area ratio, for example, as shown in Fig. 11. The second sorting unit 32 may further sort the third polysilicon sorted based on the total area and the area ratio based on the total area and the number of depressions, as shown in Fig. 12, to sort the third polysilicon.

[0122] The second sorting unit 32 may transmit a control signal to a physical mechanism for sorting the third polysilicon. Such a physical mechanism may be included in the sorting control device 30, or may be included in a sorting device (not shown) for the third polysilicon that is different from the sorting control device 30.

[0123] According to the method for producing polysilicon according to the present embodiment having the above-described configuration, the third polysilicon suitable for the chipping step S3 can be efficiently separated. Therefore, it is possible to reduce both the labor cost and the production loss required to obtain polysilicon that is less likely to generate fine powder.

[0124] As an example, the weight percentage of polysilicon before the rough crushing step S1 is 100%. At this time, the polysilicon to be subjected to the chipping step S3 is, for example, about 15%, which is a relatively large percentage, depending on the polysilicon deposition conditions. Therefore, reducing the loss percentage of polysilicon by the chipping step S3 is important for improving manufacturing efficiency. Therefore, the method for manufacturing polysilicon according to this embodiment, which can improve the efficiency of the chipping step S3, greatly contributes to improving manufacturing efficiency.

[0125] [Software implementation example] The functions of the generation device 10, the estimation device 20 or the separation control device 30 (hereinafter referred to as the "device") can be realized by a program for causing a computer to function as the device, and a program for causing a computer to function as each control block of the device.

[0126] In this case, the device includes a computer having at least one control device (e.g., a processor) and at least one storage device (e.g., a memory) as hardware for executing the program. The control device and storage device execute the program to realize each function described in each of the above embodiments.

[0127] The program may be non-transitory and may be recorded in one or more computer-readable recording media. The recording media may or may not be included in the device. In the latter case, the program may be provided to the device via any wired or wireless transmission medium.

[0128] In addition, some or all of the functions of each of the control blocks can be realized by a logic circuit. For example, an integrated circuit in which a logic circuit that functions as each of the control blocks is formed is also included in the scope of the present invention. In addition, the functions of each of the control blocks can be realized by, for example, a quantum computer.

[0129] [Additional Notes] The present invention is not limited to the above-described embodiments, and various modifications are possible within the scope of the claims. Embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included in the technical scope of the present invention. [Explanation of symbols]

[0130] 10 Generator 11 1st Acquisition Department (Acquisition Department) 12 1st generation section (generation section) 20 Estimation device 22 Estimation part 23 Area ratio calculation section 24. Depression Count Calculation Section

Claims

1. A method for estimating the likelihood of fine powder generation in polysilicon, comprising an estimation step of estimating the likelihood of fine powder generation in second polysilicon when vibration is applied from second imaging data including an image of second polysilicon that is an estimation target, using a first trained model generated using as training data determination data as to whether an amount of fine powder generation due to vibration in a first polysilicon, which is polysilicon after vibration under predetermined conditions is applied, is equal to or less than a predetermined first threshold value, and first imaging data including an image of the first polysilicon.

2. the first imaging data includes a plurality of images of the first polysilicon taken from different directions, The method of claim 1 , wherein the second imaging data includes a plurality of images of the second polysilicon taken from different directions.

3. the first imaging data includes an image of a first surface of the first polysilicon and an image of a second surface including a backside of the first surface; The estimation method according to claim 1 or 2, wherein the second imaging data includes an image of a third surface of the second polysilicon and an image of a fourth surface including a back side of the third surface.

4. an estimation step of estimating the likelihood of fine powder generation in the second polysilicon when vibration is applied from second imaging data including an image of the second polysilicon to be estimated, using a first trained model generated using as teacher data determination data on whether or not an amount of fine powder generation due to vibration in a first polysilicon, which is polysilicon after vibration under a predetermined condition is applied, and first imaging data including an image of the first polysilicon; a first sorting step of sorting the second polysilicon based on the likelihood of fine powder generation estimated in the estimation step.

5. an area ratio calculation step of calculating, using a second trained model generated using the surface state of the first polysilicon in which the amount of fine powder generation exceeds the first threshold as teacher data, an area ratio of at least a part of the surface of the third polysilicon, which is the second polysilicon in which the tendency for fine powder generation to exceed a predetermined second threshold is estimated by the estimation step, to a total area of ​​the surface of the third polysilicon, from the second imaging data of the third polysilicon, the third polysilicon being the second polysilicon in which the tendency for fine powder generation to exceed a predetermined second threshold is estimated; 5. The method for producing polysilicon according to claim 4, further comprising: a second classification step of classifying the third polysilicon based on the total area and the area ratio.

6. The method further includes a dent number calculation step of calculating a dent number, which is the number of portions estimated to be dents formed on a surface of the third polysilicon, from the second imaging data of the third polysilicon by using the second trained model; 6. The method for producing polysilicon according to claim 5, wherein the second classification step further classifies the third polysilicon based on the number of recesses in addition to the total area and the area ratio.

7. An estimation device for estimating the likelihood of fine powder generation in polysilicon, the device including an estimation unit that estimates the likelihood of fine powder generation in the second polysilicon when vibration is applied from second imaging data including an image of the second polysilicon that is an estimation target, using a first learned model generated using as teacher data determination data as to whether an amount of fine powder generation due to vibration in the first polysilicon, which is polysilicon after vibration under predetermined conditions is applied, is equal to or less than a predetermined first threshold value, and first imaging data including an image of the first polysilicon.

8. an acquisition unit that acquires, as teacher data, determination data on whether an amount of fine powder generated by vibration in a first polysilicon, which is polysilicon after vibration under a predetermined condition is applied, is equal to or less than a first threshold value, and first imaging data including an image of the first polysilicon; a generation unit that generates a first trained model from input second imaging data including an image of the second polysilicon that is the estimation target, and outputs an estimation result of the likelihood of the fine powder being generated in the second polysilicon when vibration is applied.

9. an acquisition step of acquiring, as teacher data, determination data on whether or not an amount of fine powder generated by vibration in a first polysilicon, which is polysilicon after vibration under a predetermined condition is applied, is equal to or less than a first threshold value, and first imaging data including an image of the first polysilicon; A method for generating a trained model, comprising: a generation step of generating a first trained model from input second imaging data including an image of the second polysilicon that is the subject of estimation, the first trained model outputting an estimation result of the likelihood of the fine powder being generated in the second polysilicon when vibration is applied.

Images