Learning method, learning device, learning program, control method, control device, and control program

Abstract

This learning method comprises: a step in which, from among a plurality of items of information an observer can perceive, an observer is made to perceive items of information that the observer will be caused to compare; a step in which an evaluation of the items of information, which was made by the observer who compared the items of information, is acquired; and a step in which the plurality of items of information are rated on the basis of evaluation of the plurality of items of information by the observer, acquired through a plurality of repetitions of the perception step and the evaluation acquisition step .

Description

Technical Field

[0005]

[0001] The present disclosure relates to a technology for reflecting human intentions and knowledge in control and the like.

Background Art

[0002] As a new material for semiconductor for power devices, silicon carbide (SiC) has attracted attention. The inventors have established a technology for growing high-quality SiC crystals and are proceeding with research and development for growing larger crystals towards practical application (for example, see Non-Patent Document 1).

Prior Art Documents

Non-Patent Documents

[0006] These circumstances are not limited to the control of SiC crystal growth. In various fields, the masterful skills of experienced engineers are highly individualized and difficult to express objectively using numerical values ​​or mathematical formulas, which hinders the transfer of technology and the automation of control processes.

[0007] This disclosure is made in light of these challenges, and its purpose is to provide technology that allows human intentions and knowledge to be reflected in control and other processes. [Means for solving the problem]

[0008] To solve the above problems, a learning method according to one aspect of this disclosure , knowledge Multiple memorable The feelings Report By paired comparison Steps to make the observer perceive and , watch By repeating the steps of obtaining an evaluation of the information by the observer, and the steps of making it perceived and obtaining an evaluation multiple times, Multiple The step of rating numerical information, The process involves generating a latent space by compressing the dimensionality of multiple pieces of information, and searching for latent variables in the latent space that meet predetermined conditions. of include . Note that pairwise comparisons can involve not only two choices, but also three or more choices.

[0009] Another aspect of this disclosure is a learning device. , knowledge Multiple memorable The feelings Report By paired comparison Information provision unit that allows the observer to perceive and , watch The evaluation unit obtains evaluations of the information by the observer, and by repeatedly providing information and obtaining evaluations multiple times. Multiple A rating unit that rates numerical information, A latent space generation unit generates a latent space by compressing the dimensions of multiple pieces of information, and a search unit searches for latent variables that meet predetermined conditions in the latent space. It is equipped with.

[0010] Another aspect of this disclosure is a control method, which uses information representing the state of the controlled object. By paired comparison A step to be perceived by the observer, information representing the state of the controlled object at the first time point, and information representing the state of the controlled object at a second time point different from the first time point. against by the observer Review Steps to obtain value and based on evaluation According to the control parameters corresponding to latent variables that meet predetermined conditions, which are searched in a latent space generated by dimensionality reduction of multiple rated pieces of information, The steps of controlling the object to be controlled, include .

[0011] Another aspect of this disclosure is a control device, which provides information representing the state of the controlled object. By paired comparison Information provision unit for the observer to perceive; information representing the state of the controlled object at a first time point; information representing the state of the controlled object at a second time point different from the first time point. against by the observer Review A valuation unit that obtains a value, and based on the valuation According to the control parameters corresponding to latent variables that meet predetermined conditions, which are searched in a latent space generated by dimensionality reduction of multiple rated pieces of information, It comprises a control unit that controls the object to be controlled.

[0012] Another aspect of this disclosure is a learning program. This program uses a computer to , knowledge Multiple memorable The feelings Report By paired comparison Information provision unit that allows the observer to perceive and , watch The evaluation unit obtains evaluations of the information by the observer, and by repeatedly providing information and obtaining evaluations multiple times. Multiple A rating unit that rates numerical information, A latent space generation unit generates a latent space by compressing the dimensions of multiple pieces of information, and a search unit searches for latent variables that meet predetermined conditions in the latent space. To make it function as such.

[0013] Another aspect of this disclosure is a control program. This program controls a computer to provide information representing the state of the controlled object. By paired comparison Information provision unit for the observer to perceive; information representing the state of the controlled object at a first time point; information representing the state of the controlled object at a second time point different from the first time point. against by the observer Review A valuation unit that obtains a value, and based on the valuation According to the control parameters corresponding to latent variables that meet predetermined conditions, which are searched in a latent space generated by dimensionality reduction of multiple rated pieces of information, It functions as a control unit that controls the object being controlled.

[0014] In addition, any combination of the above components, as well as those obtained by converting the expressions of the present disclosure among methods, apparatuses, systems, recording media, computer programs, etc., are also effective as aspects of the present disclosure.

Advantages of the Invention

[0015] According to the present disclosure, it is possible to provide a technology for reflecting human intentions and knowledge in control and the like.

Brief Description of the Drawings

[0016] [Figure 1] It is a diagram showing the configuration of a crystal manufacturing system for manufacturing a SiC crystal. [Figure 2] It is a diagram showing an example of the temperature distribution and flow distribution of a high-temperature metal liquid in a crucible. [Figure 3] It is a diagram showing an example of an image presented to an operator. [Figure 4] It is a diagram showing the concept of a variational autoencoder. [Figure 5] It is a diagram showing an example of a latent space generated by a variational autoencoder. [Figure 6] FIGS. 6(a) and 6(b) are diagrams plotting the scores of images. [Figure 7] FIGS. 7(a) and 7(b) are diagrams showing an example of a latent space generated by a variational autoencoder and an example of images arranged close to each other in the latent space. [Figure 8] It is a flowchart showing the procedure of the learning method according to the embodiment. [Figure 9] It is a flowchart showing the procedure of the control method according to the embodiment. [Figure 10] It is a diagram showing the configuration of a learning apparatus according to an embodiment of the present disclosure. [Figure 11] It is a diagram showing an example of the result analyzed by the determination basis analysis unit. [Figure 12]This figure shows an example of the results obtained when an image is divided into three sets based on its score, and each set is analyzed by the decision-making analysis unit. [Figure 13] This figure shows the configuration of the control device according to the embodiment of the present disclosure. [Figure 14] This figure shows examples of images with high scores among those rated by the learning device according to the embodiment. [Figure 15] This figure shows the optimal temperature distribution and flow distribution obtained by the learning device according to the embodiment. [Figure 16] This figure shows SiC crystals produced by the crystal manufacturing apparatus according to the embodiment. [Figure 17] This figure shows an example of an image to be presented to the operator. [Modes for carrying out the invention]

[0017] As an embodiment of this disclosure, we will describe a technology that enables the quantification of human intentions and knowledge and the automatic search for the optimal solution. In this embodiment, as an example, we will describe the case of optimizing the conditions for manufacturing SiC crystals.

[0018] Figure 1 shows the configuration of a crystal manufacturing system for producing SiC crystals. The crystal manufacturing system 100 comprises a crystal manufacturing apparatus 10, a control device 20, and a learning device 50. The arrows in the figure represent the flow of information, and the operator controls the crystal manufacturing system 100 via the control device 20.

[0019] The crystal manufacturing apparatus 10 comprises a carbon crucible 1, a carbon stick 3, and a heating coil 5. The crucible 1 contains the high-temperature metal liquid 2, which is the raw material. The carbon stick 3 grows SiC crystals at its tip. The heating coil 5 heats the crucible 1. The crystal manufacturing apparatus 10 also includes a configuration (not shown) for rotating or moving the crucible 1, the carbon stick 3, and the heating coil 5.

[0020] The control device 20 controls the conditions for growing SiC crystals in the crystal manufacturing apparatus 10. The control device 20 controls the rotation and movement of the crucible 1, carbon sticks 3, and heating coil 5 at least one of the following times: before growing the SiC crystals and during crystal growth. The control device 20 also controls the current flowing through the heating coil 5 and the position of the heating coil 5 so that the inside of the crucible 1 reaches a set temperature.

[0021] The operator places a predetermined amount of high-temperature metallic liquid 2 into a cylindrical crucible 1 and sets a seed crystal on the tip of a carbon stick 3. The control device 20 moves the carbon stick 3 so that its tip is immersed in the high-temperature metallic liquid 2 and heats the inside of the crucible 1 to a set temperature by passing an electric current through the heating coil 5. As a result, a SiC crystal 4 grows on the tip of the carbon stick 3.

[0022] While growing SiC crystals, the control device 20 predicts or observes the temperature and flow distribution of the high-temperature metal liquid 2 in the crucible 1 and presents the operator with an image visualizing the temperature and flow distribution of the high-temperature metal liquid 2. While observing the temperature and flow distribution of the high-temperature metal liquid 2, the operator inputs instructions to the control device 20, such as rotation and movement of the crucible 1, carbon sticks 3, and heating coils 5, and setting temperatures, in order to produce larger and higher-quality SiC crystals.

[0023] The learning device 50 searches for the optimal conditions for growing SiC crystals in the crystal manufacturing apparatus 10. In this embodiment, the learning device 50 optimizes the conditions defined by seven control parameters: the rotation speed of the seed crystal, the rotation speed of the crucible 1, the relative position of the crucible 1 and the coil, the position of the coil, the height of the high-temperature metal liquid 2, the height of the meniscus of the high-temperature metal liquid 2, and the set temperature.

[0024] Figure 2 shows an example of the temperature distribution and flow distribution of the high-temperature metal liquid 2 in the crucible 1. As described in Non-Patent Literature 1, our invention makes it possible to predict and visualize the temperature distribution and flow distribution of the high-temperature metal liquid 2 in the crucible 1 in real time, so that the operator can control the conditions by inputting instructions to the control device 20 while observing the high-temperature metal liquid 2 in the crucible 1.

[0025] Experienced operators possess knowledge of temperature and flow distributions necessary for obtaining large, high-quality SiC crystals. Based on this knowledge, they can suitably control the temperature and flow distributions of the solution in crucible 1 while growing SiC crystals, thereby enabling the growth of large, high-quality SiC crystals.

[0026] However, it is difficult to accurately express all of these insights using numbers and formulas. Furthermore, even if we can partially express these insights using numbers and formulas, it is impossible to reflect the tacit knowledge, which is difficult to express objectively, in those numbers and formulas.

[0027] To solve these problems, the learning device 50 of this embodiment prepares numerous images representing the temperature distribution and flow distribution of the high-temperature metal liquid 2 in the crucible 1 under different conditions. Two images are selected from these images and presented to the operator, for example, as shown in Figure 3, and the operator is asked which of these images is preferable for obtaining larger, higher-quality SiC crystals. In the embodiments described later, 200 to 300 pairs of images were evaluated. This type of evaluation is known as a paired comparison method, which compares two samples in pairs rather than evaluating many samples at once. However, after the inventors diligently devised and investigated this method, they found that the paired comparison method is the optimal means of learning the knowledge and tacit knowledge of skilled engineers. Based on the evaluation results of good or bad obtained from the operator, the learning device 50 assigns a score to each of the prepared images. This makes it possible to generate an objective function for optimizing the conditions for growing SiC crystals.

[0028] Elo rating is a suitable method for assigning scores to each image based on the operator's evaluation of its quality. Elo rating is used in competitive games such as chess as an index to represent skill in relative evaluation. The learning device 50 sets initial ratings for each image and updates the rating values ​​each time the operator evaluates them. The learning device 50 compares the probability of quality calculated based on the respective rating values ​​of two images with the actual evaluation results and updates the respective rating values ​​based on the difference. By repeating the evaluation by the operator, the rating values ​​of each image gradually converge to values ​​that reflect the actual relative evaluation.

[0029] It is difficult and impractical for an operator to comprehensively review a large number of images and then determine their quality ranking or the absolute value of each image's score. However, it is relatively easy to evaluate the quality of two images by comparing them. By obtaining the operator's evaluation through repeated comparisons of two images, the operator's intentions and knowledge can be quantified more accurately. Furthermore, insights that the operator themselves may not be aware of or cannot clearly express can also be reflected in the numerical values.

[0030] In some fields, if a large amount of information representing the state of the controlled object is rated as described above, then it may be sufficient to simply achieve the state with the highest score. However, this is not always the case. In the crystal manufacturing apparatus 10 of this embodiment, when growing SiC crystals, it is necessary to achieve both outward and inward flow of the high-temperature metal liquid 2. Therefore, it is necessary to select the optimal temperature distribution and flow distribution for the outward and inward flows, respectively. Here, of the seven control parameters mentioned above, the position of the crucible 1, the position of the heating coil 5, the height of the high-temperature metal liquid 2, and the set temperature cannot be changed midway through the process, so there is a constraint that similar values ​​must be selected. In such cases, when there are constraints when selecting the optimal solution, the learning device 50 performs optimization of the objective function that has been rated and scored as described above.

[0031] Through the inventor's ingenuity and research, a Variational Auto-Encoder (VAE) was discovered as a learning model to obtain the optimal solution that matches the operator's intentions and perceptions from the image and score information obtained using the paired comparison method, among various deep learning models.

[0032] Figure 4 illustrates the concept of a variational autoencoder. In the encoder section of the variational autoencoder, a high-dimensional image is compressed into low-dimensional features. In this embodiment, since the conditions defined by seven control parameters are optimized, it is desirable to compress the features to seven dimensions or less. In the decoder section, a high-dimensional image is reconstructed from the low-dimensional features. In a semi-supervised variational autoencoder, the scores assigned to the images may be used as labels. In this case, a latent space can be generated in which images with similar scores are placed close together.

[0033] Figure 5 shows an example of a latent space generated by a variational autoencoder. In Figure 5, the scores corresponding to each latent variable in the 7-dimensional latent space are schematically depicted as surfaces in 3-dimensional space. The learning device 50 can search for the latent variable that maximizes the score in the latent space. The learning device 50 can also search for paths in the latent space that transition from a low-score state to a high-score state.

[0034] Figure 6(a) plots the scores of each image on a coordinate plane obtained by compressing the 7-dimensional space of the parameters described above (rotation speed of the seed crystal, rotation speed of crucible 1, relative position of crucible 1 and coil, position of coil, height of high-temperature metallic liquid 2, height of the meniscus of high-temperature metallic liquid 2, and set temperature) into 2 dimensions. In the control parameter space, the distribution of image scores is scattered. Figure 6(b) plots the scores of each image on a coordinate plane obtained by further compressing the 7-dimensional space obtained by compressing the image data using a semi-supervised variational autoencoder with scores into 2 dimensions. The semi-supervised variational autoencoder constructs a latent space in which images with similar scores are placed close together. Therefore, searching for the optimal solution in the latent space rather than in the control parameter space yields an optimal solution that better matches the operator's intentions and intuition.

[0035] Figure 7(a) plots the scores on a coordinate plane obtained by further compressing the 7-dimensional space, which is the dimensionality of image data compressed using a semi-supervised variational autoencoder with scores, to 2 dimensions. Figure 7(b) shows the images corresponding to A, B, and C in Figure 7(a). It can be seen that the flow distribution and temperature distribution of images A, B, and C, which are placed close together in the latent space generated by the semi-supervised variational autoencoder, are similar.

[0036] The learning device 50 generates a latent space by compressing the dimensionality of images using a semi-supervised variational autoencoder with scores. The learning device 50 learns a model that predicts latent variables from control parameters based on the correspondence between the values ​​of control parameters corresponding to each image and the values ​​of latent variables in the latent space. For example, a model may be learned in which the values ​​of control parameters are input to the input layer of a neural network and the values ​​of latent variables are output from the output layer.

[0037] The learning device 50 searches for latent variables with high scores for both outward and inward flows in the generated latent space, under the constraints of fixing the position of the crucible 1, the position of the heating coil 5, the height of the high-temperature metallic liquid 2, and the set temperature, using algorithms such as a genetic algorithm. The learning device 50 outputs control parameters corresponding to the searched latent variables to the control device 20.

[0038] The control device 20 controls the crystal manufacturing apparatus 10 using control parameters acquired from the learning device 50. This makes it possible to achieve an ideal temperature distribution and flow distribution of the high-temperature metal liquid 2 that reflects the operator's intentions and knowledge, thereby enabling the growth of larger and higher-quality SiC crystals.

[0039] If, during the growth of the SiC crystal, the temperature distribution or flow distribution of the high-temperature metal liquid 2 deviates from the ideal state due to some factor, the control device 20 may instruct the learning device 50 to search for the optimal control to return to the ideal state. The learning device 50 acquires an image from the control device 20 that visualizes the current temperature distribution and flow distribution of the high-temperature metal liquid 2, and converts the acquired image into latent variables by dimensionality reduction. In the latent space, the learning device 50 determines a path to change from the current latent variables to the optimal latent variables under the above constraints, converts it into control parameters, and outputs it to the control device 20. The control device 20 controls the crystal manufacturing apparatus 10 using the control parameters acquired from the learning device 50. This makes it possible to achieve the ideal temperature distribution and flow distribution of the high-temperature metal liquid 2.

[0040] While growing SiC crystals using the crystal manufacturing apparatus 10, control and learning may be performed in parallel. The operator compares past and current images while viewing images that visualize the temperature and flow distribution presented in a time series, and evaluates their quality. The operator may also be presented with images predicting the future temperature and flow distribution of the high-temperature metal liquid 2 and asked to compare them with the current images. The learning apparatus 50 acquires the operator's evaluation and rates each of the compared images. This allows the image ratings to be updated in real time during crystal manufacturing. If the temperature and flow distribution are evaluated as deteriorating, the learning apparatus 50 searches for latent variables in the latent space to achieve higher-scoring temperature and flow distributions and outputs the corresponding control parameters to the control device 20. If the temperature and flow distribution are evaluated as improving, the learning apparatus 50 does not need to perform optimization, or it may search for latent variables in the latent space to achieve even higher-scoring temperature and flow distributions and output the corresponding control parameters to the control device 20. With this technology, operators can grow larger, higher-quality SiC crystals simply by evaluating the quality of the process while viewing visualized images of temperature and flow distributions.

[0041] Figure 8 is a flowchart showing the steps of the learning method according to the embodiment.

[0042] In step S10, the learning device 50 prepares images that visualize the temperature distribution and flow distribution of the high-temperature metallic liquid 2 in the crucible 1 under a number of different conditions. These images may be generated by predicting the temperature distribution and flow distribution of the high-temperature metallic liquid 2 in the crucible 1, or they may be obtained from observations made when SiC crystals are actually grown.

[0043] In step S12, the learning device 50 assigns an initial score to all images. When using Elo rating, the initial score may be, for example, 1500.

[0044] In step S20, the learning device 50 selects two images for the operator to compare. The learning device 50 may select images in descending order of score (rating value), or it may select images randomly. The learning device 50 may select images from a set of images that represent similar images. For example, the latent space may be divided into meshes, and a representative image may be selected from each mesh. The learning device 50 may select images from a set after screening images based on scores, etc. For example, the images may be divided into multiple sets based on scores, and images may be selected from the same set. The learning device 50 may perform Bayesian optimization aimed at obtaining high scores in the latent space. The learning device 50 may select images based on orthogonal arrays. For example, the latent space may be divided into meshes, and images corresponding to grid points may be selected. The learning device 50 may select three or more images.

[0045] In step S22, the learning device 50 presents the image selected in step S20 to the operator. The learning device 50 may present two or more images simultaneously, or it may present multiple images in chronological order. The learning device 50 may also present a reference image that serves as the basis for judgment. The learning device 50 may present the selected image with random perturbations added.

[0046] In step S24, the learning device 50 obtains an evaluation from the operator. The learning device 50 may obtain the evaluation by input to an input device, or by the operator's voice, movements, facial expressions, gaze, etc.

[0047] In step S26, the learning device 50 updates the ratings of the two images presented to the operator based on the evaluation obtained from the operator. The learning device 50 may rate the images using Elo rating or any other rating technique. The learning device 50 may also be configured to allow manual adjustment of the scores for individual images.

[0048] The learning device 50 repeats steps S20 to S26 until all images have been rated (N in S28). Once all images have been rated (Y in S28), it proceeds to step S30.

[0049] In step S30, the learning device 50 constructs a latent space by compressing the image dimensions. The method for constructing the latent space may be the conditional variational autoencoder described above, or it may be a variational autoencoder, principal component analysis, singular value decomposition, eigendecomposition, QZ decomposition, Takagi decomposition, non-negative matrix factorization, T-SNE, U-MAP, multidimensional scaling, Locally Linear Embedding (LLE), ISOMAP, Laplacian Eigenmap, etc. It is desirable that the latent space be constructed by a coordinate transformation, dimensionality reduction, etc., such that images with similar scores are placed close together and images with significantly different scores are placed far apart.

[0050] In step S32, the learning device 50 learns an AI to predict latent variables from control parameters.

[0051] In step S34, the learning device 50 searches for the optimal latent variables in the latent space.

[0052] In step S36, the learning device 50 outputs control parameters corresponding to the explored latent variables to the control device 20.

[0053] Figure 9 is a flowchart showing the procedure of the control method according to the embodiment.

[0054] In step S40, the control device 20 predicts the temperature distribution and flow distribution of the high-temperature metal liquid 2.

[0055] In step S42, the control device 20 presents the predicted temperature distribution and flow distribution of the high-temperature metal liquid 2.

[0056] In step S44, the control device 20 obtains an operator's evaluation by comparing past and current images of the temperature distribution and flow distribution of the high-temperature metal liquid 2.

[0057] In step S46, the learning device 50 rates the past and current images of the high-temperature metallic liquid 2 based on the operator's evaluation.

[0058] In step S48, the control device 20 controls the crystal manufacturing apparatus 10 based on the operator's evaluation so that an ideal temperature distribution and flow distribution of the high-temperature metal liquid 2 is achieved.

[0059] Steps S40 to S48 are repeated until the crystal production by the crystal production apparatus 10 is completed (N in S50). When the crystal production by the crystal production apparatus 10 is completed (Y in S50), the control method is terminated.

[0060] Figure 10 shows the configuration of a learning device 50 according to an embodiment of the present disclosure. The learning device 50 comprises a communication device 51, a storage device 70, and a processing device 60. The learning device 50 may be a server device, a personal computer or other device, or a mobile terminal such as a mobile phone terminal, smartphone, or tablet terminal.

[0061] The communication device 51 controls communication with other devices. The communication device 51 may communicate using any communication method, whether wired or wireless.

[0062] The storage device 70 stores programs, data, etc., used by the processing device 60. The storage device 70 may be a semiconductor memory, a hard disk, or the like.

[0063] The processing unit 60 comprises an image acquisition unit 61, an image selection unit 62, an image presentation unit 63, an evaluation acquisition unit 64, a rating unit 65, a latent space generation unit 66, a predictive AI learning unit 67, an optimal solution search unit 68, and a judgment basis analysis unit 69. These configurations can be realized in hardware terms by arbitrary circuits, a computer's CPU, memory, or other LSIs, and in software terms by programs loaded into memory, but here we are describing functional blocks realized through the cooperation of these components. Therefore, it will be understood by those skilled in the art that these functional blocks can be realized in various forms, such as hardware only or a combination of hardware and software.

[0064] The image acquisition unit 61 acquires images that visualize the temperature distribution and flow distribution of the high-temperature metallic liquid 2 inside the crucible 1 under a number of different conditions.

[0065] The image selection unit 62 selects two images for the operator to compare. The image presentation unit 63 presents the images selected by the image selection unit 62 to the operator. The evaluation acquisition unit 64 obtains an evaluation from the operator. The rating unit 65 updates the ratings of the two images presented to the operator based on the evaluation obtained from the operator.

[0066] The latent space generation unit 66 constructs a latent space by compressing the image's dimensions. The predictive AI learning unit 67 learns an AI to predict latent variables from control parameters.

[0067] The optimal solution search unit 68 searches for the optimal latent variables in the latent space. The optimal solution search unit 68 outputs control parameters corresponding to the searched latent variables to the control device 20.

[0068] The decision-making basis analysis unit 69 analyzes which parts of the image the operator emphasized when evaluating the image. The decision-making basis analysis unit 69 learns a model to predict a score from the rated image. For example, the decision-making basis analysis unit 69 may learn a model in which the image is input to the input layer of a neural network and a score is output from the output layer. In the neural network, the decision-making basis analysis unit 69 may obtain the gradient of the last convolutional layer that has the greatest influence on the score prediction and visualize the parts of the image that the operator emphasized using a method such as Grad-CAM (Gradient-weighted Class Activation Mapping).

[0069] Figure 11 shows an example of the results analyzed by the decision basis analysis unit 69. Figure 11(a) shows the results of analyzing data obtained when operator evaluations were acquired to obtain the optimal conditions for manufacturing a 3-inch SiC crystal. Figure 11(b) shows the results of analyzing data obtained when operator evaluations were acquired to obtain the optimal conditions for outward flow when manufacturing a 6-inch SiC crystal. In both cases, the upper part is emphasized, and in the case of the 6-inch crystal, the side part is also emphasized.

[0070] Figure 12 shows an example of the results obtained by the judgment basis analysis unit 69 after dividing the image into three sets based on the score. Figure 12(a) shows the results for the top 1 / 3 set, Figure 12(b) shows the results for the middle 1 / 3 set, and Figure 12(c) shows the results for the bottom 1 / 3 set. In the middle and bottom sets, emphasis is placed on the crystal surface and the area near the wall of crucible 1. In the top set, emphasis is also placed on the bottom of crucible 1.

[0071] By having the operator confirm the analysis results from the judgment basis analysis unit 69, it is possible to confirm that the rating was performed correctly. In addition, useful information can be provided to the operator for manually controlling the crystal manufacturing apparatus 10.

[0072] Figure 13 shows the configuration of a control device 20 according to an embodiment of the present disclosure. The control device 20 comprises a communication device 21, a display device 22, an input device 23, a storage device 40, and a processing device 30. The control device 20 may be a server device, a device such as a personal computer, or a mobile terminal such as a mobile phone terminal, smartphone, or tablet terminal.

[0073] The communication device 21 controls communication with other devices. The communication device 21 may communicate using any communication method, such as wired or wireless. The display device 22 displays a screen generated by the processing device 30. The display device 22 may be a liquid crystal display device, an organic EL display device, or the like. The input device 23 transmits instructions from the operator to the processing device 30. The input device 23 may be a mouse, keyboard, touchpad, or the like. The display device 22 and the input device 23 may be implemented as a touch panel. The storage device 40 stores programs, data, etc., used by the processing device 30. The storage device 40 may be a semiconductor memory, a hard disk, or the like.

[0074] The processing unit 30 comprises a prediction unit 31, a prediction image presentation unit 32, an instruction reception unit 33, a control parameter acquisition unit 34, and a control unit 35. These configurations are implemented hardware-wise by the CPU, memory, and other LSIs of any computer, and software-wise by programs loaded into memory, but here we are describing functional blocks that are realized through the cooperation of these components. Therefore, it will be understood by those skilled in the art that these functional blocks can be implemented in various forms, such as hardware only or a combination of hardware and software.

[0075] The prediction unit 31 predicts the temperature distribution and flow distribution of the high-temperature metal liquid 2. The prediction image display unit 32 displays the temperature distribution and flow distribution of the high-temperature metal liquid 2 predicted by the prediction unit 31 on the display device 22.

[0076] The instruction receiving unit 33 receives instructions from the operator via the input device 23 or the like. The control parameter acquisition unit 34 acquires the optimal control parameter values ​​from the learning device 50. The control unit 35 controls the crystal manufacturing apparatus 10 according to the instructions received by the instruction receiving unit 33 and the control parameters acquired by the control parameter acquisition unit 34.

[0077] The learning device 50 may function as part of the control device 20. That is, while growing SiC crystals, the evaluation acquisition unit 64 of the learning device 50 may acquire evaluations from an operator viewing images presented by the prediction image presentation unit 32, the rating unit 65 may rate the images, and the optimal solution search unit 68 may search for optimal control parameters and notify the control device 20. When the control parameter acquisition unit 34 acquires control parameters from the learning device 50, the control unit 35 controls the crystal manufacturing apparatus 10 according to those control parameters. This makes it possible to automatically control the crystal manufacturing apparatus 10 while updating the image rating.

[0078] [Examples] Three operators were presented with images visualizing the temperature and flow distribution of high-temperature metallic liquid 2, and their evaluations were obtained. The images were then rated and the average was calculated. To verify the accuracy of the ratings, two images were selected from the rated images, with a difference of at least a certain threshold in their scores. The operators were then tested to determine whether the image with the higher score was considered "good." Two operators underwent 200 tests each, resulting in an accuracy rate of over 95%. These scores were used to explore the optimal conditions for growing SiC crystals.

[0079] Figure 14 shows examples of high-scoring images among those rated by the learning device 50. It was confirmed that images with desirable temperature and flow distribution characteristics, such as a small temperature difference between the top and bottom, a sufficiently low temperature at the top, a high temperature at the sides, outward and unidirectional flow at the top, and a clean separation of flow between the top and bottom, were assigned high scores.

[0080] Figure 15 shows the optimal temperature and flow distribution obtained by the learning device 50 according to the embodiment. Ideal temperature and flow distribution conditions that reflect the operator's intentions and knowledge were obtained in a short period of time.

[0081] Figure 16 shows SiC crystals produced by the crystal production apparatus 10 according to the embodiment. Figure 16(a) shows SiC crystals produced under conventional experimental conditions. The surface of the crystal is rough. Figure 16(b) shows SiC crystals produced under the optimal conditions shown in Figure 11. The thickness was 1.7 to 2.6 mm. The crystal surface is smooth, and stable growth over a long period of time is possible, demonstrating that larger and higher-quality SiC crystals can be produced.

[0082] The present disclosure has been explained above based on examples. These examples are illustrative, and it will be understood by those skilled in the art that various modifications are possible in combinations of their components and processing processes, and that such modifications are also within the scope of the present disclosure.

[0083] In this embodiment, multiple images were presented to the operator for comparison, but the information provided to the observer may be any type of information that the observer can perceive through their five senses, such as moving images, text, symbols, sounds, smells, and tastes.

[0084] While the embodiments described the case of rating images related to the manufacturing of SiC crystals, the technology of this disclosure can also be applied to rating videos related to the manufacturing of SiC crystals, videos or images related to the manufacturing of other semiconductor crystalline materials, videos or images related to the manufacturing of other crystalline materials, and videos or images related to the manufacturing of other materials. Images related to material manufacturing may include, for example, pattern images of material temperature, flow, concentration, etc., crystal surface photographs, crystal surface microscope images, crystal shape images, waveform pattern images, etc.

[0085] The technology disclosed herein can also be applied to rating other images. Other images may include, for example, color photographs of crops or printed pattern images. Figure 17 shows an example of a color photograph of a crop presented to an operator. The image selection unit 62 of the learning device 50 selects two photographs for comparison from among many color photographs of crops. The image presentation unit 63 presents the color photographs of crops selected by the image selection unit 62 to the operator. The evaluation acquisition unit 64 obtains evaluations from the operator regarding the quality of the crops, such as taste, quality, and growth status. The rating unit 65 updates the ratings of the two color photographs of crops presented to the operator based on the evaluations obtained from the operator. Each color photograph of a crop may be associated with the growth conditions of that crop. The latent space generation unit 66 constructs a latent space by dimensionality reduction of the color photographs of crops. The prediction AI learning unit 67 learns an AI to predict latent variables from the growth conditions of crops. The optimal solution search unit 68 searches for the optimal latent variables in the latent space. The optimal solution search unit 68 outputs the growth conditions corresponding to the searched latent variables. This makes it possible to obtain growth conditions suitable for crops.

[0086] The control device 20 or the learning device 50 may be implemented as a server device, and the crystal manufacturing apparatus 10 may be implemented as a client terminal. The learning device 50 may be implemented as a server device, and the control device 20 may be implemented as a client terminal. The server device may be implemented by multiple devices, and the multiple devices may be provided by multiple entities. The functions of the control device 20 or the learning device 50 may be distributed among multiple devices. Some of the functions of the control device 20 or the learning device 50 may be provided in the crystal manufacturing apparatus 10. [Industrial applicability]

[0087] This disclosure relates to technologies for incorporating human intentions and knowledge into control and other processes. [Explanation of Symbols]

[0088] 1 Crucible, 2 High-temperature metallic liquid, 3 Carbon stick, 4 Crystal, 5 Heating coil, 10 Crystal manufacturing apparatus, 20 Control device, 31 Prediction unit, 32 Prediction image presentation unit, 33 Instruction reception unit, 34 Control parameter acquisition unit, 35 Control unit, 50 Learning device, 61 Image acquisition unit, 62 Image selection unit, 63 Image presentation unit, 64 Evaluation acquisition unit, 65 Rating unit, 66 Latent space generation unit, 67 Prediction AI learning unit, 68 Optimal solution search unit, 69 Judgment basis analysis unit, 100 Crystal manufacturing system.

Claims

1. A computer, The steps include: making the observer perceive multiple pieces of perceptible information through pairwise comparison; The steps include obtaining the observer's evaluation of the information, A step of rating the multiple pieces of information by repeating the steps of perceiving and obtaining evaluation multiple times, The steps include generating a latent space by dimensionality reduction of the aforementioned multiple pieces of information, The steps include: searching for latent variables that satisfy predetermined conditions in the aforementioned latent space; Execute Learning methods.

2. The step of rating the aforementioned multiple pieces of information includes the step of rating the aforementioned multiple pieces of information by Elo rating. The learning method according to claim 1.

3. The step of generating the latent space includes generating the latent space by compressing the dimensions of the plurality of pieces of information so that similar information is placed close together. The learning method according to any one of claims 1 to 2.

4. The step of generating the latent space includes generating the latent space using a variational autoencoder that uses the scores assigned to each of the plurality of pieces of information in the rating step. The learning method according to claim 3.

5. The aforementioned search step includes a step of searching for the latent variable that maximizes the score. The learning method according to claim 4.

6. The aforementioned information is an image or a video. The learning method according to any one of claims 1 to 2.

7. The aforementioned information is an image related to the manufacturing of the material. The learning method according to claim 6.

8. The aforementioned information is an image that visualizes the temperature distribution and flow distribution of the raw materials used to manufacture the aforementioned material. The learning method according to claim 7.

9. The aforementioned material is a SiC crystal. The learning method according to claim 8.

10. An information provision unit that allows the observer to perceive multiple pieces of perceptible information through pairwise comparison, An evaluation acquisition unit that acquires the observer's evaluation of the information, A rating unit that rates the multiple pieces of information by repeatedly providing the information and obtaining the evaluation multiple times, A latent space generation unit generates a latent space by compressing the dimensions of the aforementioned multiple pieces of information, A search unit that searches for latent variables that meet predetermined conditions in the aforementioned latent space, Equipped with Learning device.

11. A computer, A step of making the observer perceive information representing the state of the controlled object through pairwise comparison, A step of obtaining information representing the state of the controlled object at a first point in time, and an evaluation by the observer of information representing the state of the controlled object at a second point in time different from the first point in time, The steps include controlling the controlled object according to control parameters corresponding to latent variables that match predetermined conditions, which are searched in a latent space generated by dimensionality reduction of multiple pieces of information rated based on the above evaluation, and Execute Control method.

12. The system includes a step of rating the state of the controlled object based on the evaluation, The aforementioned control step includes a step of controlling the controlled object based on its rating. The control method according to claim 11.

13. The aforementioned information is an image or a video. The control method according to any one of claims 11 to 12.

14. The controlled object is an apparatus for manufacturing materials, The aforementioned information is an image related to the manufacturing of the material. The control method according to claim 13.

15. The aforementioned information is an image that visualizes the temperature distribution and flow distribution of the raw materials used to manufacture the aforementioned material. The control method according to claim 14.

16. The aforementioned material is a SiC crystal, The control step involves controlling at least one of the following: the rotation speed of the seed crystal, the rotation speed of the crucible, the relative position of the crucible and the coil, the position of the coil, the height of the high-temperature metal liquid, the height of the meniscus of the high-temperature metal liquid, and the set temperature. The control method according to claim 15.

17. An information provision unit that allows the observer to perceive information representing the state of the controlled object through pairwise comparison, An evaluation acquisition unit that acquires information representing the state of the controlled object at a first point in time, and an evaluation by the observer of information representing the state of the controlled object at a second point in time different from the first point in time, A control unit controls the controlled object according to control parameters corresponding to latent variables that match predetermined conditions, which are searched in a latent space generated by dimensionality reduction of multiple pieces of information rated based on the above evaluation. Equipped with Control device.

18. Computers, An information provision unit that allows the observer to perceive multiple pieces of perceptible information through pairwise comparison, An evaluation acquisition unit that acquires the observer's evaluation of the information, A rating unit that rates the multiple pieces of information by repeatedly providing the information and obtaining the evaluation multiple times, A latent space generation unit generates a latent space by compressing the dimensions of the aforementioned multiple pieces of information, A search unit that searches for latent variables that meet predetermined conditions in the aforementioned latent space, A learning program designed to function as such.

19. Computers, An information provision unit that allows the observer to perceive information representing the state of the controlled object through pairwise comparison, An evaluation acquisition unit that acquires information representing the state of the controlled object at a first point in time, and an evaluation by the observer of information representing the state of the controlled object at a second point in time different from the first point in time, A control unit controls the controlled object according to control parameters corresponding to latent variables that match predetermined conditions, which are searched in a latent space generated by dimensionality reduction of multiple pieces of information rated based on the above evaluation. A control program to enable it to function as such.

Images