Design support method and design support device for microstructure material

Abstract

To efficiently optimize a hyperparameter set in learning of AI for generating a model of a three dimensional structure from a two dimensional image of microstructure material.SOLUTION: A design support method for microstructure material supports design of microstructure material by adjusting a hyperparameter set during learning of AI for generating a model of a three dimensional structure from a two dimensional image of the microstructure material. In this designing support method, an evaluation indicator relating to geometric shape difference between an actual structure of the microstructure material and a structure of a model generated by AI is calculated by a computer, and a hyperparameter is adjusted based on the evaluation indicator (steps S10, S20, S30, and S50 ).SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to a method and apparatus for aiding in the design of microstructured materials. [Background technology]

[0002] Conventionally, microstructured materials with various microstructures have been used depending on the application, such as porous materials with many pores arranged three-dimensionally within the material, crystalline materials with a specific crystal structure, etc. When designing such microstructured materials, it is necessary to determine the microstructure so that appropriate properties can be obtained depending on the application.

[0003] Conventional design methods for microstructured materials include, for example, creating several prototypes based on the designer's experience and intuition, measuring the three-dimensional structure and properties of each of these prototypes, and then determining the final three-dimensional structure of the microstructured material. However, because this design method requires measuring the three-dimensional structure of the prototype, it has several drawbacks, including the fact that the measurement requires more time and effort than two-dimensional observation methods, the obtained measurement results have lower resolution, and the structural features may not be fully captured due to the limited field of view during measurement.

[0004] On the other hand, in recent years, as described in Non-Patent Document 1, for example, it has been proposed to use machine learning to train AI (artificial intelligence) to learn the relationship between two-dimensional images and three-dimensional structures of various microstructure materials, and to use this trained AI to generate a three-dimensional structure model from two-dimensional images of a specific microstructure. [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] Steve Kench and Samuel J. Cooper,"GENERATING 3D STRUCTURES FROM A 2D SLICE WITH GAN-BASED DIMENSIONALITY EXPANSION", arXiv:2102.07708 [cs.CV], 10 Feb 2021 Summary of the Invention [Problem to be solved by the invention]

[0006] The method in Non-Patent Document 1 requires the appropriate setting of hyperparameters to control the behavior of the AI ​​during learning, but currently there is no established method for setting hyperparameters, so designers must set the hyperparameters blindly.

[0007] In view of the above-mentioned problems, the present invention aims to provide a technology that can efficiently optimize the hyperparameters set during learning of an AI that generates a three-dimensional structure model from a two-dimensional image of a microstructure material. [Means for solving the problem]

[0008] The design support method for microstructured materials according to the present invention is a method for supporting the design of microstructured materials by adjusting hyperparameters that are set during learning of an AI that generates a three-dimensional structure model from a two-dimensional image of the microstructured material. The computer calculates an evaluation index related to the geometric shape difference between the actual structure of the microstructured material and the structure of the model generated by the AI, and the computer adjusts the hyperparameters based on the evaluation index. The design support device for microstructured materials according to the present invention is a device that supports the design of microstructured materials by adjusting hyperparameters that are set during learning of an AI that generates a three-dimensional structure model from a two-dimensional image of the microstructured material, and is equipped with an evaluation index calculation unit that calculates an evaluation index related to the geometric shape difference between the actual structure of the microstructured material and the structure of the model generated by the AI, and a parameter adjustment unit that adjusts the hyperparameters based on the evaluation index calculated by the evaluation index calculation unit. [Effects of the Invention]

[0009] According to the present invention, it is possible to provide a technology that can efficiently optimize the hyperparameters that are set during learning of an AI that generates a three-dimensional structure model from a two-dimensional image of a microstructure material. [Brief explanation of the drawings]

[0010] [Figure 1] 1 is a block diagram showing the configuration of a design support apparatus according to an embodiment of the present invention; [Figure 2] 3 is a flowchart showing a processing flow of a design support apparatus according to an embodiment of the present invention. [Figure 3] 10 is a flowchart showing details of a generator learning rate adjustment process. [Figure 4] 10 is a flowchart showing details of a classifier learning rate adjustment process. [Figure 5] 10 is a flowchart showing details of a learning image number adjustment process. [Figure 6] 10 is a flowchart showing details of a learning image number adjustment process. [Figure 7] 10 is a flowchart showing details of a learning rate adjustment process. [Figure 8] 10 is a flowchart showing details of a learning process. [Figure 9] 10 is a flowchart showing details of a learning process. [Figure 10] 10 is a flowchart showing details of a generated structure evaluation process. DETAILED DESCRIPTION OF THE INVENTION

[0011] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. In the following embodiment, a design support device will be described that supports the design of microstructured materials by efficiently learning AI used in designing microstructured materials such as porous bodies and crystalline bodies having various microstructures by adjusting hyperparameters to appropriate values ​​when learning this AI.

[0012] The design support system of the present invention can be applied to a variety of microstructured materials. For example, it can be applied to structures with internal grain boundaries, structures formed by contact between particles or fibers, and composite materials in which contrast differences between materials can be clearly observed using observation techniques such as SEM. The present invention can also be applied to porous materials such as inorganic materials (ceramics, glass, carbon, metals, etc.), organic materials (polyethylene, epoxy, acrylic, urethane, polyimide, ion exchange resin, polyvinyl butyral, cellulose, etc.), and inorganic-organic hybrid materials (CFRP, films containing inorganic fillers, etc.). For example, in the ceramic manufacturing process, a molded body corresponds to an inorganic-organic hybrid material, and a degreased body corresponds to a structure formed by contact between particles. In addition, the present invention can be applied to any material as long as the structure has a microstructure containing an air layer inside.

[0013] Fig. 1 is a block diagram showing the configuration of a design support device according to one embodiment of the present invention. The design support device 100 shown in Fig. 1 is a computer including a control unit 1, a storage unit 2, a memory 3, an operation input device 4, a display device 5, and a network device 6, and is configured by connecting these devices to each other via a bus 7.

[0014] The control unit 1 is configured using, for example, a CPU (Central Processing Unit) and performs various processes and calculations to operate the design support device 100. The control unit 1 executes programs stored in the storage unit 2 to realize the functional blocks of the AI ​​calculation unit 11 and the AI ​​learning unit 12. Details of these functional blocks will be described later. Note that some or all of the functions of the control unit 1 may be realized using devices other than a CPU, such as a GPU (Graphic Processing Unit), FPGA (Field Programmable Gate Array), or ASIC (Application Specific Integrated Circuit).

[0015] The storage unit 2 is configured using a large-capacity, non-volatile storage device, such as a magnetic storage device such as an HDD (Hard Disk Drive) or an SSD (Solid State Drive), and stores programs executed by the control unit 1 and various information used in the processing of the control unit 1. The information stored in the storage unit 2 includes generator data 21, classifier data 22, teacher image data 23, generative structure data 24, hyperparameter data 25, and evaluation index data 26. Details of these data will be described later.

[0016] The memory 3 is configured using a high-speed and volatile storage device such as a DRAM (Dynamic Random Access Memory), and is used as a work area when the control unit 1 executes a program.

[0017] The operation input device 4 is a device for detecting operation input by the user, and is configured using, for example, a keyboard or a mouse. The display device 5 is a device for displaying the processing results of the design support device 100 on a screen to present them to the user, and is configured using, for example, a liquid crystal display or an organic EL display. The network device 6 is a device for communicating with other computers via a network (not shown). By using the network device 6, the design support device 100 can transmit various data stored in the storage unit 2 to other computers, and can receive various data transmitted from other computers and store it in the storage unit 2. Note that other computers that can communicate with the design support device 100 may be used as the operation input device 4 or the display device 5.

[0018] 1 may be physically constructed on one computer, or may be distributed across multiple computers. Furthermore, the design support device 100 may be realized by a cloud computer installed on a cloud, a virtual machine operating on a virtual environment, or the like.

[0019] Next, we will explain the functional blocks of the AI ​​calculation unit 11 and the AI ​​learning unit 12 in the control unit 1. In the design support device 100, the control unit 1 operates as these functional blocks, thereby enabling learning of the AI ​​used to design a microstructured material having a desired microstructure.

[0020] The AI ​​calculation unit 11 has a generator 110 and a discriminator 111. The generator 110 is an AI that generates a model of a microstructure material having a desired three-dimensional structure according to input conditions, and performs AI calculations using a predetermined algorithm with a neural network based on generator data 21 stored in the memory unit 2. Information on the three-dimensional structure model of the microstructure material generated by the generator 110 is stored in the memory unit 2 as generated structure data 24. The discriminator 111 is an AI that determines whether an input cross-sectional image of the three-dimensional structure is a cross-sectional image of the actual microstructure material or a cross-sectional image of the three-dimensional structure model generated by the generator 110, and performs AI calculations using a predetermined algorithm with a neural network based on discriminator data 22 stored in the memory unit 2.

[0021] In the AI ​​computing unit 11, the generator 110 and the discriminator 111 have opposing objectives, and AI learning is performed by the generator 110 and the discriminator 111 competing with each other to achieve their respective objectives. That is, the generator 110 is trained to generate a three-dimensional structural model of a microstructure material that is closer to the actual object. On the other hand, the discriminator 111 is trained to more accurately discriminate that a cross-sectional image of the three-dimensional structural model generated by the generator 110 is not a cross-sectional image of the actual object. Such a configuration of the AI ​​computing unit 11 is called a GAN (Generative Adversarial Network). The learning state of the generator 110 is reflected in generator data 21 as weight information of the neural network that constitutes the generator 110. Similarly, the learning state of the discriminator 111 is reflected in discriminator data 22 as weight information of the neural network that constitutes the discriminator 111.

[0022] The AI ​​learning unit 12 is a functional block for AI learning, and includes a learning processing unit 120, an evaluation index calculation unit 121, and a parameter adjustment unit 122. The learning processing unit 120 performs learning processing of the generator 110 and the discriminator 111 using the above-described learning method, using multiple cross-sectional images of the actual microstructure material represented by the teacher image data 23, or images obtained by cutting out portions of these cross-sectional images, and multiple cross-sectional images of the three-dimensional structure model represented by the generated structure data 24, or images obtained by cutting out portions of these cross-sectional images. At this time, the learning processing unit 120 sets conditions for the learning processing based on various hyperparameters represented by the hyperparameter data 25. The evaluation index calculation unit 121 calculates an evaluation index related to the geometric shape difference between the actual structure of the microstructure material represented by the teacher image data 23 and the three-dimensional structure model represented by the generated structure data 24, and stores the calculation result in the memory unit 2 as evaluation index data 26. The parameter adjustment unit 122 adjusts the hyperparameters used by the learning processing unit 120 in the learning process, based on the evaluation index calculated by the evaluation index calculation unit 121.

[0023] The generated structure data 24 is transmitted from the design support device 100 to a computer that designs the microstructure material, for example, via the network device 6. The computer designs the microstructure material based on the generated structure data 24 received from the design support device 100. In this way, the design support device 100 can support the design of the microstructure material.

[0024] As described above, the learning processing unit 120 uses a plurality of cross-sectional images of the actual microstructure material represented by the teacher image data 23, or images obtained by cutting out portions of these cross-sectional images, and a plurality of cross-sectional images of the three-dimensional structure model represented by the generated structure data 24, or images obtained by cutting out portions of these cross-sectional images, when training the generator 110 and the discriminator 111. In the following description, among the images used when training the generator 110 and the discriminator 111, the cross-sectional images of the actual microstructure material represented by the teacher image data 23, or images obtained by cutting out portions of these cross-sectional images, are referred to as "first learning images," and the cross-sectional images of the three-dimensional structure model represented by the generated structure data 24, or images obtained by cutting out portions of these cross-sectional images, are referred to as "second learning images." The image sizes of these first learning images and second learning images can be determined arbitrarily depending on the image size of the cross-sectional image of the actual microstructure material represented by the teacher image data 23 and the size of the three-dimensional structure model represented by the generated structure data 24.

[0025] Moreover, of the first training images, the number of first training images used for training the classifier 111 will be referred to as "number of images A" below. On the other hand, of the second training images, the number of second training images used for training the classifier 111 will be referred to as "number of images B" below. Furthermore, of the second training images, the number of second training images used for training the generator 110 will be referred to as "number of images C" below.

[0026] Next, details of the processing performed by the control unit 1 will be described with reference to Fig. 2. Fig. 2 is a flowchart showing the processing flow of a design support device according to one embodiment of the present invention. In response to a user's operational input, the design support device 100 performs training of the generator 110 and the discriminator 111 while adjusting hyperparameters to appropriate values ​​by having the control unit 1 execute the processing shown in the flowchart of Fig. 2. This effectively trains the AI ​​used in the design of microstructured materials, thereby supporting the design of microstructured materials using AI.

[0027] When the user instructs the design support device 100 to start processing via the operation input device 4, in step S10, the control unit 1 uses the AI ​​calculation unit 11 and the AI ​​learning unit 12 to perform a generator learning rate adjustment process for adjusting the learning rate, which represents the update width of the weight during learning of the generator 110, among the hyperparameters represented by the hyperparameter data 25. Details of this generator learning rate adjustment process will be described later with reference to FIG.

[0028] In step S20, the control unit 1 uses the AI ​​calculation unit 11 and the AI ​​learning unit 12 to perform a classifier learning rate adjustment process to adjust the learning rate, which represents the update width of the weight during learning of the classifier 111, among the hyperparameters represented by the hyperparameter data 25. Details of this classifier learning rate adjustment process will be described later with reference to FIG.

[0029] In step S30, the control unit 1 uses the AI ​​calculation unit 11 and the AI ​​learning unit 12 to perform a training image number adjustment process to adjust the number of training images, which indicates the number of images used in training the generator 110 and the classifier 111, among the hyperparameters represented by the hyperparameter data 25. Details of this training image number adjustment process will be described later with reference to FIGS. 5 and 6.

[0030] In step S50, the control unit 1 uses the AI ​​calculation unit 11 and the AI ​​learning unit 12 to perform a learning ratio adjustment process for adjusting a learning ratio, which represents the ratio of the number of times the classifier 111 has learned to the number of times the generator 110 has learned, among the hyperparameters represented by the hyperparameter data 25. Details of this learning ratio adjustment process will be described later with reference to FIG. 7.

[0031] After adjusting the hyperparameters of the learning rate and number of learning images of the generator 110 and the classifier 111, and the learning ratio of the classifier 111 to the generator 110 through the processing of steps S10 to S50, the control unit 1 ends the processing shown in the flowchart of Figure 2.

[0032] FIG. 3 is a flowchart showing the details of the generator learning rate adjustment process executed in step S10 of FIG.

[0033] In step S100, among the various hyperparameters represented by the hyperparameter data 25, initial parameters for adjusting the learning rate of the generator 110 are set. Here, the search range and update width of the learning rate when searching for an appropriate value for the learning rate of the generator 110 are set as initial parameters. For example, if the search range is set to 1×10 -7 From 1×10 -1 The initial parameters can be set to a range of 10x the update rate.

[0034] In step S110, the parameter adjustment unit 122 sets the value of the learning rate of the generator 110 to the lower limit of the search range based on the initial parameters set in step S100. Here, the lower limit of the search range predetermined for the learning rate of the generator 110, for example, 1×10 in the above example search range, is set. -7 is set as the learning rate of the generator 110 when the learning process is performed in the next step S120.

[0035] In step S120, the learning processing unit 120 performs a learning process on the generator 110 and the discriminator 111. Here, the learning process is performed on the generator 110 and the discriminator 111 according to the learning rate of the generator 110 set to the lower limit value of the search range in step S110 or the learning rate of the generator 110 updated in step S140, which will be described later. Details of the learning process performed in step S120 will be described later with reference to FIGS. 8 and 9.

[0036] In step S130, the evaluation index calculation unit 121 performs a generated structure evaluation process to evaluate the three-dimensional structure model finally generated by the generator 110 in the learning process performed in step S120. Here, using the cross-sectional image acquired from the three-dimensional structure model represented by the generated structure data 24 and the teacher image data 23, the difference in feature amounts of these images is calculated as an evaluation index related to the geometric shape difference between the structure of the microstructure material represented by the three-dimensional structure model generated by the generator 110 after learning and the actual structure of the microstructure material. Details of the generated structure evaluation process performed in step S130 will be described later with reference to FIG. 10.

[0037] In step S140, the parameter adjustment unit 122 updates the setting value of the learning rate of the generator 110 based on the initial parameters set in step S100. Here, the learning rate value set when the learning process was performed in the most recent step S120 is updated to a different value. At this time, starting from the lower limit value set in step S110, the learning rate value of the generator 110 is gradually increased each time the process of step S140 is performed in accordance with the search range and update width set as the initial parameters in step S100. Note that the update width of the learning rate at this time is set to 10 times in the above example, but can be set to any other value.

[0038] In step S150, it is determined whether the learning rate of the generator 110 updated in step S140 exceeds the upper limit of a predetermined search range. If the updated learning rate of the generator 110 does not exceed the upper limit of the search range and is within the search range, the process returns to step S120, and steps S120 to S140 are repeated using the setting value of the learning rate of the generator 110 updated in step S140. This allows the generator 110 and the discriminator 111 to learn for multiple learning rate values ​​of the generator 110 set for each update width within the search range, and an evaluation index related to the geometric shape difference between the structure of the microstructure material represented by the three-dimensional structure model generated by the learned generator 110 and the actual structure of the microstructure material can be obtained. On the other hand, if the updated learning rate of the generator 110 exceeds the upper limit of the search range, the loop process of steps S120 to S140 ends and the process proceeds to step S160.

[0039] In step S160, the parameter adjustment unit 122 acquires the learning rate value of the generator 110 that has obtained the best evaluation index in the generated structure evaluation process of step S130 as the best value of the learning rate of the generator 110. Here, the evaluation index that has the best value among the evaluation indexes obtained for each loop process of steps S120 to S140, that is, the learning rate value of the generator 110 when a three-dimensional structure model is obtained in which the difference between the feature amounts of the cross-sectional image acquired from the three-dimensional structure model represented by the generated structure data 24 and the feature amounts of the cross-sectional image of the actual structure represented by the teacher image data 23 is smallest, is acquired as the best value of the learning rate of the generator 110. In this way, the learning rate of the generator 110 is adjusted so as to minimize the difference between the actual cross-sectional image of the microstructure material and the cross-sectional image of the three-dimensional structure model.

[0040] After performing the process of step S160, the parameter adjustment unit 122 updates the hyperparameter data 25 using the acquired best value of the learning rate of the generator 110, and ends the generator learning rate adjustment process shown in the flowchart of Fig. 3. This adjusts the learning rate of the generator 110 to an optimal value.

[0041] FIG. 4 is a flowchart showing the details of the classifier learning rate adjustment process executed in step S20 of FIG.

[0042] In step S200, among the various hyperparameters represented by the hyperparameter data 25, initial parameters for adjusting the learning rate of the classifier 111 are set. Here, similar to step S100 in FIG. 3, the search range and update width of the learning rate when searching for an appropriate value for the learning rate of the classifier 111 are set as initial parameters. For example, if the search range is set to 1×10 -7 From 1×10 -1 The initial parameters can be set to a range of 10x the update rate.

[0043] In step S210, the value of the learning rate of the generator 110 is set to the best value obtained by the above-mentioned generator learning rate adjustment process among the various hyperparameters represented by the hyperparameter data 25. Here, the best value of the learning rate of the generator 110 obtained in step S160 of Fig. 3 is set as the learning rate of the generator 110 when performing the learning process in the later step S230.

[0044] In step S220, the parameter adjustment unit 122 sets the value of the learning rate of the classifier 111 to the lower limit of the search range based on the initial parameters set in step S100. Here, the lower limit of the search range predetermined for the learning rate of the classifier 111, for example, 1×10 -7 is set as the learning rate of the classifier 111 when the learning process is carried out in the next step S230.

[0045] In step S230, the learning processing unit 120 performs a learning process on the generator 110 and the classifier 111. Here, the learning process is performed on the generator 110 and the classifier 111 in the same manner as in step S120 in Fig. 3, in accordance with the best value of the learning rate of the generator 110 set in step S210, the learning rate of the classifier 111 set to the lower limit of the search range in step S220, or the learning rate of the classifier 111 updated in step S250, which will be described later. Details of the learning process performed in step S230 will be described later with reference to Figs. 8 and 9.

[0046] In step S240, the evaluation index calculation unit 121 performs a generated structure evaluation process for evaluating the three-dimensional structure model finally generated by the generator 110 in the learning process performed in step S230. Here, similar to step S130 in Fig. 3, using the cross-sectional image acquired from the three-dimensional structure model represented by the generated structure data 24 and the teacher image data 23, the difference in feature amounts of these images is calculated as an evaluation index related to the geometric shape difference between the structure of the microstructure material represented by the three-dimensional structure model generated by the generator 110 after learning and the actual structure of the microstructure material. Details of the generated structure evaluation process performed in step S240 will be described later with reference to Fig. 10.

[0047] In step S250, the parameter adjustment unit 122 updates the setting value of the learning rate of the discriminator 111 based on the initial parameters set in step S200. Here, the learning rate value set when the most recent learning process was performed in step S230 is updated to a different value. In this case, similar to step S140 in FIG. 3, starting from the lower limit value set in step S220, the learning rate value of the discriminator 111 is gradually increased each time the process of step S250 is performed in accordance with the search range and update width set as the initial parameters in step S200. Note that the update width of the learning rate at this time is set to 10 times in the above example, but can be set to any other value.

[0048] In step S260, it is determined whether the learning rate of the discriminator 111 updated in step S250 exceeds the upper limit of a predetermined search range. If the updated learning rate of the discriminator 111 does not exceed the upper limit of the search range and is within the search range, the process returns to step S230, and steps S230 to S250 are repeated using the set value of the learning rate of the discriminator 111 updated in step S250. This allows the generator 110 and the discriminator 111 to learn for multiple learning rate values ​​of the discriminator 111 set for each update width within the search range, and an evaluation index related to the geometric shape difference between the structure of the microstructure material represented by the three-dimensional structure model generated by the learned generator 110 and the actual structure of the microstructure material can be obtained. On the other hand, if the updated learning rate of the discriminator 111 exceeds the upper limit of the search range, the loop process of steps S230 to S250 ends and the process proceeds to step S270.

[0049] In step S270, the parameter adjustment unit 122 acquires the learning rate value of the discriminator 111 that obtained the best evaluation index in the generated structure evaluation process of step S240 as the best value of the learning rate of the discriminator 111. Here, the best value of the learning rate of the discriminator 111 is acquired as the best value of the learning rate of the discriminator 111, which is the evaluation index obtained for each loop process of steps S230 to S250, that is, the learning rate value of the discriminator 111 when a three-dimensional structure model is obtained in which the difference between the feature amounts of the cross-sectional image acquired from the three-dimensional structure model represented by the generated structure data 24 and the feature amounts of the cross-sectional image of the actual structure represented by the teacher image data 23 is smallest. In this way, the learning rate of the discriminator 111 is adjusted so as to minimize the difference between the actual cross-sectional image of the microstructure material and the cross-sectional image of the three-dimensional structure model.

[0050] After performing the process of step S270, the parameter adjustment unit 122 updates the hyperparameter data 25 using the acquired best value of the learning rate of the classifier 111, and ends the classifier learning rate adjustment process shown in the flowchart of Fig. 4. As a result, the learning rate of the classifier 111 is adjusted to an optimal value.

[0051] 5 and 6 are flowcharts showing the details of the training image number adjustment process executed in step S30 of FIG.

[0052] 5, initial parameters are set for adjusting the aforementioned number of images A, B, and C among the various hyperparameters represented by the hyperparameter data 25. Here, the search range and update width when searching for appropriate values ​​for each of the number of images A, B, and C are set as initial parameters. For example, the initial parameters can be set so that the number of images A is adjusted to at least 50 or more, and the number of images B and C are adjusted to at least 150 or more.

[0053] In step S310, the value of the learning rate of the generator 110 is set to the best value obtained by the above-mentioned generator learning rate adjustment process among the various hyperparameters represented by the hyperparameter data 25. Here, similar to step S210 in Fig. 4, the best value of the learning rate of the generator 110 obtained in step S160 in Fig. 3 is set as the learning rate of the generator 110 when performing the learning process in the later step S350.

[0054] In step S320, the value of the learning rate of the classifier 111 is set to the best value obtained by the above-mentioned classifier learning rate adjustment process among the various hyperparameters represented by the hyperparameter data 25. Here, the best value of the learning rate of the classifier 111 obtained in step S270 of Fig. 4 is set as the learning rate of the classifier 111 when performing the learning process in the later step S350.

[0055] In step S330, based on the initial parameters set in step S300, the parameter adjustment unit 122 sets the aforementioned number of images A, i.e., the number of first training images used in training the classifier 111, the aforementioned number of images B, i.e., the number of second training images used in training the classifier 111, and the aforementioned number of images C, i.e., the number of second training images used in training the generator 110, to the lower limit values ​​of their respective predetermined search ranges. Here, the lower limit values ​​of the predetermined search ranges for each of the numbers of images A, B, and C are set as the numbers of first training images and second training images to be input to the classifier 111 and the generator 110, respectively, when the training process is performed in a later step S350.

[0056] In step S340, it is determined whether the current set numbers of images A, B, and C satisfy the relational expression A≦B≦C. If the relational expression A≦B≦C is satisfied, the process proceeds to step S350; if not, the process proceeds to step S370. Here, the number of images A is the lower limit of the search range set in step S330 or steps S400 and S430 described below, or the value updated in step S370 described below. The number of images B is the lower limit of the search range set in step S330 or step S430 described below, or the value updated in step S390 described below. The number of images C is the lower limit of the search range set in step S330 or the value updated in step S420 described below.

[0057] In step S350, the learning processing unit 120 performs a learning process on the generator 110 and the classifier 111. Here, the learning process is performed on the generator 110 and the classifier 111 in the same manner as in step S120 of FIG. 3 and step S230 of FIG. 4, based on the best values ​​of the learning rates of the generator 110 and the classifier 111 set in steps S310 and S320, respectively, and the combination of the currently set numbers of images A, B, and C. Note that the value of the number of images A used in this learning process is the lower limit of the search range set in any of steps S330, S400, and S430, or a value updated in step S370, which will be described later. Similarly, the value of the number of images B is the lower limit of the search range set in any of steps S330 and S430, or a value updated in step S390, which will be described later, and the value of the number of images C is the lower limit of the search range set in step S330, or a value updated in step S420, which will be described later. The learning process performed in step S350 will be described in detail below with reference to FIGS.

[0058] In step S360, the evaluation index calculation unit 121 performs a generated structure evaluation process for evaluating the three-dimensional structure model finally generated by the generator 110 in the learning process performed in step S350. Here, similar to step S130 in Fig. 3 and step S240 in Fig. 4, using the cross-sectional image acquired from the three-dimensional structure model represented by the generated structure data 24 and the teacher image data 23, the difference in feature amounts of these images is calculated as an evaluation index related to the geometric shape difference between the structure of the microstructure material represented by the three-dimensional structure model generated by the generator 110 after learning and the actual structure of the microstructure material. Details of the generated structure evaluation process performed in step S360 will be described later with reference to Fig. 10.

[0059] In step S370, parameter adjustment unit 122 updates the set value of number of images A based on the initial parameters set in step S300. Here, the value of number of images A that was set when the learning process was performed in the most recent step S350 is updated to a different value. In this case, starting from the lower limit value set in one of steps S330, S400, or S430, the value of number of images A is gradually increased each time the process of step S370 is performed. Note that the update interval for number of images A at this time can be set arbitrarily.

[0060] In step S380, it is determined whether the number of images A updated in step S370 exceeds the upper limit of the predetermined search range. If the number of images A does not exceed the upper limit of the search range and is within the search range, the process returns to step S340, and the processing of steps S340 to S370 is repeated using the set value of the number of images A updated in step S370. On the other hand, if the updated number of images A exceeds the upper limit of the search range, the loop processing of steps S340 to S370 ends, and the process proceeds to step S390 in FIG. 6.

[0061] In step S390, parameter adjustment unit 122 updates the set value of number of images B based on the initial parameters set in step S300. Here, the value of number of images B that was set when the learning process was performed in the most recent step S350 is updated to a different value. In doing so, starting from the lower limit value set in either step S330 or S430, the value of number of images B is gradually increased each time the process of step S390 is performed. Note that the update interval for number of images B at this time can be set arbitrarily.

[0062] In step S400, the parameter adjustment unit 122 sets the number of images A to the lower limit of a predetermined search range based on the initial parameters set in step S300. Here, the number of images A, which has been updated to the upper limit by the loop processing of steps S340 to S370, is returned to the lower limit value before the loop processing set in step S330.

[0063] In step S410, it is determined whether the number of images B updated in step S390 exceeds the upper limit of the predetermined search range. If the number of images B does not exceed the upper limit of the search range and is within the search range, the process returns to step S340, and the processing of steps S340 to S400 is repeated using the set value of the number of images B updated in step S390. On the other hand, if the updated number of images B exceeds the upper limit of the search range, the process proceeds to step S420.

[0064] In step S420, parameter adjustment unit 122 updates the set value of the number of images C based on the initial parameters set in step S300. Here, the value of the number of images C that was set when the learning process was performed in the most recent step S350 is updated to a different value. In this case, starting from the lower limit value set in step S330, the value of the number of images C is gradually increased each time the process of step S420 is performed. Note that the update interval for the number of images C at this time can be set arbitrarily.

[0065] In step S430, the parameter adjustment unit 122 sets the numbers of images A and B to the lower limit values ​​of a predetermined search range based on the initial parameters set in step S300. Here, the numbers of images A and B, which have been updated to the upper limit values ​​by the loop processing of steps S340 to S410, are returned to the lower limit values ​​set before the loop processing in step S330.

[0066] In step S440, it is determined whether the number of images C updated in step S420 exceeds the upper limit of the predetermined search range. If the number of images C does not exceed the upper limit of the search range and is within the search range, the process returns to step S340, and the processing of steps S340 to S430 is repeated using the set value of the number of images C updated in step S420. On the other hand, if the updated number of images C exceeds the upper limit of the search range, the process proceeds to step S450.

[0067] By processing steps S310 to S440 described above, the generator 110 and the discriminator 111 are trained for each combination of the values ​​of the number of images A, B, and C, each of which is set multiple times within the search range, and an evaluation index can be obtained regarding the geometric shape difference between the structure of the microstructure material represented by the three-dimensional structure model generated by the trained generator 110 and the actual structure of the microstructure material.

[0068] In step S450, the parameter adjustment unit 122 acquires the combination of the values ​​of the numbers of images A, B, and C that yields the best evaluation index in the generated structure evaluation process of step S360 as the best value for the number of training images to be used for training the generator 110 and the discriminator 111. Here, the combination of the values ​​of the numbers of images A, B, and C that yields the best value among the evaluation indexes obtained for each of the above loop processes, i.e., the combination of the values ​​of the numbers of images A, B, and C that yields the three-dimensional structure model that minimizes the difference between the feature amounts of the cross-sectional image acquired from the three-dimensional structure model represented by the generated structure data 24 and the feature amounts of the cross-sectional image of the actual structure represented by the teacher image data 23, is acquired as the best value for the number of first training images and the second training images to be used for training the generator 110 and the discriminator 111, respectively. In this way, the number of images input to the generator 110 and the discriminator 111 as the first training images and the second training images is adjusted so as to minimize the difference between the cross-sectional image of the actual microstructure material and the cross-sectional image of the three-dimensional structure model.

[0069] After performing the processing of step S450, the parameter adjustment unit 122 updates the hyperparameter data 25 using the best value of the combination of the acquired values ​​of the numbers of images A, B, and C, and ends the processing for adjusting the number of training images shown in the flowcharts of Figures 5 and 6. This adjusts the numbers of first training images and second training images used for training the generator 110 and the classifier 111, respectively, to optimal values.

[0070] FIG. 7 is a flowchart showing the details of the learning rate adjustment process executed in step S50 of FIG.

[0071] In step S500, initial parameters are set for adjusting the learning rates of the generator 110 and the discriminator 111, among the various hyperparameters represented by the hyperparameter data 25. Here, the search range and update step for searching for an appropriate value for the learning rate are set as the initial parameters. For example, the initial parameters can be set so that the search range for the learning rate is from 0.01 to 100 and the update step is 10 times.

[0072] In step S510, the value of the learning rate of the generator 110 is set to the best value obtained by the above-described generator learning rate adjustment process among the various hyperparameters represented by the hyperparameter data 25. Here, similar to step S210 in Fig. 4 and step S310 in Fig. 5, the best value of the learning rate of the generator 110 obtained in step S160 in Fig. 3 is set as the learning rate of the generator 110 when performing the learning process in the later step S550.

[0073] In step S520, the value of the learning rate of the classifier 111 is set to the best value obtained by the above-mentioned classifier learning rate adjustment process among the various hyperparameters represented by the hyperparameter data 25. Here, similar to step S320 in Fig. 5, the best value of the learning rate of the classifier 111 obtained in step S270 in Fig. 4 is set as the learning rate of the classifier 111 when performing the learning process in the later step S550.

[0074] In step S530, the values ​​of the numbers of images A, B, and C, among the various hyperparameters represented by the hyperparameter data 25, are set to the best values ​​found by the above-mentioned training image number adjustment process. Here, the best value of the combination of the values ​​of the numbers of images A, B, and C obtained in step S450 of Fig. 6 is set as the number of first training images and the number of second training images used, respectively, for training the generator 110 and the classifier 111 when the training process is performed in a later step S550.

[0075] In step S540, the parameter adjustment unit 122 sets the learning ratio of the generator 110 and the discriminator 111, i.e., the ratio of the number of times the discriminator 111 has learned to the number of times the generator 110 has learned, to the lower limit of a predetermined search range based on the initial parameters set in step S500. Here, the lower limit of the search range predetermined for the learning ratio, for example, 0.01 in the above example search range, is set as the learning ratio of the generator 110 and the discriminator 111 when performing the learning process in the next step S550.

[0076] In step S550, the learning processing unit 120 performs a learning process on the generator 110 and the classifier 111. Here, the learning process is performed on the generator 110 and the classifier 111 in the same manner as in step S120 in Fig. 3, step S230 in Fig. 4, and step S350 in Fig. 5, in accordance with the best values ​​of the learning rates of the generator 110 and the classifier 111 set in steps S510 and S520, respectively, the best values ​​of the numbers of images A, B, and C set in step S530, and the learning ratio set as the lower limit of the search range in step S540 or the learning ratio updated in step S570, which will be described later. Details of the learning process performed in step S550 will be described later with reference to Figs. 8 and 9.

[0077] In step S560, the evaluation index calculation unit 121 performs a generated structure evaluation process to evaluate the three-dimensional structure model finally generated by the generator 110 in the learning process performed in step S550. Here, similar to step S130 in Fig. 3, step S240 in Fig. 4, and step S360 in Fig. 5, using the cross-sectional image acquired from the three-dimensional structure model represented by the generated structure data 24 and the teacher image data 23, the difference in the feature amounts of these images is calculated as an evaluation index related to the geometric shape difference between the structure of the microstructure material represented by the three-dimensional structure model generated by the generator 110 after learning and the actual structure of the microstructure material. Details of the generated structure evaluation process performed in step S560 will be described later with reference to Fig. 10.

[0078] In step S570, parameter adjustment unit 122 updates the learning ratio setting based on the initial parameters set in step S500. Here, the learning ratio value set when the most recent learning process was performed in step S550 is updated to a different value. Starting from the lower limit set in step S540, the learning ratio value is gradually increased each time the process of step S570 is performed. Note that the learning ratio update step size is set to 10 times in the example described above, but can be set to any other value.

[0079] In step S580, it is determined whether the learning ratio updated in step S570 exceeds the upper limit of the predetermined search range. If the updated learning ratio does not exceed the upper limit of the search range and is within the search range, the process returns to step S550, and steps S550 to S570 are repeated using the learning ratio setting updated in step S570. This allows the generator 110 and the discriminator 111 to learn for multiple learning ratio values ​​set for each update width within the search range, thereby determining an evaluation index related to the geometric shape difference between the structure of the microstructure material represented by the three-dimensional structure model generated by the learned generator 110 and the actual structure of the microstructure material. On the other hand, if the updated learning ratio exceeds the upper limit of the search range, the loop process of steps S550 to S570 ends and the process proceeds to step S590.

[0080] In step S590, the parameter adjustment unit 122 acquires the learning ratio value that yields the best evaluation index in the generated structure evaluation process of step S560 as the best value of the learning ratio. Here, the evaluation index that yields the best value among the evaluation indexes calculated for each loop process of steps S550 to S570, i.e., the learning ratio value when a three-dimensional structure model is obtained that minimizes the difference between the feature amounts of the cross-sectional image acquired from the three-dimensional structure model represented by the generated structure data 24 and the feature amounts of the cross-sectional image of the actual structure represented by the teacher image data 23, is acquired as the best value of the learning ratio of the generator 110 and the discriminator 111. In this way, the learning ratio is adjusted so as to minimize the difference between the actual cross-sectional image of the microstructure material and the cross-sectional image of the three-dimensional structure model.

[0081] After performing the process of step S590, the parameter adjustment unit 122 updates the hyperparameter data 25 using the acquired best value of the learning rate, and ends the learning rate adjustment process shown in the flowchart of Fig. 7. As a result, the learning rates of the generator 110 and the classifier 111 are adjusted to optimal values.

[0082] 8 and 9 are flowcharts showing the details of the learning process executed in step S120 in FIG. 3, step S230 in FIG. 4, step S350 in FIG. 5, and step S550 in FIG. 7, respectively.

[0083] In step S600, various hyperparameters are set. Here, among the various hyperparameters represented by the hyperparameter data 25, those that have not been set in the previous processing are set to predetermined default values. That is, when the learning processing is performed in step S120 of FIG. 3, each hyperparameter other than the learning rate of the generator 110, which was set to the lower limit of the search range in step S110 or updated in step S140, is set to a predetermined default value. Also, when the learning processing is performed in step S230 of FIG. 4, each hyperparameter other than the learning rate of the generator 110, which was set to the best value in step S210, and the learning rate of the discriminator 111, which was set to the lower limit of the search range in step S220 or updated in step S250, is set to a predetermined default value. When the learning process is performed in step S350 of FIG. 5, all hyperparameters except the learning rates of the generator 110 and the classifier 111, which were set to their best values ​​in steps S310 and S320, and the image counts A, B, and C, which were set to their best values ​​in steps S330, S400, and S430 or updated in steps S370, S390, and S420, are set to predetermined default values. Furthermore, when the learning process is performed in step S550 of FIG. 7, all hyperparameters except the learning rates of the generator 110 and the classifier 111, which were set to their best values ​​in steps S510 and S520, the image counts A, B, and C, which were set to their best values ​​in step S530, and the learning rate, which was set to their lower limit values ​​in the search range in step S540 or updated in step S570, are set to predetermined default values. At this time, the number of learning iterations in the learning process is also set to a default value. By the process of step S600, all hyperparameters have been set, including those that were set before the learning process was performed in the adjustment process of each hyperparameter.

[0084] In step S610, the number of times D the classifier 111 learns and the number of times G the generator 110 learns are determined based on the set hyperparameters. Here, the number of times D and G the learns are determined so as to satisfy the set learning ratio.

[0085] In step S620, the teacher image data 23 stored in the storage unit 2 is read into the learning processing unit 120.

[0086] In step S630, the value of the variable d for counting the number of times the classifier 111 has learned is set to an initial value of 0.

[0087] In step S640, based on the value of the number of images A among the set hyperparameters, A cross-sectional images of a predetermined size are randomly cut out as the above-mentioned first training images from the multiple training images represented by the training image data 23 read in step S620. Then, the A cut-out first training images are input to the classifier 111, and AI calculations are performed by the classifier 111. Note that the number A of first training images at this time is preferably adjusted to 50 or more in the above-mentioned training image number adjustment process.

[0088] In step S650, structure generation is performed by the generator 110. Here, a three-dimensional structure model of a microstructure material according to predetermined conditions is generated by AI calculations performed by the generator 110, and information on the generated three-dimensional structure model is stored in the memory unit 2 as generated structure data 24.

[0089] In step S660, based on the value of the number of images B among the set hyperparameters, B cross-sectional images of a predetermined size are randomly cut out as the above-mentioned second training images from the three-dimensional structure model of the microstructure material generated in step S650. Then, the B cut-out second training images are input to the classifier 111, and AI calculation is performed by the classifier 111. Note that the number B of second training images at this time is preferably adjusted to 150 or more in the above-mentioned training image number adjustment process.

[0090] In step S670, the weights of the classifier 111 represented by the classifier data 22 are updated based on the value of the learning rate of the classifier 111, which is one of the set hyperparameters. Here, the classification results obtained in step S640 for the A first learning images and the classification results obtained in step S660 for the B second learning images are fed back to the weights of the classifier 111 according to the learning rate of the classifier 111 so as to increase the difference between these two results. That is, the weights of the classifier 111 are adjusted and the values ​​of the classifier data 22 are updated so that the classification results in step S640 are more likely to be determined as cross-sectional images of the real object and the classification results in step S660 are less likely to be determined as cross-sectional images of the real object.

[0091] In step S680, the value of the aforementioned variable d is incremented by 1, thereby counting up the variable d.

[0092] In step S690, it is determined whether the value of variable d counted up in step S680 exceeds the number of times D of learning of classifier 111 determined in step S610. If the value of variable d is equal to or less than the number of times D of learning of classifier 111, the process returns to step S640, and the processes of steps S640 to S680 are repeated using classifier data 22 updated in step S670. On the other hand, if the value of variable d exceeds the number of times D of learning, the loop process of steps S640 to S680 ends, and the process proceeds to step S700 in FIG. 9.

[0093] In step S700, the value of the variable g for counting the number of times the generator 110 has learned is set to an initial value of 0.

[0094] In step S710, structure generation is performed by the generator 110. Here, similar to step S650 in Fig. 8, a three-dimensional structure model of a microstructure material according to predetermined conditions is generated by AI calculations performed by the generator 110, and information on the generated three-dimensional structure model is stored in the memory unit 2 as generated structure data 24.

[0095] In step S720, based on the value of the number of images C among the set hyperparameters, C cross-sectional images of a predetermined size are randomly cut out as the second training images from the three-dimensional structure model of the microstructure material generated in step S710. Then, the C cut-out second training images are input to the classifier 111, and AI calculations are performed by the classifier 111. Note that the number C of second training images at this time is preferably adjusted to 150 or more in the above-mentioned training image number adjustment process.

[0096] In step S730, the weights of the generator 110 represented by the generator data 21 are updated based on the value of the learning rate of the generator 110, which is one of the set hyperparameters. Here, the discrimination results of step S720 are fed back to the weights of the generator 110 according to the learning rate of the generator 110 so that the discrimination results obtained for the C second training images in step S720 approach the discrimination result that the images are cross-sectional images of the real object. That is, the weights of the generator 110 are adjusted and the values ​​of the generator data 21 are updated so that the discrimination results of step S720 are more likely to be determined to be cross-sectional images of the real object.

[0097] In step S740, the value of the aforementioned variable g is incremented by 1, thereby counting up the variable g.

[0098] In step S750, it is determined whether the value of variable g counted up in step S740 exceeds the number of times G the generator 110 has learned, which was determined in step S610 of Fig. 8. If the value of variable g is equal to or less than the number of times G the generator 110 has learned, the process returns to step S710, and the processes of steps S710 to S740 are repeated using the generator data 21 updated in step S730. On the other hand, if the value of variable g exceeds the number of times G the generator 110 has learned, the process ends the loop of steps S710 to S740, and proceeds to step S760.

[0099] In step S760, structure generation is performed by the generator 110. Here, similar to step S650 in Fig. 8 and step S710 in Fig. 9, a three-dimensional structure model of the microstructure material according to predetermined conditions is generated by AI calculations performed by the generator 110, and information on the generated three-dimensional structure model is stored in the memory unit 2 as generated structure data 24. As a result, information on the three-dimensional structure model of the microstructure material generated by the generator 110 after learning is recorded in the generated structure data 24. The generated structure data 24 recorded here is used in the generated structure evaluation process executed in step S130 in Fig. 3, step S240 in Fig. 4, step S360 in Fig. 5, and step S560 in Fig. 7, respectively.

[0100] In step S770, it is determined whether learning has been performed the number of times specified as the default value in step S600. Here, the number of times the series of processes in steps S630 to S760 has been executed is compared with the number of times learning has been performed specified as the default value, and if the number of times the series of processes has been executed is less than the number of times learning has been performed, the process returns to step S630 in Figure 8 and the above-mentioned process is repeated. On the other hand, if the number of times the series of processes has been executed reaches the number of times learning has been performed, the learning process shown in the flowcharts of Figures 8 and 9 ends.

[0101] FIG. 10 is a flowchart showing the details of the generated structure evaluation process executed in step S130 of FIG. 3, step S240 of FIG. 4, step S360 of FIG. 5, and step S560 of FIG.

[0102] In step S800, cross-sectional images of the generated structure obtained in the immediately preceding learning process are acquired. Here, the generated structure data 24 stored in the storage unit 2 is read, a large number of cross sections are set for the three-dimensional structure model of the microstructure material represented by the generated structure data 24, and images of a predetermined image size are randomly cut out from each cross section, thereby making it possible to acquire a large number of cross-sectional images from the generated structure obtained in the immediately preceding learning process.

[0103] In step S810, feature quantities of the geometric shape of the generated structure are calculated from the cross-sectional images acquired in step S800. Here, feature quantities of the geometric shape of the generated structure can be calculated by calculating a probability function called, for example, tppf (Two-point probability function) based on a large number of cross-sectional images cut out from the generated structure. Note that tppf is a probability function that represents the distribution of pores in the three-dimensional structure of a porous body, and is expressed as a function of distance r that represents the probability that two points separated by an arbitrary distance r within the three-dimensional structure of the porous body are both pores. This can be calculated using well-known calculation methods.

[0104] In step S820, the feature quantities of the geometric shape of the actual structure are calculated from the actual cross-sectional image of the microstructure material represented by the teacher image data 23. Here, similar to step S810, the feature quantities of the geometric shape of the actual structure can be calculated by calculating, for example, tppf for the actual structure.

[0105] In step S830, the difference between the feature amounts of the geometric shapes of the generated structure and the real structure calculated in steps S810 and S820, respectively, is calculated, and the calculation result is output as the evaluation index described above. Here, for example, by calculating the sum of squares S of the errors of the first to third quartiles of the tppf of the real structure and the generated structure, which is expressed by the following formula (1), it is possible to calculate an evaluation index that represents the difference in the geometric shapes between the real structure and the generated structure.

number

[0106] In formula (1), N represents the image size of the cross-sectional image used when calculating the feature quantities of the geometric shapes of the generated structure and the real structure in steps S810 and S820, and is, for example, 64. q (r) and tppf_synthetic_2d_iso q(r) represents the q-th quartile (q=1 to 3) of the geometric shape features of the generated structure and the real structure calculated in steps S810 and S820, respectively, and is defined, for example, by the following equations (2) and (3), respectively.

number

[0107] In equations (2) and (3), J original and J. synthetic represents a set of cross-sectional images of a predetermined size randomly cut out from the real structure and the generated structure. q [ ] represents the q-th quartile of the data group in the brackets, and tppf_2d(r;I,dir) represents the tppf in the dir (v: vertical, h: horizontal) direction of the two-dimensional cross-sectional image I.

[0108] After the evaluation index calculated in step S830 is output, the value of the evaluation index is stored as evaluation index data 26 in the storage unit 2 together with the values ​​of each hyperparameter set in the immediately preceding learning process. As a result, the value of the calculated evaluation index for each combination of hyperparameter values ​​for which the learning process was performed is recorded in the evaluation index data 26. After step S830 is performed, the generative structure evaluation process shown in the flowchart of FIG. 10 ends.

[0109] According to the embodiment of the present invention described above, the following advantageous effects are achieved.

[0110] (1) A design support method for microstructured materials using a design support device 100 supports the design of microstructured materials by adjusting hyperparameters set during learning by an AI calculation unit 11, which is an AI that generates a three-dimensional structure model from a two-dimensional image of the microstructured material. In this design support method, the design support device 100, which is a computer, calculates an evaluation index related to the geometric shape difference between the actual structure of the microstructured material and the structure of the model generated by the AI ​​calculation unit 11, and adjusts the hyperparameters based on this evaluation index (steps S10, S20, S30, S50). This allows for efficient optimization of the hyperparameters set during learning by an AI that generates a three-dimensional structure model from a two-dimensional image of the microstructured material.

[0111] (2) The above evaluation index represents the magnitude of the difference between the actual cross-sectional image (first two-dimensional image) of the microstructure material and the cross-sectional image (second two-dimensional image) of the model generated by the AI ​​calculation unit 11. The design support device 100 adjusts the hyperparameters based on this evaluation index so as to minimize the difference between the first two-dimensional image and the second two-dimensional image. This allows the hyperparameters to be adjusted so as to maximize the learning effect of the AI.

[0112] (3) The AI ​​calculation unit 11 is a GAN having a generator 110 that generates a three-dimensional structure model from a two-dimensional image and a classifier 111 that determines whether a cross-sectional image of the model generated by the generator 110 is a first two-dimensional image or a second two-dimensional image. The hyperparameters include a learning rate that represents the update width of the weights of the generator 110 and the classifier 111, a number of training images that represents the number of images used in training the generator 110 and the classifier 111, and a learning ratio that represents the ratio of the number of training times of the classifier 111 to the number of training times of the generator 110. The design support device 100 adjusts at least one of the learning rate, the number of training images, and the learning ratio in the hyperparameters. At this time, it is preferable to adjust the learning rate (steps S10 and S20), the number of training images (step S30), and the learning ratio (step S50) in this order. This allows for reliable adjustment of each of these hyperparameters required for training the AI ​​to generate a three-dimensional structural model from a two-dimensional image of a microstructured material.

[0113] (4) Adjusting the learning rate: The learning rate of the generator 110 and the discriminator 111 is set to 1 × 10 -7 From 1×10 -1 (Steps S100, S200) In this way, the learning rates of the generator 110 and the classifier 111 can be efficiently adjusted within appropriate ranges.

[0114] (5) Images used for training by the generator 110 or the classifier 111 include first training images cut out from a first two-dimensional image and second training images cut out from a second two-dimensional image. Adjusting the number of training images involves adjusting the number of training images for the classifier 111 so that the number of first training images (number of images A) is 50 or more and the number of second training images (number of images B) is 150 or more, and adjusting the number of training images for the generator 110 so that the number of second training images (number of images C) is 150 or more (step S300). Here, it is preferable to adjust the numbers of training images so that A, B, and C satisfy A≦B≦C, where A is the number of first training images used for training by the classifier 111, B is the number of second training images used for training by the classifier 111, and C is the number of second training images used for training by the generator 110 (step S340). In this way, the number of training images used for training the generator 110 and the classifier 111 can be adjusted efficiently within an appropriate range.

[0115] (6) In the adjustment of the learning ratio, the learning ratio can be adjusted within a range of 0.01 to 100 (step S500). In this way, the learning ratio of the classifier 111 to the generator 110 can be efficiently adjusted within an appropriate range.

[0116] In the embodiment described above, the hyperparameters of the learning rate of the generator 110, the learning rate of the classifier 111, the number of learning images of the generator 110 and the classifier 111, and the learning ratio of the generator 110 and the classifier 111 are adjusted by performing the processes of steps S10, S20, S30, and S50 in the flowchart of Fig. 2, but it is not necessary to adjust all of these hyperparameters, and some of the processes may be omitted. Any process that adjusts at least one of these hyperparameters by performing at least one of the processes of steps S10, S20, S30, and S50 is within the scope of the present invention.

[0117] The present invention is not limited to the above-described embodiment, and can be implemented using any components without departing from the spirit of the present invention.

[0118] The above-described embodiments and modifications are merely examples, and the present invention is not limited to these details as long as the features of the invention are not impaired. Furthermore, although various embodiments and modifications have been described above, the present invention is not limited to these details. Other aspects that can be considered within the scope of the technical idea of ​​the present invention are also included within the scope of the present invention. [Explanation of symbols]

[0119] 1...control unit, 2...storage unit, 3...memory, 4...operation input device, 5...display device, 6...network device, 7...bus, 11...AI calculation unit, 110...generator, 111...discriminator, 12...AI learning unit, 120...learning processing unit, 121...evaluation index calculation unit, 122...parameter adjustment unit, 21...generator data, 22...discriminator data, 23...teacher image data, 24...generated structure data, 25...hyperparameter data, 26...evaluation index data, 100...design support device

Claims

1. A method for supporting the design of a microstructured material by adjusting hyperparameters set during learning of an AI that generates a three-dimensional structure model from a two-dimensional image of the microstructured material, comprising: Calculating an evaluation index relating to the geometric shape difference between the actual structure of the microstructure material and the structure of the model generated by the AI ​​using a computer; The method for supporting design of microstructured materials includes adjusting the hyperparameters based on the evaluation index by the computer.

2. 2. The method for designing a microstructured material according to claim 1, the evaluation index represents a magnitude of a difference between a first two-dimensional image, which is a cross-sectional image of the actual microstructure material, and a second two-dimensional image, which is a cross-sectional image of the model; A design support method for microstructured materials, which adjusts the hyperparameters so as to minimize the difference based on the evaluation index.

3. 3. The method for designing a microstructured material according to claim 2, The AI ​​is a GAN having a generator that generates the model from the two-dimensional image and a discriminator that discriminates whether a cross-sectional image of the model generated by the generator is the first two-dimensional image or the second two-dimensional image, the hyperparameters include a learning rate representing an update width of each weight of the generator and the classifier, a number of training images representing the number of images used for training the generator and the classifier, respectively, and a learning ratio representing a ratio of the number of times the classifier has been trained to the number of times the generator has been trained; A design support method for microstructured materials, which adjusts at least one of the learning rate, the number of learning images, and the learning ratio.

4. 4. The method for designing a microstructured material according to claim 3, A design support method for microstructured materials, wherein the learning rate, the number of learning images, and the learning ratio are adjusted in this order.

5. 5. The method for designing a microstructured material according to claim 3, further comprising: The learning rate of the generator and the classifier is set to 1×10 －７ From 1 x 10 －１ A design support method for microstructured materials, adjusting each within the range.

6. 5. The method for designing a microstructured material according to claim 3, further comprising: the images used for training the generator or the classifier include a first training image cut out from the first two-dimensional image and a second training image cut out from the second two-dimensional image; adjusting the number of learning images for the classifier so that the number of first learning images is 50 or more and the number of second learning images is 150 or more; A design support method for microstructured materials, comprising adjusting the number of training images of the generator so that the number of second training images is 150 or more.

7. 7. The method for designing a microstructured material according to claim 6, A design support method for microstructured materials, wherein when the number of first training images used for training the classifier is A, the number of second training images used for training the classifier is B, and the number of second training images used for training the generator is C, the numbers of training images are adjusted so that A, B, and C satisfy A≦B≦C.

8. 5. The method for designing a microstructured material according to claim 3, further comprising: The learning ratio is adjusted within a range of 0.01 to 100.

9. An apparatus for assisting in the design of a microstructure material by adjusting hyperparameters set during learning of an AI that generates a three-dimensional structure model from a two-dimensional image of the microstructure material, an evaluation index calculation unit that calculates an evaluation index relating to a geometric shape difference between the actual structure of the microstructure material and the structure of the model generated by the AI; a parameter adjustment unit that adjusts the hyperparameters based on the evaluation index calculated by the evaluation index calculation unit.

10. 10. The microstructure material design support system according to claim 9, the evaluation index represents a magnitude of a difference between a first two-dimensional image, which is a cross-sectional image of the actual microstructure material, and a second two-dimensional image, which is a cross-sectional image of the model; The parameter adjustment unit adjusts the hyperparameters so as to minimize the difference based on the evaluation index.

11. 11. The microstructure material design support system according to claim 10, The AI ​​is a GAN having a generator that generates the model from the two-dimensional image and a discriminator that discriminates whether a cross-sectional image of the model generated by the generator is the first two-dimensional image or the second two-dimensional image, the hyperparameters include a learning rate representing an update width of each weight of the generator and the classifier, a number of training images representing the number of images used for training the generator and the classifier, respectively, and a learning ratio representing a ratio of the number of times the classifier has been trained to the number of times the generator has been trained; The parameter adjustment unit adjusts at least one of the learning rate, the number of learning images, and the learning ratio.

12. 12. The microstructure material design support system according to claim 11, The parameter adjustment unit adjusts the learning rate, the number of learning images, and the learning ratio in this order.

13. 13. The computer-aided design system for microstructured materials according to claim 11 or 12, The parameter adjustment unit adjusts the learning rate of the generator and the classifier to 1×10 －７ From 1 x 10 －１ A design support device for microstructured materials that adjusts each within the range.

14. 13. The computer-aided design system for microstructured materials according to claim 11 or 12, the images used for training the generator or the classifier include a first training image cut out from the first two-dimensional image and a second training image cut out from the second two-dimensional image; The parameter adjustment unit adjusting the number of learning images for the classifier so that the number of first learning images is 50 or more and the number of second learning images is 150 or more; The design support device for microstructured materials adjusts the number of training images of the generator so that the number of second training images is 150 or more.

15. 15. The microstructure material design support system according to claim 14, The parameter adjustment unit adjusts the number of training images so that A, B, and C satisfy A≦B≦C, where A is the number of first training images used to train the classifier, B is the number of second training images used to train the classifier, and C is the number of second training images used to train the generator.

16. 13. The computer-aided design system for microstructured materials according to claim 11 or 12, The parameter adjustment unit adjusts the learning ratio within a range of 0.01 to 100.

Images