Data processing device for composite materials

Abstract

The present invention provides a data processing apparatus for composite materials, which can improve the prediction accuracy of various data related to composite materials manufactured using multiple compounding agents. The data processing apparatus (1) for composite materials that predicts various data related to composite materials manufactured using multiple compounding agents includes: a first model generation processing unit (23) that generates a first learned model (32) representing the correlation between a composite material image after image processing and physical properties as one of the various data; a virtual image generation processing unit (24) that generates multiple virtual images that mimic the composite material image, i.e., virtual composite material images; and a physical property prediction processing unit (25) that uses the first learned model (32) and multiple virtual composite material images to predict physical properties.

Description

Technical Field

[0001] This invention relates to AI prediction technology for predicting various data (physical properties, manufacturing conditions, etc.) of composite materials used as sheathing materials for wires, etc., through AI (artificial intelligence), and to a data processing apparatus for composite materials. Background Technology

[0002] Previously, regarding composite materials manufactured by mixing multiple compounding agents (polymers, fillers, etc.) in a mixing mill, methods using SEM images of the composite material obtained by scanning electron microscopy (SEM) are known, for example, as a method for predicting physical properties through AI (artificial intelligence). However, in cases where multiple polymers are mixed, this method sometimes fails to distinguish the dispersion states of the various polymers from each other using SEM images.

[0003] As another method, it is known to use SPM images of composite materials obtained by scanning probe microscopy (SPM) (for example, see Patent Document 1). In this method, for example, even when multiple polymers are mixed with fillers, not only the dispersion state of the polymers and fillers, but also the dispersion state of the multiple polymers relative to each other can be distinguished by the SPM image. Therefore, compared with the method using SEM images, it has the characteristic of high prediction accuracy of the physical properties of the composite material.

[0004] As described in the background section above, there are various AI prediction techniques for various data related to composite materials. However, there is still room for improvement in AI prediction techniques for various data (physical properties, manufacturing conditions, etc.) related to composite materials manufactured using multiple compounding agents (e.g., multiple polymers).

[0005] Patent Document 1: Chinese Patent Application Publication No. 113408188 Summary of the Invention

[0006] Therefore, the present invention has been made in view of the above circumstances, and its object is to provide a data processing apparatus for composite materials that can improve the prediction accuracy of various data of composite materials manufactured using a variety of compounding agents.

[0007] To address the aforementioned issues, this invention provides a data processing apparatus for composite materials that predicts various data related to composite materials manufactured using multiple compounding agents. The apparatus comprises: a first model generation processing unit that generates a first learned model representing the correlation between a composite material image obtained by imagerizing the composite material and a physical property, which is one of the various data; a virtual image generation processing unit that generates multiple virtual images, i.e., virtual composite material images, that mimic the composite material image; and a physical property prediction processing unit that uses the first learned model and the multiple virtual composite material images to predict the physical property.

[0008] According to the present invention, a data processing apparatus for composite materials is provided that can improve the prediction accuracy of various data related to composite materials manufactured using a variety of compounding agents. Attached Figure Description

[0009] Figure 1 This is a schematic structural diagram of a data processing device for composite materials according to one embodiment of the present invention.

[0010] Figure 2 This is a diagram representing an example of an SPM (Secondary Amplitude Image).

[0011] Figure 3 This is a diagram representing an example of a virtual SPM image.

[0012] Figure 4 This is a flowchart illustrating the control process of a data processing device for composite materials.

[0013] Figure 5 This is a flowchart of data acquisition and processing.

[0014] Figure 6 It is a flowchart for learning how to append data.

[0015] Figure 7 This is the flowchart for the first model generation process.

[0016] Figure 8 This is a flowchart of the virtual image generation and processing.

[0017] Figure 9 This is a flowchart of the physical property prediction process.

[0018] Figure 10 This is a flowchart of virtual image extraction and processing.

[0019] Figure 11 This is a flowchart of the second model generation process.

[0020] Figure 12 This is a flowchart of the manufacturing condition prediction process.

[0021] Figure 13 This is an example of a screen displaying prediction results.

[0022] Figure 14 It is a graph representing the predicted results of physical properties.

[0023] Figure 15 (a) is an example of a second-amplitude image, (b) is an example of a first-phase difference image, (c) is an example of a second-phase difference image, (d) is an example of an adhesive force mapping image, and (e) is an example of an elastic force mapping image.

[0024] Figure 16 This is a graph representing the calculated average absolute percentage error when using each SPM image.

[0025] Figure 17 This is a diagram representing the input and output of the hyperparameter optimization processing unit;

[0026] Figure 18 This is a diagram illustrating hyperparameter optimization.

[0027] Figure 19 This is a flowchart of the hyperparameter optimization process;

[0028] Figure 20 This is an example graph showing the relationship between the number of rounds and batch size and the average value of MAPE. Detailed Implementation

[0029] [Implementation Method]

[0030] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.

[0031] Traditionally, the development of materials and raw materials has relied heavily on the experience and knowledge of technicians and researchers. However, due to the development of predictive technologies based on AI (Artificial Intelligence), including machine learning and deep learning (AI prediction technology), this reliance has decreased, and efficiency in materials and raw material development has improved dramatically in recent years. Furthermore, the "AI prediction technology" used in this specification is specifically for materials and raw material development and is also known as materials informatics (MI).

[0032] In this technological trend, the inventors have conducted repeated and in-depth research to effectively utilize AI prediction technology and improve its prediction accuracy in the development of composite materials manufactured using multiple compounding agents (polymers, fillers, etc.). In this in-depth research, the inventors focused on improving the prediction accuracy of "various data (physical properties, manufacturing conditions, etc.) of composite materials manufactured using multiple compounding agents (e.g., multiple polymers)," specifically on the importance of "data on the dispersion state of the multiple compounding agents constituting the composite material (compounding agent dispersion data)," and more specifically, "composite material images (e.g., SPM images, etc.) reflecting information on the dispersion state of multiple compounding agents," and thus completed this invention.

[0033] Therefore, the present invention has been made in view of the above circumstances, and its object is to provide a data processing apparatus or method for composite materials that can improve the prediction accuracy of "various data of composite materials manufactured using multiple compounding agents".

[0034] In addition, the aim is to provide a data processing device or method for composite materials that can improve the prediction accuracy of "the dispersion state of various compounding agents constituting composite materials with good physical properties".

[0035] Furthermore, the present invention aims to provide a data processing apparatus or method for composite materials that can improve the prediction accuracy of "manufacturing conditions for composite materials with good physical properties".

[0036] In addition, the objective is to provide a composite material data processing device or method that improves the visualization of prediction results for various data of composite materials manufactured using multiple compounding agents.

[0037] Figure 1 This is a schematic structural diagram of the composite material data processing device 1 of this embodiment. The composite material data processing device 1 is a device for processing composite material data. Specifically, it is a device with an AI prediction function, which uses AI (artificial intelligence) to predict various data (physical properties, manufacturing conditions, etc.) related to composite materials manufactured using various compounding agents (polymers, fillers, etc.).

[0038] Furthermore, in this embodiment, the composite material is, for example, a coating material for electrical wires, and is a material manufactured by at least mixing a polymer and a filler using a mixing machine. Additionally, the various data related to the composite material that becomes the prediction target in the AI ​​prediction function include at least data related to "physical properties" and "manufacturing conditions." More specifically, the former, "physical properties," includes, for example, elongation and tensile strength, while the latter, "manufacturing conditions," includes, for example, the rotational speed (screw speed), mixing time, and mixing temperature of the mixing machine used for mixing. Regarding this description, other elements may be included as long as they do not depart from the spirit of the invention.

[0039] The composite material data processing device 1 has at least a control unit 2 and a storage unit 3. Specifically, it is composed of a computer such as a personal computer or a server device, and has arithmetic elements such as a CPU, memory such as RAM and ROM, storage devices such as hard disks, various multiple ports (USB port, HDMI port, Ethernet port, DisplayPort, etc.), wireless / wired LAN card and other communication interfaces, software, etc.

[0040] The control unit 2 comprises a data acquisition and processing unit 21, a learning data addition processing unit 22, a first model generation processing unit 23, a virtual image generation processing unit 24, a property prediction processing unit 25, a virtual image extraction processing unit 26, a second model generation processing unit 27, a manufacturing condition prediction processing unit 28, and a prediction result prompting processing unit 29. These data acquisition and processing units 21, learning data addition processing unit 22, first model generation processing unit 23, virtual image generation processing unit 24, property prediction processing unit 25, virtual image extraction processing unit 26, second model generation processing unit 27, manufacturing condition prediction processing unit 28, and prediction result prompting processing unit 29 are implemented by appropriately combining the aforementioned computing elements, memory, storage device, communication interface, software, etc. Details of each unit will be described later. The storage unit 3 is implemented through a predetermined storage area of ​​a memory or storage device.

[0041] Furthermore, a display 4 and an input device 5 are connected to the composite material data processing device 1. The display 4 is, for example, a liquid crystal display, and the input device 5 is, for example, a keyboard or a mouse. In addition, the display 4 and the input device 5 in this embodiment are part of the composite material data processing device 1.

[0042] Alternatively, a touch panel can be used to form the display 4, which can also serve as the input display device for the input device 5.

[0043] This configuration improves convenience, and as a result, it contributes to increased efficiency in the development of materials and raw materials (the development of composite materials).

[0044] Alternatively, the display 4 and input device 5 (including input display devices such as touch panels) can be configured separately from the composite material data processing device 1, enabling them to communicate with the composite material data processing device 1 via wireless LAN or the like. That is, the display 4 and / or input device 5 can be configured as portable terminals such as tablet computers or smartphones.

[0045] This configuration makes it easy to transport the portable terminal to the manufacturing site, pilot plant, etc., thus helping to improve the efficiency of material and raw material development (composite material development).

[0046] Additionally, an example of the content displayed on monitor 4 (see...) Figure 13 (As will be described later, but based on the development of composite materials, it is possible to efficiently obtain prediction results of various data for composite materials manufactured using multiple compounding agents. That is, the composite material data processing device 1 of this embodiment improves the visualization of prediction results of various data for composite materials.)

[0047] (Data Acquisition and Processing Department 21)

[0048] The data acquisition and processing unit 21 performs data acquisition processing on data acquired from an external device for learning purposes, 31. Figure 4 Specifically, the data acquisition and processing unit 21 acquires the data required for machine learning from an external device such as a server device via a network, and stores the acquired data as learning data 31 in the storage unit 3.

[0049] Furthermore, the learning data 31 is the data used as training data for machine learning in the first model generation process and the second model generation process described later. Moreover, regarding the specific method of obtaining the data, other methods may be used as long as they do not depart from the spirit of the invention; for example, the data required for machine learning may be obtained by inputting data into the composite material data processing device 1 via a medium such as a USB memory.

[0050] In this embodiment, the data acquisition processing unit 21 acquires, for example, the data described in (a) to (c) below:

[0051] (a) SPM image of the composite material

[0052] (b) Physical properties (property values) of composite materials

[0053] (c) Manufacturing conditions of composite materials,

[0054] The data from (a) to (c) are then linked together and stored in storage unit 3 as learning data 31.

[0055] (a) is an example of a composite material image, obtained by imaging the composite material (the cross-section or surface of the composite material) using a scanning probe microscope (SPM). Preferably, the SPM image used is a first-order phase difference image measured at a resonant frequency of 100 kHz or higher and 200 kHz or lower, a second-order phase difference image measured at a resonant frequency of 700 kHz or higher and 900 kHz or lower, or a second-order amplitude image measured at a resonant frequency of 700 kHz or higher and 900 kHz or lower. Furthermore, Figure 2 An example of an SPM image is shown. Figure 2 The SPM image shown is a secondary amplitude image measured at a resonant frequency above 700 kHz and below 900 kHz.

[0056] More specifically, in primary phase difference images, secondary phase difference images, or secondary amplitude images, contrast between complexing agents is easily revealed, so a correlation can easily be obtained between the brightness of each pixel in the image and the complexing agent. That is, by using primary phase difference images, secondary phase difference images, or secondary amplitude images, a great deal of information related to the dispersion state of the complexing agent can be reflected in the image.

[0057] By constructing it in this way, for example in composite materials made using multiple compounding agents (e.g., multiple polymers, etc.), it is easier to distinguish the compounding agents from each other, and thus it is easier to grasp the "dispersion state of the multiple compounding agents constituting the composite material" as one of the data for composite materials. As a result, it is possible to improve the prediction accuracy of various data (physical properties, manufacturing conditions, etc.) related to the composite material.

[0058] Furthermore, regarding the SPM image input from an external device to the composite material data processing device 1, if the input data is larger than the size (width and height) of the data used for learning 31, a predetermined size can be cut out from the SPM image related to the input data to use as an SPM image (image cropping function). In this case, for example, the central portion of the SPM image can be cropped, and the information of the cropping position can be preset.

[0059] By configuring it in this way, namely by having an image cropping function, the constraints on the data related to the SPM image input from an external device are reduced, thus improving the convenience of the composite material data processing device 1. Furthermore, as a result, it also helps to improve the prediction accuracy of various data related to composite materials (physical properties, manufacturing conditions, etc.).

[0060] The physical properties of the composite material in (b) include one or more of elongation and tensile strength, but may also include both of these two properties. Furthermore, other elements may be included as long as they do not depart from the spirit of the invention. For example, deterioration characteristics of elongation and / or tensile strength may also be included (e.g., elongation, tensile strength, etc. after a heat load test).

[0061] The manufacturing conditions for the composite material in (c) are at least one of the following: the rotational speed (screw speed) of the mixer used in the mixing process, the mixing time, and the mixing temperature; or all of these three manufacturing conditions.

[0062] Furthermore, regarding these three manufacturing conditions, the inventors believe they have the greatest impact on the physical properties of the composite material. However, other conditions may also be used as long as they do not depart from the spirit of the invention. For example, manufacturing conditions other than the speed of the mixer, mixing time, or mixing temperature may be used. Conversely, manufacturing conditions other than the speed of the mixer, mixing time, or mixing temperature may also be used.

[0063] (Study Data Addition Processing Department 22)

[0064] When the amount of data in the learning data 31 is small, the learning data appending processing unit 22 performs learning data appending processing to add data to the learning data 31. Figure 4 (S2). Furthermore, the data appending processing unit 22 is not necessary and can be omitted.

[0065] The learning data addition processing unit 22 performs learning data addition processing, for example, when the number of data in the learning data 31 is less than the predetermined number, or when the user determines that the number of data in the learning data 31 is less than the predetermined number.

[0066] Furthermore, the learning data appending processing unit 22 generates a new SPM image, for example, by processing the SPM image (referred to as the original image) contained in the learning data 31. More specifically, it generates a new SPM image by performing at least one of the following processing: inversion processing or rotation processing on the original image. In the inversion processing, the original image is inverted vertically or horizontally to generate a new SPM image. In the rotation processing, the original image is rotated clockwise. , or A new SPM image is generated. Alternatively, both inversion and rotation processing can be performed to generate a new SPM image. The new SPM image generated by the learning data appending processing unit 22 is appended to the learning data 31 with the same physical properties and manufacturing conditions as the original image.

[0067] Furthermore, the rotation angle in rotation processing is not limited to , or Furthermore, in cases where the material properties are anisotropic, it is preferable that the rotation angle in the rotation process does not include... , And only Here, "the properties of matter are anisotropic" means that the properties of matter differ depending on the direction.

[0068] Furthermore, the processing performed in the learning data appending processing unit 22 can be preset by the user or randomly selected. Additionally, the method for selecting the SPM image that becomes the original image can be randomly selected from the learning data 31 or appropriately selected by the user.

[0069] By constructing it in this way, even when the number of data points in the learning data 31 is small, the number of data points can be increased, thus helping to improve the prediction accuracy of the physical properties of composite materials.

[0070] In addition, during the data acquisition processing of the data acquisition processing unit 21, if the input data is an SPM image of a larger size than the size used as learning data 31, the learning data addition processing unit 22 can be configured to appropriately change the cut-out portion from the input SPM image to increase the amount of data in the learning data 31.

[0071] This configuration allows for a further increase in the amount of data used in the learning process 31, which in turn helps to improve the accuracy of predictions regarding the physical properties of composite materials.

[0072] (First Model Generation and Processing Unit 23)

[0073] The first model generation processing unit 23 performs a first model generation process that uses the learning data 31 as training data to perform machine learning on the correlation between the SPM image and the physical properties (physical property values) of the composite material to generate a first learned model 32. Figure 4 (S3). Furthermore, the first learned model 32 generated in the first model generation process is stored in the storage unit 3 as a learned model representing the relationship between the SPM image and the physical property value (i.e., the correlation between the SPM image and the physical properties of the composite material).

[0074] Furthermore, other elements may be used as long as they do not depart from the spirit of the present invention. For example, the machine learning algorithm used in the first model generation process may be configured as a first learned model 32, which is a combination of a learned convolutional neural network that extracts a predetermined number of features (e.g., more than 1,000 features) from the SPM image and a regression model that has learned the correlation between the features extracted by the learned convolutional neural network and the physical properties of the composite material.

[0075] (Virtual Image Generation and Processing Unit 24)

[0076] The virtual image generation processing unit 24 performs virtual image generation processing to generate multiple virtual images that mimic SPM images, i.e., virtual SPM images (an example of composite material images). Figure 4 (S4). Specifically, the virtual image generation processing unit 24 uses the SPM images contained in the learning data 31 as training data to generate a virtual SPM image, i.e., a virtual SPM image, that is similar to the SPM images used as training data. The SPM images used as training data can be randomly selected and extracted from the learning data 31, or they can be appropriately selected and extracted by the user. In addition, the virtual image generation processing unit 24 can generate virtual SPM images using so-called image generation AI. Here, for example, the virtual image generation processing unit 24 is configured to use an image generation AI capable of generating 1024 images. A method called PG-GAN (Processive Growing-Generative Adversarial Network) is used to generate virtual SPM images from high-resolution images of 1024 pixels.

[0077] This configuration allows for the easy and rapid generation of multiple virtual SPM images (more virtual SPM images), thereby improving the prediction accuracy of "various data related to composite materials manufactured using multiple compounding agents".

[0078] Furthermore, the number of virtual SPM images generated by the virtual image generation processing unit 24 is not particularly limited, but it is preferable to configure the virtual image generation processing unit 24 to generate at least 5,000, more preferably 10,000 or more virtual SPM images. Here, the virtual image generation processing unit 24 is configured to generate approximately 20,000 virtual SPM images. The virtual SPM images generated in the virtual image generation process are stored in the storage unit 3 as virtual SPM image data 33. Figure 3 This represents an example of a generated virtual SPM image.

[0079] (Physical Property Prediction and Processing Department 25)

[0080] The property prediction processing unit 25 performs property prediction processing to predict the properties of the composite material using the first learned model 32 generated in the first model generation processing and multiple virtual SPM images generated in the virtual image generation processing. Figure 4(S5). More specifically, the property prediction processing unit 25 applies multiple (partial or all) virtual SPM images stored as virtual SPM image data 33 in the storage unit 3 to the first learned model 32 to predict the properties of the composite material. Furthermore, the predicted properties of the composite material are stored in the storage unit 3 as predicted property data 34.

[0081] This configuration allows for improved prediction accuracy of various data related to composite materials manufactured using multiple compounding agents. Furthermore, in this embodiment, only property prediction using virtual SPM images is performed during property prediction processing; however, property prediction using actual SPM images is also possible.

[0082] (Virtual Image Extraction and Processing Unit 26)

[0083] The virtual image extraction and processing unit 26 performs virtual image extraction processing. Figure 4 (S6) The virtual image extraction process extracts virtual SPM images that are correlated with physical properties that are higher than those of the SPM images contained in the learning data 31 (i.e., physical properties predicted by the physical property prediction processing unit 25 and which are good physical properties) as extracted virtual SPM images. Specifically, after extracting the best physical property value (optimal physical property value) from the data contained in the learning data 31, the virtual image extraction processing unit 26 extracts virtual SPM images that have physical property values ​​higher than the optimal physical property value based on the predicted physical property data 34. The extracted virtual SPM images (extracted virtual SPM images) are stored in the storage unit 3 as extracted virtual SPM image data 35.

[0084] This configuration allows for improved prediction accuracy of the dispersion state of various compounding agents that constitute composite materials with good physical properties.

[0085] Furthermore, regarding the extraction criteria for the virtual SPM images in the virtual image extraction processing unit 26, the optimal physical property value in the learning data 31 is used as the basis here. However, the extraction criteria for the virtual SPM images can be appropriately changed as long as it does not depart from the spirit of the present invention. For example, the virtual image extraction processing unit 26 may be configured to: refer to the predicted physical property (physical property value) predicted using the first learned model 32 and multiple virtual SPM images, i.e., the predicted physical property data 34, and extract a predetermined number (which can be changed by the setting change unit) of physical property values ​​in descending order of the predicted physical property values ​​contained in the predicted physical property data 34, and extract virtual SPM images that are the extracted physical property values.

[0086] By configuring it in this way, that is, by setting a setting change unit having a setting of the number of extracted property values ​​contained in the change prediction property data 34, the control load in the data processing of the composite material data processing device 1 can be reduced or adjusted.

[0087] (Second Model Generation Processing Unit 27)

[0088] The second model generation processing unit 27 performs a second model generation process that uses the learning data 31 as training data to perform machine learning on the correlation between the SPM image and the manufacturing conditions of the composite material to generate a second learned model 36. Figure 4 (S7). Furthermore, the second learned model 36 generated in the second model generation process is stored in the storage unit 3 as a learned model representing the relationship between the SPM image and the manufacturing conditions (i.e., the correlation between the SPM image and the manufacturing conditions of the composite material). In addition, the first learned model 32 described above is a model that predicts the "physical properties of the composite material" from the SPM image, while the second learned model 36 is a model that predicts the "manufacturing conditions of the composite material" from the SPM image.

[0089] Furthermore, other elements may be used as long as they do not depart from the spirit of the present invention. For example, the second model generation processing unit 27 may also adopt a structure that is substantially the same as that of the first model generation processing unit 23 described above. That is, it may also be configured as a second learned model 36 consisting of a learned convolutional neural network that extracts predetermined feature quantities (e.g., more than 1,000 feature quantities) from the SPM image and a regression model obtained by learning the correlation between the feature quantities extracted by the learned convolutional neural network and the physical properties of the composite material.

[0090] Furthermore, as mentioned above, one or more of the following can be used as manufacturing conditions: the speed of the mixing mill, the mixing time, and the mixing temperature. In the case of predicting multiple manufacturing conditions, a second learned model 36 can be generated according to each manufacturing condition.

[0091] (Manufacturing Condition Prediction and Processing Department 28)

[0092] The manufacturing condition prediction processing unit 28 performs manufacturing condition prediction processing by utilizing the second learned model 36 generated in the second model generation processing and the virtual SPM image extracted by the virtual image extraction processing unit 26 to predict manufacturing conditions. Figure 4 (S8). More specifically, the manufacturing condition prediction processing unit 28 applies multiple (partial or complete) virtual SPM images stored as extracted virtual SPM image data 35 to the second learned model 36 to predict the manufacturing conditions of the composite material. Furthermore, the predicted manufacturing conditions of the composite material are stored in the storage unit 3 as predicted manufacturing condition data 37.

[0093] This configuration enables the prediction of manufacturing conditions for virtual SPM images that are presumed to have good physical properties. In other words, it improves the accuracy of predictions regarding "manufacturing conditions that result in composite materials with good physical properties."

[0094] Furthermore, regarding the aforementioned manufacturing condition prediction processing unit 28, it uses a virtual SPM image extracted by the virtual image extraction processing unit 26 to predict the manufacturing conditions of the composite material. However, it is also possible to use any SPM image input from an external device to predict the manufacturing conditions of the composite material. That is, by using both the method of predicting the manufacturing conditions of the composite material based on a "virtual SPM image" and the method of predicting the manufacturing conditions of the composite material based on an "arbitrary SPM image," as a result, it is possible to provide clues for discovering more appropriate manufacturing conditions in material / raw material development (composite material development) for deeper research.

[0095] (Prediction results are displayed to the processing department 29)

[0096] The prediction result prompting processing unit 29 performs prediction result prompting processing on the prediction results of the material property prediction processing unit 25, the extraction results of the virtual image extraction processing unit 26, the prediction results of the manufacturing condition prediction processing unit 28, etc. Figure 4 (S9). The prediction result prompting processing unit 29 prompts the user with the prediction result, for example, by displaying the prediction result on the display 4. For example, the prediction result prompting processing unit 29 can be configured to display a virtual SPM image with good prediction results for the material properties, and display the virtual SPM image, the predicted material properties (material property values), and the predicted manufacturing conditions on the display 4. The specific display screen displayed on the display 4 during the prediction result prompting processing will be described later.

[0097] (Control process)

[0098] Figure 4 This is a flowchart illustrating the control flow of the data processing device 1 for composite materials.

[0099] exist Figure 4 In the control flow shown, step S1 is the data acquisition processing performed by the data acquisition processing unit 21. In the data acquisition processing S1, as... Figure 5 As shown, in step S11, data on the SPM image, physical properties, and manufacturing conditions are acquired. In step S12, the acquired data is stored in the storage unit 3 as learning data 31 (or appended to the learning data 31). Afterwards, the process returns to... Figure 4 Step S2. Furthermore, the data acquisition process S1 does not need to be performed before the prediction of physical properties, etc. For example, it can be performed each time new data is added, or it can be performed at predetermined intervals (e.g., every few days).

[0100] Next, step S2 is the learning data appending processing performed by the learning data appending processing unit 22. In the learning data appending processing S2, as... Figure 6 As shown, in step S21, it is determined whether the number of data points in the learning data 31 is less than a preset data point threshold. If the determination in step S21 is negative (N), the process returns to the previous step and proceeds to the next step. Figure 4 Step S3.

[0101] If the determination is "yes" in step S21, after randomly extracting an SPM image from the learning data 31 in step S22, the extracted SPM image (original image) is appropriately inverted / rotated in step S23 to generate a new SPM image. Furthermore, while an SPM image is randomly extracted from the learning data 31 as the original image, it can also be configured so that the user can select the original image. Additionally, the processing performed on the original image in step S23 can be preset by the user or randomly selected. Furthermore, multiple new SPM images with different processing can be generated in step S23. Then, in step S24, the new SPM image generated in step S23 is appended to the learning data 31 in association with data on material properties and manufacturing conditions (the same data as the original image). Then, the process returns to step S21.

[0102] Next, step S3 is the first model generation process performed by the first model generation processing unit 23. In the first model generation process S3, as... Figure 7 As shown, in step S31, the learning data 31 is used as training data to generate a first learned model 32 representing the correlation between the SPM image and the material properties. Then, in step S32, the first learned model 32 generated in step S31 is stored in the storage unit 3. After that, the process returns to... Figure 4 Step S4.

[0103] Next, step S4 is the virtual image generation process performed by the virtual image generation processing unit 24. In the virtual image generation process S4, as... Figure 8As shown, in step S41, 1 is substituted into the variable i representing the number of iterations as an initial value. Then, in step S42, SPM images are randomly extracted from the training data 31. In step S43, using PG-GAN, the extracted SPM images are used as training data to generate multiple (e.g., 100) virtual SPM images similar to the training data. Then, in step S44, the multiple virtual SPM images generated in step S43 are stored in the storage unit 3 as virtual SPM image data 33 (or appended to the virtual SPM image data 33). Then, in step S45, it is determined whether variable i is greater than or equal to 200. If the determination is "no" in step S45, the process returns to step S42. If the determination is "yes" in step S45, the process returns to the previous step. Figure 4 Step S5. Furthermore, the number of loops is set to 200, but the number of loops can be changed appropriately. With the number of loops set to 200, and 100 virtual SPM images generated in step S43, a total of 20,000 virtual SPM images are generated.

[0104] Next, step S5 is the property prediction processing performed by the property prediction processing unit 25. In property prediction processing S5, in step S51, the virtual SPM images contained in the multiple virtual SPM image data 33 are applied to the first learned model 32 to predict the properties of the composite material. Then, in step S52, the properties predicted in step S51 are stored in the storage unit 3 as predicted property data 34. After that, it returns to the previous step and proceeds to the next step. Figure 4 Step S6.

[0105] Next, step S6 is the virtual image extraction processing performed by the virtual image extraction processing unit 26. In the virtual image extraction processing S6, as... Figure 10 As shown, in step S61, the best physical property value with the best physical property value is extracted from the data contained in the learning data 31. Then, in step S62, based on the predicted physical property data 34, a virtual SPM image with a physical property value greater than or equal to the best physical property value is extracted. Then, in step S63, the virtual SPM image extracted in step S62 is stored in the storage unit 3 as extracted virtual SPM image data 35. Then, the process returns to... Figure 4 Step S7.

[0106] Next, step S7 is the second model generation process performed by the second model generation processing unit 27. In the second model generation process S7, as... Figure 11 As shown, in step S71, the learning data 31 is used as training data to generate a second learned model 36 representing the correlation between the SPM image and manufacturing conditions. Then, in step S72, the second learned model 36 generated in step S71 is stored in the storage unit 3. Afterwards, the process returns to... Figure 4 Step S8.

[0107] Next, step S8 is the manufacturing condition prediction processing performed by the manufacturing condition prediction processing unit 28. In the manufacturing condition prediction processing S8, as... Figure 12 As shown, in step S81, the virtual SPM images extracted from the virtual SPM image data 35, i.e., the virtual SPM images presumed to yield good physical properties, are applied to the second learned model 36 to predict manufacturing conditions. Then, in step S82, the manufacturing conditions predicted in step S81 are stored in the storage unit 3 as predicted manufacturing condition data 37. Afterwards, the process returns to... Figure 4 Step S9.

[0108] Next, step S9 is the prediction result prompting process performed by the prediction result prompting processing unit 29. In the prediction result prompting process S9, for example, the prediction result prompting processing unit 29 performs prediction result prompting processing. Figure 13 The prediction result display screen 100 shown is displayed on the monitor 4. The prediction result display screen 100 has a virtual SPM image display unit 101 that displays a virtual SPM image, a prediction property display unit 102 that displays predicted property, and a prediction manufacturing condition display unit 103 that displays predicted manufacturing conditions, and displays the virtual SPM image, predicted property, and predicted manufacturing conditions together.

[0109] With this configuration, users can easily and intuitively grasp the required information and easily determine how to set the actual manufacturing conditions. That is, it is an improved data processing device 1 for composite materials that can provide visualization of the prediction results for various data related to the manufacture of composite materials using multiple compounding agents.

[0110] More specifically, in this embodiment, the virtual SPM image allows for a visual understanding of the dispersion state of various compounding agents that contribute to the good physical properties of the composite material. Furthermore, it enables the simultaneous study and prediction of physical properties and manufacturing conditions while visually understanding the dispersion state of these compounding agents. Consequently, deeper research in material / raw material development (composite material development) is possible. Additionally, in the illustrated example, the order of physical properties is also displayed in the predicted physical property display unit 102, providing a more intuitive and easier-to-understand display. Moreover, Figure 13 The displayed content and layout are merely examples, and can be appropriately modified as long as they do not depart from the scope of the present invention.

[0111] (Predictions regarding physical properties)

[0112] Using the composite material data processing apparatus 1 of this embodiment, 20,000 virtual SPM images are generated, and the physical properties of the generated 20,000 virtual SPM images are predicted. Furthermore, the composite material with the formulation in Table 1 below is used here as the composite material.

[0113] [Table 1]

[0114]

[0115] The prediction results of physical properties are shown in Figure 14 .like Figure 14 As shown, it can be seen that compared with the best physical property value contained in the learning data 31 (i.e., the best physical property value among composite materials that can be achieved under the current conditions), there exists a virtual SPM image that exceeds the predicted physical property value. By determining such a virtual SPM image and predicting its manufacturing conditions, it is possible to expect to achieve composite materials with better physical properties.

[0116] (Preferred SPM image)

[0117] Scanning probe microscopes (SPMs) offer various imaging modes, enabling the acquisition of diverse images. For instance, in Bimodal Dual AC mode, shape images, primary amplitude images, secondary amplitude images, primary phase difference images, and secondary phase difference images can be obtained. Furthermore, in Fast Force Mapping mode, adhesive force mapping images and elastic force mapping images can be acquired. The inventors have conducted research and found that in shape images, primary amplitude images, and images acquired in other modes, it is difficult to demonstrate contrast between complexing agents contained within the image, making them unsuitable for predicting physical properties.

[0118] Therefore, as a preferred SPM image, it can be said that any one of the following is preferred: secondary amplitude image, primary phase difference image, secondary phase difference image, adhesive force mapping image, and elastic force mapping image. Figure 15 (a) to (e) represent examples of these secondary amplitude images, primary phase difference images, secondary phase difference images, adhesive force mapping images, and elastic force mapping images.

[0119] Furthermore, by using these secondary amplitude images, primary phase difference images, secondary phase difference images, adhesive force mapping images, and elastic force mapping images as SPM images respectively, a first learned model 32 is generated, and the mean absolute percentage error is calculated on the generated first learned model 32. The calculation results are as follows: Figure 16 As shown. In Figure 16 The figure shows the mean absolute percentage error and the average value obtained by averaging the predicted object's physical properties as elongation and tensile strength.

[0120] like Figure 16As shown, when predicting elongation, a single-phase difference image, a double-phase difference image, or an adhesive force mapping image is preferably used as the SPM image. Similarly, when predicting tensile strength, a double-amplitude image or an elastic force mapping image is preferably used as the SPM image. Furthermore, when predicting both elongation and tensile strength, a double-amplitude image, a single-phase difference image, or a double-phase difference image with good average values ​​is preferred. Additionally, since adhesive force mapping and elastic force mapping images take time to acquire, it is more preferable to use a double-amplitude image, a single-phase difference image, or a double-phase difference image, which can be acquired in a shorter time.

[0121] (Regarding the tuning of hyperparameters in machine learning)

[0122] like Figure 17 As shown, the composite material data processing apparatus 1 may include a hyperparameter optimization processing unit 231 that performs hyperparameter optimization processing to optimize the hyperparameters in the machine learning of the first learned model 32. Here, hyperparameters refer to parameters such as weights required to control the behavior of various algorithms during the learning process in machine learning, and need to be preset before machine learning. Examples of hyperparameters include the number of rounds, batch size (also called mini-batch size), learning rate, threshold, number of layers, and number of neurons in each layer.

[0123] In this embodiment, the hyperparameter optimization processing unit 231 optimizes the number of epochs and the batch size. Here, the number of epochs is the number of learning iterations, a parameter representing how many times the learning process using the training data is repeated. Additionally, in machine learning, training data is typically segmented, and the parameter representing the number of data points in each batch after this segmentation is called the batch size. For example, if the number of epochs is too small, sufficient learning may not be possible, leading to decreased prediction accuracy. Conversely, if the number of epochs is too large, prediction accuracy may decrease due to overlearning, and the learning time also increases. Furthermore, if the batch size is too small, the model is prone to getting trapped in local solutions, thus reducing prediction accuracy and increasing the number of model updates, thereby increasing learning time. Moreover, when the batch size is too large, the features of each data point are averaged, easily losing their unique characteristics and reducing prediction accuracy. As described above, the number of epochs and the batch size are important parameters affecting prediction accuracy, and they need to be set to the optimal number of epochs and batch size based on the learning data 31 to be used.

[0124] The hyperparameter optimization processing unit 231 uses the learning data 31 for machine learning used in the first learned model 32 to find the combination of the number of rounds and batch size that minimizes the average of the average error rate (MAPE) obtained through cross-validation. Alternatively, the average of the median error rate (APE) of the predicted values ​​can be used instead of the average of the average error rate (MAPE). The learning data 31 and setting data 38, including various settings required for hyperparameter optimization processing, are input to the hyperparameter optimization processing unit 231. The setting data 38 includes settings for data splitting in cross-validation and settings for the range of batch size and number of rounds to be searched. These settings are input by a user, for example, using the input device 5. Here, the number of splits in cross-validation is set to 9, the range of batch size to be searched is set to 4, 8, 16, 32, and 64, and the range of number of rounds to be searched is set to 1 to 100.

[0125] like Figure 18 As shown, the hyperparameter optimization processing unit 231 divides the learning data 31 into multiple (here, 9) segments and prepares 9 segmentation patterns. In each of these 9 segmentation patterns, the learning data after each segmentation is set as test data, and the other learning data are set as training data. Then, the hyperparameter optimization processing unit 231 calculates the average error rate (MAPE) in each segmentation pattern. That is, in each segmentation pattern, machine learning is performed using the training data to generate a learned model, and the average error rate (MAPE) when the test data is applied to the generated learned model is calculated. Then, the average of the average error rates calculated in each of the 9 segmentation patterns is calculated (the average of the MAPE). The above calculation is performed for all combinations of batch size and number of rounds. Here, since the batch size is 5 patterns and the number of rounds is 100 patterns, the calculation is performed for 5... The average value of MAPE for each of the 100 = 500 combinations.

[0126] Among the 500 combinations obtained by calculating the average MAPE, the combination of batch size and number of rounds that minimizes the average MAPE is the optimal hyperparameter. The determined optimal hyperparameter is stored as optimal hyperparameter data 39 in storage unit 3 (see reference). Figure 17 In the first model generation processing unit 23, machine learning is performed using the number of rounds and batch size of the determined optimal hyperparameters to generate a first learned model 32. This improves the prediction accuracy of property predictions using the first learned model 32.

[0127] Figure 19 This is a flowchart of the hyperparameter optimization process S10. For example, the hyperparameter optimization process S10 can be performed in... Figure 4 The process is performed between steps S2 and S3. For example... Figure 19As shown, firstly, in step 100, setting data 38 is input. In step 100, the settings for data splitting in cross-validation, as well as the range of batch size and round number to be searched, are input. Then, in step 101, 1 is used as the initial value and substituted into the variables ne (representing the round number) and nb (representing the batch size). Then, in step 102, based on the setting data 38 input in step 100, data splitting processing for cross-validation is performed. Here, the process of generating 9 splitting patterns is performed (see...). Figure 18 Then, in step 103, the number of rounds is set to ne, and the batch size is set to 2. （i+1） The MAPE for each segmentation pattern is calculated. Then, in step 104, the average MAPE of the nine segmentation patterns is calculated. The calculated average MAPE is stored in storage unit 3 in relation to the number of rounds and batch size.

[0128] Next, in step 105, it is determined whether the variable ne is greater than or equal to 100. If the determination is "no" in step 105, the variable ne is incremented in step 106, and the process returns to step 103. If the determination is "yes" in step 105, in step 107, it is determined whether the variable nb is greater than or equal to 5. If the determination is "no" in step 107, 1 is substituted into the variable ne in step 108, the variable nb is incremented, and the process returns to step 103. If the determination is "yes" in step 107, in step 109, the combination of round number and batch size with the smallest average value among the 500 MAPE values ​​calculated in step 104 is calculated and stored as optimal hyperparameter data 39 in storage unit 3. Then, the process returns, for example, to... Figure 4 Step S3. In the first model generation process of step S3, machine learning is performed using a combination of round number and batch size stored as the optimal hyperparameter data 39.

[0129] Figure 20 This is a graph illustrating an example of the relationship between the number of rounds and batch size obtained through hyperparameter optimization and the average MAPE. Figure 20 In the examples, the average MAPE was the smallest (20.7%) in the combination of batch size 16 and number of rounds 5. Subsequently, the average MAPE was found to be the second smallest in the combination of batch size 8 and number of rounds 19, and the third smallest in the combination of batch size 4 and number of rounds 73.

[0130] Furthermore, in this process of hyperparameter optimization, hyperparameters in the machine learning of the first learned model 32 are optimized, but it is not limited to this; hyperparameters in the machine learning of the second learned model 36 can also be optimized.

[0131] (Summary of implementation methods)

[0132] Next, the technical concept learned from the embodiments described above will be described by reference to the accompanying reference numerals and the like. However, the reference numerals and the like in the following description do not limit the constituent elements within the scope of patent protection to the components specifically shown in the embodiments.

[0133] [1] A composite material data processing apparatus 1 predicts various data related to a composite material manufactured using a variety of compounding agents. The composite material data processing apparatus 1 includes: a first model generation processing unit 23 that generates a first learned model 32 representing the correlation between a composite material image after the composite material is visualized and a physical property as one of the various data; a virtual image generation processing unit 24 that generates a plurality of virtual images that mimic the composite material image, i.e., virtual composite material images; and a physical property prediction processing unit 25 that uses the first learned model 32 and the plurality of virtual composite material images to predict the physical property.

[0134] [2] According to the data processing device 1 for composite materials described in [1], the physical property includes one or more of elongation and tensile strength.

[0135] [3] According to the composite material data processing apparatus 1 described in [1] or [2], wherein the composite material image is an SPM image obtained by imaging with a scanning probe microscope (SPM).

[0136] [4] According to the composite material data processing device 1 described in [3], the SPM image is a primary phase difference image, a secondary phase difference image, or a secondary amplitude image.

[0137] [5] According to the composite material data processing device 1 described in [1], the virtual image generation processing unit 24 uses the composite material image as training data to generate the virtual composite material image.

[0138] [6] According to the composite material data processing apparatus 1 described in [1], the composite material data processing apparatus includes: a virtual image extraction processing unit 26, which extracts the virtual composite material image that is correlated with a property that is more correlated with the composite material image.

[0139] [7] According to the composite material data processing apparatus 1 described in [6], the composite material data processing apparatus includes: a second model generation processing unit 27 that generates a second learned model 36 representing the correlation between the composite material image and manufacturing conditions, which are one of the various data; and a manufacturing condition prediction processing unit 28 that extracts the virtual composite material image, i.e., extracts the virtual composite material image, using the second learned model 36 and the virtual image extraction processing unit 26, and predicts the manufacturing conditions.

[0140] [8] According to the data processing device 1 for composite materials described in [7], the manufacturing conditions include one or more of the following: the rotational speed of the mixer, the mixing time, and the mixing temperature.

[0141] [9] According to the composite material data processing apparatus 1 described in [1], the composite material data processing apparatus includes: a hyperparameter optimization processing unit 231, which optimizes the hyperparameters in the machine learning of the first learned model 32.

[0142]

[10] According to the composite material data processing apparatus 1 described in [9], wherein the hyperparameters are the number of rounds and the batch size, the hyperparameter optimization processing unit 231 uses the learning data 31 used in the machine learning of the first learned model 32 to calculate the average error rate for all preset combinations of rounds and batch sizes through cross-validation, and calculates the number of rounds and batch size with the smallest average error rate as the optimal hyperparameters, and the first model generation processing unit 23 performs machine learning with the number of rounds and batch size of the optimal hyperparameters to generate the first learned model 32.

[0143] The embodiments of the present invention have been described above, but the embodiments described above do not limit the scope of the invention to which patent protection is claimed. Furthermore, it should be noted that the combinations of features described in the embodiments are not necessarily all necessary means to solve the problems of the invention.

[0144] Furthermore, appropriate modifications can be made to the invention as long as they do not depart from the spirit of the invention. An example is given below.

[0145] [1] In a composite material data processing apparatus 1 that predicts various data related to composite materials manufactured using multiple compounding agents, the apparatus includes: a second model generation processing unit 27 that generates a second learned model 36 representing the correlation between the composite material image and manufacturing conditions, which are one of the various data; and a manufacturing condition prediction processing unit 28 that uses the second learned model 36 and the virtual composite material image extracted by the virtual image extraction processing unit 26 to predict the manufacturing conditions.

[0146] [2] As an embodiment of the present invention, a data processing device 1 for composite materials is described, but it may also be a data processing method for composite materials.

[0147] Explanation of reference numerals in the attached figures

[0148] 1. Data processing device for composite materials

[0149] 2. Control Department

[0150] 21 Data Acquisition and Processing Department

[0151] 22 Learning Data Appending Processing Department

[0152] 23 First Model Generation and Processing Unit

[0153] 24 Virtual Image Generation and Processing Unit

[0154] 25. Property Prediction and Processing Department

[0155] 26 Virtual Image Extraction and Processing Department

[0156] 27 Second Model Generation Processing Unit

[0157] 28 Manufacturing Condition Prediction and Processing Department

[0158] 29. Prediction results prompt processing department

[0159] 3. Storage Unit

[0160] 31. Learning to use data

[0161] 32 First learned model

[0162] 33 Virtual SPM image data,

[0163] 34 Predictive physical property data,

[0164] 35. Extracting virtual SPM image data

[0165] 36 Second learned model

[0166] 37. Predictive manufacturing condition data,

[0167] 4 monitors

[0168] 5. Input device.

Claims

1. A data processing device for composite materials, which predicts various data related to composite materials manufactured using multiple compounding agents, characterized in that, The data processing device for composite materials includes: A first model generation processing unit generates a first learned model, which represents the correlation between a composite material image obtained by imagerizing the composite material and physical properties, which are one of the various data. A virtual image generation processing unit generates multiple virtual composite material images, which are virtual images obtained by imitating the composite material images; as well as The property prediction processing unit uses the first learned model and multiple virtual composite material images to predict the property.

2. The data processing device for composite materials according to claim 1, characterized in that, The physical properties include one or more of elongation and tensile strength.

3. The data processing apparatus for composite materials according to claim 1 or 2, characterized in that, The composite material image is an SPM image obtained by imaging with a scanning probe microscope.

4. The data processing device for composite materials according to claim 3, characterized in that, The SPM image is a primary phase difference image, a secondary phase difference image, or a secondary amplitude image.

5. The data processing device for composite materials according to claim 1, characterized in that, The virtual image generation and processing unit uses the composite material image as training data to generate the virtual composite material image.

6. The data processing device for composite materials according to claim 1, characterized in that, The composite material data processing device includes a virtual image extraction and processing unit, which extracts virtual composite material images that are correlated with physical properties that are highly correlated with the composite material image.

7. The data processing apparatus for composite materials according to claim 6, characterized in that, The data processing device for composite materials includes: The second model generation processing unit generates a second learned model, which represents the correlation between the composite material image and the manufacturing conditions, which are one of the various data. as well as The manufacturing condition prediction processing unit uses the second learned model and the virtual composite material image extracted by the virtual image extraction processing unit to predict the manufacturing conditions.

8. The data processing apparatus for composite materials according to claim 7, characterized in that, The manufacturing conditions include one or more of the following: the speed of the mixing mill, the mixing time, and the mixing temperature.

9. The data processing device for composite materials according to claim 1, characterized in that, The data processing device for composite materials includes a hyperparameter optimization processing unit, which optimizes the hyperparameters in the machine learning of the first learned model.

10. The data processing apparatus for composite materials according to claim 9, characterized in that, The hyperparameters are the number of rounds and the batch size. The hyperparameter optimization processing unit uses the learning data from the machine learning process of the first learned model to calculate the average error rate through cross-validation for all pre-defined combinations of number of rounds and batch size, and then selects the number of rounds and batch size with the minimum average error rate as the optimal hyperparameters. The first model generation processing unit uses the number of rounds and batch size of the optimal hyperparameters to perform machine learning and generate the first learned model.

Images