Physical property prediction device, physical property prediction method, and program

Abstract

Provided is a physical property prediction device comprising: an operation unit that generates physical property information for a composition on the basis of input data; and an output unit that outputs the physical property information. Using first material information as the input data, the operation unit generates, on the basis of a first neural network, physical structure information regarding a composition for which the physical properties are to be predicted, and using second material information of a type differing from the first material information as the input data, generates, on the basis of a second neural network, chemical structure information regarding the composition. The operation unit then generates integrated information in which the physical structure information and the chemical structure information are integrated, and generates the physical property information of the composition on the basis of the integrated information.

Description

Physical property prediction device, physical property prediction method, and program

[0001] The present invention relates to a property prediction device, a property prediction method, and a program for predicting property information of compositions such as composite materials. This application claims priority to Japanese Patent Application No. 2022-081299, filed on May 18, 2022, the contents of which are incorporated herein by reference.

[0002] In recent years, research has been progressing on methods for utilizing computer-based information processing technologies, such as chemoinformatics, materials informatics, and process informatics, in chemical research, with the aim of searching for new materials and shortening the time required for material development. Among these, a method has been proposed for the development of new materials such as carbon nanotubes, in which images obtained through experiments are used as training data, machine learning is performed based on a multilayer neural network, and unknown physical properties of the composition are predicted (see, for example, Non-Patent Document 1).

[0003] Honda, T., Muroga, S., Nakajima, H. et al. Virtual experiments by deep learning on tangible materials. Communications Materials, 2, 88 (2021). https: / / doi.org / 10.1038 / s43246-021-00195-2 August 30, 2021

[0004] The composition prediction method described in Non-Patent Document 1 was limited to capturing information only in images showing the structure of the composition. For general compositions, images do not necessarily contain information showing their physical properties; physical properties such as chemical properties, optical properties, and molecular structure are also present. Furthermore, composite materials have various combinations of structural elements, but no clear definition exists, making it difficult to handle structural information based on images or numerical values. Therefore, the composition prediction method described in Non-Patent Document 1 was unable to predict physical property information for compositions with complex structures, such as composite materials formed from multiple materials or structures.

[0005] The present invention aims to provide a property prediction device, a property prediction method, and a program that can predict the properties of unknown compositions such as composite materials based on machine learning that integrates complex material information.

[0006] One aspect of the present invention is a physical property prediction device comprising: a calculation unit that generates physical property information of a composition based on input data; and an output unit that outputs the physical property information, wherein the calculation unit generates physical structure information of a composition to be predicted based on a first neural network using first material information as the input data; generates chemical structure information of the composition based on a second neural network using second material information of a type different from the first material information as the input data; generates integrated information that integrates the physical structure information and the chemical structure information; and generates the physical property information of the composition based on the integrated information.

[0007] According to the present invention, it is possible to predict the physical properties of unknown compositions such as composite materials based on machine learning that integrates complex material information.

[0008] 1 is a block diagram showing the configuration of a physical property prediction system; FIG. 2 is a flowchart showing the processing flow of a machine learning process executed in a physical property prediction device; FIG. 3 is a flowchart showing the processing flow of a virtual experiment process executed in a physical property prediction device; FIG. 4 is a diagram showing experimental results using a physical property prediction device; FIG. 5 is a diagram showing an example of image data generated by a physical property prediction device; and FIG. 6 is a diagram showing an example of spectral data generated by a physical property prediction device.

[0009] Hereinafter, an embodiment of a physical property prediction apparatus according to the present invention will be described with reference to the drawings. The physical property prediction apparatus predicts new complex materials having multiple physical property information based on machine learning using a neural network.

[0010] As shown in FIG. 1 , the physical property prediction system 1 includes, for example, a physical property prediction device 10 communicatively connected to a communication network NW, and a server device 20. The server device 20 is, for example, a storage server that stores a huge amount of information, such as experimental data. The server device 20 is communicatively connected to the physical property prediction device 10 via the communication network NW. The server device 20 may also be directly connected to the physical property prediction device 10. The server device 20 includes, for example, a storage unit 22 that stores experimental data. The storage unit 22 is, for example, a storage device such as a hard disk drive (HDD) or flash memory. The experimental data is stored in the storage unit 22 via the communication network NW. The contents of the experimental data will be described later. The server device 20 provides the experimental data to the physical property prediction device 10 via the communication network NW.

[0011] The physical property prediction apparatus 10 is configured, for example, by a terminal device for information processing, such as a personal computer. The physical property prediction apparatus 10 may be configured by two or more terminal devices capable of cooperative processing. The physical property prediction apparatus 10 includes, for example, a calculation unit 12 that performs calculations necessary for information processing, a storage unit 14 that stores data and programs necessary for the calculations, an output unit 16 that outputs information such as the calculation results of the calculation unit 12, and an acquisition unit 18 that acquires data necessary for the calculations via a communication network NW.

[0012] The calculation unit 12 may be realized by hardware (including circuitry) such as an LSI (Large Scale Integration), an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or a GPU (Graphics Processing Unit), or may be realized by a combination of software and hardware. The program may be stored in advance in a storage device (a storage device having a non-transitory storage medium) such as an HDD or flash memory of the storage unit 14, or may be stored in a removable storage medium such as a DVD or CD-ROM, and installed in the HDD or flash memory of the storage unit 14 by inserting the storage medium (non-transitory storage medium) into a drive device.

[0013] The memory unit 14 is, for example, a storage device such as a hard disk drive or flash memory. The memory unit 14 stores, for example, programs necessary for the calculations of the calculation unit 12 and various other information. The output unit 16 is, for example, a display device configured with an LCD (Liquid Crystal Display), an organic EL (Electro Luminescence) display, or the like. The output unit 16 is configured integrally with or separately from the property prediction device 10.

[0014] The acquisition unit 18 includes a communication interface capable of communicating with the communication network NW. The acquisition unit 18 communicates via the communication network NW using, for example, a cellular network, a Wi-Fi network, Bluetooth (registered trademark), DSRC (Dedicated Short Range Communication), etc. The acquisition unit 18 communicates with the server device 20 via the communication network NW, for example, and acquires the experiment data stored in the server device 20.

[0015] Next, the processing executed by the calculation unit 12 will be described. In the following description, the composition is, for example, a polymer composite material. The composition is not limited to a polymer composite material and may be something else. The calculation unit 12 executes a learning process in which deep learning based on a neural network is performed using experimental data acquired from the server device 20, and a virtual experiment process in which a virtual experiment is performed using input data and physical property information of the composition is output. The calculation unit 12 performs machine learning using, for example, different types of measurement data, and improves its ability to calculate predicted values ​​of physical property information of the composition for the input data.

[0016] The calculation unit 12, for example, performs machine learning using multimodal AI (artificial intelligence) based on input data, and calculates physical property information of the composition based on the multimodal AI after machine learning. Multimodal AI is a computer program that executes a process that uses multiple different types of information to output advanced determination results. Multimodal AI can output calculated values ​​according to the purpose using different types of input data, such as images and numerical data. The calculation unit 12, for example, performs a virtual experiment using image data of the composition and measured values ​​of the substances, and calculates predicted values ​​of the physical property information of the composition.

[0017] The calculation unit 12 is configured to generate physical structure information of the composition to be predicted based on a first learning model based on a first neural network using first material information including image data from the input data, for example. The physical structure information is information representing the mixing and dispersion state of substances in the composition, composition distribution, etc. The calculation unit 12 is also configured to generate chemical structure information of the composition based on a second learning model based on a second neural network using second material information including different types of numerical values ​​from the first material information. The chemical structure information is information representing the chemical structure of the substances that make up the composition, their types, number, interactions, etc. The calculation unit 12 is configured to generate physical property information of the composition by integrating the generated physical structure information and chemical structure information.

[0018] The calculation unit 12 uses different types of input data, such as image data showing the structure of the composition, measurements of the composition based on infrared absorption spectrometry (IR), and measurements including numerical values ​​based on Raman spectroscopy, to calculate predicted values ​​of at least three types of physical property information among physical property indices of the composition, such as Young's modulus, strength, breaking elongation, glass transition temperature, density, electrical resistivity, storage modulus, and loss tangent. The physical property information is not limited to the above-mentioned eight, and may include other physical property indices.

[0019] The calculation unit 12 executes a learning process based on machine learning to improve the accuracy of generating physical property information of the composition. The calculation unit 12 selects data belonging to a predetermined category from the input data input to the server device 20 and generates first material information. The first material information includes, for example, image data indicating physical structure information of the composition. The first material information is, for example, data based on images compiled into a database from past experiments. The first material information is, for example, image data indicating the structure of a polymer composite material generated from a matrix, additives, and fillers blended based on a predetermined blending ratio.

[0020] The first material information may be, for example, image data including physical structure information captured by an optical microscope, a laser microscope, a scanning electron microscope (SEM), a scanning probe microscope (SPM), an atomic force microscope (AFM), a transmission electron microscope (TEM), birefringence imaging, or a laser scanner. The first material information may be two-dimensional pixel data obtained by thermography, three-dimensional voxel data obtained by tomography such as X-ray computed tomography (CT) or TEM-CT, or hyperspectral data including wavelength information. Furthermore, the first material information may be replaced with data such as light scattering, X-ray diffraction, ultrasonic attenuation, gas adsorption testing, dielectric relaxation spectrum, or viscoelastic spectrum, as long as it indirectly includes physical structure information. Furthermore, the first material information may be generated by combining the above data.

[0021] The calculation unit 12 executes first machine learning based on a first neural network using pre-stored first material information. The first neural network is configured with a first learning model used in a calculation to calculate an output value. The first learning model is configured, for example, to calculate a predicted value of physical structure information of a composition when the first material information is input. Initial parameters for calculating the predicted value are set in the first learning model.

[0022] The calculation unit 12 repeatedly executes a process of calculating an output value for an input value using the first neural network and adjusting parameters included in the first learning model based on a comparison between the output value and a correct answer value. The calculation unit 12 improves the accuracy of the first learning model based on the process of executing the first machine learning. The first learning model is, for example, a model capable of performing supervised learning and unsupervised learning using the first material information.

[0023] When there is a small amount of input data, the calculation unit 12 performs unsupervised learning using the first material information with the first neural network. When there is a large amount of input data, the calculation unit 12 may perform supervised learning using the first material information with the first neural network. The calculation unit 12 is configured to execute processing based on generative adversarial networks (GAN) in unsupervised learning using the first neural network, for example. The GAN is composed of two networks: a generator network that generates data and a discriminator network that discriminates whether the generated data is true or false.

[0024] The generative network generates image data of the composition based on the first material information. The discriminant network discriminates the authenticity of the generated image data. The generative network generates image data with added noise and learns to deceive the discriminant network. The discriminant network learns to more accurately discriminate the authenticity of the generated image data. According to the GAN, the learning accuracy is improved by competition between the generative network and the discriminant network. According to the GAN, unsupervised learning can be performed to learn the features of an object even when correct answer data is not given. According to the GAN, by learning features from data, non-existent data can be generated and transformed according to the features of existing data.

[0025] The calculation unit 12 adjusts parameters of the first learning model by first machine learning using a first neural network in the above-mentioned first machine learning step. The calculation unit 12 may perform one or more first machine learning processes based on different types of first material information. The calculation unit 12 calculates a predicted value of physical structure information of the composition using the first learning model whose parameters have been adjusted.

[0026] The calculation unit 12 performs the first machine learning for calculating the physical structure information of the composition described above, and also performs the second machine learning for calculating the chemical structure information of the composition. The second machine learning may be performed sequentially with respect to the first machine learning, or may be performed in parallel.

[0027] The calculation unit 12 selects data belonging to a predetermined category from the input data input to the server device 20, and generates second material information of a type different from the first material information. The second material information includes, for example, numerical data of measured values ​​indicating chemical structure information of the composition. The second material information is, for example, data based on measured values ​​compiled into a database in past experiments. The second material information is, for example, numerical data indicating the compounding ratio and manufacturing conditions of a polymer composite material generated from a matrix, additives, and fillers compounded based on a predetermined compounding ratio. The second material information includes, for example, spectral data that quantifies the characteristics of the chemical structure of the composition.

[0028] The second material information may be data including chemical structure information that quantifies the characteristics of the chemical structure of the composition, such as infrared absorption spectroscopy (IR), Raman spectroscopy, ultraviolet-visible absorption spectroscopy (UV-VIS), near infrared (NIR) spectroscopy, far infrared (FIR) spectroscopy, terahertz (THz) spectroscopy, nuclear magnetic resonance (NMR), X-ray photoelectron spectroscopy (XPS), energy dispersive X-ray spectroscopy (EDS), electron energy loss spectroscopy (EELS), fluorescence spectroscopy, mass spectrometry (MS), electron spin resonance (ESR), or the like. The second material information may be data such as molecular structures, such as the arrangement of atoms and chemical bonds, their text representations (Simplified Molecular Input Line Entry System: SMILES), molecular descriptors, fingerprints, etc. The second material information may also include information from chromatography, thermal analysis, odor measurement, etc., as long as it indirectly includes chemical structure information. The second material information may also be generated by combining the above data.

[0029] The calculation unit 12 performs second machine learning based on a second neural network using the second material information. The second neural network is configured with a second learning model used in a calculation to calculate an output value. The second learning model is configured, for example, to calculate a predicted value of chemical structure information of a composition when the second material information is input. Initial parameters for calculating the predicted value are set in the second learning model.

[0030] The calculation unit 12 repeatedly executes a process of calculating an output value for an input value using the second neural network and adjusting parameters included in the second learning model based on a comparison between the output value and a correct answer value. The calculation unit 12 improves the accuracy of the second learning model based on the process of executing the second machine learning. The second learning model is, for example, a model capable of performing supervised learning and unsupervised learning using the second material information.

[0031] When there is a small amount of input data, the calculation unit 12 performs unsupervised learning based on the second material information using the second neural network. When there is a large amount of input data, the calculation unit 12 may perform supervised learning using the second material information using the second neural network. The calculation unit 12 is configured to, for example, execute processing based on a generative adversarial network in unsupervised learning using the second neural network.

[0032] The calculation unit 12 adjusts parameters of the second learning model by second machine learning using a second neural network in the above-mentioned second machine learning step. The calculation unit 12 may perform one or more second machine learning processes based on different types of second material information. The calculation unit 12 calculates chemical structure information of the composition using the second learning model whose parameters have been adjusted.

[0033] The calculation unit 12 generates integrated information by integrating the physical structure information generated by the first neural network and the chemical structure information generated by the second neural network. The calculation unit 12 uses the integrated information to perform third machine learning based on a third neural network. The third neural network is configured with a third learning model that calculates physical property information of the composition based on the integrated information.

[0034] The third learning model is configured to, for example, calculate a predicted value of the physical property information of the composition when integrated information is input. Initial parameters for calculating the predicted value are set in the third learning model. The calculation unit 12 adjusts the parameters of the third learning model through third machine learning using a third neural network. The calculation unit 12 digitizes the physical structure information including the image data in generating the integrated information and converts it into numerical data. The calculation unit 12 repeatedly performs machine learning using, for example, a convolutional neural network included in the third neural network, extracts features included in the image data, and classifies the image data by feature. The calculation unit 12 converts the classified image data into one-dimensional vector data in which the feature amounts are digitized.

[0035] When generating integrated information based on multiple pieces of chemical structure information such as spectra, the calculation unit 12 repeatedly executes machine learning using a third neural network to extract features contained in the numerical data and classify the data by feature. The calculation unit 12 converts the classified numerical data into one-dimensional vector data in which the features are quantified. The calculation unit 12 integrates the physical structure information based on the one-dimensional vector data and the chemical structure information based on the one-dimensional vector data to generate integrated information based on the one-dimensional vector data.

[0036] The calculation unit 12 adjusts parameters of the third learning model by third machine learning using a fully connected neural network included in the third neural network using the integrated information based on the one-dimensional vector data. The calculation unit 12 generates physical property information of the composition based on the one-dimensional vector data based on the integrated information using the adjusted third learning model. After performing the machine learning, a virtual experiment can be performed to predict the physical properties of a new composition using the physical property prediction device 10.

[0037] In generating the integrated information, the calculation unit 12 may not only perform machine learning using a third neural network through deep learning, but also extract features contained in multiple pieces of chemical structure information based on numerical data using analytical methods such as multiple linear regression (MLR), principal component regression (PCR), partial least squares regression (PLS), support vector machine (SVM), decision tree, random forest, logistic regression, elastic net regression, lasso regression, ridge regression, Gaussian process regression (GPR), and gradient boosting regression, and classify the features. Furthermore, before performing machine learning in generating the integrated information, the calculation unit 12 may select variables for multiple pieces of physical structure information and chemical structure information based on numerical data using computational methods such as forward method, backward method, stepwise method, and genetic algorithm (GA).

[0038] 2 is a flowchart showing the flow of the machine learning process of the physical property prediction method executed by the physical property prediction apparatus 10. The calculation unit 12 acquires input data from the server device 20 via the communication network NW (step S100). The calculation unit 12 performs first machine learning and second machine learning based on the input data. When performing the first machine learning, the calculation unit 12 acquires first material information including image data (step S102).

[0039] The calculation unit 12 repeatedly performs a first machine learning process using the first neural network with the learning image data (step S104). The calculation unit 12 adjusts parameters of the first learning model constituting the first neural network through the first machine learning process, and generates physical structure information based on the adjusted first learning model (step S106).

[0040] When performing the second machine learning, the calculation unit 12 acquires second material information including spectral data (step S108). The calculation unit 12 repeatedly performs the second machine learning using a second neural network with the training spectral data (step S110). The calculation unit 12 adjusts parameters of a second learning model constituting the second neural network through the second machine learning, and generates chemical structure information based on the adjusted second learning model (step S112).

[0041] The calculation unit 12 integrates the generated physical structure information and chemical structure information (step S114). The calculation unit 12 performs third machine learning using a third neural network using the integrated information (step S116). The calculation unit 12 adjusts parameters of a third learning model constituting the third neural network using the integrated information, and repeatedly performs third machine learning to generate physical property information based on the adjusted third learning model.

[0042] 3 is a flowchart showing the processing flow of a virtual experiment of the physical property prediction method executed by the physical property prediction device 10. Input data related to a new composition to be used in the virtual experiment is created (step S200). The input data may be data generated in the machine learning process, or may be newly created. The composition is formed by mixing at least two of the matrix, filler, and additive in unknown proportions. The calculation unit 12 generates first material information based on image data from the input data. The calculation unit 12 acquires the generated first material information (step S202).

[0043] The calculation unit 12 repeatedly performs a first virtual experiment using a first learning model in which parameters of a first neural network have been adjusted based on the first material information (step S204). The calculation unit 12 generates physical structure information based on the first virtual experiment (step S206). The calculation unit 12 generates second material information based on the spectral data of the input data. The calculation unit 12 acquires the generated second material information (step S208). The calculation unit 12 repeatedly performs a second virtual experiment using a second learning model in which parameters of a second neural network have been adjusted based on the second material information (step S210).

[0044] The calculation unit 12 generates chemical structure information based on the second virtual experiment (step S212). The calculation unit 12 integrates the physical structure information and the chemical structure information to generate integrated information (step S214). The calculation unit 12 generates physical property information of the composition based on a third neural network using the integrated information (step S216). The calculation unit 12 generates physical property information based on an adjusted third learning model of the third neural network using the integrated information. The calculation unit 12 repeats the virtual experiment by changing the input data to generate physical property information of the unknown composition. The calculation unit 12 stores the generated physical property information in the memory unit 14 and creates a database. The calculation unit 12 controls the output unit 16 to output the physical property information (step S218).

[0045] FIG. 4 shows the results of an experiment verifying the accuracy of the physical property information generated by the physical property prediction device 10. In the experiment, the accuracy of eight physical property information was predicted when a weight fraction was given to the composition of a certain composition. The eight properties include, for example, Young's modulus, tensile strength, elongation at break, logarithmic electrical resistivity, density, room-temperature storage modulus, and room-temperature loss tangent (see FIG. 4(B)). The accuracy of the physical property information was evaluated based on the eight properties. The input data used were a combination of image data (optical image) of first material information including physical structure information and spectral data (IR) of second material information including chemical structure information, and a combination of image data (optical image) of the first material information and spectral data (IR) and spectral data (Raman) of the second material information. For comparison, calculated values ​​of physical property information based on a neural network using only image data (optical image), calculated values ​​of physical property information based on a neural network using only spectral data (IR), and calculated values ​​of physical property information based on a neural network using only spectral data (Raman) were calculated. The accuracy was compared between physical property information calculated using training data and physical property information calculated using newly generated test data.

[0046] 4A, when training data was used, the accuracy of both the physical property information calculated based on a single neural network and the physical property information calculated based on multiple neural networks of the physical property prediction device 10 was high. On the other hand, when test data was used, the accuracy of the physical property information calculated based on a single neural network was low, whereas the accuracy of the physical property information calculated based on multiple neural networks of the physical property prediction device 10 was high.

[0047] As shown in Figure 4(B), the predicted values ​​of the eight properties calculated by the property prediction device 10 were stable when calculated based on multiple neural networks, whereas the values ​​of the property information calculated based on a single neural network varied widely. The weight fractions of the compositions and the eight properties described above may be changed, increased or decreased, without being limited to these. Furthermore, the property information may be changed to other conditions to include different parameters, or other conditions may be added.

[0048] As described above, the physical property prediction device 10 can accurately predict the physical properties of unknown compositions by performing machine learning that integrates different complex material information such as image data and spectral data.

[0049] In the above experiments, the input data was subjected to data expansion to generate a large amount of training data, enabling deep learning to generate more accurate predictions. For example, when expanding image data, if the original image is approximately 1000 x 1000 pixels, it is difficult to perform machine learning at that size due to the relationship between the performance of the computer constituting the property prediction device 10 and the amount of data.

[0050] Image data obtained through normal observations typically has a data size exceeding 1,000 x 1,000 pixels per image. Therefore, performing deep learning using the entire image as is would strain computer memory and increase the size of the neural network. Furthermore, the amount of experimental data obtained through actual experiments is limited, making it difficult to obtain sufficient training data for deep learning. Therefore, the image data was set to an appropriate scale (observation magnification) and pixel count that reflects the physical structure of the composition, so that the first machine learning process in the property prediction device 10 could be smoothly performed. Furthermore, the image data was subjected to processing such as rotation, tilting, and noise addition, thereby enhancing the data.

[0051] Image data was acquired based on observations at low magnification (100x), which allows for a broad understanding of the physical structure of the material for deep learning. The images were then divided into 128 x 128 pixel sub-images. The sub-images were divided to capture sufficient information about the physical structure of fillers, such as particles, to enable feature classification in deep learning. The 128 x 128 pixel sub-images were then rotated in four directions at 90° intervals for data expansion. This avoided straining computer memory and prevented the neural network from expanding in size. This enabled the acquisition of sufficient training data from limited experimental data, enabling deep learning training, resulting in stable neural network learning and improved accuracy.

[0052] The image data was also preprocessed to perform first machine learning based on the first neural network. The data for one pixel in the image data is a 1-byte (8-bit) unsigned integer, type uint8, with a value range of 0-255. With this data as is, training of the first neural network, which uses the material information of the composition as input, to generate an image would not proceed correctly. The image data was preprocessed by dividing the brightness value of each pixel by 127.5 and subtracting it by 1 to ensure that the data range fell within the range of -1 to 1. Based on the preprocessing of the image data, the learning process was stabilized in the generative model of the first neural network.

[0053] As shown in Figure 5, the physical property prediction device 10 can output generated image data equivalent to the measured image data by performing second machine learning based on the data-augmented image data.

[0054] Neural networks that generate images using material information about a composition as input have traditionally been trained using one-hot vectors as input. In general object recognition, training is performed by inputting vectors such as cat = 1 - dog = 0, or cat = 0 - dog = 1. In the case of materials compositions, the input vector becomes longer depending on the sample types used for training, posing a problem of straining computer memory when the number of sample types increases. Therefore, by training a neural network using material information about the composition (continuous variables such as blending ratios and manufacturing conditions) as input, it is possible to generate data directly from material information, thereby reducing computer memory requirements. In this study, a generative model was constructed using the blending ratio of an acrylic cured resin composite material converted to a weight fraction in the 0-1 range as input.

[0055] Unlike image data, no data extension method has been established for spectral data. Therefore, the spectral data was extended by applying Gaussian noise to ensure smooth execution of the second machine learning in the property prediction device 10. For example, the spectral data was expanded to a data length of 1024 by selecting a wavenumber range of 1800 cm-1 or less. The spectral data was preprocessed so that the absorbance values ​​were normalized by the maximum and minimum values ​​within one sample and the data fell within the range of 0-1.

[0056] The spectral data is normalized absorbance using minimum and maximum values, and is therefore numerical data in the range of 0 to 1. Here, in order to train the generative model, the normalized absorbance for each wavenumber was divided by 0.5 and then subtracted by 1, thereby performing preprocessing so that the data range fell within the range of -1 to 1. This stabilized the learning process of the second neural network in the second learning model.

[0057] When training was performed using models such as Conditional Table GAN as a generative model, the model's training and generation behavior was unstable. Therefore, by transforming the spectral data with a data length of 1024 into a 32x32 two-dimensional data array and training the neural network of the generative model in the same way as for images, the training process of the neural network in the generative model became stable.

[0058] The spectral data was augmented by generating Gaussian noise with a mean of 0 and a standard deviation of 0.001 and adding different patterns of noise to the original spectral data, resulting in training data that was augmented 128 times.

[0059] As shown in Figure 6, based on the expanded spectral data, generated spectral data (chemical structure information) close to the actual measured values ​​of the spectral data could be generated based on the second neural network. In the above experiment, by performing data expansion, the number of deep learning processes performed in the property prediction device 10 increased, and the learning process of the second neural network became more stable and the estimation accuracy improved compared to when data expansion was not performed.

[0060] In spectrum generation, if a one-hot vector is used as input to the generative model, the input vector becomes longer depending on the sample types used for learning, which poses a problem of straining computer memory when the number of sample types increases. Therefore, by training the second neural network using material information of the composition (continuous variables such as blending ratios and manufacturing conditions) as input, it becomes possible to generate data directly from the material information, resulting in reduced computer memory. In this study, a generative model was constructed using the blending ratio of an acrylic cured resin composite material in the range of 0-1 as input, as a weight fraction.

[0061] The spectral data used included a numerical array of corresponding blending ratios, images as physical structure information, and IR and Raman spectra as chemical structure information for 75 acrylic cured resin composite materials with different blending ratios of the components.

[0062] Furthermore, in the above experiments, the number of layers and layer structure of the machine learning model were adjusted by determining the impact on the processing behavior in the learning process and the accuracy of the calculation results when the number of layers, layer structure, etc. were changed. Furthermore, in GAN learning, if learning is not performed in a way that is appropriate for the model structure, the learning process will not be executed smoothly, so the GAN learning process was adjusted according to the model structure.

[0063] In the process of generating integrated information, it is necessary to combine information of different dimensions, such as images and spectra. For example, it is possible to perform machine learning by converting 128 x 128 pixel image data into a one-dimensional vector of 16,384 pixels and processing it in parallel with one-dimensional spectral data. In this case, spatial features contained in the image data, such as the relative arrangement of materials, are lost, making it impossible to fully train the neural network on physical structure information. Therefore, features were extracted from the two-dimensional image data using a neural network including a convolutional layer that incorporates information on the relative arrangement of neighboring pixels. This was then integrated with features obtained from the spectral data, and the physical property information of the composition was calculated via a fully connected neural network. This resulted in the effect of training the data without losing physical and chemical structure information.

[0064] The calculated physical property information of the composition includes eight physical property indicators: Young's modulus, tensile strength, elongation at break, glass transition temperature, density, room temperature storage modulus, room temperature loss tangent, and surface electrical resistivity. The physical property information used was preprocessed by standardizing the average and standard deviation. In the physical property prediction device 10, a neural network was constructed that integrates physical structure information and chemical structure information using this training data. By inputting the numerical values ​​of the material composition using the trained neural network, physical property information including indicators of the eight characteristics was obtained as output.

[0065] The matrices used in this experiment were, for example, phenyl glycidyl ether acrylate hexamethylene diisocyanate urethane prepolymer (AH-600, Kyoeisha Chemical Co., Ltd.), 2-ethylhexyl methacrylate (2-EH, Tokyo Chemical Industry Co., Ltd.), 2-hydroxyethyl methacrylate ethylene glycol methacrylate (2-HEMA, Tokyo Chemical Industry Co., Ltd.), benzyl methacrylate (Light Ester BZ, Kyoeisha Co., Ltd.), and bifunctional aliphatic urethane acrylate (EBECRYL230, Daicel Allnex Corporation). The matrices were prepared by varying the loading ratios between ~50% AH-600, ~90% 2-EH, ~30% 2-HEMA, ~30% Light Ester BZ, and ~90% EBECRYL230. The total amount of matrix was taken as 100, and the percentages of additives and fillers (unit: phr, per hundred resin) were added externally to prepare composite materials.

[0066] The additives used in this experiment were trimethylolpropane trimethacrylate (Light Ester TMP, Kyoeisha Chemical Co., Ltd.) as a crosslinker, 1,3,5-tris(2-(3-sulfanylbutanoyloxy)ethyl)-1,3,5-triazinane-2,4,6-trione (Karenz MT-NR1, Showa Denko K.K.) as a chain transfer agent, and t-amylperoxy-2-ethylhexanoate (Luperox 575, Arkema Yoshitomi Co., Ltd.) as a reaction initiator. Light Ester TMP was added at a rate of ~10 phr, Karenz MT-NR1 at a rate of ~5 phr, and the reaction initiator, Luperox 575, was fixed at 1 phr in all samples.

[0067] The fillers used in this experiment were, for example, spherical alumina particles (Alnabeads CB-A20S, Showa Denko K.K.) as the spherical filler, chopped carbon fiber (DIALEAD K223HE, Mitsubishi Chemical Corporation) as the rigid rod-like filler, and single-walled carbon nanotubes (SG101, Zeon Corporation) as the flexible fibrous filler. The alumina particles were added at ~30 phr, the chopped carbon fiber at ~30 phr, and the single-walled carbon nanotubes at ~0.5 phr. The filler dispersion conditions were fixed at 10,000 rpm for 20 minutes, and the thermal curing conditions for the acrylic cured resin composite material were 100°C for 1 hour for all samples.

[0068] The matrix is ​​not limited to the above, but may be a thermoplastic resin such as polyethylene, polypropylene, polyamide, polystyrene, acrylic resin, vinyl resin, or polyester; a biodegradable polymer such as polycaprolactone, polybutylene succinate, polylactic acid, polyvinyl alcohol, or polyhydroxybutyrate; an elastomer; an epoxy resin; a thermosetting polyester; a polyurethane; a polyimide; a heat- or light-curable resin having a vinyl group, a (meth)acrylic group, an epoxy group, or an oxetane; or a polymer alloy containing multiple high-molecular polymers. The matrix may also contain a solvent or the like made of a low-molecular-weight compound. The matrix may be in a solid or liquid state, but may also be in a slurry, paste, electrolyte, or other state. The state and efficacy of the matrix may be adapted to change based on environmental conditions.

[0069] The filler is not limited to the above, and may be carbon fiber, glass fiber, single-walled carbon nanotube, multi-walled carbon nanotube, graphene, graphene oxide, reduced graphene oxide, carbon black, cellulose nanofiber, cellulose nanocrystal, wood fiber, talc, montmorillonite, mica, silica, alumina, titania, calcium carbonate, hexagonal boron nitride, aluminum nitride, zeolite, metal oxide, metal powder, magnetic material, etc. The filler may be a single substance or a substance containing a combination of at least two or more substances. The state and efficacy of the filler may be subject to change based on environmental conditions.

[0070] The additives are not limited to those listed above, and may include crosslinking agents, crosslinking accelerators, scorch inhibitors, chain transfer agents, antioxidants, thermal decomposition inhibitors, antihydrolysis agents, antiozonants, weather resistant agents, light stabilizers, UV absorbers, antistatic agents, compatibilizers, tackifiers, plasticizers, lubricants, sliding agents, surfactants, crystal nucleating agents, crystallization accelerators, crystallization retarders, flame retardants, foaming agents, foaming aids, antibacterial agents, fungicides, pigments, fluorescent agents, fragrances, etc. The additives may be not only a single substance but also a substance containing a combination of at least two or more substances. The state and efficacy of the additive may be subject to change based on environmental conditions.

[0071] As described above, the property prediction device 10 can predict the properties of an unknown composition based on machine learning that integrates complex material information. The property prediction device 10 can generate property information of a composition based on integrated information that integrates physical structure information of the composition generated using image data and chemical structure information of the composition generated using the image data and a different type of spectral data. The property prediction device 10 performs machine learning based on multimodal AI that can integrate and determine different types of information, thereby enabling the calculation of property information of the composition with higher accuracy than when machine learning based on a single type of information is performed to calculate property information of the composition.

[0072] Although one embodiment of the present invention has been described above, the present invention is not limited to the above embodiment and can be modified as appropriate without departing from the spirit of the present invention. Although the physical property prediction device 10 of the embodiment exemplifies a polymer composite material as a composition, the present invention is not limited to this and may also handle other composite materials having complex structures.

[0073] 10 Physical property prediction device 12 Calculation unit 16 Output unit

Claims

1. A calculation unit that generates physical property information of a composition based on input data; an output unit that outputs the physical property information, The calculation unit is generating physical structure information of the composition to be predicted based on a first neural network using the first material information as the input data; generating chemical structure information of the composition based on a second neural network using second material information of a type different from the first material information as the input data; generating integrated information by integrating the physical structure information and the chemical structure information, and generating the physical property information of the composition based on the integrated information; Physical property prediction device.

2. the first material information includes image data indicative of a physical structure of the composition; The calculation unit generates the physical structure information based on the image data. The physical property prediction device according to claim 1 .

3. the second material information includes data that quantifies characteristics of a chemical structure of the composition; The calculation unit generates the chemical structure information based on the data. The physical property prediction device according to claim 2 .

4. The calculation unit generates the physical property information including indicators of at least three types of physical properties of the composition. The physical property prediction device according to claim 3 .

5. The calculation unit is performing a first machine learning process using the first neural network by using the image data for learning; Adjusting a first learning model constituting the first neural network by the first machine learning; generating the physical structure information based on the adjusted first learning model; The physical property prediction device according to claim 4 .

6. The calculation unit is performing a second machine learning process using the second neural network by using the spectral data for training; Adjusting a second learning model constituting the second neural network by the second machine learning; generating the chemical structure information based on the adjusted second learning model; The physical property prediction device according to claim 5 .

7. The calculation unit is performing a third machine learning using a third neural network using the integrated information for learning that is an integration of the calculation result of the first machine learning and the calculation result of the second machine learning; Adjusting a third learning model constituting the third neural network by the third machine learning; generating the physical property information based on the adjusted third learning model; The physical property prediction device according to claim 6 .

8. The calculation unit generates the physical property information of the composition including at least two of a matrix, a filler, and an additive for constituting the composition. The physical property prediction device according to claim 7 .

9. A computer constituting a property prediction device for generating property information of a composition based on input data, generating physical structure information of the composition to be predicted based on a first neural network using the first material information as the input data; generating chemical structure information of the composition based on a second neural network using second material information of a type different from the first material information as the input data; generating integrated information by integrating the physical structure information and the chemical structure information, and generating physical property information of the composition based on the integrated information; Executing a process of outputting the generated physical property information. Methods for predicting physical properties.

10. A computer constituting a property prediction device for generating property information of a composition based on input data, generating physical structure information of the composition to be predicted based on a first neural network using the first material information as the input data; generating chemical structure information of the composition based on a second neural network using second material information of a type different from the first material information as the input data; generating integrated information by integrating the physical structure information and the chemical structure information, and generating physical property information of the composition based on the integrated information; Executing a process of outputting the generated physical property information; program.