WO2023224012A1 - Physical property prediction device, physical property prediction method, and program - Google Patents

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

The present invention relates to a physical property prediction device, a physical property prediction method, and a program for predicting physical property information of a composition such as a composite material.
This application claims priority based on Japanese Patent Application No. 2022-081299 filed in Japan on May 18, 2022, the contents of which are incorporated herein.

In recent years, research has been progressing on methods to utilize 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 material development period. Among these, in the development of new materials such as carbon nanotubes, a method has been proposed that uses images obtained through experiments as training data and performs machine learning based on a multilayer neural network to predict unknown physical properties of compositions. (For example, see Non-Patent Document 1).

In the composition prediction method described in Non-Patent Document 1, the information to be captured is limited to images showing the structure of the composition. For general compositions, information indicating physical properties is not necessarily included in the image; physical properties such as chemical properties, optical properties, and molecular structure also exist. Furthermore, although there are various combinations of elements that make up the structure of composite materials, there is no clear definition, and it is difficult to handle structural information based on images or numerical values. Therefore, according to the composition prediction method described in Non-Patent Document 1, physical property information is predicted for compositions with complex structures such as composite materials formed based on multiple materials and structures. I couldn't deal with that.

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

One aspect of the present invention includes a calculation unit that generates physical property information of the composition based on input data, and an output unit that outputs the physical property information, and the calculation unit includes first material information as the input data. to generate physical structure information of a composition to be predicted based on a first neural network; a physical property prediction device that generates chemical structure information of the composition, 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. It is.

According to the present invention, the physical properties of an unknown composition such as a composite material can be predicted based on machine learning that integrates complex material information.

FIG. 1 is a block diagram showing the configuration of a physical property prediction system. It is a flowchart which shows the process flow of the machine learning process performed in a physical property prediction device. It is a flowchart which shows the process flow of the virtual experiment process performed in a physical property prediction device. FIG. 3 is a diagram showing experimental results using a physical property prediction device. It is a figure showing an example of image data generated by a physical property prediction device. It is a figure showing an example of spectrum data generated by a physical property prediction device.

Hereinafter, embodiments of a physical property prediction device according to the present invention will be described with reference to the drawings. The physical property prediction device predicts a new complex material having multiple physical property information based on machine learning using a neural network.

As shown in FIG. 1, the physical property prediction system 1 includes, for example, a physical property prediction device 10 communicably 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 communicably connected to the physical property prediction device 10 via the communication network NW. The server device 20 may 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 a flash memory. 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 experimental data to the physical property prediction device 10 via the communication network NW.

The physical property prediction device 10 is configured by, for example, a terminal device such as a personal computer for the purpose of information processing. The physical property prediction device 10 may be composed of two or more terminal devices that can cooperate to perform processing. The physical property prediction device 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 calculation, and an output unit that outputs information such as calculation results of the calculation unit 12. 16, and an acquisition unit 18 that acquires data necessary for calculation via the communication network NW.

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

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

The acquisition unit 18 includes a communication interface that can communicate with the communication network NW. The acquisition unit 18 performs communication via the communication network NW using, for example, a cellular network, a Wi-Fi network, Bluetooth (registered trademark), DSRC (Dedicated Short Range Communication), or the like. The acquisition unit 18 communicates with the server device 20 via the communication network NW, for example, and acquires the experimental data stored in the server device 20.

Next, the processing executed in the calculation unit 12 will be explained. In the following description, the composition is, for example, a polymeric composite. The composition is not limited to polymer composite materials, but may be other materials. The calculation unit 12 performs a learning process of performing deep learning based on a neural network using experimental data acquired from the server device 20, and a virtual experiment process of performing a virtual experiment using input data and outputting physical property information of the composition. Execute. For example, the calculation unit 12 performs machine learning using different types of measurement data to improve the ability to calculate predicted values of physical property information of the composition based on input data.

The calculation unit 12, for example, performs machine learning using multimodal AI (Artificial Intelligence) based on the 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 processing that outputs advanced determination results using multiple different types of information. According to multimodal AI, it is possible to output calculated values depending on the purpose using different types of input data such as images and numerical data. The calculation unit 12 performs a virtual experiment using, for example, image data of the composition and measured values of the substance, and calculates predicted values of physical property information of the composition.

For example, the calculation unit 12 is configured to generate physical structure information of a composition to be predicted based on a first learning model based on a first neural network using first material information including image data among input data. It is composed of Physical structure information is information representing the mixing/dispersion state of substances, composition distribution, etc. in a composition. Further, the calculation unit 12 is 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 a different type of numerical value from the first material information. has been done. Chemical structure information is information representing the chemical structure of substances constituting the composition, their types, numbers, 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.

The calculation unit 12 calculates different types of data, such as image data showing the structure of the composition, measured values based on infrared absorption spectrometry (IR) of the composition, and measured values including numerical values based on Raman spectroscopy. Using the input data, predict the predicted values of at least three types of physical property indicators of the composition, such as Young's modulus, strength, elongation at break, glass transition temperature, density, electrical resistance, storage modulus, and loss tangent. calculate. The physical property information is not limited to the eight indicators described above, and may include other physical property indicators.

The calculation unit 12 executes a learning process based on machine learning in order to improve the accuracy of generating physical property information of the composition. The calculation unit 12 selects data belonging to a predetermined type of category from among 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 in past experiments. The first material information is, for example, image data showing the structure of a polymer composite material created from a matrix, additives, and fillers blended based on a predetermined blending ratio.

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

The calculation unit 12 executes first machine learning based on a first neural network using first material information stored in advance. The first neural network is configured with a first learning model used for calculations to calculate output values. The first learning model is configured, for example, to calculate a predicted value of the physical structure information of the composition upon input of the first material information. Initial parameters for calculating predicted values are set in the first learning model.

The calculation unit 12 repeatedly executes a process of calculating an output value for the 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 the correct value. do. The calculation unit 12 improves the accuracy of the first learning model based on the process of performing the first machine learning described above. The first learning model is, for example, a model that can perform supervised learning and unsupervised learning using the first material information.

When the input data is small, the calculation unit 12 performs unsupervised learning using the first neural network based on the first material information. When there is a large amount of input data, the calculation unit 12 may perform supervised learning using the first material information using 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. A GAN is composed of two networks: a generation network (generator) that generates data and a discrimination network (discriminator) that identifies the authenticity of the generated data.

The generation network generates image data of the composition based on the first material information. The identification network identifies the authenticity of the generated image data. The generative network generates noise-added image data that learns to fool the identification network. The identification network learns to make the generated image data more accurate. According to GAN, learning accuracy is improved by competition between a generation network and an identification network. According to GAN, it is possible to perform unsupervised learning to learn the characteristics of an object even if correct data is not provided. According to GAN, by learning features from data, it is possible to generate data that does not exist, and it is possible to transform data that does not exist in accordance with the features of the data.

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

In addition to executing the first machine learning for calculating the physical structure information of the composition described above, the calculation unit 12 executes 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.

The calculation unit 12 selects data belonging to a predetermined type of category from among the input data input to the server device 20, and generates second material information of a different type 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 blending ratio and manufacturing conditions of a polymer composite material produced from a matrix, additives, and fillers blended based on a predetermined blending ratio. The second material information includes, for example, spectral data quantifying the characteristics of the chemical structure of the composition.

The second material information includes, for example, infrared spectroscopy (IR), Raman spectroscopy, ultraviolet/visible absorption spectroscopy (UV-VIS), near infrared (NIR) spectroscopy, and far Infrared (FIR) spectroscopy, terahertz (THz) spectroscopy, nuclear magnetic resonance (NMR), X-ray photoelectron spectroscopy (XPS), energy dispersive X-ray spectroscopy ( Chemical structure of compositions for energy dispersive spectrometry (EDS), electron energy loss spectroscopy (EELS), fluorescence spectroscopy, mass spectrometry (MS), electron spin resonance measurement (ESR), etc. The data may include chemical structure information that quantifies the properties of. The second material information may be data such as a molecular structure such as the arrangement of atoms and chemical bonds, its text representation (Simplified Molecular Input Line Entry System: SMILES), and molecular descriptors and fingerprints. Further, the second material information may include information on chromatography, thermal analysis, odor measurement, etc., as long as it indirectly includes chemical structure information. The second material information may be generated by combining the above data.

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

The calculation unit 12 repeatedly executes a process of calculating an output value for the 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 the correct value. do. The calculation unit 12 improves the accuracy of the second learning model based on the process of performing the second machine learning described above. The second learning model is, for example, a model that can perform supervised learning and unsupervised learning using the second material information.

When the input data is small, 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 execute processing based on a generative adversarial network in unsupervised learning using a second neural network, for example.

The calculation unit 12 adjusts the parameters of the second learning model through second machine learning using the second neural network in the second machine learning step described above. The calculation unit 12 may perform one or more types of second machine learning based on different types of second material information. The calculation unit 12 calculates chemical structure information of the composition using the second learning model with adjusted parameters.

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 the third neural network. The third neural network is configured by a third learning model that calculates physical property information of the composition based on the integrated information.

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

When multiple pieces of chemical structure information based on numerical data such as spectra exist in the generation of integrated information, the calculation unit 12 repeatedly performs machine learning using a third neural network to extract features included in the numerical data. , classified by characteristics. The calculation unit 12 converts the classified numerical data into one-dimensional vector data in which feature amounts are digitized. The calculation unit 12 integrates physical structure information based on one-dimensional vector data and chemical structure information based on one-dimensional vector data, and generates integrated information based on one-dimensional vector data.

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

The calculation unit 12 not only executes machine learning using a third neural network based on deep learning in generating integrated information, but also performs 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 Features included in a plurality of pieces of chemical structure information based on numerical data may be extracted using an analysis method such as Regression (GPR) or gradient boosting regression, and classified for each feature. Furthermore, before performing machine learning in generating integrated information, the calculation unit 12 processes numerical data using calculation methods such as forward method, backward method, stepwise method, and genetic algorithms (GA). Variables may be selected based on a plurality of pieces of physical structure information and chemical structure information.

FIG. 2 shows a flowchart of the process flow of the machine learning process of the physical property prediction method executed in the physical property prediction device 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 input data. The calculation unit 12 acquires first material information including image data when performing the first machine learning (step S102).

The calculation unit 12 repeatedly performs first machine learning using the first neural network using the image data for learning (step S104). The calculation unit 12 adjusts the parameters of the first learning model constituting the first neural network by first machine learning, and generates physical structure information based on the adjusted first learning model (step S106).

The calculation unit 12 acquires second material information including spectral data when executing the second machine learning (step S108). The calculation unit 12 repeatedly performs second machine learning using the second neural network using the learning spectrum data (step S110). The calculation unit 12 adjusts the parameters of the second learning model constituting the second neural network by second machine learning, and generates chemical structure information based on the adjusted second learning model (step S112).

The calculation unit 12 integrates the generated physical structure information and chemical structure information (step S114). The calculation unit 12 uses the integrated information to perform third machine learning using the third neural network (step S116). The calculation unit 12 uses the integrated information to adjust parameters of a third learning model that constitutes the third neural network, and repeatedly executes third machine learning that generates physical property information based on the adjusted third learning model. .

FIG. 3 shows a flowchart of a process flow of a virtual experiment of a physical property prediction method executed in the physical property prediction device 10. Input data regarding a new composition 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 of at least two of the matrix, filler, and additive in an unknown blending ratio. The calculation unit 12 generates first material information based on image data of the input data. The calculation unit 12 acquires the generated first material information (step S202).

The calculation unit 12 repeatedly performs the first virtual experiment using the first learning model in which the parameters of the first neural network are 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 spectrum data of the input data. The calculation unit 12 acquires the generated second material information (step S208). The calculation unit 12 repeatedly performs the second virtual experiment using the second learning model in which the parameters of the second neural network are adjusted based on the second material information (step S210).

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 uses the integrated information to generate physical property information of the composition based on the third neural network (step S216). The calculation unit 12 generates physical property information based on the adjusted third learning model of the third neural network using the integrated information. The calculation unit 12 changes the input data and repeats the virtual experiment to generate physical property information of the unknown composition. The calculation unit 12 stores the generated physical property information in the storage unit 14 and creates a database. The calculation unit 12 controls the output unit 16 to output physical property information (step S218).

FIG. 4 shows the results of an experiment that verified the accuracy of the physical property information generated by the physical property prediction device 10. In experiments, the accuracy of physical property information for eight properties was predicted when weight fractions were given to the composition of a certain composition. The 8 properties are, for example, Young's modulus, tensile strength, elongation at break, logarithmic electrical resistance, 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 eight characteristics. The input data includes 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 image data (optical image) of first material information. ) and the combination of spectral data (IR) and spectral data (Raman) of the second material information were used. Calculated values of physical property information based on a neural network using only image data (optical images) as comparison targets, calculated values of physical property information based on a neural network using only spectral data (IR), and only spectral data (Raman) Calculated values of physical property information based on a neural network were calculated. For accuracy comparison, physical property information calculated using learning data and physical property information calculated using newly generated test data were used.

As shown in FIG. 4(A), when learning data is used, 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 are both Accuracy 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 physical property information calculated based on multiple neural networks of the physical property prediction device 10 was Accuracy was high.

As shown in FIG. 4(B), the physical property predicted values for the eight properties by the physical property predicting device 10 are stable when calculated based on multiple neural networks, whereas the numerical values are stable when calculated based on multiple neural networks. The physical property information calculated based on the network varied in numerical values. The weight fraction and eight characteristics of the composition described above are not limited to these, and may be changed and increased or decreased. Further, the physical property information may be changed to other conditions to include different parameters, or other conditions may be added.

As described above, according to the physical property prediction device 10, the physical properties of an unknown composition can be accurately predicted by performing machine learning that integrates different and complex material information such as image data and spectral data.

In the above experiment, the input data was expanded to generate a large amount of learning data, making it possible to perform deep learning that generates more accurate predicted values. For example, when data augmenting image data, if the original image is approximately 1000 x 1000 pixels, machine learning may be performed depending on the performance of the computer configuring the physical property prediction device 10 and the amount of data if the original image is of approximately 1000 x 1000 pixels. is difficult.

Image data obtained through normal observation has a data size exceeding 1000 x 1000 pixels per piece, so if deep learning is performed using the entire image as it is, it will overwhelm the computer's memory and increase the size of the neural network. do. Furthermore, the amount of experimental data that can be obtained through actual experiments is limited, and a sufficient number of learning data cannot be obtained before performing deep learning. Therefore, the image data was set to an appropriate scale (observation magnification) and pixel number that reflected the physical structure of the composition so that the first machine learning in the physical property prediction device 10 was executed smoothly. In addition, the image data was processed by rotation, tilting, adding noise, etc., and data expansion was performed.

The image data was acquired based on observation at a low magnification (100x) that allows the physical structure of the material to be broadly captured in deep learning, and was divided into small images of 128 x 128 pixels. The small images were divided to include enough information on the physical structure of fillers such as particles to enable deep learning to classify the features. Furthermore, data expansion was performed on a small image of 128 x 128 pixels by rotating it in four directions every 90 degrees. This does not put pressure on the computer's memory, prevents the size of the neural network from increasing, and enables deep learning education by obtaining a sufficient number of learning data from limited experimental data, thereby stabilizing the learning of the neural network. , the effect of improving accuracy was obtained.

The image data was also pre-processed for first machine learning based on the first neural network. One pixel of image data is a 1-byte (8-bit) unsigned integer of type uint8, and has a value range of 0-255. If this data remains as it is, the training of the first neural network, which generates an image by inputting the material information of the composition, will not proceed correctly. The image data was preprocessed by dividing the luminance value of each pixel by 127.5 and subtracting by 1 so that the data range was within the range of -1 to 1. Based on the pre-processing of the image data, the learning process was stabilized in the generative model of the first neural network.

As shown in FIG. 5, according to the physical property prediction device 10, by performing the second machine learning based on the data-enhanced image data, it is possible to output generated image data equivalent to the measured image data. .

In a neural network that generates an image by inputting material information of a composition, learning has conventionally been performed by inputting one-hot vectors. In general object recognition, learning is performed by inputting vectors such as 1 for cat - 0 for dog, and 0 for cat - 1 for dog. In the case of compositions in the materials field, the input vector becomes longer depending on the sample types used for learning, and when the number of sample types increases, there is a problem in that the memory of the computer is overwhelmed. Therefore, by training a neural network using the composition's material information (continuous variables such as blending ratios and manufacturing conditions) as input, we were able to generate data directly from the material information, thereby reducing computer memory. This time, we constructed a generative model using as input the blending ratio of the acrylic cured resin composite material converted to a weight fraction in the 0-1 range.

Unlike image data, there is no established data expansion method for spectral data. Therefore, the spectral data was expanded by applying Gaussian noise so that the second machine learning in the physical property prediction device 10 could be executed smoothly. The spectral data had a data length of 1024 by selecting, for example, a wave number range of 1800 cm-1 or less. The spectral data was preprocessed so that the absorbance value was normalized by the maximum value and minimum value within one sample so that the data fell within the range of 0-1.

The spectral data is numerical data in the range of 0-1 because it is normalized absorbance using the minimum and maximum values. Here, in order to train the generative model, the normalized absorbance of each wavenumber is divided by 0.5 and subtracted by 1, so that the preprocessing is performed to keep the data range within the range of -1 to 1. It was conducted. This stabilized the learning process of the second neural network in the second learning model.

When learning was performed using a model such as Conditional table GAN as a generative model, the learning and generation behavior of the model was unstable. Therefore, by transforming the spectral data with a data length of 1024 into a 32 x 32 two-dimensional data array and having the neural network of the generative model learn it in the same way as images, the learning process of the neural network in the generative model was stabilized.

The spectral data was expanded by generating Gaussian noise with an average of 0 and a standard deviation of 0.001, and adding noise with a different pattern to the original spectral data, resulting in 128 times expanded learning data. was generated.

As shown in FIG. 6, based on the expanded spectrum data, it was possible to generate generated spectrum data (chemical structure information) that was close to the actually measured value of the spectrum data based on the second neural network. In the above experiment, by performing data expansion, the number of times deep learning is executed in the physical property prediction device 10 increases, and the learning process of the second neural network is stabilized compared to the case without data expansion, and the estimation Accuracy increased.

In spectrum generation, if a one-hot vector is used as an input in a generative model, the input vector becomes longer by the number of sample types used for learning, and when the number of sample types increases, there is a problem in that the memory of the computer is overwhelmed. Therefore, by training a second neural network using the composition's material information (continuous variables such as blending ratios and manufacturing conditions) as input, we were able to generate data directly from the material information, resulting in the effect of reducing computer memory. . This time, we constructed a generative model using as input the blending ratio of the acrylic cured resin composite material in the weight fraction range of 0-1.

The spectral data used included numerical arrays consisting of corresponding compounding ratios, images as physical structure information, and IR spectra and Raman spectra as chemical structure information for 75 acrylic cured resin composite materials with different compounding ratios. .

In addition, in the above experiment, the machine learning model was able to adjust the number of layers and layer structure by determining the effect of changing the number of layers, layer structure, etc. on the processing behavior in the learning process and the accuracy that appears in the calculation results. It was done. In addition, in GAN learning, the learning process cannot be executed smoothly unless learning appropriate to the model structure is performed, so the GAN learning process was adjusted according to the model structure.

In the process of generating integrated information, it is necessary to combine information with different dimensions, such as images and spectra. For example, it is possible to perform machine learning by converting 128x128 pixel image data into a 16384 pixel one-dimensional vector and processing it in parallel with one-dimensional spectral data. In this case, spatial features such as the relative arrangement of materials included in the image data are impaired, so that the neural network cannot sufficiently learn physical structure information. Therefore, from two-dimensional image data, features are extracted based on a neural network that includes a convolution layer that captures information about relative placement with neighboring pixels, and then integrated with features obtained from spectral data. Physical property information of the composition was calculated via a fully connected neural network. This resulted in the effect of learning processing without losing physical structure information and chemical structure information.

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, normal temperature storage modulus, normal temperature loss tangent, and surface electrical resistance. For the physical property information, values were used that were preprocessed by standardization using the average and standard deviation. In the physical property prediction device 10, a neural network that integrates physical structure information and chemical structure information was constructed using these learning data. By inputting numerical values of material composition using a trained neural network, physical property information including indicators of eight properties was obtained as output.

The matrices used in this experiment include, for example, phenyl glycidyl ether acrylate hexamethylene diisocyanate urethane prepolymer (AH-600, Kyoeisha Chemical Co., Ltd.), 2-ethylhexyl methacrylate (2-EH, Tokyo Kasei Kogyo Co., Ltd.), methacrylate Acid 2-hydroxyethyl ethylene glycol methacrylate (2-HEMA, Tokyo Kasei Kogyo Co., Ltd.), benzyl methacrylate (Light Ester BZ, Kyoeisha Co., Ltd.), difunctional aliphatic urethane acrylate (EBECRYL230, Daicel Ornex Co., Ltd.), be. The matrix was prepared by changing the addition ratio in the range of ~50% for AH-600, ~90% for 2-EH, ~30% for 2-HEMA, ~30% for Light Ester BZ, and ~90% for EBECRYL230. Created. The total amount of matrix at this time was set as 100, and the proportion of additives and fillers was added externally (unit: phr, per hundred resin) to produce a composite material.

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

The fillers used in this experiment include, for example, spherical alumina particles (Alunabeads CB-A20S, Showa Denko K.K.) as a spherical filler, carbon fiber chopped fiber (DIALEAD K223HE, Mitsubishi Chemical Corporation) as a rigid rod filler, and flexible fibrous fillers. Single-walled carbon nanotubes (SG101, Nippon Zeon Co., Ltd.) were used as fillers. Alumina particles were added at ~30 phr, chopped carbon fibers at ~30 phr, and single-walled carbon nanotubes at ~0.5 phr. In addition, the filler dispersion conditions were fixed at a rotation speed of 10,000 rpm for 20 minutes, and the thermosetting conditions for the acrylic cured resin composite material were 100°C for 1 hour for all samples.

The matrix is not limited to the above, but may include thermoplastic resins such as polyethylene, polypropylene, polyamide, polystyrene, acrylic resin, vinyl resin, polyester, polycaprolactone, polybutylene succinate, polylactic acid, polyvinyl alcohol, polyhydroxybutyrate, etc. It may be a biodegradable polymer, an elastomer, an epoxy resin, a thermosetting polyester, a polyurethane, a polyimide, a thermo/photocurable resin having a vinyl group, a (meth)acrylic group, an epoxy group, an oxetane, etc. It may also be a polymer alloy containing a high molecular weight polymer. Further, the matrix may contain a solvent made of a low molecular compound, or the like. The matrix is not limited to a solid or liquid state, but may be in a slurry, paste, electrolyte, or the like. The state and efficacy of the matrix may be changed based on environmental conditions.

Fillers include, but are not limited to, carbon fibers, glass fibers, single-walled carbon nanotubes, multi-walled carbon nanotubes, graphene, graphene oxide, reduced graphene oxide, carbon black, cellulose nanofibers, cellulose nanocrystals, wood fibers, talc, and montmorillonite. , mica, silica, alumina, titania, calcium carbonate, hexagonal boron nitride, aluminum nitride, zeolite, metal oxide, metal powder, magnetic material, and the like. The filler may be a substance containing not only a single substance but also a combination of at least two or more substances. The state and efficacy of the filler may be changed based on environmental conditions.

Additives are not limited to the above, but include crosslinking agents, crosslinking accelerators, scorch inhibitors, chain transfer agents, antioxidants, thermal decomposition inhibitors, anti-hydrolysis agents, ozone deterioration inhibitors, anti-climate agents, and photoprotectants. Stabilizers, ultraviolet absorbers, antistatic agents, compatibilizers, tackifiers, plasticizers, lubricants, sliding agents, surfactants, crystal nucleating agents, crystallization promoters, crystallization retarders, flame retardants, It may be a foaming agent, a foaming aid, an antibacterial agent, a fungicide, a pigment, a fluorescent agent, a perfume, or the like. Further, the additive may be a substance containing not only a single substance but also a combination of at least two or more substances. The condition and efficacy of the additive may be varied based on environmental conditions.

As described above, according to the physical property prediction device 10, the physical properties of an unknown composition can be predicted based on machine learning that integrates complex material information. According to the physical property prediction device 10, integrated information is obtained by integrating the physical structure information of the composition generated using image data and the chemical structure information of the composition generated using the image data and different types of spectral data. Physical property information of the composition can be generated based on the following. According to the physical property prediction device 10, by performing machine learning based on multimodal AI that can integrate and determine different types of information, machine learning based on a single type of information can be performed to obtain physical property information of a composition. It is possible to calculate physical property information of the composition with higher accuracy than when calculating .

Although one embodiment of the present invention has been described above, the present invention is not limited to the above-described one embodiment, and can be modified as appropriate without departing from the spirit thereof. Although the physical property prediction device 10 of the embodiment uses a polymer composite material as an example of the composition, the composition is not limited to this, and other composite materials having a complicated structure may be handled.

10 Physical property prediction device 12 Arithmetic unit 16 Output unit

Claims (10)

1. a calculation unit that generates physical property information of the composition based on input data;
   an output unit that outputs the physical property information,
   The arithmetic unit is
   generating physical structure information of a composition to be predicted based on a first neural network using 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 different type 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 indicating the 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 spectral data quantifying the characteristics of the chemical structure of the composition,
   the calculation unit generates the chemical structure information based on the spectral data;
   The physical property prediction device according to claim 2.
4. the calculation unit generates the physical property information including at least three types of physical property indicators of the composition;
   The physical property prediction device according to claim 3.
5. The arithmetic unit is
   Performing first machine learning using the first neural network 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 arithmetic unit is
   performing second machine learning using the second neural network using the spectrum data for learning;
   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 arithmetic unit is
   performing third machine learning using a third neural network using the integrated information; 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 containing at least two of a matrix, a filler, and an additive for forming the composition.
   The physical property prediction device according to claim 7.
9. A computer constituting a physical property prediction device that generates physical property information of the composition based on input data,
   generating physical structure information of a composition to be predicted based on a first neural network using 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 different type from the first material information as the input data;
   generating integrated information that integrates 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;
   Physical property prediction method.
10. A computer constituting a physical property prediction device that generates physical property information of the composition based on input data,
    Generating physical structure information of a composition to be predicted based on a first neural network using 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 different type from the first material information as the input data;
    generating integrated information that integrates 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.

Images