Method for generating trained prediction model, method for predicting physical quantity, program, and computer-readable recording medium having trained prediction model recorded thereon

Abstract

To provide a learned prediction model generation method capable of highly accurately predicting the physical amount of an inhomogeneous material from a small number of molecule dynamic simulation results through short-time calculation.SOLUTION: A learned prediction model for corresponding to a prediction target part region of an inhomogeneous material to be predicted is generated by setting a plurality of learning part regions included in the entire region of a learning numerical simulation model including arrangement information of two or more different material qualities of the learning inhomogeneous material, executing numerical simulation for the entire region of a learning numerical simulation model to acquire a partial region physical amount for each of a plurality of learning partial regions, generating a learning data set including partial region arrangement information and the partial region physical amount for each of the learning partial regions, and causing a prediction model to execute machine learning by using the learning data sets corresponding to the learning partial regions.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to a method for generating a trained prediction model for physical quantities of heterogeneous materials, a physical quantity prediction method, a program for causing a computer to execute the trained prediction model generation method, a program for causing a computer to execute the physical quantity prediction method, and a computer-readable recording medium on which is recorded a trained prediction model created by the trained prediction model generation method. [Background technology]

[0002] To accelerate the development of fuel-efficient tires, it will be helpful to clarify the relationship between the energy loss (hysteresis) that accompanies deformation and nanostructure. Filler-filled rubber is a composite material of filler and polymer, and since filler morphology is one of the factors that control the material properties, large-scale molecular dynamics simulations that model the aggregate structure of the filler have been carried out to elucidate the mechanism. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2006-313522 A [Patent Document 2] JP 2019-91237 A [Patent Document 3] JP 2020-74095 A [Patent Document 4] International Publication No. 99 / 007543 Summary of the Invention [Problem to be solved by the invention]

[0004] Patent Document 1 discloses an equivalent material constant calculation system that divides a structure into a plurality of small regions and includes a means for calculating an equivalent material constant in each of the small regions, but this system does not use machine learning.

[0005] Patent Document 2 discloses a time series data generating device having a model generating unit that generates a probabilistic time development model by learning teacher time series data acquired from a deterministic time development model, and a time series data generating unit that generates time series data based on the probabilistic time development model.

[0006] Patent Document 3 discloses a prediction device that accurately predicts the physical properties of a polymer, but the regression model used by the prediction device is not obtained by using a simulation.

[0007] Patent Document 4 discloses a tire design method using a neural network, but the neural network is not constructed by learning using the results of a simulation.

[0008] In recent years, machine learning has been used to discover new knowledge and improve the efficiency of development. Machine learning is also effective in improving the efficiency of research and development using molecular dynamics simulations, but in large-scale simulations that model filler morphology, it is difficult to create sufficient learning data due to the computation time involved.

[0009] Therefore, the present invention aims to provide a trained prediction model generation method, a physical quantity prediction method, a program, and a computer-readable recording medium having the trained prediction model recorded thereon, which enable prediction of physical quantities of heterogeneous materials with high accuracy from a small number of molecular dynamics simulation results obtained by short-term calculations. [Means for solving the problem]

[0010] According to the present invention, A trained prediction model generation method for generating a trained prediction model for predicting a physical quantity of a heterogeneous material in which two or more different materials are arranged in a heterogeneous manner, comprising: A learning partial domain setting step of setting a plurality of learning partial domains included in the entire domain of the learning numerical simulation model, the learning partial domains including the arrangement information of two or more different materials in the learning heterogeneous material; a learning partial region physical quantity acquisition step of performing a numerical simulation on the entire region of the learning numerical simulation model to acquire partial region physical quantities for each of the plurality of learning partial regions; a learning dataset creation step of creating a learning dataset including partial region arrangement information and partial region physical quantities for each of the plurality of learning partial regions; a trained prediction model creation step of creating a trained prediction model for corresponding to a prediction target partial region of a heterogeneous material to be predicted by performing machine learning on a prediction model using a plurality of the training data sets corresponding to the plurality of training partial regions, respectively; A method for generating a trained predictive model having the following structure is provided: Further, according to the present invention, A physical quantity prediction method for predicting a physical quantity of a heterogeneous material in which two or more different materials are arranged heterogeneously, comprising the steps of: A prediction target partial region setting step of setting a plurality of prediction target partial regions in a heterogeneous material to be predicted; a partial region physical quantity prediction step of predicting a partial region physical quantity for each prediction target partial region based on partial region arrangement information of two or more different materials for each of the plurality of prediction target partial regions while using a trained prediction model for corresponding to each prediction target partial region; an entire region physical quantity acquisition step of acquiring a physical quantity of an entire region of the heterogeneous material of the prediction target based on the partial region physical quantity for each of the plurality of prediction target partial regions; having A method for predicting physical quantities is provided. Furthermore, according to the present invention, there is provided a program for causing a computer to execute the above-mentioned trained predictive model generation method. Furthermore, according to the present invention, there is provided a program for causing a computer to execute the above physical quantity prediction method. Furthermore, according to the present invention, there is provided a computer-readable recording medium having recorded thereon a trained prediction model created by the above-mentioned trained prediction model generation method. Effect of the Invention

[0011] According to the present invention, it is possible to predict the physical quantities of a heterogeneous material with high accuracy from the results of a small number of molecular dynamics simulations performed in a short time. [Brief description of the drawings]

[0012] [Figure 1] FIG. 1 is a conceptual diagram for explaining an overview of a trained prediction model generating method and a physical quantity predicting method according to an embodiment of the present invention. [Diagram 2] 1 is a flowchart for explaining a method for generating a trained predictive model according to an embodiment of the present invention. [Diagram 3] 13A shows an example in which learning partial regions are set from the entire region without overlapping, and FIG. 13B shows an example in which learning partial regions are set from the entire region so as to overlap. [Figure 4] FIG. 11 is a conceptual diagram showing a relationship between a regional range for acquiring learning partial region arrangement information and a regional range for acquiring learning partial region physical quantities in an embodiment of the present invention, or a relationship between a regional range for acquiring prediction target partial region arrangement information and a regional range for predicting prediction target partial region physical quantities. [Diagram 5] 1 is a flowchart illustrating a learned physical quantity prediction method according to an embodiment of the present invention. [Figure 6] 1 is a table for explaining the effects of the trained prediction model generation method and physical quantity prediction method according to an embodiment of the present invention. [Figure 7] 11 is a table for explaining the effect of using corresponding regions instead of partial regions in an embodiment of the present invention. [Figure 8] 1 is a table for explaining the effect of using layout information of an interface region in an embodiment of the present invention. [Figure 9] 1 is a table for explaining the effect of using image data to create a numerical simulation model in an embodiment of the present invention. [Figure 10]1 is a table for explaining the effect of increasing the number of training data sets in an embodiment of the present invention. [Figure 11] 13 is a table for comparing prediction errors and simulation times of physical quantities according to the embodiment of the present invention with prediction errors and simulation times of physical quantities according to Comparative Example 1 and Comparative Example 2. [Figure 12] 1 is a flowchart illustrating a simulation method according to an embodiment of the present invention. [Figure 13] FIG. 1 is a conceptual perspective view showing a simulation model used in an embodiment of the present invention. [Figure 14] FIG. 2 is a diagram for explaining a main chain and a cross-linked chain. [Figure 15] (a) is a schematic diagram for explaining bond stretch potential energy, (b) is a schematic diagram for explaining bending potential energy, (c) is a schematic diagram for explaining torsion potential energy, and (d) is a schematic diagram for explaining Lennard-Jones potential energy. [Figure 16] FIG. 1 is a conceptual diagram to explain non-bonded potential energy, such as the Lennard-Jones potential energy, between filler particles. [Figure 17] A conceptual diagram to explain non-bonded potential energy, such as the Lennard-Jones potential energy between polymer particles. [Figure 18] This figure shows that there is nonbonded potential energy other than the nonbonded potential energy shown in Figure 17. [Figure 19] FIG. 1 is a conceptual diagram to explain non-bonded potential energy, such as the Lennard-Jones potential energy, between a filler particle and a polymer particle. [Figure 20] FIG. 1 is a conceptual diagram to explain non-bonded potential energy, such as the Lennard-Jones potential energy between other filler particles and polymer particles. [Figure 21] FIG. 1 is a schematic diagram for illustrating non-bonded potential energy such as the Lennard-Jones potential energy between further filler particles and polymer particles. [Figure 22] FIG. 1 is a schematic diagram for illustrating non-bonded potential energy such as the Lennard-Jones potential energy between further filler particles and polymer particles. [Diagram 23] FIG. 14 is a conceptual perspective view for explaining a response analysis of the simulation model shown in FIG. [Figure 24] FIG. 1 is a functional block diagram showing a configuration of a trained prediction model generation device A according to an embodiment of the present invention. [Diagram 25] FIG. 1 is a functional block diagram showing a configuration of a learned physical quantity prediction device A according to an embodiment of the present invention. [Figure 26] FIG. 2 is a functional block diagram showing a configuration of a trained prediction model generation device B according to an embodiment of the present invention. [Figure 27] FIG. 2 is a functional block diagram showing a configuration of a learned physical quantity prediction device B according to an embodiment of the present invention. [Figure 28] FIG. 2 is a functional block diagram showing a configuration of a trained prediction model generation device C according to an embodiment of the present invention. [Figure 29] FIG. 2 is a functional block diagram showing a configuration of a learned physical quantity prediction device C according to an embodiment of the present invention. [Diagram 30] 1 is a functional block diagram of an analysis device according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0013] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings.

[0014] In an embodiment of the present invention, as shown in Fig. 1, a plurality of learning partial regions are set from a learning numerical simulation model corresponding to a learning heterogeneous material, and a learned prediction model is created to be applied to each prediction target partial region of the prediction target heterogeneous material by using a learning data set for each learning partial region, which includes arrangement information and physical quantities corresponding to each of the plurality of learning partial regions. Then, a plurality of prediction target partial regions are set from a prediction target simulation model corresponding to a prediction target heterogeneous material, and the physical quantities corresponding to each prediction target partial region are predicted by using the arrangement information of the prediction target partial region and the learned prediction model to be applied to each prediction target partial region. Then, the total value of the physical quantities corresponding to each prediction target partial region over the entire region of the prediction target simulation model is set as the physical quantity for the entire region of the prediction target heterogeneous material.

[0015] [First embodiment] The trained prediction model generation method according to the first embodiment is intended to generate a trained prediction model for predicting physical quantities of a heterogeneous material in which two or more different materials are arranged heterogeneously, and has the following steps as shown in Figure 2.

[0016] (Step A1) Creating a numerical simulation model for learning A learning numerical simulation model is created that includes arrangement information of two or more different materials in a learning heterogeneous material.

[0017] The learning numerical simulation model has a structure as shown in FIG. 13, as will be described later.

[0018] The configuration information represents a structure such as that shown in Fig. 13. The configuration information also represents a fine structure (morphology).

[0019] In the learning numerical simulation model, the heterogeneous material is represented by particles and the bonds connecting the particles. Generally, each particle has multiple bonds and multiple types of bonds connecting the particles. These are represented by configuration information. There is configuration information that corresponds to the entire region and that corresponds to a partial region, but this is the same for both types of configuration information.

[0020] As will be described later, the simulation that handles the learning numerical simulation model uses a molecular dynamics method.

[0021] By using the learning numerical simulation model, it is possible to carry out simulations of large deformations of filler-filled rubber.

[0022] Specifically, the maximum strain that can be handled by continuum simulation is about 50%, whereas the maximum strain that can be handled by molecular dynamics is about 500%.

[0023] Additionally, as will be described later, interfaces between different materials can be modeled by the thickness of the interacting region and the strength of the interaction.

[0024] [interface] The training numerical simulation model may include arrangement information regarding an interface region between two or more different materials. Here, an interface region is defined as a region that exists between two or more different materials. In this way, the training partial region arrangement information acquired from the training numerical simulation model can include arrangement information of the interface region. By digitizing the interface in this way, it is possible to investigate the effect of the filler type.

[0025] (Step A2) Learning sub-region setting step A plurality of learning partial regions are set, and learning partial region definition data for defining the set learning partial regions is generated for each learning partial region. This learning partial region definition data is data that defines which portion of the entire region the learning partial region occupies.

[0026] Referring to Fig. 3, the learning partial region is a part of the whole region. In the example of Fig. 3(a), when the size of the whole region is 1 x 1, the size of the learning partial region is 0.5 x 0.5. Therefore, as shown in Fig. 3(a), if there is no overlap between the learning partial regions, four partial regions can be set.

[0027] [Increase in number of training subregions] The plurality of learning partial regions may be set so as to include two or more learning partial regions that have different partial region arrangement information while sharing at least a portion of the area of ​​the entire region.

[0028] For example, a plurality of learning partial regions may be set so as to partially overlap each other. As shown in Fig. 3(b), in addition to the four learning partial regions shown in Fig. 3(a), a learning partial region that partially overlaps with these learning partial regions may be provided. In this way, even if the same learning numerical simulation model is used, the number of learning data sets each corresponding to each learning partial region can be increased.

[0029] In addition, the number of learning partial regions corresponding to the entire region of the same learning numerical simulation model may be increased by changing the image data creation direction (the slice direction of the structure in the case of three-dimensional data, or rotation or inversion in the case of two-dimensional data) or by using mirror conditions. In this case, two or more learning partial regions are generated that share a common area in the entire region but have different partial region arrangement information.

[0030] (Step A3) Partial region arrangement information acquisition step For each learning partial region, the layout information (learning partial region layout information) is extracted from the layout information over the entire region of the learning numerical simulation model based on the learning partial region definition data. The learning partial region arrangement information includes the coordinates of the polymer particles 21a and the coordinates of the filler particles 11a shown in FIG. 13, and these coordinates are represented by text data or CSV data indicating the numerical values ​​of the coordinates.

[0031] [Image data] For each training partial region, location information (training partial region location information) may be extracted from the training image data of the heterogeneous material. In this case, the arrangement information (learning partial region arrangement information) may be obtained directly from the image data of the learning heterogeneous material. Alternatively, a learning numerical simulation model may be created from the image data of the learning heterogeneous material, and the arrangement information (learning partial region arrangement information) may be obtained from this. In this case, the coordinates of the polymer particles 21a and the center coordinates of the filler particles 11a are obtained from the image data, and text data or CSV data indicating the numerical values ​​of the obtained coordinates are generated. The image data is obtained from a learning heterogeneous material using, for example, an electron microscope.

[0032] In this case, a learning numerical simulation model may be created from image data of the learning heterogeneous material, for example, from image data of an electron microscope photograph or the like. When there are two types of constituent materials, such as filler and polymer, grayscale image data may be used. Here, the image data is in accordance with a predetermined format. Also, instead of the image data, other types of data obtained from the image data (for example, numerical data, character data, etc.) may be used.

[0033] (Step A4) Partial domain physical quantity acquisition step A numerical simulation is performed on the entire domain of the learning numerical simulation model to obtain partial domain physical quantities (learning partial domain physical quantities) corresponding to each of a plurality of learning partial domains. The partial region physical quantity acquisition step includes the following two steps.

[0034] (Step A4-1) Acquisition of physical quantities for the entire region For the entire region of the learning numerical simulation model, for example, a numerical simulation according to the third embodiment is performed to obtain an overall physical quantity corresponding to the entire learning numerical simulation model. At this time, the distribution of the overall physical quantity in the entire region is also obtained.

[0035] Here, the distribution of the physical quantity over the entire region corresponds to each particle and each bond connecting the particles in the learning numerical simulation model. The particles referred to here are, for example, the filler particles 11a contained in the filler models 11A, 11B, 11C, and 11D shown in Fig. 13, and the filler particles 11a contained in the polymer model 121, and the bonds referred to here are, for example, bonds corresponding to the bond chains such as the main chain 21b and the cross-linked bond chain 21c in the polymer model, bonds corresponding to the interactions acting between the filler particles 11a, and bonds corresponding to the interactions acting between the filler particles 11a and the polymer particles 21a.

[0036] (Step A4-2) Extraction of partial domain physical quantities for learning For each learning partial region, a physical quantity (learning partial region physical quantity) corresponding to the learning partial region defined by the learning partial region definition information is extracted from the distribution of the physical quantity over the entire region. Also, the distribution of the learning partial region physical quantity in the learning partial region is extracted.

[0037] Here, the distribution of the physical quantities over the training partial domain corresponds to each particle and each bond connecting the particles in the training numerical simulation model, i.e., each particle is associated with a physical quantity related to various bonds involving the particle.

[0038] Here, the physical quantity of the learning subdomain to be calculated is either or both of energy and stress.

[0039] The shapes of the learning partial regions and the prediction target partial region described below are preferably cubic or rectangular parallelepiped.

[0040] It is preferable that the learning partial regions and the prediction target partial regions described below have the same size.

[0041] It is preferable that the shapes of the learning partial regions and the prediction target partial regions described below are all the same.

[0042] Since the smaller the statistical error is, the higher the accuracy is, it is preferable that the learning subregion and the prediction target subregion described below have each side length of 100 nm or an equivalent length or more.

[0043] If the difference between the entire region and the learning partial region or the prediction target partial region described below is small, the variation in physical quantities of these partial regions will be small, so it is preferable that the dimensions of these partial regions be 80% or less of the entire region.

[0044] When the learning partial region and the prediction target partial region described later include inclusions with extremely high rigidity, such as fillers, it is preferable to set 10 or more learning partial regions and prediction target partial regions from the entire region.

[0045] (Step A5) Learning dataset creation step Next, a learning data set is created, which includes arrangement information (learning partial region arrangement information) corresponding to each learning partial region and partial region physical quantities (learning partial region physical quantities, including distribution in the learning partial region).

[0046] Note that steps A1 to A4 may be performed for a plurality of pieces of heterogeneous training material, and a plurality of training data sets each corresponding to a respective piece of heterogeneous training material may be collected across the plurality of pieces of heterogeneous training material.

[0047] [Area range for acquiring learning subdomain arrangement information and area range for acquiring learning subdomain physical quantities] For at least some of the training data sets, the regional range for acquiring partial region arrangement information (training partial region arrangement information) included in each training data set is made the same as the regional range for acquiring the partial region physical quantity (training partial region physical quantity) included in the training data set. Specifically, the regional range for acquiring the training partial region defined by the training partial region definition data 357 is selected as a common range for the training partial region arrangement information and the training partial region physical quantity. In other words, for at least some of the training partial regions, the regional range for acquiring the partial region arrangement information for each training partial region is made the same as the regional range for acquiring the partial region physical quantity corresponding to the prediction target partial region.

[0048] Also, for example, as shown in FIG. 4, for at least some of the training data sets, the regional range for acquiring partial region arrangement information (training partial region arrangement information) included in each training data set is set to include and be wider than the regional range for acquiring the partial region physical quantity (training partial region physical quantity) included in the training data set. Specifically, the range of the training partial region defined by the training partial region definition data 357 is selected as the regional range for acquiring the training partial region physical quantity. Then, a range including and wider than that range is selected as the regional range for acquiring the training partial region arrangement information. In other words, for at least some of the training partial regions, the regional range for acquiring the partial region arrangement information for each training partial region is set to be wider than the regional range for acquiring the partial region physical quantity corresponding to the training partial region.

[0049] In the case of the latter combination of selections as shown in Figure 4, when obtaining a training partial region physical quantity corresponding to the training partial region from the simulation results, the influence of the configuration information in the vicinity of the training partial region can be taken into account. Furthermore, when creating data from a large-scale simulation, the size effect of the large-scale model can be taken into account.

[0050] Generally, the range of the learning partial region itself is selected as the regional range for acquiring the learning partial region physical quantity.

[0051] Therefore, typically, when the area range for acquiring the learning partial area placement information and the area range for acquiring the learning partial area physical quantity are the same, the range of the learning partial area itself becomes the area range for acquiring the learning partial area physical quantity and the area range for acquiring the learning partial area placement information.

[0052] Furthermore, in general, when the area range for acquiring the learning partial region arrangement information is wider than the area range for acquiring the learning partial region physical quantity, the area range of the learning partial region itself becomes the area range for acquiring the learning partial region physical quantity, and a range including the area range of the learning partial region and wider than it becomes the area range for acquiring the learning partial region arrangement information.

[0053] [interface] The arrangement information on the interface region between two or more different materials may not be included in the training numerical simulation model, and instead, the arrangement information on the interface region between two or more different materials may be directly included in the training partial region arrangement information. In other words, the training data set itself may include the arrangement information on the interface region between two or more different materials. An interfacial region is defined herein as existing between two or more dissimilar materials. By digitizing the interface in this way, it is possible to investigate the effect of the filler type.

[0054] (Step A6) Trained predictive model creation step By subjecting the predictive model to machine learning using multiple learning data sets corresponding to multiple learning sub-domains, a trained predictive model is created to correspond to the prediction target sub-domains of the heterogeneous material to be predicted, as described below.

[0055] It should be noted that the partial region to be predicted here is a partial region having the same size and shape as the partial region for learning.

[0056] [Second embodiment] The physical quantity prediction method according to the second embodiment predicts the physical quantities of a heterogeneous material in which two or more different materials are arranged heterogeneously, and includes the following steps as shown in FIG.

[0057] (Step B1) Creating a numerical simulation model for the prediction target A numerical simulation model of a prediction target for numerical simulation is created, which includes arrangement information of two or more different materials in a heterogeneous material of a prediction target.

[0058] Like the learning numerical simulation model, the prediction target numerical simulation model has a structure as shown in FIG. 13, as will be described later. The heterogeneous material to be predicted is represented in the numerical simulation model by particles and bonds connecting the particles. Furthermore, the partial domain physical quantity prediction step described later may be performed on the premise that the simulation handling the prediction target numerical simulation model uses a molecular dynamics method. By using a numerical simulation model for the prediction target, it is possible to predict physical quantities based on simulations up to large deformations of filler-filled rubber.

[0059] Specifically, the maximum strain that can be handled by continuum simulation is about 50%, whereas the maximum strain that can be handled by molecular dynamics is about 500%.

[0060] Additionally, as will be described later, interfaces between different materials can be modeled by the thickness of the interacting region and the strength of the interaction. The prediction target numerical simulation model has a similar configuration to the learning numerical simulation model in the first embodiment, and can be used as a simulation target. In the second embodiment, other numerical models may be used instead of the prediction target numerical simulation model.

[0061] [interface] An interface region between two or more different materials may be defined, and the layout information of the interface region may also be included in the prediction target numerical simulation model. In this way, the layout information of the interface can be included in the prediction target partial region layout information acquired from the prediction target numerical simulation model.

[0062] [Image data] A numerical simulation model of the prediction target may be created based on image data of the heterogeneous material of the prediction target.

[0063] (Step B2) Prediction target partial region setting step A plurality of partial regions to be predicted are set, and prediction target partial region definition data for defining the set prediction target partial regions is generated for each prediction target partial region. This prediction target partial region definition data is data that defines which part of the entire region the prediction target partial region occupies.

[0064] The multiple prediction target partial regions are set, for example, so that all the prediction target partial regions occupy the entire region of the prediction target numerical simulation model when collected together. However, this is not limited to the above, and for example, the prediction target partial regions may be set so as to partially overlap each other, so that all the prediction target partial regions occupy the entire region of the prediction target numerical simulation model in a double or multiple manner when collected together.

[0065] (Step B3) Partial region arrangement information acquisition step Layout information (prediction target partial region layout information) corresponding to each of the plurality of prediction target partial regions is obtained.

[0066] For each prediction target partial region, the placement information (prediction target partial region placement information) is extracted from the placement information over the entire region of the prediction target numerical simulation model based on the prediction target partial region definition data. The prediction target partial region placement information includes the coordinates of the polymer particles 21a and the coordinates of the filler particles 11a shown in FIG. 13, and these coordinates are represented by text data or CSV data indicating the numerical values ​​of the coordinates.

[0067] [Image data] For each prediction target partial region, configuration information (prediction target partial region configuration information) may be extracted from image data of the heterogeneous material to be predicted. In this case, the placement information (prediction target partial region placement information) may be obtained directly from image data of the heterogeneous material to be predicted. Alternatively, a prediction target numerical simulation model may be created from image data of the heterogeneous material to be predicted, and the placement information (prediction target partial region placement information) may be obtained from this. In this case, the coordinates of the polymer particles 21a and the center coordinates of the filler particles 11a are obtained from the image data, and text data or CSV data indicating the numerical values ​​of the obtained coordinates are generated. The image data is obtained from a learning heterogeneous material using, for example, an electron microscope.

[0068] [interface] An interface region between two or more different materials may be defined, and the layout information of the interface region may be directly included in the layout information of the prediction target partial region. By digitizing the interface in this way, it is possible to investigate the effect of the filler type.

[0069] (Step B4) Partial domain physical quantity prediction step Using the trained prediction model, a partial region physical quantity (prediction target partial region physical quantity) corresponding to each of a plurality of prediction target partial regions is predicted based on the arrangement information (prediction target partial region arrangement information) corresponding to the prediction target partial region. Also, a distribution of the prediction target partial region physical quantity in the prediction target partial region is extracted.

[0070] Here, since the physical quantities are predicted based on the particles constituting the numerical simulation model to be predicted and the bonds connecting the particles, the distribution of the physical quantities over the prediction target partial region is on a particle-by-particle basis at the bond destination. Each particle is associated with a physical quantity corresponding to the various bonds in which the particle is involved.

[0071] The trained prediction model used here for matching with the prediction target partial region is generated by the trained prediction model generation method according to the first embodiment. However, this is not limiting, and a trained prediction model generated by another method may be used for matching with the prediction target partial region.

[0072] For each prediction target partial region, the layout information (prediction target partial region layout information) is extracted from the layout information over the entire region of the prediction target numerical simulation model based on the prediction target partial region definition data.

[0073] [Area range for obtaining the prediction target sub-domain layout information and the area range for predicting the prediction target sub-domain physical quantities] For each partial region to be predicted, a prediction target data set can be defined as including partial region to be predicted definition data, partial region to be predicted arrangement information, and physical quantity to be predicted. At this time, for example, for at least some of the prediction target data sets, the regional range for acquiring the partial region arrangement information (prediction target partial region arrangement information) included in each prediction target data set is made the same as the regional range for predicting the partial region physical quantity (prediction target partial region physical quantity) included in the prediction target data set. Specifically, the range of the prediction target partial region defined by the prediction target partial region definition data is selected as the regional range for acquisition or the regional range for prediction common to the partial region arrangement information and the partial region physical quantity. In other words, for at least some of the prediction target partial regions, the range of the partial region arrangement information corresponding to each prediction target partial region is made the same as the regional range for acquiring the partial region physical quantity corresponding to the prediction target partial region.

[0074] Also, for example, as shown in Fig. 4, for at least some of the prediction target data sets, the regional range for acquiring the partial region arrangement information (prediction target partial region arrangement information) included in each prediction target data set is set to include the regional range for predicting the partial region physical quantity (prediction target partial region physical quantity) included in the prediction target data set and is set to be wider than the regional range. Specifically, the range of the prediction target partial region defined by the prediction target partial region definition data is selected as the regional range for predicting the prediction target partial region physical quantity. Then, a range including the range and wider than the range is selected as the regional range for acquiring the prediction target partial region arrangement information. In other words, for at least some of the prediction target partial regions, the acquisition range of the partial region arrangement information for each prediction target partial region is set to be wider than the regional range for predicting the partial region physical quantity for the prediction target partial region.

[0075] In the latter combination of selections as shown in FIG. 4, when obtaining a prediction target partial region physical quantity corresponding to the prediction target partial region from the prediction target partial region layout information and the learned prediction model, the influence of layout information in the vicinity of the prediction target partial region can be taken into account.

[0076] In addition, when creating data from a large-scale simulation, the size effect of the large-scale model can be taken into account, i.e., it can be made to correspond to a trained model created using a large-scale simulation.

[0077] Generally, the range of the prediction target partial domain itself is selected as the domain range for predicting the prediction target partial domain physical quantity.

[0078] Therefore, in general, when the areal range for acquiring the prediction target partial region layout information and the areal range for predicting the prediction target partial region physical quantity are the same, the range of the prediction target partial region itself becomes the areal range for predicting the prediction target partial region physical quantity and the areal range for acquiring the prediction target partial region layout information.

[0079] Furthermore, in general, when the areal range for acquiring the prediction target partial region arrangement information is wider than the areal range for predicting the prediction target partial region physical quantity, the range of the prediction target partial region itself becomes the areal range for predicting the prediction target partial region physical quantity. Then, a range including the prediction target partial region and wider than it becomes the areal range for acquiring the prediction target partial region arrangement information.

[0080] (Step B5) Whole-domain physical quantity prediction step The physical quantities of the entire region of the heterogeneous material to be predicted are predicted based at least on the prediction target partial region physical quantities and prediction target partial region definition data corresponding to each of a plurality of prediction target partial regions included in the heterogeneous material to be predicted.

[0081] In step B4, the physical quantities of each prediction target partial region are predicted based on the prediction target partial region arrangement information while using the trained prediction model corresponding to each prediction target partial region. In step B5, the predicted prediction target partial region physical quantities are summed for all required prediction target partial regions to obtain the physical quantities of the entire region of the heterogeneous material to be predicted.

[0082] For example, if the entire region can be filled without any excess or deficiency by connecting all of the prediction target partial regions, it is possible to predict the physical quantities for the entire prediction target by adding up the prediction target partial region physical quantities corresponding to all of the prediction target partial regions.

[0083] Alternatively, the prediction target partial region physical quantities corresponding to all prediction target partial regions may be summed up, with the prediction target partial region being taken as a unit. At this time, definition data of each partial region to be predicted is also used in order to know where each partial region to be predicted is located in the entire region of the numerical simulation model to be predicted.

[0084] Alternatively, the physical quantities of all the partial regions to be predicted may be summed up while taking into consideration the distribution of the physical quantities in each partial region to be predicted. In this case, for each particle indicated by the placement information of the partial region to be predicted, it is necessary to sum up the physical quantities corresponding to each bond involving that particle, so the definition data, placement information, and physical quantities (for each bond) of each partial region to be predicted are used.

[0085] It is also possible to convert discrete continuous data calculated from different times / deformation states, or the difference between them, into data, which makes it possible to obtain stress-strain curves, hysteresis curves, and hysteresis data.

[0086] Discrete numerical data may be used as the output values, or a curve connecting these may be converted into image data and used as the output data.

[0087] [Example 1] <Effect of using subregions on prediction error> As shown in Fig. 6, for example, when the stress value calculated by simulation with a strain of 10% is used as a reference, the prediction error is 20% when the entire region is treated as one piece of data as in the conventional case. In contrast, when the region is divided into partial regions according to the embodiment of the present invention, the prediction error is 10%, which is an improvement of 50 points compared to the conventional case.

[0088] [Example 2] <Effects of expanding the area range for acquiring placement information> As shown in FIG. 7, when the regional range for acquiring partial region arrangement information is the same as the regional range for acquiring partial region physical quantities, the prediction error of the physical quantities is 10%.

[0089] In contrast, when the area range for obtaining sub-area placement information is made wider than the area range for obtaining sub-area physical quantities so that there is a surplus of about 5 percent per side for the input to the trained prediction model creation unit, the prediction error of the physical quantities is 8%.

[0090] Therefore, by making the area range for obtaining configuration information wider than the area range for obtaining physical quantities for input to the learned prediction model creation unit (machine learning unit), the prediction error was improved by 20 points.

[0091] [Example 3] <Effects of adding interface area placement information> Referring to Fig. 8, the relative numerical value (e.g., the relative numerical value of strain) of the simulation results using the molecular dynamics method is 1 for filler 1, which is less affected by the interface, and 1.2 for filler 2, which is more affected by the interface. In other words, the numerical value becomes larger when the interface is more affected. Here, the relative numerical value is based on the numerical value for filler 1, which is less affected by the interface. The same applies to Fig. 8 below.

[0092] Then, when a prediction model is generated without including the arrangement information of the interface region of the learning heterogeneous material in the learning partial region arrangement information, and further, when the physical quantity prediction method according to the second embodiment is executed without including the arrangement information of the interface region of the heterogeneous material to be predicted in the prediction target partial region arrangement information, the value is 1 for filler 1, which is less affected by the interface, and also 1 for filler 2, which is more affected by the interface.

[0093] In contrast, when a prediction model is generated by including the arrangement information of the interface region of the learning heterogeneous material in the learning partial region arrangement information, and further including the arrangement information of the interface region to be predicted in the prediction target partial region arrangement information, and a physical quantity prediction method according to the second embodiment is executed, the result is 1 for filler 1, which is less affected by the interface, and 1.2 for filler 2, which is more affected by the interface.

[0094] Therefore, it can be seen that by including the arrangement information of the interface region of the heterogeneous material to be learned in the learning partial region arrangement information, and by including the arrangement information of the interface region of the heterogeneous material to be predicted in the prediction target partial region arrangement information, more accurate prediction results for physical quantities can be obtained.

[0095] [Example 4] <Effects of using image data for location information> Referring to Figure 9, when the learned prediction model generation method is executed using the center coordinates of the filler as the coordinates of the filler in a learning numerical simulation model, and further when the physical quantity prediction method according to the second embodiment is executed using the center coordinates of the filler as the coordinates of the filler in a prediction target numerical simulation model, the prediction error of the physical quantity is about 20%. In contrast, when image data of the learning heterogeneous material is used as the learning partial region arrangement information and image data of the prediction target heterogeneous material is used as the prediction target partial region arrangement information, the prediction error of the physical quantity is approximately 8%.

[0096] Therefore, the prediction error is improved by 60 points by using image data.

[0097] [Example 5] <Effect of increasing the number of training datasets corresponding to the same heterogeneous training materials> As explained above in the section [Increase in the number of training subregions], it is possible to increase the number of training data sets corresponding to the same training heterogeneous material.

[0098] 10, according to one example, when the ratio of the number of training partial domains acquired from the entire domain of the same training numerical simulation model is 1:12, the prediction error becomes 1.2:1. In other words, when the number of training partial domains is increased by 12 times, the prediction error when the trained prediction model generating method according to the first embodiment and the physical quantity prediction method according to the second embodiment are executed can be improved by 17 points.

[0099] [Example 6] <Effect of using subdomains on prediction error and simulation time> Referring to Fig. 11, the stress at 10% elongation was calculated by simulation using 10 learning numerical simulation models. Here, molecular dynamics was performed with a 50 million particle model for the calculation. The calculation time for each model, including model creation, took one month.

[0100] In the method of Comparative Example 1, when creating a prediction model, the entire area of ​​the learning numerical simulation model was set as the learning area.

[0101] In addition, when predicting physical quantities, the entire area of ​​the prediction target numerical simulation model was set as the prediction target area.

[0102] For the heterogeneous material to be predicted, in which the stress at 10% elongation was maximum, the predicted stress had an error of 60% based on the stress calculated by simulation.

[0103] In the method of Comparative Example 2, similarly to Comparative Example 1, when creating a prediction model, the entire region of the learning numerical simulation model was set as the learning region.

[0104] However, it differs from Comparative Example 1 in that the number of data points was increased from 10 to 80 by expanding the data using a mirror model. Moreover, similarly to Comparative Example 1, when predicting the physical quantity, the entire region of the prediction target numerical simulation model was set as the prediction target region. For the heterogeneous material to be predicted, in which the stress at 10% elongation is maximized, the predicted stress had an error of 50% based on the stress calculated by simulation.

[0105] In the method of this embodiment, when creating a prediction model, a partial region of the learning numerical simulation model was used as the learning region. As a result, the number of data items increased from 10 to 10,000. In other words, the number of data items increased 1,000 times. Note that 1 / 8 of the entire region was used as the learning partial region. Also, a learning data set was created from 125 learning partial regions.

[0106] In addition, when predicting the physical quantities, each partial domain included in the entire domain of the prediction target numerical simulation model was set as the prediction target domain.

[0107] For the heterogeneous material to be predicted, in which the stress at 10% elongation is maximized, the predicted stress had an error of 20% based on the stress calculated by simulation.

[0108] Therefore, when the error in the predicted stress is taken as the standard for the method of Comparative Example 1, the method of Comparative Example 2 improves by 17 points, and the method of the present embodiment improves by 67 points.

[0109] In addition, the simulation time required to create the same number of learning data (here, as an example, 10,000 pieces) was 1,000 in the method of Comparative Example 1 and 125 in the method of Comparative Example 2, when compared with the time required in the method of the embodiment.

[0110] Therefore, in terms of the simulation time required to create the same amount of learning data, when using Comparative Example 1 as a reference, Comparative Example 2 improved by 87.5 points, and the method of the embodiment improved by 99.9 percent.

[0111] [Third embodiment] Fig. 12 is a diagram showing an example of a flow of a simulation method for a composite material according to an embodiment. The simulation method shown in Fig. 12 is for analyzing a composite material by a molecular dynamics method using a computer. That is, the simulation method is executed by a computer.

[0112] The method is applied to several different types of composite materials.

[0113] As shown in FIG. 12, this simulation method for a composite material mainly includes a step (S101) of creating a simulation model of the composite material, a step (S103) of providing input to the simulation model and performing response analysis using molecular dynamics, and a step (S105) of calculating values ​​of predetermined physical properties based on the response analysis.

[0114] Here, the composite material to be simulated includes a first substance and a second substance. The first substance is, for example, a polymer (polymer material). The second substance is, for example, filler particles added to a raw material material. The composite material preferably has a structure in which the first substance is a base material and the second substance is distributed in the base material. In the following description, the first substance is a polymer and the second substance is filler particles as an example. In addition to the form of the composite material described below, the composite material may be in the form of a blend polymer composed of multiple types of polymers. For example, a blend polymer having a sea-island structure or a lamellar structure may be used. In the case of the sea-island structure, the first substance may be treated as a polymer that constitutes the sea part of the sea-island structure, and the second substance may be treated as a polymer that constitutes the island part of the sea-island structure. The blend polymer may be composed of a crystalline polymer and a non-crystalline polymer. In this case, the first substance may be treated as a non-crystalline polymer and the second substance as a crystalline polymer. Furthermore, the composite material may have a hard segment phase and a soft segment phase in one molecule, like a thermoplastic elastomer, in which case the first substance can be treated as the soft segment phase and the second substance as the hard segment phase.

[0115] The types of composite materials differ not only by the types of molecules that compose them and the number of each molecule per unit volume, but also by the microstructure (morphology) of the filler, such as its dispersion state. The microstructure (morphology) of the filler, such as its dispersion state, can be defined by factors such as the filler radius, filler concentration, filler dispersion radius, boundary layer thickness, covariance, and cumulative particle size distribution.

[0116] (Creating a simulation model) In step S101 of creating a simulation model of a composite material, a simulation model of the composite material is created that includes a first substance model that models a first substance in the composite material and a second substance model that models a second substance in the composite material. For this purpose, for example, a method described in JP 2017-220137 A is used.

[0117] FIG. 13 is a conceptual diagram showing an example of a simulation model of a composite material. As shown in FIG. 13, the simulation model 1 is, for example, a particle model created in a model creation area, which is a virtual space of a substantially cubic shape. The model creation area is a three-dimensional space extending in the X-axis, Y-axis, and Z-axis directions perpendicular to each other. The simulation model 1 includes four filler models 11A, 11B, 11C, and 11D modeled with a plurality of filler particles 11a, and four polymer models 21 modeled with a plurality of polymer particles 21a and a main chain 21b. The filler models 11A, 11B, 11C, and 11D are collectively described as the filler model 11. Note that in the example shown in FIG. 13, an example in which the simulation model 1 is modeled with four filler models 11A, 11B, 11C, and 11D will be described, but there is no limit to the number of filler models to be modeled. The simulation model 1 may include three or less filler models 11, or may include more than four filler models 11. 13 shows only four polymer models 21, but a plurality of polymer models 21 exist throughout the entire model creation region in the simulation model 1. In the example shown in Fig. 13, the model creation region is a virtual space having a substantially rectangular parallelepiped shape, but the model creation region may have any shape, such as a sphere, an ellipse, a rectangular parallelepiped, or a polyhedron.

[0118] The filler model 11 is modeled in a state where a plurality of filler particles 11a are aggregated into approximately spherical bodies. The filler models 11 are arranged at a predetermined distance from each other. The plurality of filler models 11 may be mutually aggregated and connected to each other at their outer edges by a covalent bond chain (not shown).

[0119] The filler particles to be modeled include, for example, carbon black particles, silica particles, and alumina particles. The filler particles 11a are modeled as an aggregate of a plurality of atoms that constitute the filler. A filler particle group in which a plurality of filler particles 11a are aggregated is formed as the filler models 11A, 11B, 11C, and 11D.

[0120] The relative positions of the filler particles 11a are determined by a bond chain (not shown) between the filler particles 11a. The bond chain functions as a spring with a defined equilibrium length, which is the bond distance between the filler particles 11a, and a defined spring constant, and binds the filler particles 11a. The bond chain defines the relative positions of the filler particles 11a and the potential energy that generates a force due to twisting, bending, etc. The filler model 11 is defined by numerical data including the mass, volume, diameter, and initial coordinates of the filler particles 11a, the number of aggregated particles, etc., for treating the filler by molecular dynamics. The numerical data of the filler model 11 is input to a computer.

[0121] Examples of polymers modeled in the polymer model 21 include rubber, resin, and elastomer. The polymer particle 21a is a model of an assembly of multiple polymer atoms. A group of polymer particles in which multiple polymer particles 21a are linked by a bond chain is formed as the polymer model 21. That is, the polymer model 21 has a configuration in which multiple polymer atoms and polymer particles 21a, which are an assembly of multiple polymer atoms, are linked to each other by a bond chain, and this polymer model 21 is arranged at a predetermined density in the model creation area. The bond chain functions as a spring with, for example, a defined equilibrium length and spring constant. The polymer particles 21a are linked by the main chain 21b between the multiple polymer particles 21a to specify their relative positions. The main chain 21b functions as a spring with a defined equilibrium length, which is the bond distance between the polymer particles 21a, and a defined spring constant, and restrains each polymer particle 21a. The main chain 21b has the relative positions of the polymer particles 21a and the potential energy that generates a force due to twisting, bending, etc. defined.

[0122] Fig. 14 is a diagram showing an example of a cross-linking bond used in the embodiment. As shown in Fig. 14, cross-linking chains 21c are provided between polymer particles 21a of three polymer models 21. The cross-linking chains 21c function as springs with equilibrium length, which is the bond distance between the polymer particles 21a, and spring constants defined, and bind the polymer particles 21a.

[0123] In addition, a modifier that enhances the affinity with the filler is mixed into the polymer as necessary. Examples of the modifier include a hydroxyl group, a carbonyl group, and a functional group of an atomic group. In correspondence with the modifier, a particle model that models the modifier particles and a bond chain (not shown) are arranged between the polymer model 21 and the filler model 11.

[0124] This polymer model 21 is defined by numerical data (including the mass, volume, diameter, and initial coordinates of polymer particles 21a) for treating the polymer by molecular dynamics. The numerical data of the polymer model 21 is input to the computer as parameters.

[0125] In the simulation model 1, interactions are provided between at least some of the filler particles 11a, between the polymer particles 21a, and between the filler particles 11a and the polymer particles 21a. In some cases, an interaction force that exchanges force may be provided between all of the particles. The interaction between the filler particles 11a and the polymer particles 21a may be a chemical interaction (attractive force) or a physical interaction (bond bond).

[0126] By applying potential energy (described later) to the bond chains including the main chain 21b and the cross-linked bond chains 21c shown in the figure, and also to particle models not connected by bond chains, interactions are imparted. As a result, a force determined by the interaction acts between the particle models.

[0127] In the composite material, the polymer may be composed of a plurality of types of polymers, and in this case, an interaction may be given between different types of polymer particles 21a in the simulation model 1. In this case, the interaction between the filler particles 11a and the polymer particles 21a may be made different between the different types of polymer particles 21a.

[0128] The interaction between particles is defined, for example, by the Lennard-Jones potential energy shown in the following formula. At this time, the values ​​of σ and ε in the following formula are appropriately adjusted. By increasing the upper limit distance (cutoff distance) for calculating the potential energy, the force acting over a long distance can be adjusted. It is preferable to gradually decrease the parameters of the interaction between the filler particles 11a and the interaction between the polymer particles 21a until the interaction between the filler particles 11a and the interaction between the polymer particles 21a reach a constant value. By gradually approaching the σ and ε of the Lennard-Jones potential energy from a large value to the original value, the particles can approach each other at a gentle speed that does not lead the molecules to an unnatural state. In addition, by gradually decreasing the cutoff distance, the force in the interaction can be adjusted within an appropriate range.

[0129] U(r) = 4 ε [(σ / r) p -(σ / r) q ] (p and q are positive numbers) The above potential energies can be summarized in the molecular structure model shown in Figure 15.

[0130] Referring to Figure 15, the molecular structure model has a three-dimensional structure, and as shown in Figures 15(a) to (c), the bond length r, which is the length of the bond (also called "bond") between each particle 21a, 21a, the bond angle θ, which is the angle between three adjacent particles 3, as shown in Figure 15(b), and the torsion φ, which is the angle between a first plane P1 formed by three adjacent particles and a second plane P2 formed by three particles 3 that are common to two of the particles, are defined.

[0131] In the molecular dynamics calculation, the interactions of the bond stretch potential energy (abbreviation: Ebs) between connected particles shown in Figure 15(a), the bending potential energy (abbreviation: Ebe) consisting of three consecutive particles 3 shown in Figure 15(b), the torsion potential energy (abbreviation: Eto) shown in Figure 15(c), and the Lennard-Jones potential energy (abbreviation: Evw) between particles that are not connected to each other as shown in Figure 14(d) are taken into account for the molecular structure model.

[0132] Bond stretching potential energy, bending potential energy and torsion potential energy are all bond potential energies.

[0133] In FIG. 16, the dashed lines indicate the Lennard-Jones potential energies, which are non-bonded potential energies between the filler particles 11A, 11B, 11C and 11D.

[0134] In Fig. 17, the broken line indicates the Lennard-Jones potential energy, which is the non-bonded potential energy between the polymer particles 21a. Fig. 17 is a schematic diagram, and in reality, non-bonded potential energy occurs between each of the polymer particles 21a, as shown in Fig. 18.

[0135] In FIG. 19, the broken line indicates the Lennard-Jones potential energy, which is the non-bonded potential energy between the filler particle 11A and each polymer particle 21a.

[0136] In FIG. 20, the broken line indicates the Lennard-Jones potential energy, which is the non-bonded potential energy between the filler particle 11C and each polymer particle 21a.

[0137] In FIG. 21, the broken line indicates the Lennard-Jones potential energy, which is the non-bonded potential energy between the filler particle 11D and each polymer particle 21a.

[0138] In FIG. 22, the broken line indicates the Lennard-Jones potential energy, which is the non-bonded potential energy between the filler particle 11B and each polymer particle 21a.

[0139] (Response analysis) Response analysis by molecular dynamics performed by providing an input to the simulation model 1 is an analysis for investigating how the response of the simulation model 1 changes over time, for example, it is a response analysis related to the behavior of the simulation model 1 when an input is provided to extend the simulation model 1. Fig. 23 is a diagram for explaining an example of response analysis of the simulation model 1 performed in one embodiment.

[0140] In the example of response analysis shown in FIG. 23, an input is given to the simulation model 1 so as to stretch it in the vertical direction, and at this time, a time-series analysis is performed to see how the polymer model 21 and filler models 11A to 11D move over time.

[0141] Since the polymer model 21 and the filler models 11A to 11D each have mass, when subjected to a force resulting from an input given to the simulation model 1, they start to move according to the equation of motion, but at this time, their movement is restricted by interactions and bond chains. By calculating such movements at predetermined time intervals, the temporal response is calculated.

[0142] For example, when a stepwise displacement is applied to the simulation model 1 as an input and the stretched state is maintained, the polymer model 21 and the filler models 11A to 11D move with time and eventually come to a certain stationary state. When the stepwise displacement is applied, a large stretching speed is applied. The displacement is applied to the simulation model 1 so as to achieve, for example, 200% or 300% stretching. Therefore, in the response analysis, the process in which the polymer model 21 and the filler models 11A to 11D move and then come to a stationary state can be analyzed in time series. The stress generated in the simulation model 1 can be calculated by calculating the force acting on the polymer model 21 and the filler models 11A to 11D at this time. In addition, the energy stored in the simulation model 1 can be calculated. Furthermore, the amount of energy dissipated from the simulation model 1 can be calculated by subtracting the stored energy from the input energy corresponding to the applied input. That is, the stress relaxation process can be calculated.

[0143] Such elongation includes uniaxial or biaxial elongation in the simulation model 1. In addition, the response analysis in the above embodiment is an analysis of extension deformation, but is not limited to the analysis of extension deformation. For example, as long as the response analysis is possible, an analysis in which the simulation model 1 is compressed or sheared may be performed. Also, an analysis in which at least two of the extension, compression, and shear deformations are combined may be performed.

[0144] Furthermore, the response analysis is not limited to a form in which a step-like input is applied and the relaxation response is analyzed, but may be a form in which a triangular wave or sine wave input (displacement) is applied and the vibration of the simulation model 1 at that time is analyzed (cyclic extension analysis). In the case of a triangular wave or sine wave input (displacement), it is preferable that the input frequency is set so as to correspond to the frequency in the actual use situation of the structure in which the composite material is used, and it is also preferable that the input level is set so as to correspond to the maximum strain, strain at the crack tip, or apparent strain in the actual use situation of the structure in which the composite material is used, from the viewpoint of evaluating the fracture characteristics of the composite material in the actual use situation of the actual structure.

[0145] In response analysis, the bond chains including the bond chains 21b and the cross-linked bond chains 21c of the polymer model 21 may have a length equal to or greater than a predetermined threshold value for breaking. When the interparticle distance is equal to or greater than the threshold value, according to one embodiment, a bond breaking calculation function that reduces at least one of the bond energy and bond strength of the interparticle bond in the case where the interparticle distance is less than the threshold value may be applied to the bond chain. Also, the bond breaking calculation function may not be applied.

[0146] Data may be generated to interpolate between discrete data constituting the calculated physical characteristic. For example, if discrete time responses are calculated, data may be generated to interpolate between them, or if discrete hysteresis curves are calculated, data may be generated to interpolate between them.

[0147] [Fourth embodiment] (Trained predictive model generation device A) Referring to FIG. 24, a trained prediction model generation device 300A according to the fourth embodiment of the present invention includes a training numerical simulation model creation unit 301, a training partial region setting unit 303, a training partial region arrangement information acquisition unit 305, a simulation execution unit (entire region physical quantity acquisition unit) 307, a training partial region physical quantity extraction unit 309, a training dataset creation unit 311 and a trained prediction model creation unit 313.

[0148] The learning numerical simulation model creation unit 301 generates a learning numerical simulation model 353 corresponding to the learning heterogeneous material.

[0149] The learning numerical simulation model creation unit 301 may generate a learning numerical simulation model 353 based on image data 351 of the learning heterogeneous material.

[0150] Furthermore, the learning numerical simulation model creation unit 301 may generate the learning numerical simulation model 353 not based on image data of the learning heterogeneous material. For example, the learning numerical simulation model 353 may be generated by a computer program. In this case, the image data 351 in FIG. 24 may be omitted.

[0151] The arrangement information of the interface region may be included in the learning numerical simulation model 353. The arrangement information of the interface region here may be based on the image data 351 of the learning heterogeneous material or on other data.

[0152] The learning partial region setting unit 303 is for setting a plurality of learning partial regions. To this end, the learning partial region setting unit 303 generates a plurality of learning partial region definition data 357 corresponding to the plurality of learning partial regions, respectively. Each of the learning partial region definition data 357 defines each learning partial region.

[0153] The learning partial region arrangement information acquisition unit 305 acquires a plurality of pieces of learning partial region arrangement information 359 each corresponding to each learning partial region, based on the learning numerical simulation model 353 and a plurality of pieces of learning partial region definition data 357 .

[0154] The simulation execution unit (entire area physical quantity acquisition unit) 307 executes a numerical simulation for the entire area of ​​the learning numerical simulation model 353 and acquires physical quantities (including distribution) 355 throughout the entire area of ​​the learning numerical simulation model.

[0155] The training partial region physical quantity extraction unit 309 acquires multiple training partial region physical quantities 361 each corresponding to each training partial region, based on physical quantity (including distribution) 355 over the entire region of the training numerical simulation model 353 and multiple training partial region definition data 357. The training partial region physical quantities 361 corresponding to each training partial region are one or a set for the training partial region, but may include a distribution of these quantities over the training partial region.

[0156] The training data set creation unit 311 creates a plurality of training data sets 363 based on a plurality of pieces of training partial region arrangement information 359 and a plurality of training partial region physical quantities 361. Each training data set 363 includes training partial region arrangement information 359B and a training partial region physical quantity 361B related to one training partial region. Here, for each training partial region, the training partial region arrangement information 359 and the training partial region arrangement information 359B are the same, and the training partial region physical quantity 361 and the training partial region physical quantity 361B are the same.

[0157] Note that, instead of including the placement information of the interface region in the learning numerical simulation model 353, the placement information of the interface region may be included in the learning partial region placement information 359, so that the placement information of the interface region is included in the learning partial region placement information 359B. Also, the placement information of the interface region may be directly included in the learning partial region placement information 359B.

[0158] The trained prediction model creation unit 313 trains the prediction model 365 to correspond to the prediction target partial region of the heterogeneous material to be predicted (i.e., creates the trained prediction model 365) by performing machine learning on the prediction model 365 using multiple training data sets 363 each corresponding to a multiple training partial region.

[0159] Next, a description will be given of the correspondence between each step of the trained prediction model generating method shown in FIG. 2 and each functional block of the trained physical quantity prediction device 300A shown in FIG.

[0160] The learning numerical simulation model creation unit 301 creates a learning numerical simulation model (step A1).

[0161] The learning partial region setting unit 303 executes a learning partial region setting step of setting a plurality of learning partial regions (step A2), thereby generating a plurality of learning partial region definition data 357.

[0162] The learning partial region arrangement information acquisition unit 305 executes a learning partial region arrangement information acquisition step of acquiring arrangement information (learning partial region arrangement information) 359 corresponding to each learning partial region from the learning numerical simulation model 353 (step A3).

[0163] The simulation execution unit (whole region physical quantity acquisition unit) 307 executes the whole region physical quantity acquisition step (step A4-1).

[0164] The learning partial region physical quantity extraction unit 309 executes a learning partial region physical quantity extraction step (step A4-2).

[0165] The learning data set creation unit 311 executes a learning data set creation step (step A5).

[0166] The trained prediction model creation unit 313 executes a trained prediction model creation step (step A6).

[0167] [Fifth embodiment] (Physical quantity prediction device A) With reference to FIG. 25 , a physical quantity prediction device 400A according to the fifth embodiment of the present invention includes a prediction target numerical simulation model creation unit 401, a prediction target partial region setting unit 403, a prediction target partial region arrangement information acquisition unit 405, a partial region physical quantity prediction unit 407, a prediction target data set creation unit 409, and an entire region physical quantity acquisition unit 411.

[0168] The prediction target numerical simulation model creation unit 401 generates a prediction target numerical simulation model 453 corresponding to the heterogeneous material of the prediction target.

[0169] The prediction target numerical simulation model creation unit 401 may generate a prediction target numerical simulation model 453 based on image data 451 of the heterogeneous material of the prediction target.

[0170] Furthermore, the prediction target numerical simulation model creation unit 401 may generate the prediction target numerical simulation model 453 without being based on image data of the heterogeneous material of the prediction target.

[0171] The numerical simulation model 453 to be predicted may include information on the location of the interface region. The information on the location of the interface region may be based on image data 451 of the heterogeneous material to be predicted, or may be based on other data.

[0172] The prediction target partial region setting unit 403 is for setting a plurality of prediction target partial regions. To this end, the prediction target partial region setting unit 403 generates a plurality of prediction target partial region definition data 455 corresponding to the plurality of prediction target partial regions, respectively. Each prediction target partial region definition data 455 defines each prediction target partial region.

[0173] The prediction target partial region arrangement information acquisition unit 405 acquires a plurality of prediction target partial region arrangement information 457 each corresponding to each prediction target partial region, based on the prediction target numerical simulation model 453 and a plurality of prediction target partial region definition data 455 .

[0174] The partial region physical quantity prediction unit 407 predicts the prediction target partial region physical quantity (including distribution) 459 corresponding to the prediction target partial region based on the prediction target partial region arrangement information corresponding to each prediction target partial region while using the trained prediction model 365 to be applied to each prediction target partial region (obtaining a predicted value of the prediction target partial region physical quantity (including distribution) 459 as the prediction target partial region physical quantity). By performing this for a plurality of prediction target partial regions, the prediction target partial region physical quantity corresponding to each of the plurality of prediction target partial regions is predicted (obtaining a predicted value of the prediction target partial region physical quantity as the prediction target partial region physical quantity).

[0175] The prediction target dataset creation unit 409 creates a plurality of prediction target datasets 461 based on a plurality of pieces of prediction target partial region definition data 455, a plurality of pieces of prediction target partial region arrangement information 457, and a plurality of prediction target partial region physical quantities (including distribution) 459. Each prediction target dataset 461 includes prediction target partial region definition data 455B, prediction target partial region arrangement information 457B, and prediction target partial region physical quantity (including distribution) 459B related to one prediction target partial region. Here, for each prediction target partial region, the prediction target partial region definition data 455 and the prediction target partial region definition data 455B are the same, the prediction target partial region arrangement information 457 and the prediction target partial region arrangement information 457B are the same, and the prediction target partial region physical quantity (including distribution) 459 and the prediction target partial region physical quantity (including distribution) 459B are the same.

[0176] The entire region physical quantity acquisition unit 411 acquires a physical quantity 463 for the entire region of the heterogeneous material to be predicted based on a plurality of prediction target data sets 461. Here, the physical quantity 463 may include a distribution of the physical quantity over the entire region.

[0177] Next, the correspondence between each step of the learned physical quantity prediction method shown in FIG. 5 and each functional block of the physical quantity prediction device 400A shown in FIG. 25 will be described.

[0178] The prediction target numerical simulation model creation unit 401 executes a prediction target numerical simulation model creation step to create a prediction target numerical simulation model 453 for numerical simulation, which includes arrangement information of two or more different materials in the heterogeneous material to be predicted (step B1).

[0179] The prediction target partial region setting unit 403 executes a prediction target partial region setting step of setting a plurality of prediction target partial regions (step B2), thereby generating a plurality of prediction target partial region definition data 455.

[0180] The prediction target partial region arrangement information acquisition unit 405 executes a prediction region corresponding arrangement information acquisition step of acquiring arrangement information (prediction target partial region arrangement information) 457 corresponding to each prediction target partial region from the prediction target numerical simulation model 453 (step B3).

[0181] Note that, instead of including the layout information of the interface region in the prediction target numerical simulation model 453, the layout information of the interface region may be included in the prediction target partial region layout information 457. The layout information of the interface included in the prediction target partial region layout information 457 is directly included in the prediction target partial region layout information 457B in the prediction target data set 461.

[0182] The partial region physical quantity prediction unit 407 executes a partial region physical quantity prediction step of predicting a partial region physical quantity (including distribution) 459 corresponding to the prediction target partial region, based on arrangement information (prediction target partial region arrangement information) 457 corresponding to each of the multiple prediction target partial regions, while using the learned prediction model 365 for association with the prediction target partial region (step B4).

[0183] The entire region physical quantity acquisition unit 411 executes an entire region physical quantity acquisition step of acquiring a physical quantity 463 of the entire region of the heterogeneous material to be predicted based on partial region physical quantities (including distribution) 459 corresponding to each of a plurality of partial regions to be predicted (step B5). The physical quantity 463 of the entire region is one or a set of physical quantities for the entire region, but may include a distribution over the entire region.

[0184] [Sixth embodiment] (Trained predictive model generation device B) Referring to FIG. 26, a trained prediction model generation device 300B according to the sixth embodiment of the present invention includes a training numerical simulation model creation unit 301B, a training partial region setting unit 303, a training partial region arrangement information acquisition unit 305B, a simulation execution unit (entire region physical quantity acquisition unit) 307, a training partial region physical quantity extraction unit 309, a training dataset creation unit 311 and a trained prediction model creation unit 313.

[0185] The learned prediction model generation device 300B differs from the learned prediction model generation device 300A in that the learning numerical simulation model creation unit 301 is replaced with a learning numerical simulation model creation unit 301B, and the learning partial region placement information acquisition unit 305 is replaced with a learning partial region placement information acquisition unit 305B.

[0186] The learning numerical simulation model creation unit 301B generates the learning numerical simulation model 353 based on the image data 351 of the learning heterogeneous material. However, the learning numerical simulation model creation unit 301B may generate the learning numerical simulation model 353 without based on the image data 351 of the learning heterogeneous material. For example, the learning numerical simulation model 353 may be generated by a computer program. The image data 351 may be created by a computer so as to represent, for example, a morphology to be used for learning.

[0187] The learning numerical simulation model 353 may include layout information of the interface region. The layout information of the interface region here may be based on the learning image data 351 of the heterogeneous material or on other data.

[0188] The learning partial region arrangement information acquisition unit 305B acquires a plurality of pieces of learning partial region arrangement information 359, each of which corresponds to each learning partial region, based on the learning image data 351 and a plurality of pieces of learning partial region definition data 357. The learning partial region arrangement information 359 may be a part of the learning image data 351, or may be data generated based on the part of the data.

[0189] Furthermore, the arrangement information of the interface region may be included in the multiple pieces of learning partial region arrangement information 359. The arrangement information of the interface region here is related to the learning heterogeneous material, but it does not matter on what basis it is acquired or generated.

[0190] When partial image data acquired from the image data 351 is used as the multiple pieces of learning partial region arrangement information 359, the learning numerical simulation model creation unit 301B may generate the learning numerical simulation model 353 based on the image data 351, or may generate the learning numerical simulation model 353 without based on the image data 351. In either case, however, the image data 351 and the learning numerical simulation model 353 are associated with each other.

[0191] The other functional units are similar to the corresponding functional units of the trained prediction model generation device A according to the fourth embodiment, so duplicated explanations will be omitted.

[0192] The learning numerical simulation model creation unit 301B executes creation of a learning numerical simulation model (step A1).

[0193] The learning partial region setting unit 303 executes a learning partial region setting step of setting a plurality of learning partial regions (step A2), thereby generating a plurality of learning partial region definition data 357.

[0194] The learning partial region arrangement information acquisition unit 305B executes a learning partial region arrangement information acquisition step of acquiring the arrangement information 359 corresponding to each learning partial region from the image data 351 (step A3).

[0195] The simulation execution unit (whole region physical quantity acquisition unit) 307 executes the whole region physical quantity acquisition step (step A4-1).

[0196] The learning partial region physical quantity extraction unit 309 executes a learning partial region physical quantity extraction step (step A4-2).

[0197] The learning data set creation unit 311 executes a learning data set creation step (step A5).

[0198] The trained prediction model creation unit 313 executes a trained prediction model creation step (step A6).

[0199] [Seventh embodiment] (Physical quantity prediction device B) With reference to FIG. 27 , a physical quantity prediction device 400B according to the seventh embodiment of the present invention includes a prediction target partial region setting unit 403, a prediction target partial region arrangement information acquisition unit 405B, a partial region physical quantity prediction unit 407, a prediction target data set creation unit 409, and an entire region physical quantity acquisition unit 411.

[0200] The physical quantity prediction device 400B differs from the physical quantity prediction device 400A in that the prediction target numerical simulation model creation unit 401 is deleted, and the prediction target partial region arrangement information acquisition unit 405 is replaced with a prediction target partial region arrangement information acquisition unit 405B.

[0201] The prediction target partial region arrangement information acquisition unit 405B acquires a plurality of pieces of prediction target partial region arrangement information 457, each of which corresponds to each prediction target partial region, based on prediction target image data 451 and a plurality of prediction target partial region definition data 455. The image data 451 may be created by a computer so as to represent, for example, a morphology to be used for prediction.

[0202] The other functional units are similar to the corresponding functional units of the physical quantity prediction device A according to the fifth embodiment, and therefore a duplicated description will be omitted.

[0203] Step B1 is omitted. A numerical simulation model of the prediction target is not created.

[0204] The prediction target partial region setting unit 403 executes a prediction target partial region setting step of setting a plurality of prediction target partial regions (step B2), thereby generating a plurality of prediction target partial region definition data 455.

[0205] The partial region to be predicted arrangement information acquisition unit 405B executes a prediction region corresponding arrangement information acquisition step of acquiring arrangement information (partial region to be predicted arrangement information) 457 corresponding to each partial region to be predicted from the image data 451 (step B3).

[0206] The partial region physical quantity prediction unit 407 executes a partial region physical quantity prediction step of predicting a partial region physical quantity (including distribution) 459 corresponding to the prediction target partial region, based on the arrangement information 457 corresponding to each of the multiple prediction target partial regions, while using the learned prediction model 365 for matching the prediction target partial region (step B4).

[0207] The entire region physical quantity acquisition unit 411 executes an entire region physical quantity acquisition step of acquiring a physical quantity 463 of the entire region of the heterogeneous material to be predicted based on the partial region physical quantity (including distribution) 459 corresponding to each of the plurality of partial regions to be predicted (step B5). The physical quantity 463 of the entire region is one or a set of physical quantities for the entire region, but may include a distribution over the entire region.

[0208] [Eighth embodiment] (Trained predictive model generating device C) Referring to Figure 28, a trained prediction model generation device 300C according to the eighth embodiment of the present invention includes a training numerical simulation model creation unit 301B, a training partial region setting unit 303, a simulation execution unit (entire region physical quantity acquisition unit) 307, a training partial region physical quantity extraction unit 309, a training dataset creation unit 311 and a trained prediction model creation unit 313.

[0209] The trained prediction model generation device 300C differs from the trained prediction model generation device 300B in that the training partial region arrangement information acquisition unit 305B is omitted.

[0210] Image data may be used to represent the plurality of learning partial region arrangement information 359. The image data here is image data of a region of the same size as the learning partial region of the learning heterogeneous material, data generated based on the image data, or data corresponding to the image data. The image data may be obtained by photographing the actual learning heterogeneous material, or may be created by a computer so as to represent, for example, a morphology to be used for learning.

[0211] The training data set creation unit 311 includes, for each training partial region, training partial region arrangement information 359 and training partial region physical quantity 361 represented by image data or other data as they are in the training data set 363.

[0212] [Ninth embodiment] (Physical quantity prediction device C) 29, a physical quantity prediction device 400C according to the ninth embodiment of the present invention includes a prediction target partial region setting unit 403, a partial region physical quantity prediction unit 407, a prediction target data set creation unit 409, and an entire region physical quantity acquisition unit 411.

[0213] The physical quantity prediction device 400C differs from the physical quantity prediction device 400B in that the prediction target partial region arrangement information acquisition unit 405B is omitted.

[0214] The plurality of prediction target partial region arrangement information 457 may be expressed using image data. The image data here is image data of a region having the same size as the prediction target partial region of the prediction target heterogeneous material, or image data obtained by modifying the image data. Such image data may be obtained by photographing the actual prediction target heterogeneous material, or may be created by a computer. For example, such image data may be created by a computer so as to represent a morphology to be used for prediction.

[0215] Note that, for each partial region to be predicted, the prediction target dataset creation unit 409 includes, as is, the partial region to be predicted arrangement information 457, the partial region to be predicted physical quantity (including distribution) 459, and the partial region to be predicted definition data 455, which are represented by image data or other data, in the prediction target dataset 461.

[0216] [Tenth embodiment] As the trained prediction model generation device, it is up to the user to freely select any one of the trained prediction model generation device A according to the fourth embodiment, the trained prediction model generation device B according to the sixth embodiment, and the trained prediction model generation device C according to the eighth embodiment. Furthermore, as the physical quantity prediction device, any one of the physical quantity prediction device A according to the fifth embodiment, the physical quantity prediction device B according to the seventh embodiment, and the physical quantity prediction device C according to the ninth embodiment may be freely used. In addition, it is also possible to freely select which trained prediction model generation device and which physical quantity prediction device to combine.

[0217] [Eleventh embodiment] (Analysis equipment) FIG. 30 is a functional block diagram of an analysis device that performs the composite material analysis method of one embodiment.

[0218] As shown in Fig. 30, the analysis device 50 is composed of a computer including a processing unit 52 and a storage unit 54. The analysis device 50 is electrically connected to an input operation system 53 equipped with a mouse and a keyboard, and a monitor 55. The input operation system 53 sets data such as information on the polymer and filler for which a simulation model of a composite material is to be created, the type of response analysis, boundary conditions in the response analysis, and input conditions to be given to the simulation model 1. The input data is sent to the processing unit 52 or the storage unit 54.

[0219] The processing unit 52 includes, for example, a central processing unit (CPU) and a memory. When executing various processes, the processing unit 52 reads and starts a computer program from the storage unit 54. The computer program executes various processes. For example, the processing unit 52 loads data related to various processes stored in advance from the storage unit 54 into an area in the memory allocated to the processing unit 52 as necessary, and executes various processes related to the creation of a simulation model 1 of a composite material and response analysis of a composite material using the simulation model 1 based on the loaded data.

[0220] The processing unit 52 includes a model creating unit 52a, a condition setting unit 52b, an analyzing unit 52c, and an evaluating unit 52d.

[0221] The model creation unit 52a creates a simulation model 1 suitable for the molecular dynamics method based on the data stored in the storage unit 54 in advance and various input conditions. When creating a simulation model 1 modeling a composite material such as a filler and a polymer as shown in FIG. 13, the model creation unit 52a sets the arrangement and setting of components such as the number of molecules, molecular weight, molecular chain length, number of molecular chains, branching, shape, size, and the target number of molecules contained in the simulation model 1 to be created, and the number of calculation steps. In addition, the model creation unit 52a sets the initial conditions of various calculation parameters such as hydrogen bonds between the filler particles 11a, between the polymer particles 21a, and between the filler and polymer particles, interactions such as intermolecular forces, etc. In addition, the model creation unit 52a creates cross-linked chains 21c shown in FIG. 14 as necessary.

[0222] As calculation parameters for adjusting interactions between particles, including interactions between filler particles 11a and interactions between polymer particles 21a, the values ​​of σ and ε are set in the case of the above Lennard-Jones potential energy.

[0223] The condition setting unit 52b sets various conditions used for response analysis such as tension analysis, vibration analysis, shear analysis, etc. For example, in the case of tension analysis, the conditions include conditions such as the elongation rate, uniaxial elongation, biaxial elongation, and elongation speed of the simulation model 1.

[0224] The analysis unit 52c executes a numerical analysis of the simulation model 1 based on the analysis conditions set by the condition setting unit 52b. The analysis unit 52c also executes a numerical analysis by molecular dynamics using the simulation model 1 of the composite material created by the model creation unit 52a to obtain physical quantities. Here, the analysis unit 52c executes deformation analysis such as extension analysis and shear analysis and vibration analysis as the numerical analysis. The analysis unit 52c also calculates physical quantities such as values ​​such as displacements in the polymer particles 21a and the filler particles 11 obtained as a result of the numerical analysis or strains obtained by performing a predetermined arithmetic process on the obtained values, the amount of energy accumulated in the simulation model 1, and the amount of energy dissipated from the simulation model 1.

[0225] The analysis unit 52c may also obtain various physical quantities such as nominal strain obtained by calculating the motion displacement and nominal stress obtained from the results of the numerical analysis. This makes it possible to obtain the relationship between the strain and numerical values ​​representing changes in the state of the entire simulation model, such as the bond length of the polymer molecules of the entire simulation model, which change every analysis time, the polymer particle speed, the speed or bond length between the crosslinking points and the free end, and physical quantities such as orientation. In addition, the relationship between the pressure or analysis time and numerical values ​​representing changes in state, such as the bond length of the polymer particles 21a and the moving speed of the polymer particles 21a, which change every analysis time, may also be obtained. Furthermore, the relationship between the temperature or analysis time and numerical values ​​representing changes in state, such as the bond length of the polymer particles 21a and the speed of the polymer particles 21a, which change every analysis time, may also be obtained. This makes it possible to perform a more detailed analysis of changes in the local molecular state of the polymer particles 21a.

[0226] The analysis unit 52c stores the analysis results of the composite material in the storage unit . The evaluation unit 52d evaluates the fracture properties of the composite material based on information regarding the degree of dissipation of energy dissipated from the simulation model 1, obtained by the numerical solution (response analysis) of the analysis unit 52c. The method of evaluating the fracture properties is as described above.

[0227] The storage unit 54 is an appropriate combination of non-volatile memory, which is a recording medium that can only be read, such as a hard disk device, an optical magnetic disk device, a flash memory, and a CD-ROM, and volatile memory, which is a recording medium that can be read and written, such as a RAM (Random Access Memory).

[0228] The storage unit 54 stores data for creating a simulation model of a composite material to be analyzed via the input operation system 53, such as data on fillers such as carbon black, silica, and alumina, and data on polymers such as rubber, resin, and elastomer. The storage unit 54 also stores a computer program for implementing a composite material analysis method. This computer program may be capable of implementing the composite material analysis method according to the present embodiment by combining it with a computer program already stored in the computer or computer system. The term "computer system" as used herein includes an OS (Operating System) and hardware such as peripheral devices.

[0229] The monitor 55 is, for example, a display device such as a liquid crystal display device. The monitor 55 displays a setting screen for setting conditions for executing the above-mentioned numerical solution (response analysis) and inputs to be given to the simulation model 1, and also displays the state of the simulation model 1 during or at the end of the analysis in the analysis unit 52c, and further displays the evaluation of the fracture properties obtained by the evaluation unit 52. The storage unit 54 may be located in another device such as a database server. For example, the analysis device 50 may access the processing unit 52 and the storage unit 54 by communication from a terminal device equipped with the input operation system 53 and the monitor 55.

[0230] In this manner, the computer program can cause the computer to carry out a method for analyzing composite materials.

[0231] The trained prediction model generating device, physical quantity prediction device, and analysis device can be realized by hardware, software, or a combination of these. The trained prediction model generating method, physical quantity prediction method, and simulation method performed by the trained prediction model generating device, physical quantity prediction device, and analysis device can also be realized by hardware, software, or a combination of these. Here, being realized by software means being realized by a computer reading and executing a program.

[0232] The program can be stored and provided to the computer using various types of non-transitory computer readable media. The non-transitory computer readable media includes various types of tangible storage media. Examples of the non-transitory computer readable media include magnetic recording media (e.g., flexible disks, magnetic tapes, hard disk drives), magneto-optical recording media (e.g., magneto-optical disks), CD-ROM (Read Only Memory), CD-R, CD-R / W, and semiconductor memory (e.g., mask ROM, PROM (Programmable ROM), EPROM (Erasable PROM), flash ROM, RAM (random access memory)). The program may also be provided to the computer by various types of transitory computer readable media. Examples of the transitory computer readable media include electric signals, optical signals, and electromagnetic waves. The transitory computer readable media can provide the program to the computer via a wired communication path such as an electric wire and an optical fiber, or via a wireless communication path.

[0233] The predictive model may be stored and provided to a computer using various types of non-transitory computer readable media.

[0234] The present invention can be implemented in various other forms without departing from its spirit or main features. Therefore, the above-described embodiments are merely examples and should not be interpreted as being limited. The scope of the present invention is defined by the claims and is not limited to the text of the specification. Furthermore, all modifications and variations within the scope of the claims are within the scope of the present invention. [Industrial Applicability]

[0235] The present invention can be used to predict the physical quantities of inhomogeneous materials. [Explanation of symbols]

[0236] 300 Trained predictive model generating device 301 Numerical Simulation Model Creation Department for Learning 303 Learning sub-region setting unit 305 Learning sub-area arrangement information acquisition unit 307 Simulation execution unit (whole area physical quantity acquisition unit) 309 Partial area physical quantity extraction unit for learning 311 Training Data Set Creation Department 313 Trained Prediction Model Creation Unit 400 Physical quantity prediction device 401 Prediction target numerical simulation model creation department 403 Prediction target region setting unit 405 Prediction target partial region arrangement information acquisition unit 407 Partial area physical quantity prediction unit 409 Prediction target data set creation unit 411 Whole area physical quantity acquisition part

Claims

1. A trained prediction model generation method for generating a trained prediction model for predicting a physical quantity of a heterogeneous material in which two or more different materials are arranged in a heterogeneous manner, comprising: A learning partial region setting step of setting a plurality of learning partial regions included in an entire region of the learning numerical simulation model, the learning partial region including the arrangement information of two or more different materials in the learning heterogeneous material; a learning partial region physical quantity acquisition step of performing a numerical simulation on the entire region of the learning numerical simulation model to acquire partial region physical quantities for each of the plurality of learning partial regions; a learning dataset creation step of creating a learning dataset including partial region arrangement information and partial region physical quantities for each of the plurality of learning partial regions; a trained prediction model creation step of creating a trained prediction model for corresponding to a prediction target partial region of a heterogeneous material to be predicted by performing machine learning on a prediction model using a plurality of the training data sets corresponding to the plurality of training partial regions, respectively; A method for generating a trained predictive model by having a computer execute the above.

2. The learning partial region physical quantity acquisition step includes: an entire region physical quantity acquisition step of performing a numerical simulation on the entire region of the learning numerical simulation model to acquire a distribution of physical quantities over the entire region of the learning numerical simulation model; a learning partial region physical quantity extraction step of extracting a partial region physical quantity for each learning partial region from a distribution of physical quantities over the entire region; Including, The method for generating a trained prediction model according to claim 1 .

3. For each of at least a portion of training data sets, a regional range for acquiring the partial region arrangement information included in the training data set is the same as a regional range of the training partial region, and a regional range for acquiring the partial region physical quantity included in the training data set is the same as a regional range of the training partial region. The method for generating a trained prediction model according to claim 1 or 2.

4. for each of at least a portion of the training data sets, a regional range for acquiring the partial region physical quantity included in the training data set is the same as a range of the training partial region, and a regional range for acquiring the partial region arrangement information included in the training data set includes the range of the training partial region and is wider than the range of the training partial region; The method for generating a trained prediction model according to claim 1 or 2.

5. The computer further executes a step of including arrangement information of an interface region between two or more different materials of the training heterogeneous material in the partial region arrangement information of the training data set. A method for generating a trained prediction model according to any one of claims 1 to 4.

6. the learning partial region physical quantity acquisition step acquires the partial region arrangement information for each of the plurality of learning partial regions from the learning numerical simulation model; A method for generating a trained prediction model according to any one of claims 1 to 5.

7. using image data of the learning heterogeneous material or data acquired from the image data as the partial region arrangement information for each of the plurality of learning partial regions; A method for generating a trained prediction model according to any one of claims 1 to 5.

8. causing the computer to further execute the step of generating the image data; The method for generating a trained prediction model according to claim 7.

9. In the learning partial region setting step, The plurality of learning partial regions are set so that at least two of the learning partial regions include a common region within an area occupied by the entire area. A method for generating a trained prediction model according to any one of claims 1 to 8.

10. the learning numerical simulation model is a model based on particles and bonds connecting the particles, The numerical simulation is a molecular dynamics method. A method for generating a trained prediction model according to any one of claims 1 to 9.

11. A program for causing a computer to execute the trained prediction model generation method according to any one of claims 1 to 10.

Images