JP2020190981A - Analysis method of composite material and computer program for analysis of composite material - Google Patents

Abstract

To appropriately reproduce a behavior of a particle model in an analysis model while shortening a time required for analysis in the analysis of a composite material by a computer using a molecular dynamics method.SOLUTION: An analysis method includes causing a prediction module to perform machine learning using learning data at least including information on a position of a particle model for learning before and after the elapse of a second time longer than a first time by repeating calculation of a behavior after the elapse of the first time by a molecular dynamics method using a learning model including the particle model for learning that models a substance in a composite material. The prediction module sets each of particle models for analysis in the analysis model of the composite material as a particle model of interest, uses information on a relative position of a vicinity particle model around the particle model of interest with respect to the particle model of interest (or adding information on moving speed), predicts information on the position of the particle model of interest after the elapse of the second time (or the information on moving speed), and obtains the information on the position of the particle model of interest.SELECTED DRAWING: Figure 1

Description

In the present invention, a computer evaluates the characteristics of a composite material by simulating the behavior of the analysis particle model using an analysis model including a plurality of analysis particle models that model a substance in the composite material. The present invention relates to a method for analyzing a composite material and a computer program for analyzing a composite material that performs this analysis method.

Conventionally, various numerical calculations by molecular dynamics have been proposed in order to elucidate the mechanism of fracture of nanostructures of composite materials such as rubber materials.
For example, there is known a method for creating an analysis model of a composite material for analyzing the effect of the bonding state of the polymer on the surface of the filler in the composite material containing the polymer and the filler on the material properties of the composite material (. Patent Document 1). In this creation method, an analysis model of a composite material is created by dispersing the filler in the polymer in the analysis model, specifying the bond position with the polymer on the dispersed filler surface, and bonding the polymer at the specified bond position. To do. Using this analytical model, the computer analyzes the behavior of the polymer particle model and the filler particle model in the analytical model by the molecular dynamics method.
According to the above method, the polymer particle model can be bonded at any point on the surface of the filler particle model in the analysis model, and the influence of the bonding state of the polymer on the surface of the filler on the material properties of the composite material can be measured by molecular dynamics. It is said that it can be analyzed using the method.

JP-A-2015-0642242

By the way, in order to accelerate the development of rubber materials that improve the wear resistance performance of tires, it is helpful to clarify the mechanism of nanostructure destruction due to deformation of rubber materials. By analyzing the fracture of the nanostructure of the rubber material, it is possible to accelerate the development of materials for improving the fracture strength of the filler-filled rubber used in actual tires.

In the above method, for example, a large number of 10,000 or more particle models (polymer particle model and filler particle model) are arranged in the analysis model, and a predetermined input is given to the analysis model to give a predetermined input to the particle model in the analysis model. The behavior of is repeatedly calculated at predetermined time intervals according to the molecular dynamics method. In this method, it is preferable to set the time interval finely and increase the number of repetitions in order to appropriately reproduce the behavior due to the interaction between the plurality of particle models. However, by making the time interval finer, the number of repetitions until the elapsed time in the simulation reaches a predetermined time increases, and the time required for analysis becomes long. For example, it takes several weeks for the elapsed time to reach a predetermined time. The long time required for such analysis is an obstacle in accelerating material development.

Therefore, the present invention provides an analysis method for a composite material capable of shortening the time required for analysis and appropriately reproducing the behavior of the particle model in the analysis model, and an analysis computer program for realizing this analysis method. The purpose is to do.

In one aspect of the present invention, a computer simulates the behavior of the analytical particle model using an analytical model that includes a plurality of analytical particle models that model substances in the composite material, thereby characteristic of the composite material. This is an analysis method for evaluating. The analysis method is
(1) A creation step in which the computer creates the analysis model including the analysis particle model, and
(2) The computer is a step of preparing a prediction module in which machine learning is performed using learning data including learning input data and learning output data, and the learning output data is a learning particle that models the substance. The learning particle model during the second time period longer than the first time by repeating the calculation of the behavior at the first time interval by the molecular dynamics method using the learning model including the model. Information on the position of the learning particle model after the lapse of the second time, or information on the position of the learning particle model after the lapse of the second time and the learning particle whose position has changed due to the movement of The learning input data includes at least information on the movement speed of the model after the lapse of the second time, and the training input data is information on the position of the learning particle model before the lapse of the second time, or the lapse of the second time. A prediction module preparation step that includes information on the previous position and information on the movement speed of the learning particle model before the second time elapses.
(3) Information on the relative position of the neighboring particle model located within a predetermined range around the attention particle model with respect to the attention particle model, using each of the analysis particle models in the analysis model as the attention particle model. Alternatively, using at least the relative position information and the moving speed information of the attention particle model and the neighboring particle model, the prediction module is provided with the position information of the attention particle model after the lapse of the second time. Alternatively, the elapsed time in the simulation is performed by performing a prediction process for predicting the position information of the focused particle model after the lapse of the second time and the moving speed information of the focused particle model after the lapse of the second time. The position acquisition step of obtaining the position information of the particle model of interest until is reached a predetermined time, and
(4) An evaluation step in which the computer evaluates the characteristics of the composite material using the position information obtained in the position acquisition step.
including.

The position of the position acquisition step is changed by the movement of the particle model for analysis during the period of the third time by the computer performing the calculation of the behavior by the molecular dynamics method using the analysis model. Further includes a calculation process for calculating the position information of the particle model for analysis after the lapse of the third time.
In the position acquisition step, the computer
A step of performing the prediction process using the information on the position of the particle model for analysis after movement obtained in the calculation process as the information on the position of the particle model of interest before movement used in the prediction process.
A step of performing the calculation process using the information on the position of the particle model of interest after movement obtained in the prediction process as information on the position of the particle model for analysis before movement used in the calculation process, and the prediction process. One of the steps of further performing the prediction process by using the position information after the movement of the attention particle model obtained in the above as the position information before the movement of the attention particle model used in the prediction process. By repeating the steps, the characteristics of the composite material can be analyzed by obtaining information on the position of the particle model for analysis in the analysis model until the elapsed time reaches a predetermined time longer than the second time. It is preferable to do it.

It is preferable that the calculation by the molecular dynamics method using the learning model and the calculation by the molecular dynamics method using the analysis model are performed under the same analysis conditions.

In the learning data, each of the learning particle models is used as a learning attention particle model, and a learning neighborhood particle model located within the predetermined range around the learning attention particle model before the lapse of the second time. The information on the position relative to the learning attention particle model is included as the position information of the learning input data, and the position information of the learning attention particle model after the lapse of the second time is the learning output data. It is preferable to include it as information on the above-mentioned position.

In addition to the relative position information, the prediction module includes information on the moving speed of the learning attention particle model before the lapse of the second time, and within the predetermined range around the learning attention particle model. Information on the relative movement speed of the located learning proximity particle model with respect to the learning attention particle model before the lapse of the second time is used as the learning input data, and the position information of the learning attention particle model is used. In addition, the information on the moving speed of the learning attention particle model after the lapse of the second time is machine-learned as the learning output data.
In the prediction process, in addition to the information on the position of the particle model of interest after the lapse of the second time, the information on the moving speed of the particle model of interest after the lapse of the second time can also be predicted. preferable.

The prediction module uses the distribution of the displacement amount of the analysis particle model before and after the lapse of the second time obtained by the prediction module and the analysis particle model of the second time by molecular dynamics. By machine learning, the difference from the displacement distribution of the particle model for analysis before and after the elapse of the second time obtained by repeating the calculation at a time interval shorter than the time interval is within an allowable range. It is preferably created by adjusting the parameters of the prediction module to be set.

The prediction module uses the distribution of the moving speed in the particle model for analysis obtained as a prediction result by the prediction module over time and the particle model for analysis by the molecular dynamics method. The machine so that the difference from the distribution of the moving speed in the particle model for analysis at the lapse of the second time obtained by repeating the calculation at a time interval shorter than the time interval of the time is within an allowable range. It is preferably created by adjusting the parameters of the prediction module set in the learning.

It is preferable that the learning model and the analysis model are created under the same model creation conditions.

Each of the analysis particle model and the learning particle model includes a plurality of types of particle models that model a plurality of substances in the composite material.
The training data includes information on the type of the particle model regarding the particle model for training.
In the prediction process, it is preferable to make a prediction using information on the type of the particle model regarding the particle model of interest and the neighboring particle model.

The learning data is the force of interaction used in the molecular dynamics method, in which each of the learning particle models is used as a learning attention particle model and acts with a learning neighborhood particle model around the learning attention particle model. Includes information about the type of
In the prediction process, it is preferable to make a prediction using information on the type of the force of the interaction with respect to the particle model of interest and the nearby particle model.

In the training data, when one of the learning particle models is a bound particle model that is bound to the learning particle model around one of the learning particle models by a bound chain, the bound particle model is combined with the bound particle model. It contains information on the presence or absence of the bound chain that can be distinguished from the unbound learning particle model that is not bound by a chain.
In the prediction process, it is preferable to make a prediction using the information on the presence or absence of the binding chain regarding the particle model of interest and the neighboring particle model.

In the training data, when one of the learning particle models is a bound particle model that is bound to the learning particle model around one of the learning particle models by a binding chain, the bound particle model is combined with the binding particle model. Contains information on the type of binding chain that can be distinguished from the bound particle model bound by a different type of binding chain from the chain.
In the prediction process, it is preferable to make a prediction using the information on the type of the binding chain regarding the particle model of interest and the neighboring particle model.

It is preferable that the prediction module obtains the displacement amount and the moving speed due to the movement of the particle model for analysis.

Another aspect of the present invention is a computer program for analyzing a composite material, which comprises causing a computer to execute the method for analyzing the composite material.

According to the above-mentioned composite material analysis method and the composite material analysis computer program, the time required for analysis can be shortened and the behavior of the particle model in the analysis model can be appropriately reproduced.

It is a figure explaining the main process in the analysis method of the composite material of one Embodiment. It is a conceptual diagram which shows an example of the simulation model of the composite material used in the analysis method of the composite material of one Embodiment. It is a figure which shows an example of the cross-linking used in the analysis method of the composite material of one Embodiment. It is a figure explaining the force and movement received by a particle model of interest. It is a figure which shows an example of the process performed by the analysis method of the composite material of one Embodiment. (A) to (d) are diagrams for explaining the behavior of the particle model for analysis. It is a figure which shows an example of the flow of the analysis method of the composite material of one Embodiment. It is a figure which shows an example which calculates the movement of a particle model by deforming a learning model used in the analysis method of a composite material of one Embodiment. (A) to (c) are diagrams showing an example of one learning particle model of interest and a learning neighborhood particle model of the learning particle model used in the method of analyzing a composite material of one embodiment. It is a figure which shows an example of the validity of the prediction result obtained by the analysis method of the composite material of one Embodiment. It is a figure which shows another example of the validity of the prediction result obtained by the analysis method of the composite material of one Embodiment. It is a functional block diagram of the analysis apparatus which performs the analysis method of the composite material of one Embodiment.

Hereinafter, the method for analyzing the composite material of the present invention and the computer program for analyzing the composite material will be described in detail.
FIG. 1 is a diagram illustrating a main process in the method for analyzing a composite material of one embodiment. The prediction module 100 shown in FIG. 1 is a machine-learned module formed in a computer. The prediction module 100 is a position after a predetermined time (second time) of a plurality of particle models for analysis in an analysis model in which a substance in a composite material can be simulated by a molecular dynamics method. Predict.
Specifically, when each of the analysis particle models in the analysis model is set as the attention particle model and the information on the relative position of the neighboring particle model with respect to the attention particle model is input to the prediction module 100, the attention particle model becomes predetermined. Predict the position information after the lapse of time (second time). The near particle model is an analytical particle model located within a predetermined range around the particle model of interest. That is, each of the particle models for analysis in the analysis model is used as the particle model of interest, and the particle model of interest after a predetermined time (second time) elapses from the information on the relative positions of the neighboring particle models surrounding the particle model of interest. Predict location information. Neighboring particle models that input information on their relative positions to the particle model of interest for prediction are, for example, those within the range in which force effectively acts between the particle models for analysis due to the interaction used in the calculation of molecular dynamics. Limited to. In this way, the prediction module 100 predicts the position information after a predetermined time (second time) elapses, one by one, using each of the particle models for analysis as the particle model of interest. Specifically, the prediction module 100 obtains the amount of displacement due to the movement of the particle model for analysis during the second time period.
In the example shown in FIG. 1, by inputting the information on the relative position of the neighboring particle model into the prediction module 100, the prediction module 100 predicts the information on the position of the particle model of interest, but in addition to the above information on the relative position, As shown in parentheses, by inputting information on the moving speed of the particle model of interest and the neighboring particle model before the elapse of a predetermined time (second time) into the prediction module 100, the prediction module 100 can be used as the particle model of interest. It is also possible to output information on the moving speed in addition to the information on the position after the lapse of a predetermined time (second time). The moving speed information includes the moving speed value and the moving direction.

The prediction module 100 performs machine learning for predicting the position of the particle model for analysis in advance by using the learning input data and the learning output data for machine learning. The training input data includes, for example, information on the position of the learning particle model in the learning model that models the substance. The training output data is learned after a predetermined time (second time) elapses, in which the position is changed by the movement of the learning particle model used as the training input data during the predetermined time (second time) period. Includes position information (position information after movement) of the particle model. Regarding the position information of the learning particle model after the elapse of the predetermined time (second time), the predetermined time (second time) is determined by the molecular dynamics method using the learning particle model used as the learning input data. By repeating the calculation of the predetermined behavior at the shorter first time interval, it is possible to obtain the information on the movement of the learning particle model after the second time period longer than the first time. The second time is an integral multiple of the first time. Hereinafter, the predetermined time is referred to as a second time in order to distinguish it from the first time.
As described above, attention is paid to the prediction module 100 by inputting the information of the relative position and the information of the movement speed of the attention particle model and the neighboring particle model before the lapse of a predetermined time (second time) into the prediction module 100. When the position information and the movement speed information of the particle model after a predetermined time (second time) elapses are output, the learning output data used by the prediction module 100 for machine learning is for learning after the second time elapses. In addition to the information on the position of the particle model after movement, as shown in parentheses in FIG. 1, information on the movement speed of the training particle model after the lapse of a second time is included, and the training input data is for training. In addition to the information on the position of the particle model before the second time elapses, the information on the movement speed of the learning particle model before the second time elapses is included. That is, the information on the moving speed of the learning particle model before the lapse of the second time is included as the learning input data, and the information on the moving speed of the learning particle model after the lapse of the second time is included as the learning input data.

The prediction module 100 includes a model using a neural network represented by well-known deep learning, a model using a well-known random forest method that "classifies" or "regresses" using a plurality of decision trees, and LASSO. Includes a model using regression. Further, a non-linear function using a polynomial, kriging, or RBF (Radial Base Function) can be used as the model of the prediction module 100.
A prediction module 100 that has undergone such machine learning is prepared.

FIG. 2 is a conceptual diagram showing an example of a simulation model of a composite material used in the method for analyzing a composite material of one embodiment. Since the learning model is also a simulation model having the same configuration as the analysis model, it will be described below using the analysis model as a representative. Further, the simulation model 1 shown in FIG. 2 is a model of a composite material having a structure in which the first substance (polymer) is used as a base material and the second substance (filler particles) are distributed in the base material. However, the simulation model 1 is not limited to a model of a composite material having a structure in which the first substance is used as a base material and the second substance is distributed in the base material. Examples of the composite material include the form of a blended polymer composed of a plurality of types of polymers. For example, a blend polymer having a sea-island structure and a lamellar structure can be mentioned. The blended polymer may consist of a crystalline polymer and a non-crystalline polymer. Further, the form of the composite material may include a form having a hard segment phase and a soft segment phase in one molecule, such as a thermoplastic elastomer.

In the simulation model 1 shown in FIG. 2, for example, a particle model for analysis is created in a model creation area which is a virtual space having a substantially cubic shape. The model creation area is a three-dimensional space extending in the X-axis, Y-axis, and Z-axis directions orthogonal to each other. The simulation model 1 is modeled by a plurality of filler particle models 11a, which are particle models for analysis, and includes four filler models 11A, 11B, 11C, and 11D in which the plurality of filler particle models 11a are spherically aggregated, and particles for analysis. Includes a plurality of model polymer particle models 21a and four polymer models 21 in which the binding chain 21b is modeled. In the example shown in FIG. 2, an example in which the simulation model 1 includes four filler models 11A, 11B, 11C, and 11D will be described, but the number of filler models to be modeled is not limited. The simulation model 1 may include three or less filler models 11 and may include more than four filler models 11. Further, in FIG. 2, only four polymer models 21 are shown, but in the simulation model 1, a plurality of polymer models 21 exist over the entire area within the model creation region. In the example shown in FIG. 2, an example in which the model creation region is a virtual space having a substantially rectangular parallelepiped shape is shown, but any shape such as a spherical shape, an elliptical shape, a rectangular parallelepiped shape, or a polyhedral shape may be used. Hereinafter, when the filler models 11A, 11B, 11C, and 11D are generically described, they are referred to as the filler model 11.

The filler model 11 is modeled in a state in which a plurality of filler particle models 11a are assembled in a substantially spherical body. Further, the filler models 11 are arranged in a state of being separated from each other with a predetermined interval. The plurality of filler models 11 may be connected to each other by covalent bonds at the outer edges in a state of being mutually aggregated.

The filler particles to be modeled include, for example, carbon black particles, silica particles, alumina particles and the like. The filler particle model 11a, which is a particle model for analysis, is a model of a collection of a plurality of atoms constituting the filler. Further, a filler particle group in which a plurality of filler particle models 11a are assembled is formed as filler models 11A, 11B, 11C, 11D.
The relative position of the filler particle model 11a is specified by a binding chain (not shown) between the plurality of filler particle models 11a. This bond chain has a function as a spring in which the equilibrium length and the spring constant, which are the bond distances between the filler particle models 11a, are defined, and constrains the filler particle models 11a. The binding chain defines the relative position of the filler particle model 11a and the potential for force to be generated by twisting, bending, or the like. The filler model 11 is defined by numerical data including the mass, volume, diameter and initial coordinates of the filler particle model 11a for handling the filler in molecular dynamics, the number of aggregates, and the like. The numerical data of the filler model 11 is input to the computer.

Polymers modeled in the polymer model 21 include, for example, rubbers, resins, and elastomers. The polymer particle model 21a, which is a particle model for analysis, is a model of an aggregate of atoms of a plurality of polymers. Further, a polymer particle group in which a plurality of polymer particle models 21a are linked by a binding chain is formed as the polymer model 21. That is, the polymer model 21 has a configuration in which polymer particle models 21a, which are an aggregate of a plurality of polymer atoms and a plurality of polymer atoms, are connected to each other by a bonding chain, and the polymer model 21 is predetermined in the model creation region. Arranged in density. The coupling chain functions as a spring in which, for example, an equilibrium length and a spring constant are defined. The polymer particle model 21a is bound by a binding chain 21b between the plurality of polymer particle models 21a, and the relative position is specified. The bond chain 21b has a function as a spring in which an equilibrium length and a spring constant, which are bond distances between the polymer particle models 21a, are defined, and constrains the polymer particle models 21a. The binding chain 21b defines the relative position of the polymer particle model 21a and the potential for force to be generated by twisting, bending, or the like. FIG. 3 is a diagram showing an example of cross-linking used in the method for analyzing a composite material of one embodiment. As shown in FIG. 3, a crosslinked bond 21c is provided between the polymer particle models 21a of the three polymer models 21. The crosslinked bond chain 21c has a function as a spring in which an equilibrium length and a spring constant, which are bond distances between the polymer particle models 21a, are defined, and constrains the polymer particle models 21a.
In addition, a modifier that enhances the affinity with the filler is added to the polymer, if necessary. Examples of the modifier include a hydroxyl group, a carbonyl group, a functional group of an atomic group, and the like. Corresponding to this modifier, a particle model that models the modifier particles and a binding chain (not shown) are arranged between the polymer model 21 and the filler model 11.
The polymer model 21 is defined by numerical data (including mass, volume, diameter, initial coordinates, etc. of the polymer particle model 21a) for handling the polymer in molecular dynamics. The numerical data of the polymer model 21 is input to the computer as a parameter.

In the simulation model 1, interactions are given between the particles of the filler particle models 11a, between the polymer particle models 21a, and between at least a part of the particles between the filler particle model 11a and the particles of the polymer particle model 21a. In some cases, an interaction may be given to exchange forces between all the particles. As the interaction between the filler particle model 11a and the polymer particle model 21a, a chemical interaction (attractive force) may be given, or a physical interaction (bond bond) may be given.
Interactions are imparted by providing the potential described below between the illustrated binding chain 21b and the binding chain including the crosslinked binding chain 21c, and further between the particle models not linked by the binding chain. As a result, a force determined by the interaction acts between the particle models.
In the composite material, the polymer may be composed of a plurality of types of polymers, in which case an interaction may be imparted between the different types of polymer particle models 21a in the simulation model 1. The interaction between the filler particle model 11a and the polymer particle model 21a in this case may be different depending on the type of the polymer particle model 21a.
The interaction between particles is defined by, for example, the Lennard-Jones potential shown in the following equation. At this time, the values of σ and ε in the following equation are adjusted as appropriate. By increasing the upper limit distance (cutoff distance) for calculating the potential, the force acting over a long distance can be adjusted.

The particle model for analysis as shown in FIG. 2 receives a force due to the interaction, and further receives a force due to adjacent coupling chains, and the particle model moves by generating a velocity according to the equation of motion.
FIG. 4 is a diagram illustrating the force and movement of the particle model of interest. The particle model P1 of interest moves by receiving a force from the neighboring particle models P2 to P4. Therefore, the moving direction and moving speed of the focused particle model P1 are determined according to the relative positions of the neighboring particle models P2 to P4 with respect to the focused particle model P1. The calculation process of the movement of the particle model of interest by such a molecular dynamics method is called MD process. That is, in the MD process, by calculating the behavior of the particle model for analysis, the position information of the particle model for analysis after the lapse of the third time is obtained by moving the particle model for analysis during the period of the third time. To calculate. The third time is preferably shorter than the second time. The third time may be the same as, for example, the first time. On the other hand, the prediction processing of the movement of the particle model of interest by the prediction module 100 shown in FIG. 1 is called AI processing.

According to one embodiment, as shown below, by repeating the AI process and the MD process a plurality of times, information on the position of the particle model for analysis in the analysis model after a predetermined time longer than the second time has elapsed. Can be sought. That is,
(A) Performing AI processing by using the information on the position of the particle model for analysis obtained by MD processing after movement as the information on the position of the particle model of interest before movement used in AI processing.
(B) Performing MD processing by using the information on the position of the particle model of interest obtained by AI processing after movement as the information on the position of the particle model for analysis before movement used in MD processing.
(C) Further performing AI processing by using the information on the position of the particle model of interest after movement obtained by AI processing as the information on the position of the particle model of interest before movement used in AI prediction processing.
Information on the position of the particle model for analysis in the analysis model after a predetermined time longer than the second time has elapsed by repeatedly performing any one of (A) to (C) by combining the three forms of Ask for.

FIG. 5 is a diagram showing an example of processing performed by the method for analyzing a composite material of one embodiment.
In the figure, MD1 to MD7 indicate 7 times of MD treatment, and AI1 and AI2 indicate 2 times of AI treatment.
In this way, the position information of the particle model for analysis in the analysis model when the elapsed time in the simulation reaches a predetermined time by repeating the AI processing and the MD processing by combining the above (A) to (C). Can be sought.
Here, the position of the particle model for analysis predicted by the AI process is the position after the lapse of the second time. On the other hand, the position of the particle model for analysis calculated by the MD process is the position after the lapse of the third time, which is shorter than the second time. In this way, the time interval of the position of the analysis particle model predicted by the prediction module 100 in the AI process can be made longer than the time interval of the position of the analysis particle model calculated by the MD process. In MD processing, setting a long time interval by the molecular dynamics method means that each of the plurality of particle models for analysis moves under the force of interaction and binding chain, so that the particle model of the vicinity to one particle model of interest Setting a long time interval based on the force received by the particle model of interest with the relative position fixed causes the inconvenience described below.

6 (a) to 6 (d) are diagrams for explaining the behavior of the particle model for analysis. 6 (a) to 6 (d) show a model in which three particle models P5 to P7 for analysis are bound by a binding chain, and particle models P8 to P11 for analysis. FIG. 6A shows a state before the analysis particle models P5 to P7 move by receiving a force from the analysis particle models P8 to P11. In the figure, the arrows indicate the initial movement directions given to the analysis particle models P5 to P7 before the movement is started by receiving the force from the analysis particle models P8 to P11. By repeating the MD treatment of the analysis particle models P5 to P7 a plurality of times at short time intervals, the analysis particle models P5 to P7 move as shown in FIG. 6B. On the other hand, when MD processing is performed at a long time interval at one time, the analysis particle models P5 to P7 collide with the analysis particle models P8 to P11 after the same time elapses, resulting in behavior that cannot occur in reality. Shown. This is because when MD processing is performed at short time intervals, repulsive force is generated due to interaction when the particle models P5 to P7 for analysis approach the particle models P8 to P11 for analysis, but a long time interval can be calculated once. This is because the repulsive force generated when the analysis particle models P5 to P7 approach the analysis particle models P8 to P11 is not taken into consideration when the execution (MD processing is performed). Therefore, in the MD process, the time interval cannot be lengthened from the viewpoint of calculating the position of the particle model for analysis with high accuracy. On the other hand, in the AI processing by the prediction module 100, the particle model of interest before and after the lapse of the second time created by joining the results of the MD processing at short time intervals like the first time (for analysis). Since the position information of the particle model) is machine-learned as training data, the movement that cannot occur in reality as shown in FIG. 6 (c) is not predicted, and as shown in FIG. 6 (d), FIG. 6 ( The same result as the result shown in b) can be obtained. That is, from the analysis particle models P5 to P7 before movement shown in FIG. 6A, the prediction module 100 predicts the positions of the analysis particle models P5 to P7 after movement after the lapse of the second time. , Information on a position substantially similar to the position of the analysis particle models P5 to P7 calculated by the MD process shown in FIG. 6B can be obtained.

In this way, the prediction module 100 is prepared, and each of the particle models for analysis in the simulation model 1 is used as the particle model of interest, and the neighboring particle model located within a predetermined range around the particle model of interest is relative to the particle model of interest. At least the position information can be used to cause the prediction module 100 to predict the position information of the particle model of interest after the second lapse of time. To predict the position of the particle model of interest after the lapse of a second time, which is longer than the short third time set so that collisions between the particle models for analysis do not occur in the MD processing, in one AI process. Therefore, it is possible to shorten the time required for analysis until the elapsed time in the simulation reaches a predetermined time. Moreover, since the prediction by AI processing is machine-learned using the movement information of the learning particle model obtained by MD processing as learning data, it does not predict the movement that cannot actually occur such as the collision of the particle model. The behavior of the particle model in the analysis model can be reproduced appropriately.

FIG. 7 is a diagram showing an example of the flow of the method for analyzing the composite material of one embodiment.
The analysis method shown in FIG. 7 is a method for analyzing a composite material using a molecular dynamics method using a computer. That is, the analysis of the composite material is performed by a computer.

First, the computer creates an analysis model including an analysis particle model, which is an analysis model of the composite material, and a learning model including a learning particle model (step ST10).
The analysis model and the learning model to be created are the simulation model 1 as shown in FIG. The analysis model and the learning model may have the same configuration or may be the same. Since the learning model is used for creating learning data, it does not have to be a large-scale simulation model 1 as large as the analysis model, and may be smaller than the analysis model. The small scale includes the small size of the model creation area of the simulation model and / or the small number of particle models for analysis. As for the training data, the information on the position of the particle model for analysis after movement in the analysis model calculated by MD processing may be updated as new learning data, and the prediction module 100 may be machine-learned each time it is updated. ..

Next, the computer calculates the learning particle model by the molecular dynamics method using the learning model and creates the learning data (step ST12). FIG. 8 is a diagram showing an example in which the learning model used in the method for analyzing the composite material of one embodiment is stretched and deformed in the uniaxial direction to calculate the movement of the learning particle model by MD processing. The learning model 2 shown in FIG. 8 is a model having the same configuration as the simulation model 1 shown in FIG.
The computer repeatedly calculates the behavior after the lapse of the first time by the molecular dynamics method using the learning model 2 including the learning particle model.
As a result, the position information of the learning particle model after the lapse of the second time due to the movement of the learning particle model during the period of the second time longer than the first time, or the information of this position and the second Information on the moving speed of the learning particle model after the lapse of time and information on the position of the learning particle model before moving, or information on this position and information on the moving speed of the learning particle model before moving, respectively. Create learning data that includes at least training output data and learning input data. Here, when the training data includes at least the position information of the learning particle model before and after the lapse of the second time, or the information of this position and the information of the movement speed, the training data further includes other information. But it means that it is okay. Other information includes information on the type of the learning particle model, information on the presence or absence of binding chains, and information on the types of binding chains, if any.

Next, the computer causes the prediction module 100 to perform machine learning (step ST14). That is, the computer prepares the machine-learned prediction module 100. The prediction module 100 is a model using a neural network represented by well-known deep learning, a model using a well-known random forest method that "classifies" or "regresses" using a plurality of decision trees, LASSO. Machine learning is performed using a model that uses regression.

Next, the computer uses the machine-learned prediction module 100 to predict the position of the analysis particle model in the analysis model (AI processing), or for analysis particles by the molecular dynamics method using the analysis model. The position of the model is calculated (MD processing) (step ST16).
In the AI processing, as shown in FIG. 1, each of the particle models for analysis in the analysis model is used as the particle model of interest, and the information on the relative position of the neighboring particle model around the particle model of interest with respect to the particle model of interest, or this relative position. Information on the position of the particle model of interest after the lapse of the second time and information of the position of the particle model of interest after the lapse of the second time and the information of the moving speed of the particle model of interest and the particle model of interest after the lapse of the second time are used in the prediction module 100. Predict movement speed information.
In the MD process, for example, the position of the particle model for analysis after the lapse of a third time shorter than the second time is calculated by the molecular dynamics method.
In this way, the computer updates the position of the particle model for analysis after AI processing or MD processing as a new position, and uses it as information on the position of the particle model for analysis used for the next AI processing or MD processing before movement (step). ST18).

Next, the computer determines whether or not the elapsed time in the simulation regarding the movement of the analysis model has reached a predetermined time (step ST20). If it is determined that the elapsed time does not reach the predetermined time, the process returns to step ST16 and steps ST16-ST20 are repeated. When it is determined that the elapsed time has reached a predetermined time, the AI process and the MD process are terminated, and the movement of the particle model for analysis is terminated. After that, the force acting on each of the particle models for analysis is calculated, and the information of the physical quantity such as the force acting on the entire analysis model is calculated. In this way, the characteristics of the composite material are evaluated using the position information of each particle model for analysis (step ST22).

As described above, in one embodiment, the information on the position of the particle model of interest after the movement of the particle model of interest that has undergone AI processing or MD processing is obtained, and the information of the position of the particle model of interest before movement when further performing AI processing or MD processing. AI treatment or MD treatment is repeated. In the AI process, the position information of the particle model for analysis in the analysis model after the lapse of the second time, which is longer than the first time, is obtained in one AI process, so that the time required for the analysis is shortened and the analysis is performed. The behavior of the particle model for analysis in the model can be reproduced appropriately. In particular, when the MD process is performed after the AI process is repeated, the position error due to the prediction of the particle model for analysis when the AI process is repeated can be corrected by the MD process, so that the particle model for analysis can be used. The behavior can be reproduced more appropriately.

According to one embodiment, the learning data for which the prediction module 100 performs machine learning is determined around the learning particle model before the lapse of the second time, using each learning particle model as a learning particle model. It is preferable to include at least the information on the position of the learning proximity particle model located within the range relative to the learning attention particle model and the information on the position of the learning attention particle model after the lapse of the second time. In this case, the predetermined range is the same as the range used to limit the neighboring particle model in the analysis model. Using such learning data, the prediction module 100 can efficiently predict the position information of the particle model of interest after the movement from the position information of the neighboring particle model before the movement.

9 (a) to 9 (c) are diagrams showing an example of one learning particle model of interest and a learning neighborhood particle model of the learning particle model. 9 (a) to 9 (c) do not have a form in which a plurality of filler particle models 11a are aggregated as in the filler model 11 shown in FIG. 1 for easy understanding. The learning particle model within the range of a circle having a predetermined radius centered on the learning attention particle model P20 in FIG. 9A is defined as the learning neighborhood particle models P21 to P26.
FIG. 9B shows an example in which the position information of the learning particle model is used as the learning input data. As shown in FIG. 9B, the learning proximity particle models P21 to P26 within a predetermined radius centered on the learning attention particle model P20 impart a force that affects the movement of the attention particle model P20. Set as to. The x and y coordinate values represented by the XY coordinate system are used as information on the relative positions of the learning neighboring particle models P21 to P26 centered on the position of the learning particle model P20. Since the examples of FIGS. 9A to 9C are represented by a learning particle model on a plane, x and y coordinate values are used as relative position information, but the learning particles in the three-dimensional space. In the case of the model, the x, y, z coordinate values represented by the XYZ coordinate system are used as the relative position information.
FIG. 9C shows an example of the positions of the learning attention particle model 20 and the learning neighborhood particle model P21-P26 after the lapse of the second time. In this case, the information on the position of the moved learning proximity particle model P21-P26 is not used, and the information on the position of the learning attention particle model P20 is used as the learning output data.
In FIG. 9A, the learning attention particle model 20 and the learning neighborhood particle model P21 to 26 are used as examples, but the learning neighborhood particle model is used for each of the learning particle models shown in FIG. 9A. Is set, and learning input data and learning output data are created.

The learning neighborhood particle models P21 to P26 are all learning particle models within a predetermined range centered on the learning attention particle model P20, but if necessary, the learning neighborhood particle model You may limit the number. In this case, it is preferable to preferentially set the learning particle model close to the learning attention particle model P20 as the learning neighborhood particle model. The closer the learning particle model is to the learning attention particle model P20, the greater the force given to the learning attention particle model P20 by the interaction, and the greater the contribution to the movement of the learning attention particle model P20.
When the setting conditions for setting such a learning neighborhood particle model are used, even when the prediction is performed by AI processing using the analysis model, the neighborhood particle model is set under the same setting conditions as the learning neighborhood particle model. It is preferable to set it.

According to one embodiment, the prediction module 100 uses each of the learning particle models as a learning attention particle model, and in addition to the relative position information, the movement speed of the learning attention particle model before the lapse of the second time. Information and the relative movement speed of the learning proximity particle model located within a predetermined range around the learning attention particle model with respect to the learning attention particle model (the movement speed of the learning attention particle model is zero). The information before the elapse of the time of 2 may be used as the learning input data, and the information of the moving speed after the elapse of the second time of the attention particle model for learning may be used as the learning output data for machine learning. At this time, in the prediction by AI processing, in addition to the information on the position of the particle model of interest after the elapse of the second time, the elapse of the second time is based on the information on the relative moving velocity of the neighboring particle model with respect to the particle model of interest. It is preferable to predict the moving velocity information of the particle model of interest later. The moving speed information includes the moving direction and the moving speed value in the moving direction. The relative movement speed of the learning attention particle model after the lapse of the second time corresponds to the change in the movement speed of the learning attention particle model before the lapse of the second time from zero. By adding the moving speed of the particle model of interest before the lapse of the second time to this relative moving speed, the moving speed of the particle model of interest after the lapse of the second time can be obtained. By adding the movement direction of the attention particle model after the lapse of the second time to the movement direction of the attention particle model before the lapse of the second time, the movement direction of the attention particle model after the lapse of the second time can be obtained. You can ask.
By considering the moving speeds of the particle model of interest and the nearby particle model, it is possible to avoid unrealistic movement such as the particle model of interest colliding with or extremely close to the neighboring particle model. Therefore, it is preferable to use the information of the relative movement speed with respect to the learning particle model as the information before the passage of the second time in the learning particle model used as the learning input data.

Further, in addition to the information on the moving speed and the moving direction, the information on the acceleration (the magnitude of the acceleration and the direction of acceleration) may be included as the learning data.

According to one embodiment, the distribution of the displacement amount of the particle model for analysis before and after the passage of the second time obtained by the prediction module 100, and the time interval of the second time for the particle model for analysis by the molecular dynamics method. The prediction module 100 is machine-learned so that the difference from the displacement distribution of the particle model for analysis before and after the lapse of the second time obtained by iterative calculation at shorter time intervals is within an allowable range. It is preferably created by adjusting the parameters of the prediction module 100 that are sometimes set. By keeping the difference between the displacement distribution of the particle model for analysis by the prediction module 100 and the displacement distribution of the particle model for analysis in MD processing within an allowable range, that is, by reducing the difference, the prediction accuracy in AI processing is improved. ..
Further, according to one embodiment, the distribution of the moving speed in the analysis particle model over time obtained as the prediction result by the prediction module 100 and the analysis particle model are obtained by the molecular dynamics method. The prediction module 100 so that the difference from the distribution of the moving speed in the particle model for analysis over the passage of time obtained by repeating the calculation at a time interval shorter than the time interval of the time is within an allowable range. Is also preferably created by adjusting the parameters of the prediction module 100 set by machine learning. By keeping the difference between the moving velocity distribution of the particle model for analysis by the prediction module 100 and the moving velocity distribution of the particle model for analysis in MD processing within an allowable range, that is, by reducing the difference, the prediction accuracy in AI processing is improved. ..
Examples of such adjustment parameters include the number of layers and the number of nodes in the case of deep learning.

According to one embodiment, the learning model and the analysis model are preferably created under the same model creation conditions. The preparation conditions preferably include, for example, at least the number density of the polymer particle model, the type of interaction, and the temperature. Since the temperature is represented by the average value of the moving speeds in the particle model, the same temperature means that the average values of the moving speeds of the particle model for analysis are substantially the same. When the learning model and the analysis model are models in which a large number of filler particle models 11a are arranged in the polymer model 21 as shown in FIG. 2, the polymer particle model 21a is attracted by the filling of the filler particle model 11a, and the polymer particles are attracted. The number density of the model 21a varies from place to place. In such a case, when performing AI treatment using the analysis model, it is preferable to predict the position by AI treatment while identifying the information on the number density of the polymer particle model 21a. Therefore, it is preferable that the training data includes information on the number density of the polymer particle model, and the prediction module 100 performs machine learning together with the information on the number density in addition to the information on the position of the polymer particle model. In this case, it is preferable that the surface area ratio or volume fraction in the filler model 11 in which the filler particle model 11a is aggregated is substantially the same between the learning model and the analysis model. In this case, it is not necessary to identify the arrangement of the filler model, the diameter of the filler model, and the like.
By using such the same model creation conditions for creating a learning model and an analysis model, the prediction accuracy by AI processing is improved.

Further, according to one embodiment, it is preferable that the calculation by the molecular dynamics method using the learning model and the calculation by the molecular dynamics method using the analysis model are performed under the same analysis conditions. For example, when the uniaxial extension as shown in FIG. 8 is reproduced by the analysis model, the prediction module 100 uses the learning data obtained by calculating by the molecular dynamics method so that the uniaxial extension is also reproduced in the learning model. Let them learn by machine. In this case, it is preferable that the extension speeds in the uniaxial extension reproduced by the learning model and the analysis data are the same. When reproducing the shear deformation in the training model and the analysis model, it is preferable that the shear rates are the same. In this case, the first time interval when calculating the movement in the learning model is shorter than the second time interval predicted by the analysis model, as described above.

In order to model a plurality of substances in the composite material, each of the analysis particle model and the learning particle model often includes a plurality of types of particle models. In this case, the training data includes information on the type of particle model related to the particle model for learning, and it is preferable to use information on the type of particle model related to the particle model of interest and the neighboring particle model in the prediction in AI processing. Since the mass applied to the attention particle model differs depending on the type of the particle model of interest, the amount of movement of the particle model of interest differs depending on the type of the particle model even if the same force is applied. Therefore, in order to improve the prediction accuracy by AI processing, the training data includes information on the type of particle model related to the particle model for training, and in the prediction in AI processing, the type of particle model related to the particle model of interest and the neighboring particle model. It is preferable to use the information of. Therefore, in the AI process, the prediction module 100 uses at least the information on the type of the particle model of interest and the information on the relative position of the neighboring particle model with respect to the particle model of interest to determine the position of the particle model of interest after a second lapse of time. It is preferable to predict.

Also, since the interaction acts between the two particle models, the type of force of the interaction is determined by the two particle models. Therefore, the training data contains information on the types of interaction forces used in molecular dynamics that act with the learning neighbor particle model around the training particle model of interest and is of interest in predictions in AI processing. It is preferable to use information on the type of interaction force for the particle model and the neighboring particle model in order to improve the prediction accuracy by AI processing. Therefore, in AI processing, the prediction module 100 uses at least the information on the type of interaction force and the information on the relative position of the neighboring particle model with respect to the particle model of interest, and the prediction module 100 is the first of the particle models of interest. It is preferable to predict the position after the lapse of time of 2.

Further, as shown in FIG. 2, the polymer particle model 21a is bound to another polymer particle model 21a by the binding chain 21b, and the degree of freedom of movement is limited. Such movement limitation is expressed by the force of interaction and the force received by the binding chain, but in order to improve the prediction accuracy, is the particle model of interest bound by the neighboring particle model and the binding chain? It is preferable to acquire information on whether or not to use it for prediction by AI processing. Therefore, the training data indicates that if one of the learning particle models is a bound particle model that is bound by a bound chain to the learning particle model that surrounds one of the learned particle models, this bound particle model is combined with the bound chain. It is preferable to include information on the presence or absence of bound chains that can be distinguished from the non-bound learning particle model that does not bind in, and to use the information on the presence or absence of bound chains for the particle model of interest and the neighboring particle model in the prediction in AI processing.
When the particle model of interest is a bound particle model, the particle model of interest also has different movement restrictions depending on whether the bound particles are in a continuous main chain or at the end of the main chain. Therefore, when the particle model of interest is a bound particle model, the training data is also the bound particle so that the prediction can be made by AI processing using the information on whether the bound particle model is in the main chain or at the end. It contains information on whether the model is in the main chain or at the end, and the prediction module 100 preferably performs machine learning in advance using this information.

Furthermore, the movement limitation of the particle model of interest differs depending on the type of binding chain. Therefore, if one of the learning particle models is a bound particle model that is bound by a bound chain to the training particle model that surrounds one of the training particle models, the training data includes this bound particle model. It contains information on the type of bound particle that can be distinguished from the bound particle model bound by different types of bound particles, and it is predicted that the prediction in AI processing will use the information on the type of bound particle related to the particle model of interest and the neighboring particle model. It is preferable from the viewpoint of improving accuracy.

FIG. 10 is a diagram showing an example of the validity of the prediction result obtained by the method for analyzing the composite material of one embodiment.
FIG. 10 shows the maximum bond length between the analysis particle models bonded by the bond chain when the time elapses in the simulation, using the information on the position of the analysis particle model at time / repetition “0” as the initial information. An example of a change in the maximum separation distance) in the combined analytical particle model is shown. Line A shows the variation in the maximum bond length at each “time / repetition” when iteratively calculated by MD processing at the first time interval. Therefore, the horizontal axis indicates the passage of time with respect to the first time. Lines B and C show the maximum bond length obtained from the position information obtained by AI processing of the particle model for analysis when a time longer than the first time elapses after the MD processing is repeated a predetermined number of times. The maximum bond length of line B at "time / repetition 44" on the horizontal axis is the position of the particle model for analysis obtained by repeating MD processing from "time / repetition 0" to "time / repetition 40". From the information, the maximum bond length obtained by predicting the position information of the particle model for analysis at the time of "time / repetition 44" on the horizontal axis by one AI process is shown. Therefore, when the value of "time / repetition" of line B is N (natural number of 5 or more), the maximum bond length is MD processing from "time / repetition 0" to "time / repetition N-4". The maximum bond obtained by predicting the position information of the particle model for analysis at the time when the horizontal axis becomes "time / repetition N" from the position information of the particle model for analysis obtained by repeating The length.
The maximum bond length of the line C at "time / repetition 42" on the horizontal axis is the position of the particle model for analysis in which MD processing is repeated from "time / repetition 0" to "time / repetition 40". From the information, the maximum bond length obtained by predicting the position information of the particle model for analysis at the time of "time / repetition 42" on the horizontal axis by one AI process is shown. Therefore, when the value of "time / repetition" of line C is M (natural number of 3 or more), the maximum bond length is MD processing from "time / repetition 0" to "time / repetition M-2". From the information on the position of the particle model for analysis obtained by repeating the above, the information on the position of the particle model for analysis at the time when the horizontal axis becomes "time / repetition M" was obtained by predicting it by one AI process. Maximum bond length.
That is, the line B shows the result of predicting the position of the particle model for analysis at one time by AI processing at a time interval four times the first time, and the line C shows the result at a time interval twice the first time. The result of predicting the position of the particle model for analysis by AI processing at one time is shown.
The lines B and C are close to the line A, and it can be seen that the prediction result by the AI processing is close to the calculation result by the MD processing. Therefore, it can be said that the prediction result by AI processing is appropriate.

FIG. 11 is a diagram showing another example of the validity of the prediction result obtained by the method for analyzing the composite material of one embodiment.
In FIG. 11, as shown in FIG. 8, about 1 million particle models for analysis are included in the analysis model, and the results of stress due to strain when uniaxial elongation is applied to the analysis model are shown. The line D in the figure shows the calculation result of the stress against the strain when the analysis model is repeatedly calculated by the MD process at the first time interval. Line E in the figure shows the position of the particle model for analysis in four times the time of the first time by AI processing after the analysis model is continuously repeated 10 times at the first time interval. The result of the stress against the strain obtained by repeating this process a plurality of times is shown. That is, the line E is the result of repeating the MD process 10 times, the AI process 1 time, the MD process 10 times, the AI process 1 time, and so on.
As shown in FIG. 11, the line E is similar to the line D, and it can be seen that even if the AI processing is performed, the same result as the MD processing can be obtained. Therefore, it can be said that the prediction result by AI processing is appropriate.

As for the time required for AI processing, a large amount of time is spent on position information processing for creating information on the relative position of the neighboring particle model with respect to the particle model of interest, using each of the particle models for analysis as the particle model of interest, while using this information. The time required for the process predicted by the prediction module 100 is extremely short. Similarly, regarding the time required for calculation in MD processing, a large amount of time is spent on position information processing for creating information on the relative position of the neighboring particle model with respect to the particle model of interest, using each of the particle models for analysis as the particle model of interest. , The time required for the process of calculating the moving speed according to the equation of motion using this information is extremely short. When the analysis of the present embodiment is performed, the MD process or the AI process is repeated until the elapsed time reaches a predetermined time. Therefore, the particle model for analysis is performed at a second time interval longer than the first time interval used in the MD process. By using the AI process that can predict the position of, the number of repetitions can be reduced, and the total time required for position information processing can be shortened accordingly. Therefore, in the case of the line E shown in FIG. 11, in the analysis in which the processing is performed until 14 times the time of the first time elapses, the MD processing is performed 10 times and the AI processing is performed once, so that the positions are 11 times. Information processing. On the other hand, in the case of the line D shown in FIG. 11, in the analysis in which the processing is performed until 14 times the time of the first time elapses, the MD processing is performed 14 times, so that the position information processing is performed 14 times. Therefore, the time required for position information processing, which occupies most of the analysis time in the case of line E, is 0.73 times (= 11/14) the time required for position information processing in the case of line D, and position information processing. The total time required for this is reduced by 27%. Therefore, by increasing the ratio of the number of AI processes to the number of MD processes, the time required for position information processing is further shortened.
In this way, the prediction by the AI process can be used to perform one prediction at a second time interval longer than the first time interval used in the MD process, so that the analysis time is shortened.

FIG. 12 is a functional block diagram of an analysis device that performs an analysis method for a composite material of one embodiment.
As shown in FIG. 12, the analysis device 50 includes 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 including a mouse and a keyboard and a monitor 55. The input operation system 53 includes information on the polymer and filler for which the model of the composite material is created, analysis conditions including the type of analysis for evaluating the composite material, boundary conditions in the analysis, and input conditions given to the analysis model. Etc. are set. These input data are sent to the processing unit 52 or the storage unit 54.

The processing unit 52 includes, for example, a central processing unit (CPU: Central Processing Unit) and a memory (not shown). The processing unit 52 reads the computer program from the storage unit 54 and starts the computer program when executing various processes. The computer program executes various processes. For example, the processing unit 52 expands the data related to various processes stored in advance from the storage unit 54 into an area allocated to itself on the memory as needed, and based on the expanded data, an analysis model of the composite material or Various processes described below regarding the creation of the learning model and the analysis for evaluating the characteristics of the composite material using the analysis model 1 are executed.

The processing unit 52 includes a model creation unit 52a, a condition setting unit 52b, a prediction module unit 52c, a calculation processing unit 52d, a prediction processing unit 52e, and an evaluation unit 52f.
The model creation unit 52a creates an analysis model and a learning model suitable for the molecular dynamics method based on the data stored in the storage unit 54 in advance and various input conditions. When creating an analysis model and a learning model that model a composite material such as a filler and a polymer as shown in FIG. 2, the model creation unit 52a determines the number of molecules, the molecular weight, the length of the molecular chain, and the number of molecular chains of the filler and the polymer. Setting of arrangement of components such as branching, shape, size, and target number of molecules included in the analysis model and learning model to be created, and the first time and second time used for MD processing and AI processing. The time, the third time, and the sequence of the MD process and the AI process when the repetitive process is performed as shown in FIG. 5 are set. In addition, the model creation unit 52a sets initial conditions for various calculation parameters of interactions such as bonds between the filler particle models 11a, between the polymer particle models 21a, and between the filler / polymer particle models, and intermolecular forces. In addition, the model creation unit 52a also creates the crosslinked chain 21c and the like shown in FIG. 3 as needed.

The calculation parameters for adjusting the interaction between the particle models including the interaction between the filler particle model 11a and the interaction between the polymer particle model 21a, which are the particle models for analysis, are σ, in the case of the Lennard-Jones potential described above. The value of ε is set.

The condition setting unit 52b sets the type of test and analysis conditions used for analysis for evaluating a composite material such as extension analysis, vibration analysis, and shear analysis. For example, in the case of extension analysis, the analysis conditions include conditions such as the elongation rate of the analysis model and the learning model, uniaxial extension, biaxial extension, and extension rate.

The calculation processing unit 52d executes numerical calculation using the molecular dynamics method of the analysis model or the learning model based on the analysis conditions set by the condition setting unit 52b. Further, the calculation processing unit 52d is created by the model creation unit 52a, and the analysis model or learning model in which the analysis particle model or the learning particle model is arranged at the position in the initial state, or the initial stage after a lapse of time. After the particle model for analysis or learning particle model of the state is moved, the physical quantity is acquired by executing the numerical calculation by the molecular dynamics method using the analysis model or learning model stored in the memory. The calculation processing unit 52d executes deformation analysis such as extension analysis and shear analysis and vibration analysis as numerical analysis. Specifically, the calculation processing unit 52d receives information on the position after the lapse of the first time or the third time in each particle model in the analysis model or the learning model obtained as a result of the numerical calculation, or information on this position. And the information of the moving speed, and the physical quantity such as the strain obtained by executing the predetermined calculation process on the value obtained from the calculation result are calculated, and these calculation results are calculated after the lapse of the first time or the third time. As a result, it is stored in the memory together with the analysis model or the learning model including the moved analytical particle model after the lapse of the first time or the learning particle model after the lapse of the third time.
When the calculation processing unit 52d next performs the calculation by the molecular dynamics method according to the sequence set by the condition setting unit 52b, the analysis model after the lapse of the first time or the third time lapse stored in the memory Calculations are performed using the later learning model.

The reason why the calculation is performed by the molecular dynamics method using the learning model is to create learning data for making the prediction module 100 perform machine learning. The reason why the MD process is performed using the analysis model is to perform an analysis for evaluating the characteristics of the composite material. Therefore, when a module constructed by machine learning is prepared in advance in the prediction module unit 52c, it is not necessary to create learning data. Therefore, in this case, it is not necessary to create a learning model in the model creation unit 52a, and further, in the calculation processing unit 52d, the molecular power using the learning model to create the learning data for causing the prediction module 100 to perform machine learning. There is no need to do academic calculations.

The prediction module unit 52c includes a prediction module 100 having the function shown in FIG.
In such a prediction module 100, in the prediction module unit 52c, information on the position of the learning particle model before and after the first time interval obtained by calculation by the molecular dynamics method and stored in the memory, or information on this position and movement. Using at least the velocity information, information on the position of the learning particle model before and after the lapse of the second time, which is longer than the first time, is created as the learning input data and the learning output data of the learning data.
The prediction module unit 52c causes the prediction module 100 to perform machine learning using the training data.

When the prediction processing unit 52e performs AI processing according to the sequence set by the condition setting unit 52b, the prediction module 100 uses the analysis model stored in the memory to analyze the particles after the second time elapses. Predict the position information of the model, or the position information and the moving speed information. Specifically, using each of the particle models for analysis as the particle model of interest, the prediction module 100 predicts the position information of the particle model of interest, or the information on this position and the information on the moving speed (AI processing). The information on the position of the analysis particle model predicted by the prediction module 100 is stored in the memory together with the analysis model including the information on the position of the analysis particle model after the lapse of the second time.
In this way, the elapsed time in the simulation reaches a predetermined time for the MD processing by the calculation processing unit 52d and the AI processing by the prediction processing unit 52e according to the sequence of the MD processing and the AI processing set by the condition setting unit 52b. Repeat until you do.

The evaluation unit 52f reads out the position information of the particle model for analysis in the analysis model until the elapsed time reaches a predetermined time, or the information of this position and the information of the moving speed from the memory, and further, the analysis model. Calculate the predetermined physical quantity in. For example, when stretching analysis is performed by uniaxial stretching, the stress against strain is calculated as a physical quantity. In addition, in the extension analysis, if the analytical particle model bound by the binding chain exceeds a predetermined distance, it is judged that the binding chain is broken and the force acting by the binding chain is reduced. The number of broken chains is calculated as an index of the breaking characteristics of the composite material while reproducing the breaking of the bound chain.

The storage unit 54 can read and write to a non-volatile memory such as a hard disk device, a magneto-optical disk device, a flash memory and a CD-ROM that can only be read, and a RAM (Random Access Memory). Volatile memory, which is a possible recording medium, is appropriately combined.

In the storage unit 54, data for creating an analysis model and a learning model of the composite material to be analyzed via the input operation system 53, for example, data of fillers such as carbon black, silica, and alumina, rubber, and resin. , And data of polymers such as elastomers are stored. Further, the storage unit 54 stores a computer program or the like for realizing a method for analyzing the composite material. This computer program may be capable of realizing the method for analyzing a composite material according to the present embodiment in combination with a computer program already recorded in a computer or a computer system. The term "computer system" as used herein includes hardware such as an OS (Operating System) and peripheral devices.

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 MD processing and AI processing and inputs for giving deformation to the analysis model and the learning model, and is in the process of predicting the solution in the prediction processing unit 52e. Alternatively, the state of the analysis model at the end of the prediction is displayed, and further, the evaluation result of the composite material using the physical quantity obtained by the evaluation unit 52f is displayed. 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 including the input operation system 53 and the monitor 55.

In this way, the computer program can cause the computer to execute the method of analyzing the composite material.

The method for analyzing the composite material of the present invention and the computer program for analyzing the composite material have been described in detail above, but the present invention is not limited to the above embodiment, and various improvements are made without departing from the gist of the present invention. Of course, you may make changes.

1 Simulation model 2 Learning model 11, 11A, 11B, 11C, 11D Filler model 11a Filler particle model 21,21A, 21B, 21C Polymer model 21a Polymer particle model 21b Bonded chain 21c Cross-linked bonded chain 50 Analyzer 52 Processing unit 52a Model creation Unit 52b Condition setting unit 52c Prediction module unit 52d Calculation processing unit 52e Prediction processing unit 52f Evaluation unit 53 Input operation system 54 Storage unit 55 Monitor 100 Prediction module

Claims (14)

An analysis method in which a computer evaluates the characteristics of a composite material by simulating the behavior of the analysis particle model using an analysis model that includes a plurality of analysis particle models that model substances in the composite material. There,
(1) A creation step in which the computer creates the analysis model including the analysis particle model, and
(2) The computer is a step of preparing a prediction module in which machine learning is performed using learning data including learning input data and learning output data, and the learning output data is a learning particle that models the substance. The learning particle model during the second time period longer than the first time by repeating the calculation of the behavior at the first time interval by the molecular dynamics method using the learning model including the model. Information on the position of the learning particle model after the lapse of the second time, or information on the position of the learning particle model after the lapse of the second time and the learning particle whose position has changed due to the movement of The learning input data includes at least information on the movement speed of the model after the lapse of the second time, and the training input data is information on the position of the learning particle model before the lapse of the second time, or the lapse of the second time. A prediction module preparation step that includes information on the previous position and information on the movement speed of the learning particle model before the second time elapses.
(3) Information on the relative position of the neighboring particle model located within a predetermined range around the attention particle model with respect to the attention particle model, using each of the analysis particle models in the analysis model as the attention particle model. Alternatively, using at least the relative position information and the moving speed information of the attention particle model and the neighboring particle model, the prediction module is provided with the position information of the attention particle model after the lapse of the second time. Alternatively, the elapsed time in the simulation is performed by performing a prediction process for predicting the position information of the focused particle model after the lapse of the second time and the moving speed information of the focused particle model after the lapse of the second time. The position acquisition step of obtaining the position information of the particle model of interest until is reached a predetermined time, and
(4) An evaluation step in which the computer evaluates the characteristics of the composite material using the position information obtained in the position acquisition step.
A method for analyzing a composite material, which comprises. The position of the position acquisition step is changed by the movement of the particle model for analysis during the period of the third time by the computer performing the calculation of the behavior by the molecular dynamics method using the analysis model. Further includes a calculation process for calculating the position information of the particle model for analysis after the lapse of the third time.
In the position acquisition step, the computer
A step of performing the prediction process using the information on the position of the particle model for analysis after movement obtained in the calculation process as the information on the position of the particle model of interest before movement used in the prediction process.
A step of performing the calculation process using the information on the position of the particle model of interest after movement obtained in the prediction process as information on the position of the particle model for analysis before movement used in the calculation process, and the prediction process. The step of further performing the prediction process by using the position information after the movement of the attention particle model obtained in the above as the position information before the movement of the attention particle model used in the prediction process.
By repeating any one of the steps, information on the position of the particle model for analysis in the analysis model is obtained until the elapsed time reaches a predetermined time longer than the second time. The method for analyzing a composite material according to claim 1, wherein the characteristics of the composite material are analyzed. The method for analyzing a composite material according to claim 2, wherein the calculation by the molecular dynamics method using the learning model and the calculation by the molecular dynamics method using the analysis model are performed under the same analysis conditions. In the learning data, each of the learning particle models is used as a learning attention particle model, and a learning neighborhood particle model located within the predetermined range around the learning attention particle model before the lapse of the second time. The information on the position relative to the learning attention particle model is included as the position information of the learning input data, and the position information of the learning attention particle model after the lapse of the second time is the learning output data. The method for analyzing a composite material according to any one of claims 1 to 3, which is included as information on the position of the above. In addition to the relative position information, the prediction module includes information on the moving speed of the learning attention particle model before the lapse of the second time, and within the predetermined range around the learning attention particle model. Information on the relative movement speed of the located learning proximity particle model with respect to the learning attention particle model before the lapse of the second time is used as the learning input data, and the position information of the learning attention particle model is used. In addition, the information on the moving speed of the learning attention particle model after the lapse of the second time is machine-learned as the learning output data.
The claim that the prediction process predicts information on the moving speed of the attention particle model after the lapse of the second time in addition to the information on the position of the attention particle model after the lapse of the second time. 4. The method for analyzing a composite material according to 4. The prediction module uses the distribution of the displacement amount of the analysis particle model before and after the passage of the second time obtained by the prediction module, and the analysis particle model of the second time by molecular dynamics. By machine learning, the difference from the displacement distribution of the particle model for analysis before and after the elapse of the second time obtained by repeating the calculation at a time interval shorter than the time interval is within an allowable range. The method for analyzing a composite material according to any one of claims 1 to 5, which is created by adjusting the parameters of the prediction module to be set. The prediction module uses the distribution of the moving speed in the analysis particle model obtained as a prediction result by the prediction module over the lapse of the second time, and the analysis particle model by the molecular dynamics method. The machine so that the difference from the distribution of the moving speed in the particle model for analysis over the lapse of the second time obtained by repeating the calculation at a time interval shorter than the time interval of the time is within an allowable range. The method for analyzing a composite material according to any one of claims 1 to 6, which is created by adjusting the parameters of the prediction module set by learning. The method for analyzing a composite material according to any one of claims 1 to 7, wherein the learning model and the analysis model are created under the same model creation conditions. Each of the analysis particle model and the learning particle model includes a plurality of types of particle models that model a plurality of substances in the composite material.
The training data includes information on the type of the particle model regarding the particle model for training.
The method for analyzing a composite material according to any one of claims 1 to 8, wherein in the prediction process, prediction is made using information on the type of the particle model regarding the particle model of interest and the neighboring particle model. The learning data is the force of interaction used in the molecular dynamics method, in which each of the learning particle models is used as a learning attention particle model and acts with a learning neighborhood particle model around the learning attention particle model. Includes information about the type of
The method for analyzing a composite material according to any one of claims 1 to 9, wherein in the prediction process, prediction is made using information on the type of interaction force with respect to the particle model of interest and the neighboring particle model. In the training data, when one of the learning particle models is a bound particle model that is bound to the learning particle model around one of the learning particle models by a bound chain, the bound particle model is combined with the bound particle model. It contains information on the presence or absence of the bound chain that can be distinguished from the non-bound learning particle model that is not bound by a chain.
The method for analyzing a composite material according to any one of claims 1 to 10, wherein in the prediction process, prediction is made using information on the presence or absence of the binding chain with respect to the particle model of interest and the neighboring particle model. In the training data, when one of the learning particle models is a bound particle model that is bound to the learning particle model around one of the learning particle models by a binding chain, the bound particle model is combined with the binding particle model. Contains information on the type of binding chain that can be distinguished from the bound particle model bound by a different type of binding chain from the chain.
The method for analyzing a composite material according to any one of claims 1 to 11, wherein the prediction process predicts using the information on the type of the binding chain regarding the particle model of interest and the neighboring particle model. The method for analyzing a composite material according to any one of claims 1 to 12, wherein the prediction module obtains a displacement amount and a moving speed due to movement of the particle model for analysis. A computer program for analyzing a composite material, which comprises causing a computer to execute the method for analyzing a composite material according to any one of claims 1 to 13.

Images