Method for predicting the physical properties of composite materials including polymer matrix phase and dispersed phase.

Abstract

This provides a novel method for predicting the physical properties of composite materials, including polymer matrix phases and dispersed phases. [Solution] A method for predicting the physical properties of a composite material including a polymer matrix phase and a dispersed phase, A method for predicting the physical properties of a composite material including a polymer matrix phase and a dispersed phase, comprising the steps of: (S1) a computer acquiring an image of the composite material; (S2) predicting the physical properties of the composite material using a predictive model trained to output the physical properties of the composite material with features extracted from the image as input; and (S3) visualizing the fixation region in the image that the predictive model fixated on during the prediction process in step (S2) using color attributes.

Description

Technical Field

[0001] The present invention relates to a method for predicting physical property values of a composite material including a polymer matrix phase and a dispersed phase.

Background Art

[0002] As a method for predicting physical property values of a composite material in which a filler or the like is blended with a polymer, various simulations using a computer are widely used.

[0003] For example, Patent Document 1 describes a fracture toughness simulation apparatus for a polymer material. In the simulation apparatus, a simulation model including a polymer matrix phase model, a filler model, and a polymer interface phase model is created, and using the simulation model, arithmetic processing is performed so as to achieve a mechanical equilibrium state when a predetermined stress or a predetermined strain is applied, and the stress and strain for each position element (pixel) of the simulation model are calculated.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] Although the simulation apparatus of Patent Document 1 is an effective method from the viewpoint of accurately predicting fracture toughness and the like, since the stress and strain for each position element of the simulation model are calculated, the time required for prediction tends to be relatively long.

[0006] Also, from the viewpoint of elucidating the structure - property correlation of an actual composite material by utilizing the results of a simulation using a simulation model of the composite material, it is important to explore the basis for judgment of the prediction by a computer.

[0007] This invention has been made in view of the above circumstances, and aims to provide a new method for predicting the physical properties of composite materials including a polymer matrix phase and a dispersed phase.

[0008] Specifically, in one embodiment of the present invention, the objective is to shorten the time required for prediction and to promote the elucidation of structural-physical-property relationships in a method for predicting the physical properties of a composite material including a polymer matrix phase and a dispersed phase. [Means for solving the problem]

[0009] The gist of this invention is as follows: [1] to [6]. [1] A method for predicting the physical properties of a composite material including a polymer matrix phase and a dispersed phase, The computer includes the steps of: acquiring an image of the composite material (S1); predicting the physical properties of the composite material using a predictive model trained on machine learning to output the physical properties of the composite material with features extracted from the image as input (S2); and visualizing the fixation region within the image that the predictive model has focused on during the prediction process in step (S2) using color attributes (S3). A method for predicting the physical properties of composite materials including a polymer matrix phase and a dispersed phase. [2] The method for predicting the physical properties of a composite material according to [1], wherein the above step (S1) is a step of acquiring an image generated from a composite material model created by a computer, the composite material model being a composite material model including a polymer matrix phase model, a dispersed phase model, and a polymer interface phase model. [3] The method for predicting the physical properties of a composite material according to [1] or [2], wherein the above step (S2) is a step of predicting the physical properties of the composite material using a predictive model trained with training data, wherein the training data is training data including training images generated from a composite material model created by a computer, and the composite material model is a composite material model including a polymer matrix phase model, a dispersed phase model, and a polymer interface phase model. [4] The above polymer interface phase model is a polymer interface phase model that displays differences in properties due to differences in color attributes, and is a method for predicting the physical properties of composite materials as described in [2] or [3]. [5] The above polymer interface phase model is a polymer interface phase model that displays differences in elastic modulus due to differences in color attributes, and the dispersed phase model is a filler particle model, a method for predicting the physical properties of composite materials as described in any of [2] to [4]. [6] The physical properties of the composite material described above are physical properties related to the stress of the composite material, and a method for predicting the physical properties of the composite material described in any of [1] to [5]. [Effects of the Invention]

[0010] According to the present invention, a novel method for predicting the physical properties of a composite material including a polymer matrix phase and a dispersed phase can be provided.

[0011] Furthermore, according to one embodiment of the present invention, for example, the time required for prediction can be reduced by predicting the physical properties of a composite material using a predictive model trained on machine learning to output the physical properties of the composite material with features extracted from an image as input.

[0012] Furthermore, according to one embodiment of the present invention, for example, by visualizing the area of ​​focus that the prediction model has focused on in the process of predicting the physical properties of a composite material using color attributes, the basis for the prediction can be clarified, and the elucidation of the structural-physical property correlation of the composite material can be promoted. [Brief explanation of the drawing]

[0013] [Figure 1] It is a flowchart showing an example of the processing procedure of the prediction method according to an embodiment of the present invention. [Figure 2] It is a diagram showing an example of the composite material model for analysis according to an embodiment of the present invention. [Figure 3] It is a partially enlarged view of FIG. 2. [Figure 4] It is a diagram showing an example for explaining the prediction method according to an embodiment of the present invention. [Figure 5] It is a diagram showing an example for explaining the prediction method according to an embodiment of the present invention. [Figure 6] It is a diagram showing an example of the prediction accuracy of the prediction method according to an embodiment of the present invention. [Figure 7] It is a diagram showing an example for explaining the prediction method according to an embodiment of the present invention.

Embodiments for Carrying Out the Invention

[0014] A method for predicting the physical property values of a composite material including a polymer matrix phase and a dispersed phase according to an embodiment of the present invention (hereinafter sometimes referred to as "this prediction method") is to input an image of the composite material into a prediction model and output the physical property values of the composite material from the prediction model, thereby predicting the physical property values of the composite material. Further, in the process of prediction, a fixation region in the image watched by the prediction model is visualized by color attributes.

[0015] This prediction method is, for example, as shown in FIG. 1, a prediction method in which a computer executes the following steps S1 to S3. Note that the flowchart of FIG. 1 shows an example of an embodiment, and the present invention is not limited to this processing. 〔Step S1〕 Step S1 of acquiring an image of a composite material including a polymer matrix phase and a dispersed phase. 〔Step S2〕 Step S2 predicts the physical properties of the composite material using a predictive model trained on machine learning to output the physical properties of the composite material, with the features extracted from the image acquired in step S1 as input. [Step S3] Step S3 visualizes the area of ​​focus within the image that the prediction model focused on during the prediction process of Step S2, using color attributes.

[0016] Steps S1 to S3 of this prediction method are executed by a computer. The computer is the same as those used in conventional CAE (Computer-Aided Engineering). In this prediction method, a personal computer or other information processing device equipped with a processing unit (CPU), ROM, working memory, storage devices such as magnetic disks, input devices such as keyboards and mice, and display devices such as displays, typically executes pre-stored processing routines through the collaborative action of software and hardware.

[0017] The composite materials targeted by this prediction method are those containing a polymer matrix phase and a dispersed phase.

[0018] Polymers that constitute the polymer matrix phase include, for example, rubber and resins. More specifically, but not limited to, ethylene-propylene-diene monomer ternary copolymer (EPDM), acrylic rubber, urethane rubber, styrene-butadiene-styrene block polymer (SBS), styrene-isobutylene-styrene block polymer (SIBS), styrene-butadiene (SB) copolymer, styrene-isoprene (SI) copolymer, styrene-isoprene-styrene (SIS) copolymer, styrene-ethylene-butylene (SEB) copolymer, styrene-ethylene-butylene-styrene (SEBS) copolymer, and styrene-ethylene-propylene Examples include pyrene (SEP) copolymer, styrene-ethylene-propylene-styrene (SEPS) copolymer, hydrogenated versions of the above copolymers, ethylene-propylene copolymer (EPR), butadiene rubber (BR), isoprene rubber (IR), styrene-butadiene rubber (SBR), liquid isoprene rubber (liquid IR), liquid butadiene rubber (liquid BR), liquid styrene-butadiene rubber (liquid SBR), liquid styrene-isoprene rubber (liquid SI), liquid styrene-ethylene-propylene rubber (liquid SEP), and liquid isoprene-butadiene rubber (liquid IR-BR).

[0019] The dispersed phase is not particularly limited as long as it is a component dispersed in the polymer matrix phase, but examples include various additives added to the polymer constituting the polymer matrix phase for a predetermined purpose, and polymers other than the polymer constituting the polymer matrix phase. The components constituting the dispersed phase may be one type alone or two or more types in combination.

[0020] The dispersed phase can include, but is not limited to, various organic and inorganic fillers. More specifically, examples include carbon black, silica, talc, calcium carbonate, carbon fibers, and carbon nanotubes.

[0021] Furthermore, the composite materials targeted by this prediction method include, for example, polymer alloys in which a polymer different from the polymer constituting the polymer matrix phase is used as the dispersed phase.

[0022] Specific examples of composite materials include, but are not limited to, materials for vibration-damping rubber used in vibration isolation applications in automobiles, materials for various types of tubes or hoses for automobiles, materials for automobile tires, and materials for various types of sealing components.

[0023] <Step S1> This prediction method includes step S1, in which a computer acquires an image of the composite material, including the polymer matrix phase and the dispersed phase. Step S1 is a step in which a computer acquires an image obtained by photographing the composite material with an electron microscope, or a step in which a computer acquires an image generated from a composite material model for analysis that is virtually created on the computer.

[0024] The image obtained by photographing the composite material with a microscope is an image obtained by photographing the composite material, which includes a polymer matrix phase and a dispersed phase, using an electron microscope such as a transmission electron microscope or a scanning electron microscope. In step S1, the computer acquires image data having a predetermined number of pixels, which is generated based on the microscope image taken with, for example, an electron microscope.

[0025] The composite material model for analysis, virtually created on a computer, is an analysis model generated by a known simulation. For example, it is image data generated from an analysis composite material model composed of multiple position elements that can be numerically analyzed by a computer within a virtual model creation domain where periodic boundary conditions are defined.

[0026] The composite material model for analysis includes at least a polymer matrix phase model and a dispersed phase model. The dispersed phase model corresponds to a dispersed phase dispersed in the polymer matrix phase, and specifically, for example, is a filler particle model dispersed in the polymer matrix phase. Dispersed phase models such as filler particle models are generated by approximately modeling actual shapes such as spheres, oblate spheres (long ellipsoids, flattened ellipsoids), plate-like bodies, and cylindrical bodies, and their size (size of the area occupied by the model creation region), such as radius, diameter, and volume, is predetermined.

[0027] Specifically, the composite material model for analysis is created, for example, by randomly arranging multiple dispersed phase models, such as filler particle models, at arbitrary coordinates (X, Y, Z) in a virtual model creation domain based on a predetermined algorithm. Specifically, the dispersed phase models are arranged by appropriately setting, for example, the particle size and quantity of the filler particle models. The polymer matrix phase model can be set, for example, as a region other than the region where the dispersed phase models are created within the virtual model creation domain. Another example of a polymer interface phase model is a polymer interface phase model formed between a filler particle model and a polymer matrix phase model. For example, a position element of the arranged filler particle model where at least one adjacent element becomes a polymer matrix phase model can be set as the polymer interface phase model.

[0028] From the viewpoint of improving prediction accuracy, it is preferable that the polymer interface phase model displays differences in properties through differences in color attributes. By displaying color attributes (brightness, hue, saturation) in the regions (assemblies of positional elements) that constitute the polymer interface phase model, for example, by displaying the intensity of brightness, various properties such as the magnitude of the elastic modulus can be included in the image information, thereby improving prediction accuracy.

[0029] <Step S2> This prediction method includes the step of predicting the physical properties of a composite material using a machine learning-trained prediction model that takes features extracted from an image as input and outputs the physical properties of the composite material.

[0030] One embodiment of the prediction model of this prediction method is, for example, a neural network (CNN: Convolutional Neural Network) that outputs information about the physical properties of a composite material in response to an input image of the composite material.

[0031] The prediction model in this prediction method is generated by machine learning a neural network using training data that associates images of composite materials with information on their physical properties. In one embodiment of this prediction method, for example, machine learning is performed using a dataset that includes training images of composite materials and corresponding training data as training data. The training images of composite materials may be either training images obtained by photographing the composite material with an electron microscope, or training images generated from a composite material model for analysis virtually created on a computer. Furthermore, the training data included in the training data may be training data on physical properties (measured values) of composite materials obtained from actual experiments, or training data on physical properties (simulated values) obtained from virtual experiments (physical property simulations) using a composite material model for analysis performed on a computer.

[0032] The specific configuration of the prediction model is not particularly limited, but for example, a prediction model having an input layer into which an image of the composite material is input, an intermediate layer, and an output layer (coupled layer) can be used. The intermediate layer has a configuration in which a convolution layer that convolves the pixel values ​​of each pixel input from the input layer and a pooling layer that maps the pixel values ​​convolved in the convolution layer are alternately connected, and image features are extracted while compressing the pixel information of the image. The output layer outputs information about the physical properties of the composite material based on the features output from the intermediate layer.

[0033] In one preferred embodiment, the prediction model in this prediction method may be a pre-trained model, and transfer learning or fine-tuning may be performed on that model using the training data mentioned above. For example, a convolutional neural network may be pre-trained using a large image dataset (such as ImageNet). For instance, by extracting layers up to those that extract higher-order features from a neural network trained on existing labeled images such as ImageNet, and combining these layers with a classifier that takes these features as input, sufficient prediction accuracy can be ensured even with machine learning using a relatively small amount of training data.

[0034] Furthermore, in one preferred embodiment of this prediction method, it is preferable to employ a ResNet network (Residual Networks) as the convolutional neural network from the viewpoint of gradient vanishing in deep networks. A ResNet, for example, has a building block structure and includes one residual branch and a short-cut branch. An identity mapping is added to the residual branch, and the current output is directly transmitted to the next layer of the network, and the gradient of the next layer of the network is also directly transmitted to the upper layer of the network during the backpropagation process.

[0035] <Step S3> This prediction method includes a step in which the gaze region within the image that the prediction model focused on during the prediction process in step S2 is visualized using color attributes. Specifically, the gaze region within the image that formed the basis for the prediction is identified, and the gaze region is visualized by processing it in such a way that it can be visually distinguished from other regions, such as by highlighting it in stages using color attributes (hue, saturation, brightness) according to the relative strength of the basis for the prediction. For example, although not limited to the following, a gaze region image is generated by extracting gaze regions that have a significant influence on the physical properties output by the prediction model for the input image, and a gaze region-enhanced image is generated by overlaying the gaze region image onto the input image to highlight the gaze region and visualize it. This makes it possible to find factors that have a significant influence on the physical properties of composite materials, and can contribute to the establishment of guidelines for material design.

[0036] The method for performing step S3 is not particularly limited, but for example, the gaze region can be extracted using CAM (Class Activation Mapping). CAM is one method for identifying the basis of decisions made by neural networks, and for example, Grad-CAM (Gradient-weighted CAM) can be used. Grad-CAM is a method that replaces the weights used in CAM calculations with the gradient during backpropagation, and can identify the basis of decisions in various types of neural networks. Examples include Grad-CAM++, Score-CAM, Ablation-CAM, Eigen-CAM, and Integrated Grad-CAM.

[0037] Step S3 is not particularly limited, but for example, it can be performed by calculating the gradient of the features output by the prediction model, performing a GAP (Global Average Pooling) operation on the gradient to generate a pooling result with reduced dimensionality of the gradient, calculating a weighted sum of the features and the pooling result, and inputting the weighted sum into an activation function (such as ReLU (Rectified Linear Unit)).

[0038] (Usefulness of this prediction method) According to one embodiment of the present invention, the time required for prediction can be reduced by using a predictive model trained on machine learning to output the physical properties of a composite material, taking features extracted from an image as input. Specifically, for example, the time required for prediction can be reduced to about one second or less per image.

[0039] Furthermore, according to one embodiment of the present invention, for example, by visualizing the region that the prediction model focuses on during the prediction of the physical properties of a composite material using color attributes, it is extremely useful in that it can facilitate the elucidation of the structural-physical property correlation of a composite material.

[0040] <Prediction device> Furthermore, one embodiment of the present invention is a prediction device that predicts the physical properties of a composite material using software that performs each of the above steps. Specifically, for example, this is a prediction device that predicts the physical properties of a composite material using software that performs the steps of acquiring an image of the composite material (S1), predicting the physical properties of the composite material using a prediction model that has been trained to output the physical properties of the composite material using features extracted from the image as input (S2), and visualizing the gaze region in the image that the prediction model gazed upon during the prediction process of step (S2) using color attributes (S3). [Examples]

[0041] The following describes embodiments of the present invention. However, the present invention is not limited to these embodiments.

[0042] (Creating training data) A composite material model for analysis was created in a virtual model creation area on a computer, including a polymer matrix phase model, a filler particle model as a dispersed phase model, and a polymer interface phase model. Specifically, the particle size of the filler particle model was set to multiple levels (3 levels: 10 nm, 15 nm, 30 nm), the number of dispersed particles to multiple levels (3 levels: 200, 600, 1000), the thickness of the polymer interface phase model to multiple levels (3 levels: 5 nm, 10 nm, 15 nm), the elastic modulus of the polymer interface phase model to multiple levels (3 levels: 1.5 MPa, 2.0 MPa, 2.5 MPa), and the number of samples (N) was set to 5, resulting in a total of 405 composite material models for analysis. Image data was generated by slicing the created composite material models for analysis. In addition, different brightness levels were set for the polymer interface phase model corresponding to the levels of elastic modulus mentioned above.

[0043] Figure 2 shows an example of image data generated from a composite material model for analysis, and Figure 3 shows a magnified view of a part of it. In Figures 2 and 3, the areas shown in black represent the filler particle model, the areas shown in white represent the polymer matrix phase model, and the areas shown in gray represent the polymer interface phase model. The difference in elastic modulus is indicated by the shades of gray (differences in brightness).

[0044] Next, a physical property simulation (a virtual experiment using a computer) was performed on a computer to calculate the physical properties of the composite material model created above for analysis. Specifically, using Phase-Field Models analysis application software, the composite material model for analysis was subjected to tensile deformation in the vertical direction (constant strain condition (strain 1.5)), and stress simulation calculations were performed to calculate the stress and strain for each position element (pixel), thereby obtaining stress distribution information for each composite material model for analysis. Furthermore, a histogram was generated based on the obtained stress distribution information, and the top 5% of stresses were divided by the mean stress to obtain physical property values ​​related to stress. The physical property values ​​related to stress obtained above are values ​​that indicate the degree of stress concentration, and the smaller the value, the less the stress is concentrated and the more dispersed it is, which is an indicator that can be evaluated as having good durability.

[0045] Figure 4 shows an example of stress distribution information obtained from an example of a composite material model for analysis. Figure 5 also shows an example of the histogram. The top 5% of stresses in the histogram in Figure 5 are represented by the region to the right of the dashed line shown in Figure 5.

[0046] As described above, we obtained training data including a set of image data generated from a composite material model for analysis, and a set of training data for the physical properties corresponding to that image data.

[0047] (Building a predictive model) A predictive model was constructed by using the parameters of a ResNet18 model pre-trained on the large-scale image dataset ImageNet as initial values, modifying the connected layers to perform regression based on the above-mentioned physical properties, and performing machine learning using the above-mentioned training data.

[0048] (Prediction time and prediction accuracy) To verify the prediction time and accuracy, images generated from a newly created composite material model for analysis were prepared separately and input into the prediction model to predict the material properties. As a result, the time required for prediction was less than 1 second per image.

[0049] Furthermore, since the time required for property simulations (virtual experiments using a computer) to calculate the physical properties of composite material models for analysis can be several hours per image, it can be seen that the prediction method according to the embodiment of the present invention can significantly reduce the time required for prediction.

[0050] Furthermore, Figure 6 shows the prediction accuracy of the prediction model according to this embodiment. The coefficient of determination in Figure 6 is approximately 0.88, indicating that good prediction accuracy was obtained.

[0051] From the above, it can be seen that the time required for prediction can be shortened according to one embodiment of the present invention. Furthermore, it can be seen that one embodiment of the present invention can provide a prediction method that can shorten the time required for prediction and also has good prediction accuracy.

[0052] (Visualization of the area of ​​focus) In a prediction model according to an embodiment of the present invention, the EigenCAM algorithm was applied to generate an attention map, and the regions that the prediction model focused on during the prediction process were visualized. An example of the results is shown in Figure 7. In Figure 7, the regions highlighted in white are the regions that were particularly focused on. By observing this attention map, we can see that the bonding structure of filler particles and the interface phase are closely observed during the prediction process. This can contribute to establishing guidelines when considering and designing highly durable structures for composite materials, for example. [Industrial applicability]

[0053] The present invention provides a novel method for predicting the physical properties of composite materials, including a polymer matrix phase and a dispersed phase. In particular, it is extremely useful because it allows for the prediction of physical properties in a relatively short time, and by visualizing the areas of focus observed during the prediction process, it can contribute to establishing guidelines when considering or designing highly durable structures for composite materials.

Claims

1. A method for predicting the physical properties of a composite material including a polymer matrix phase and a dispersed phase, The computer includes the steps of: acquiring an image of the composite material (S1); predicting the physical properties of the composite material using a predictive model trained to output the physical properties of the composite material with features extracted from the image as input (S2); and visualizing the gaze region in the image that the predictive model gazed upon during the prediction process in step (S2) using color attributes (S3). A method for predicting the physical properties of composite materials including a polymer matrix phase and a dispersed phase.

2. The method for predicting the physical properties of a composite material according to claim 1, wherein the above step (S1) is a step of acquiring an image generated from a composite material model created by a computer, the composite material model being a composite material model including a polymer matrix phase model, a dispersed phase model, and a polymer interface phase model.

3. The method for predicting the physical properties of a composite material according to claim 1, wherein step (S2) is a step of predicting the physical properties of the composite material using a predictive model trained with training data, the training data is training data including training images generated from a composite material model created by a computer, and the composite material model is a composite material model including a polymer matrix phase model, a dispersed phase model, and a polymer interface phase model.

4. The method for predicting the physical properties of a composite material according to claim 2 or 3, wherein the polymer interface phase model described above is a polymer interface phase model that displays differences in properties due to differences in color attributes.

5. The method for predicting the physical properties of a composite material according to claim 4, wherein the polymer interface phase model is a polymer interface phase model that displays differences in elastic modulus due to differences in color attributes, and the dispersed phase model is a filler particle model.

6. A method for predicting the physical properties of a composite material according to claim 1 or 2, wherein the physical properties of the composite material are physical properties related to the stress of the composite material.

Images