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

Abstract

To allow information on a pixel close to a corner of an image to be properly calculated in convolutional calculation.SOLUTION: A convolution machine learning model (30) includes: a material image acquiring unit (33) for acquiring a material image (100) indicating a target material; a window setting unit (34) for setting a window (200) in a target region (110) on a part of the material image; a window image generating unit (35) for generating a window image by, if the window includes a non-image region (300) located outside the material image, locating an image of a predetermined edge region (120) of the material image which is different from the target region on the non-image region; and an arithmetic unit (36) for subjecting convolution calculation to the window image.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to a physical property value prediction device, a physical property value prediction method, and a program. [Background technology]

[0002] It is known that the physical properties of a substance can be calculated using the finite element method or molecular simulation. For example, the elastic modulus of filler-filled rubber used in tires can be calculated using simulations. However, the reality is that obtaining physical properties using these methods requires an enormous amount of calculation time.

[0003] For example, Patent Document 1 below describes estimating the physical property values ​​of a rubber material using a convolutional machine learning model based on a CNN (Convolutional Neural Network). By inputting a microscopic image or other image of a substance into the convolutional machine learning model and having the model infer the physical property values, the physical property values ​​of the target substance can be calculated in a relatively short time. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2021-060457 Summary of the Invention [Problem to be solved by the invention]

[0005] In the convolutional layer of a convolutional machine learning model, a filter window is set in a part of the image, and convolution operations are sequentially performed on the pixel values ​​within the window while shifting the position of this window. However, with this type of conventional convolution operation, pixels near the corners of the image (edges and corners) are subject to the convolution operation less frequently. This has led to the problem that information about pixels near the corners of the image cannot be properly calculated by the convolutional machine learning model.

[0006] An object of the present disclosure is to provide a physical property value prediction device, a physical property value prediction method, and a program that can appropriately calculate information about pixels near the corners of an image. [Means for solving the problem]

[0007] The physical property prediction device according to the present disclosure includes a convolutional machine learning model including: a material image acquisition means for acquiring a material image showing a target substance; a window setting means for setting a window in a partial target region of the material image; a window image generation means for, if the window includes a non-image region outside the material image, placing an image of a predetermined edge region of the material image that is different from the target region in the non-image region to generate a window image; and a means for performing a convolution operation on the window image, and predicts the physical property value of the target substance. This allows information on pixels near the corners of the material image to be properly calculated.

[0008] Further, a physical property prediction method according to the present disclosure is a physical property prediction method for predicting a physical property of a target substance, wherein a substance image acquisition means of a convolutional machine learning model acquires a substance image representing the target substance, a window setting means of the convolutional machine learning model sets a window in a partial target region of the substance image, and if the window includes a non-image region located outside the substance image, a window image generation means of the convolutional machine learning model generates a window image by placing an image of a predetermined edge region of the substance image that is different from the target region in the non-image region, and a calculation means of the convolutional machine learning model performs a convolution calculation on the window image. This enables the convolutional machine learning model to properly calculate information about pixels near the corners of the substance image.

[0009] Furthermore, a program according to the present disclosure causes a computer that predicts the physical property values ​​of a target substance to perform the following steps: acquire a substance image showing the target substance; set a window in a partial target region of the substance image; if the window includes a non-image region outside the substance image, generate a window image by placing an image of a predetermined edge region of the substance image that is different from the target region in the non-image region; and perform a convolution operation on the window image. This allows the computer to properly calculate information on pixels near the corners of the substance image. [Brief explanation of the drawings]

[0010] [Figure 1] 1 is a diagram illustrating a configuration of a physical property value prediction device according to an embodiment of the present disclosure. [Figure 2] FIG. 2 is a functional block diagram showing an example of functions implemented in the physical property value prediction device. [Figure 3] FIG. 1 is a diagram showing an example of a three-dimensional model showing a target substance. [Figure 4] FIG. 10 is a diagram showing an example of a substance image showing a target substance. [Figure 5] FIG. 1 is a diagram illustrating an example of processing performed in a convolutional machine learning model. [Figure 6A] FIG. 1 is a diagram illustrating an example of conventional convolution processing. [Figure 6B] FIG. 1 is a diagram illustrating an example of conventional convolution processing. [Figure 6C] FIG. 1 is a diagram illustrating an example of conventional convolution processing. [Figure 7A] FIG. 10 is a diagram illustrating an example of a method for generating a window image. [Figure 7B] FIG. 10 is a diagram illustrating an example of a method for generating a window image. [Figure 7C] FIG. 10 is a diagram illustrating an example of a method for generating a window image. [Figure 8] FIG. 1 is a diagram illustrating an example of convolution processing proposed in the present disclosure. [Figure 9] FIG. 2 is a flowchart showing an example of a flow of processing performed by the physical property value prediction device. DETAILED DESCRIPTION OF THE INVENTION

[0011] Hereinafter, an embodiment of the present disclosure will be described with reference to the drawings. In this embodiment, a material image showing a material to be subjected to physical property prediction (hereinafter referred to as a target material) is input into a convolutional machine learning model using CNN, and a physical property value showing the physical property of the target material is calculated. The morphology of a material (for example, the filler morphology of filler-filled rubber) is one of the factors that control material properties. However, in order to predict the physical properties of a target material using morphology, which is difficult to quantify, as input data, it is preferable to input the material image, which is an image of the target material, into a convolutional machine learning model.

[0012] In the convolution calculation performed in a conventional convolution machine learning model, pixels near the corners (edges and corners) of a material image are rarely subjected to the convolution calculation. In this regard, in the present embodiment, the convolution calculation is designed to appropriately calculate information on pixels near the corners of a material image.

[0013] [1. Hardware configuration] Fig. 1 is a diagram showing the configuration of a physical property prediction device 10 according to an embodiment of the present disclosure. The physical property prediction device 10 according to this embodiment is a computer such as a personal computer, a general-purpose computer, or a mobile information terminal, and includes a processor 11, a storage unit 12, a communication unit 13, a display unit 14, and an operation unit 15, as shown in Fig. 1. The physical property prediction device 10 may also include an optical disc drive for reading optical discs, a USB (Universal Serial Bus) port, etc.

[0014] The processor 11 is a program-controlled device such as a CPU (Central Processing Unit) that operates according to a program installed in the physical property prediction device 10, which is, for example, a computer. The memory unit 12 is a storage element such as a ROM (Read Only Memory) or a RAM (Random Access Memory), or a hard disk drive. The memory unit 12 stores data such as programs executed by the processor 11. The communication unit 13 is a communication interface such as a network board. The display unit 14 is a display device such as a liquid crystal display, and displays various images according to instructions from the processor 11. The operation unit 15 is a user interface such as a keyboard or a mouse, and accepts user operation inputs and outputs signals indicating the contents of the inputs to the processor 11.

[0015] [2. Functional Blocks] Fig. 2 is a functional block diagram showing an example of functions implemented in the physical property prediction device 10. As shown in Fig. 2, the physical property prediction device 10 functionally includes a substance image generation unit 20, a convolutional machine learning model 30, and a result output unit 40. The substance image generation unit 20 and the result output unit 40 may be realized mainly by a processor 11. Note that the physical property prediction device 10 does not necessarily have to implement all of the functions shown in Fig. 2, and functions other than the functions shown in Fig. 2 may also be implemented.

[0016] [2-1. Material image generation section] The substance image generating unit 20 generates a substance image showing a target substance. The substance image generating unit 20 generates the substance image from a three-dimensional model showing the target substance generated by a given simulation such as a molecular simulation.

[0017] Fig. 3 is a diagram showing an example of a three-dimensional model of a target substance. Fig. 4 is a diagram showing an example of a substance image showing a target substance. In this embodiment, as shown in Figs. 3 and 4, the substance image generating unit 20 generates a plurality of substance images 100 showing the target substance by cutting out a plurality of three-dimensional models 90 of the target substance generated by simulation along a predetermined plane (for example, the front end surface 90F).

[0018] Since simulations such as the finite element method and molecular simulation cannot target regions at infinity, in this embodiment, the three-dimensional model 90 is generated as a rectangular parallelepiped model of representative volume elements in which periodic boundaries (cyclic symmetry) are set on opposing end faces so that these end faces are periodically connected. In the example shown in FIG. 3 , periodic boundaries are set on at least two pairs of end faces of the three-dimensional model 90: the top end face 90U and the bottom end face 90D, the left end face 90L and the right end face 90R, and the front end face 90F and the rear end face 90F. Therefore, in this embodiment, periodic boundaries are set on two opposing sides (ends) of the material image 100 cut out from the three-dimensional model 90. More specifically, periodic boundaries are set on the top side 100U and the bottom side 100D, and the left side 100L and the right side 100R of the material image 100. Note that in this disclosure, the term "rectangular parallelepiped" is used to include a cube.

[0019] Alternatively, for example, the substance image generating unit 20 may generate the substance image 100 of the target substance by photographing the target substance using a camera or an electron microscope. In this case, too, it is preferable that the substance image 100 is rectangular. In this disclosure, the term "rectangle" is explained as including a square.

[0020] [2-2. Machine Learning Model] The convolutional machine learning model 30 is a machine learning model based on a convolutional neural network (CNN), and is trained using training data including material images and physical property values ​​of the material represented by the material images. When a given material image is input, the convolutional machine learning model 30 outputs the physical property values ​​of the material represented by the material image. Here, the "physical property value" may be a physical quantity such as elastic modulus, a class defined by a range of physical quantities, or a class representing the type or proportion of materials (e.g., fillers or polymers) blended into the target material. Therefore, the convolutional machine learning model 30 may be a so-called classification model or a regression model.

[0021] As shown in FIG. 2, the convolutional machine learning model 30 functionally includes a parameter storage unit 31, a learning unit 32, a substance image acquisition unit 33, a window setting unit 34, a window image generation unit 35, a calculation unit 36, and a prediction unit 37. The parameter storage unit 31 may be realized primarily by the storage unit 12 of the physical property prediction device 10, or may be realized by another storage device. The learning unit 32, the substance image acquisition unit 33, the window setting unit 34, the window image generation unit 35, the calculation unit 36, and the prediction unit 37 may be realized primarily by the processor 11. Note that the convolutional machine learning model 30 does not necessarily have to implement all of the functions shown in FIG. 2, and functions other than the functions shown in FIG. 2 may be implemented.

[0022] [2-2-1. Parameter storage section] The parameter storage unit 31 stores parameters of the convolutional machine learning model 30 for outputting physical property values ​​when a given substance image is input to the convolutional machine learning model 30. More specifically, the parameter storage unit 31 stores the number of convolutional layers, the number of nodes used in each convolutional layer, the weighting of each node, and the like as parameters of the convolutional machine learning model 30.

[0023] [2-2-2. Learning Department] The learning unit 32 performs machine learning using a substance image 100 indicating a substance and a plurality of pieces of learning data including physical property values ​​of the substance, thereby updating the parameters of the convolutional machine learning model 30 stored in the parameter storage unit 31. In other words, the parameters of the convolutional machine learning model 30 stored in the parameter storage unit 31 are learned (updated) using a plurality of pieces of learning data.

[0024] [2-2-3. Substance image acquisition section] The substance image acquisition unit 33 acquires a substance image 100 that shows a target substance. In this embodiment, the substance image acquisition unit 33 acquires the substance image 100 that the substance image generation unit 20 generates from a three-dimensional model or the like. Note that, when the substance image generation unit 20 generates a plurality of substance images 100, the substance image acquisition unit 33 acquires any one of the plurality of substance images 100.

[0025] Fig. 5 is a diagram showing an example of processing performed by the convolutional machine learning model 30. As shown in Fig. 5, the convolutional machine learning model 30 has a plurality of convolutional layers C1 to C3 and a fully connected layer F. The convolution processing by the calculation unit 36 ​​in the convolutional layers C1 to C3 may generate an image showing the characteristics of the pixels of the substance image 100 as a result of the convolution calculation. The substance image acquisition unit 33 may also acquire an image showing the result of the convolution calculation calculated by the calculation unit 36 ​​as the substance image 100 showing the target substance.

[0026] [2-2-4. Window setting section, calculation section] The window setting unit 34 sets a window to serve as a filter in a partial target region of the substance image 100 acquired by the substance image acquisition unit 33. The calculation unit 36 ​​executes a convolution calculation by extracting features of pixels of the substance image 100 from the window in each of the convolution layers C1 to C3 shown in FIG.

[0027] 6A to 6C are diagrams showing an example of conventional convolution processing. First, as shown in Fig. 5A, the window setting unit 34 sets the upper left region of the substance image 100 as the initial target region 110, and sets a rectangular window 200 therein. The size of the window 200 is preferably equal to or smaller than the size of the substance image 100.

[0028] 6A, at the initial position of the window 200 before it is moved, the window 200 includes the region in the upper left corner of the material image 100 and also includes a non-image region 300 located outside the material image 100. The calculation unit 36 ​​uses the window 200 in this position to extract pixel features in the target region 110 of the material image 100.

[0029] Next, as shown in FIG. 6B , the window setting unit 34 moves the window 200 to the right from its current position in the substance image 100 by a predetermined distance. The region of the substance image 100 that overlaps with the moved window 200 becomes the target region 110. The calculation unit 36 ​​extracts pixel features in the target region 110 of the substance image 100 using the moved window 200. This process is repeated until the window 200 reaches a position that includes the right edge of the substance image 100.

[0030] Next, the window setting unit 34 moves the window 200 downward a predetermined distance from its initial position in the substance image 100, and the calculation unit 36 ​​extracts pixel features using the moved window 200. Next, the window setting unit 34 moves the window 200 to the right a predetermined distance from its current position, and the calculation unit 36 ​​extracts pixel features using the moved window 200. By repeating this process, when the window 200 reaches a position that includes the right edge of the substance image 100, the window setting unit 34 moves the window 200 downward a predetermined distance from the position before the window 200 was continuously moved rightward (the position where the window 200 includes the left edge of the substance image 100), and the calculation unit 36 ​​extracts pixel features using the moved window 200.

[0031] 6C , the window setting unit 34 moves the window 200 rightward or downward, and the calculation unit 36 ​​repeats the convolution process to extract features of pixels in the substance image 100 until the window 200 finally reaches a position including the lower right corner region of the substance image 100. Note that information on the size and initial position of the window 200 set by the window setting unit 34, the distance moved rightward or downward in one movement, and the range of movement relative to the substance image 100 may be stored in the parameter storage unit 31 as parameters of the convolutional machine learning model 30.

[0032] However, in a conventional convolution calculation performed in this manner, pixels near the corners (edges or angles) of the substance image 100 are subjected to the convolution calculation less frequently. In this regard, in this embodiment, the calculation unit 36 ​​performs the convolution calculation on a window image 400 (see FIG. 6C), which will be described later. By doing so, it is possible to appropriately calculate information on pixels near the corners of the substance image 100.

[0033] [2-2-5. Window image generation section] When the window 200 includes a non-image area 300 located outside the substance image 100, as shown in Figure 5A, the window image generation unit 35 generates a window image by placing (copying) an image of a specified edge area 120 of the substance image 100, which is different from the target area 110 in which the window 200 is set, into the non-image area 300.

[0034] 7A to 7C are diagrams showing an example of a method for generating a window image 400. First, as shown in Fig. 7A, the window image generation unit 35 places an image of a rectangular edge region 120D of the substance image 100 that abuts on a bottom side 100D of the substance image 100 in a rectangular non-image region 300U that abuts on a top side 100U of the substance image 100. The window image generation unit 35 also places an image of a rectangular edge region 120U of the substance image 100 that abuts on a top side 100U of the substance image 100 in a rectangular non-image region 300D that abuts on a bottom side 100D of the substance image 100.

[0035] It is preferable that the shape and size of the non-image region 300U and the edge region 120D are the same. Similarly, it is preferable that the shape and size of the non-image region 300D and the edge region 120U are the same. In this embodiment, the shapes and sizes of the non-image regions 300U, 300D and the edge regions 120U, 120D are all the same. Furthermore, in this embodiment, the vertical widths of the non-image regions 300U, 300D and the edge regions 120U, 120D are set to half the vertical width of the window 200.

[0036] In the material image 100, the lower edge region 120D is located on the opposite side of the upper edge region 120U (on the opposite side of a line that passes through the center of the material image 100 and is parallel to the upper and lower edges 100U and 100D). In other words, the edge region 120D will be located on the opposite side of the target region 110 when the window 200 is positioned to include at least a portion of the edge region 120U. Similarly, the edge region 120U will be located on the opposite side of the target region 110 when the window 200 is positioned to include at least a portion of the edge region 120D.

[0037] 7B , the window image generation unit 35 places an image of a rectangular edge region 120R of the substance image 100 that abuts the right side 100R of the substance image 100 in a rectangular non-image region 300L that abuts the left side 100L of the substance image 100. The window image generation unit 35 also places an image of a rectangular edge region 120L of the substance image 100 that abuts the left side 100L of the substance image 100 in a rectangular non-image region 300R that abuts the right side 100R of the substance image 100.

[0038] It is preferable that the shapes and sizes of the non-image region 300L and the edge region 120R are the same, and it is also preferable that the shapes and sizes of the non-image region 300R and the edge region 120L are the same. In this embodiment, the shapes and sizes of the non-image regions 300L, 300R and the edge regions 120L, 120R are all the same. Also, in this embodiment, the horizontal widths of the non-image regions 300L, 300R and the edge regions 120L, 120R are set to half the horizontal width of the window 200.

[0039] In the material image 100, the right edge edge region 120R is located on the opposite side of the left edge edge region 120L (on the opposite side of a line that passes through the center of the material image 100 and is parallel to the left side 100L and the right side 100R). In other words, the edge region 120R will be located on the opposite side of the target region 110 when the window 200 is positioned to include at least a portion of the edge region 120L. Similarly, the edge region 120L will be located on the opposite side of the target region 110 when the window 200 is positioned to include at least a portion of the edge region 120R.

[0040] 7C , the window image generation unit 35 places an image of a rectangular edge region 120S including the lower right corner of the substance image 100 in a rectangular non-image region 300N that contacts the upper left vertex of the substance image 100. The window image generation unit 35 also places an image of a rectangular edge region 120W including the lower left corner of the substance image 100 in a rectangular non-image region 300E that contacts the upper right vertex of the substance image 100, places an image of a rectangular edge region 120E including the upper right corner of the substance image 100 in a rectangular non-image region 300W that contacts the lower left vertex of the substance image 100, and places an image of a rectangular edge region 120S including the upper left corner of the substance image 100 in a rectangular non-image region 300N that contacts the lower right vertex of the substance image 100. In this way, the image in which the images of edge regions 120U, 120D, 120L, 120R, 120N, 120S, 120E, and 120W are placed in non-image regions 300U, 300D, 300L, 300R, 300N, 300S, 300E, and 300W respectively is window image 400.

[0041] As described above, in this embodiment, periodic boundaries are set on the top side 100U and bottom side 100D and the left side 100L and right side 100R of the substance image 100. Therefore, the window image generator 35 generates the window image 400 by placing an image of the edge region 120 that contacts one of the two sides (two ends) on which a periodic boundary is set, in the non-image region 130 that contacts the other of the two sides. In other words, the window image generator 35 generates the window image 400 with periodic boundaries set on each side of the substance image 100.

[0042] It is preferable that the shapes and sizes of the non-image region 300S and the edge region 120N are the same, and it is also preferable that the shapes and sizes of the non-image region 300N and the edge region 120S are the same. Similarly, it is preferable that the shapes and sizes of the non-image region 300E and the edge region 120W are the same, and it is also preferable that the shapes and sizes of the non-image region 300W and the edge region 120E are the same. In this embodiment, the shapes and sizes of the non-image regions 300S, 300N, 300E, 300W and the edge regions 120S, 120N, 120E, 120W are all the same. In addition, in this embodiment, the vertical widths of non-image areas 300S, 300N, 300E, 300W and edge areas 120S, 120N, 120E, 120W are set to half the vertical width of window 200, and the horizontal widths of non-image areas 300S, 300N, 300E, 300W and edge areas 120S, 120N, 120E, 120W are set to half the horizontal width of window 200.

[0043] In the material image 100, the lower right edge region 120S is located diagonally opposite the upper left edge region 120N (in other words, on the opposite side of the center point of the material image 100), and the lower left edge region 120W is located diagonally opposite the upper right edge region 120E. In other words, the edge region 120S will be located diagonally opposite the target region 110 when the window 200 is positioned to include at least a portion of the edge region 120N. Similarly, the edge region 120N will be located diagonally opposite the target region 110 when the window 200 is positioned to include at least a portion of the edge region 120S. Furthermore, the edge region 120W will be located diagonally opposite the target region 110 when the window 200 is positioned to include at least a portion of the edge region 120E, and the edge region 120E will be located diagonally opposite the target region 110 when the window 200 is positioned to include at least a portion of the edge region 120W.

[0044] Fig. 8 is a diagram showing an example of the convolution process proposed in the present disclosure. As shown in Fig. 8, the calculation unit 36 ​​performs a convolution operation on the window image 400 generated by the window image generation unit 35. The edge region of the substance image 100 that is placed (copied) in the non-image region 300 of the substance image 100 is located on the opposite side of the target region 110 that is set at the corner of the substance image 100. Therefore, the calculation unit 36 ​​can also properly calculate information on pixels near the corners (sides and corners) of the substance image 100 in the convolution process.

[0045] 5, when the calculation unit 36 ​​generates an image as a result of the convolution calculation and the substance image acquisition unit 33 acquires this image as the substance image 100, the window image generation unit 35 may generate the window image 400 based on the substance image 100 that indicates the result of the convolution calculation. By doing so, it becomes possible to appropriately calculate information on pixels near the corners of the substance image 100 not only in the first convolution layer C1 but also in the convolution calculations in the second and subsequent convolution layers C2 and C3.

[0046] [2-2-6. Prediction Section] The prediction unit 37 calculates (predicts) the physical property values ​​of the substance indicated by the substance image 100 based on the calculation result of the convolution calculation by the calculation unit 36. The prediction unit 37 outputs the physical property values ​​of the substance based on the calculation result in the fully connected layer F shown in FIG. 5, for example.

[0047] When training data including a substance image 100 and physical property values ​​is input to the convolutional machine learning model 30, the learning unit 32 compares the physical property values ​​calculated by the prediction unit 37 with the physical property values ​​of the input training data. If the physical property values ​​calculated by the prediction unit 37 differ from the physical property values ​​of the training data, the learning unit 32 updates the parameters of the convolutional machine learning model 30 stored in the parameter storage unit 31, and the calculation unit 36 ​​and the prediction unit 37 calculate the physical property values ​​by performing a convolution calculation again based on the updated parameters. In other words, the learning unit 32 repeats updating (learning) the parameters of the convolutional machine learning model 30 until the physical property values ​​calculated by the prediction unit 37 match or approximate the physical property values ​​of the training data.

[0048] [2-3. Result output section] The result output unit 40 outputs, as a prediction result, the physical property values ​​of the substance calculated by the prediction unit 37 of the convolutional machine learning model 30. In addition, for example, when learning data is input to the convolutional machine learning model 30 and the parameters of the convolutional machine learning model 30 are updated (learned), the result output unit 40 may output information indicating that the learning has been completed, information indicating the updated parameters, or the like.

[0049] [3. Flowchart] 9 is a diagram showing an example of the flow of convolution processing performed in the convolutional machine learning model 30. First, the substance image acquisition unit 33 acquires a substance image 100 indicating a target substance (step S110). When the convolution processing is performed in the first convolutional layer C1, in step S110, the substance image acquisition unit 33 acquires the substance image 100 input to the convolutional machine learning model 30. When the convolution processing is performed in the second or subsequent convolutional layers C2 and C3, in step S110, the substance image acquisition unit 33 acquires, as the target image 100, an image generated by the convolution processing in the layer immediately before that.

[0050] Next, the window setting unit 34 determines whether the window 200 set in the convolutional layer includes a non-image region 300 outside the substance image 100 (step S120). If the window 200 does not include the non-image region 300 (No in step S120), the calculation unit 36 ​​performs convolution processing on the substance image 100 acquired in step S110 (step S130), and then ends the processing.

[0051] If the window 200 includes the non-image region 300 (Yes in step S120), the window image generation unit 35 generates a window image 400 by arranging (copying) the image of the edge region 120 in the non-image region 300 of the substance image 100, as shown in Figures 7A to 7C, for example (step S140). Then, the calculation unit 36 ​​performs a convolution process on the window image 400 generated in step S140 (step S150), and ends the process.

[0052] [4. Summary] As described above, in this embodiment, when the window 200 includes a non-image region 300 located outside the substance image 100, the window image generation unit 35 generates the window image 400 by placing an image of a predetermined edge region 120 of the substance image 100, which is different from the target region 110 in which the window 200 is set, in the non-image region 300. In this way, the edge region 120 of the substance image 100 placed in the corner of the window image 400 is located on the opposite side of the target region 110 in the substance image 100, so that the calculation unit 36 ​​can also properly calculate information about pixels near the corners of the substance image 100 in the convolution process.

[0053] Furthermore, in this embodiment, the material image 100 is an image cut out from a three-dimensional model 90 in which periodic boundaries are set at two opposing end faces, and therefore a periodic boundary is also set at each edge of the material image 100. A window image 400 generated with periodic boundaries set at each edge of the material image 100 shows the interior of a continuous region of the target material. By using the window image 400 generated in this manner as an input image for a convolution operation, the accuracy of the subsequently calculated (predicted) physical property values ​​can be ensured. That is, by generating a three-dimensional model 90 in which periodic boundary conditions are set by simulation and inputting a material image 100 cut out from the three-dimensional model 90 into the convolutional machine learning model 30 proposed in the present disclosure, physical property values ​​can be calculated in a shorter time than simulation, etc., and since information about pixels near the corners of the material image 100 is also properly calculated in the convolution operation, physical property values ​​can be calculated with higher accuracy than conventional convolutional machine learning models.

[0054] [5. Modifications] The present invention is not limited to the above-described embodiments.

[0055] (1) In the embodiment, an example has been described in which the material image 100 is a two-dimensional image cut out from the three-dimensional model 90. However, the material image 100 may be a three-dimensional image showing the three-dimensional model 90. The material image 100 may be a three-dimensional image generated based on the three-dimensional model 90 (for example, an image obtained by volume rendering the three-dimensional model 90), or may be the three-dimensional model 90 itself. In this case, the material image 100 is an image of a rectangular parallelepiped, and periodic boundaries are set on two opposing faces of the material image 100.

[0056] In this case, the window setting unit 34 sets a window 200 as a filter in the target region, which is a partial spatial region of the material image 100, and the calculation unit 36 ​​uses this window 200 to extract pixel features in the target region 110 of the material image 100. Here, the window image generation unit 35 sets an edge region 120, which is a partial spatial region of the material image 100 that is in contact with one of the two surfaces (edges) on which the periodic boundary is set, and places this in a non-image region 130 that is in contact with the other of the two surfaces, thereby generating a window image 400.

[0057] The window image generating unit 35 may generate the window image 400 by, for example, determining a spatial region of the material image 100 that includes the entire area of ​​one of two surfaces on which a periodic boundary is set as the edge region 120 and disposing this in the non-image region 130 that is adjacent to the other of the two surfaces. Alternatively, the window image generating unit 35 may generate the window image 400 by determining a spatial region of the material image 100 that is adjacent to one of the two longest sides of the four parallel sides of the rectangular parallelepiped material image 100 as the edge region 120 and disposing this in the non-image region 130 that is adjacent to the other side. Alternatively, the window image generating unit 35 may generate the window image 400 by determining a spatial region of the material image 100 that is adjacent to one of two diagonally opposite corners of the material image 100 as the edge region 120 and disposing this in the non-image region 130 that is adjacent to the other corner.

[0058] The calculation unit 36 ​​performs a convolution calculation on the window image 400 thus generated. By doing so, it becomes possible to appropriately calculate information on pixels near the corners of the substance image 100 in the convolution calculation.

[0059] (2) The physical property prediction device 10 may also have the function of a region identification unit that identifies regions of the material image 100 that contribute to the calculated physical property value using the trained convolutional machine learning model 30. The region identification unit may identify regions included in the edge region 120 of the material image 100, or may identify regions included in two edge regions 120 located on opposite sides of the material image 100 (e.g., the left edge region 120L and the right edge region 120R). In this way, it is possible to understand the relationship between the morphology or phase separation structure of a material and the physical property value of the material.

[0060] (3) The physical property value prediction device 10 may also have the function of an image search unit that searches for a material image 100 for calculating a desired physical property value from among a plurality of material images 100 (for example, the plurality of material images 100 shown in FIG. 4 generated by the material image generation unit 20). In this way, it becomes possible to search for a heterogeneous structure of a material that realizes a desired physical property.

[0061] (4) Furthermore, when training data including a substance image 100 and physical properties are input to the convolutional machine learning model 30, the training unit 32 compares the physical properties calculated (predicted) by performing a convolution process on a window image 400 generated from the substance image 100 with the physical properties of the input training data, and updates the parameters of the convolutional machine learning model 30 based on the comparison result. In other words, the physical property prediction device 10 also serves as a training device. Furthermore, by using the window image 400, the training unit 32 can train the convolutional machine learning model 30 by taking into account information about pixels near the corners of the substance image 100. [Explanation of symbols]

[0062] 10 Physical property prediction device, 11 Processor, 12 Memory unit, 13 Communication unit, 14 Display unit, 15 Operation unit, 20 Material image generation unit, 30 Convolutional machine learning model, 31 Parameter memory unit, 32 Learning unit, 33 Material image acquisition unit, 34 Window setting unit, 35 Window image generation unit, 36 Calculation unit, 37 Prediction unit, 40 Result output unit, 90 3D model, 100 Material image, 110 Target region, 120, 120U, 120D, 120L, 120R, 120N, 120S, 120E, 120W Edge region, 200 Window, 300, 300U, 300D, 300L, 300R, 300N, 300S, 300E, 300W Non-image region, 400 Window image, C1,C2,C3 convolutional layers, F fully connected layer.

Claims

1. a material image acquisition means for acquiring a material image which is an image showing a target material and is an image of a rectangle or a rectangular parallelepiped having periodic boundaries set at two ends which are two opposing sides or faces; a window setting means for setting a window in a partial target region of the material image; a window image generating means for generating a window image showing the interior of a continuous region in the target material by arranging an image of an edge region of the material image, which is different from the target region and which is in contact with the other of the two ends where the periodic boundary is set, in a non-image region which is located outside the material image and in contact with the other of the two ends; means for performing a convolution operation on the window image; and a convolutional machine learning model including: Physical property prediction device.

2. 2. The physical property value prediction device according to claim 1, The edge region is located diagonally opposite the target region in the material image. Physical property prediction device.

3. 2. The physical property value prediction device according to claim 1, the substance image acquisition means acquires, as the substance image, an image showing a result of a convolution operation on an image showing the target substance; The window image generating means generates the window image based on the material image showing at least a part of the result of the convolution operation. Physical property prediction device.

4. A physical property prediction method for predicting a physical property value of a target substance, comprising: a material image acquisition means of the convolutional machine learning model acquires a material image that is an image showing the target material and is an image of a rectangle or a rectangular parallelepiped with periodic boundaries set at two ends that are two opposing sides or faces; a window setting means for setting a window in a partial target region of the material image; a window image generation means of the convolutional machine learning model generates a window image showing the interior of a continuous region in the target material by placing an image of an edge region of the material image, which is different from the target region and which is in contact with the other of the two ends where the periodic boundary is set, in a non-image region which is located outside the material image and in contact with the other of the two ends; The calculation means of the convolutional machine learning model performs a convolution calculation on the window image. Methods for predicting physical properties.

5. A computer that predicts the physical properties of a target substance a step of acquiring a material image which is an image showing the target material and is an image of a rectangle or a rectangular parallelepiped in which periodic boundaries are set at two ends which are two opposing sides or faces; setting a window over a partial region of interest of the material image; a step of generating a window image showing the interior of a continuous region in the target material by placing an image of an edge region of the material image, which is different from the target region and which is in contact with the other of the two ends where the periodic boundary is set, in a non-image region which is located outside the material image and in contact with the other of the two ends; and performing a convolution operation on the windowed image; A program that executes the following.

Images