JP2023036279A - Physical property value prediction apparatus, physical property value prediction method, and program - Google Patents

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

The present invention relates to a physical property value prediction device, a physical property value prediction method, and a program.

It is known to calculate physical property values of target substances by the finite element method or molecular simulation. For example, the elastic modulus of filler-filled rubber used in tires can be obtained by calculation such as simulation. However, the actual situation is that a huge amount of calculation time is required to obtain physical property values by such a method.

For example, Patent Literature 1 below describes estimating 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 a convolutional machine learning model and allowing the model to infer physical property values, it is possible to calculate the physical property values of a target substance in a relatively short period of time.

Japanese Patent Application Laid-Open No. 2021-060457

In the convolutional layer of the convolutional machine learning model, a window serving as a filter is set in a part of the image, and the position of the window is shifted while convolution operations are sequentially performed on the pixel value groups within the window. However, in such a conventional convolution operation, pixels close to the corners (sides and corners) of the image are subjected to the convolution operation less times. Therefore, there is a problem that the information of pixels near the corners of the image is not properly calculated by the convolutional machine learning model.

An object of the present disclosure is to provide a physical property value prediction device, a physical property value prediction method, and a program capable of appropriately calculating information on pixels near corners of an image.

A physical property value prediction apparatus according to the present disclosure includes material image acquisition means for acquiring a material image showing a target material, window setting means for setting a window in a partial target area of the material image, and the window being the material image. If it includes a non-image area located outside of the non-image area, a window image is generated by placing an image of a predetermined edge area of the material image, which is different from the target area, in the non-image area. A convolutional machine learning model including means for generating a window image and means for performing a convolution operation on the window image is included to predict a physical property value of the target material. According to this, it becomes possible to appropriately calculate information of pixels near the corners of the material image.

Further, a physical property value prediction method according to the present disclosure is a physical property value prediction method for predicting a physical property value of a target substance, wherein a substance image acquisition means of a convolutional machine learning model acquires a substance image representing the target substance, If the window setting means of the convolutional machine learning model sets a window over a partial region of interest of the material image, and the window includes a non-image area located outside the material image, the convolutional machine learning model A window image generating means generates a window image by arranging an image of a predetermined edge region of the material image, which is different from the target region, in the non-image region, and calculates the convolutional machine learning model. Means perform a convolution operation on the window image. According to this, in the convolutional machine learning model, it becomes possible to appropriately calculate information of pixels near the corners of the material image.

Further, the program according to the present disclosure includes, in a computer for predicting physical property values of a target substance, a procedure for acquiring a material image showing the target substance, a procedure for setting a window in a partial target region of the substance image, the window contains a non-image area located outside the material image, by placing an image of a predetermined edge area of the material image that is different from the target area in the non-image area, the window A procedure for generating an image and a procedure for performing a convolution operation on the window image are performed. According to this, it becomes possible to appropriately calculate information of pixels near the corners of the material image using a computer.

1 is a diagram showing the configuration of a physical property value prediction device that is an example of an embodiment of the present disclosure; FIG. It is a functional block diagram which shows an example of the function implemented in a physical-property value prediction apparatus. It is a figure which shows an example of the three-dimensional model which shows a target substance. It is a figure which shows an example of the substance image which shows a target substance. FIG. 4 is a diagram showing an example of processing performed in a convolutional machine learning model; It is a figure which shows an example of the conventional convolution process. It is a figure which shows an example of the conventional convolution process. It is a figure which shows an example of the conventional convolution process. FIG. 10 is a diagram showing an example of a method for generating a window image; FIG. 10 is a diagram showing an example of a method for generating a window image; FIG. 10 is a diagram showing an example of a method for generating a window image; FIG. 4 is a diagram illustrating an example of convolution processing proposed in the present disclosure; It is a flowchart which shows an example of the flow of the process performed with a physical-property value prediction apparatus.

Hereinafter, embodiments of the present disclosure will be described with reference to the drawings. In this embodiment, by inputting a material image representing a material whose physical properties are to be predicted (hereinafter referred to as a target material) into a convolutional machine learning model by CNN, a physical property value representing the physical properties of the target material is calculated. . Material morphology (for example, filler morphology of filler-filled rubber) is one of the factors that control material properties. It is preferable to input material images that are input to the convolutional machine learning model.

In the convolution operation performed by the conventional convolutional machine learning model, pixels close to the corners (sides and corners) of the material image are subjected to the convolution operation less frequently. In this regard, in the present embodiment, in the convolution operation, information on pixels near the corners of the material image can also be appropriately calculated.

[1. Hardware configuration]
FIG. 1 is a diagram showing the configuration of a physical property value prediction device 10 that is an example of an embodiment of the present disclosure. A physical property value prediction apparatus 10 according to the present embodiment is a computer such as a personal computer, a general-purpose computer, or a personal digital assistant, and as shown in FIG. A portion 15 is included. The physical property value prediction device 10 may include an optical disk drive for reading optical disks, a USB (Universal Serial Bus) port, and the like.

The processor 11 is a program control device such as a CPU (Central Processing Unit) that operates according to a program installed in the physical property value prediction device 10, which is a computer, for example. The storage unit 12 is a storage element such as a ROM (Read Only Memory) or a RAM (Random Access Memory), a hard disk drive, or the like. The storage unit 12 stores data such as programs to be executed by the processor 11 . The communication unit 13 is, for example, 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 and a mouse, and receives user operation input and outputs a signal indicating the content of the input to the processor 11 .

[2. Function block]
FIG. 2 is a functional block diagram showing an example of functions implemented in the physical property value prediction device 10. As shown in FIG. As shown in FIG. 2 , the physical property value prediction device 10 functionally includes a material image generator 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 implemented mainly by the processor 11 . 2 may not be implemented in the physical property value prediction device 10, and functions other than the functions shown in FIG. 2 may be implemented.

[2-1. Substance image generator]
The substance image generator 20 generates a substance image representing the target substance. The substance image generation unit 20 generates a substance image from a three-dimensional model representing a target substance generated by a given simulation such as molecular simulation.

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 generator 20 cuts 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). Thereby, a plurality of substance images 100 representing the target substance are generated.

In simulations such as the finite element method and molecular simulation, an area at infinity cannot be the object of calculation. It is generated as a rectangular parallelepiped model of a representative volume element with periodic boundaries (periodic symmetry) set on the end faces. 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 upper end face 90U and the lower 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. be. Therefore, in the present embodiment, the material image 100 cut out from the three-dimensional model 90 has periodic boundaries set on two sides (ends) facing each other. More specifically, periodic boundaries are set on the upper side 100U and the lower side 100D, and the left side 100L and the right side 100R of the material image 100, respectively. Note that, in the present disclosure, the term “rectangular parallelepiped” will be described as including a cube.

Alternatively, for example, the substance image generation unit 20 may generate the substance image 100 of the target substance by photographing the target substance using a camera or an electron microscope. Also in this case, the material image 100 is preferably rectangular. In addition, in this disclosure, the “rectangle” is described as including a square.

[2-2. machine learning model]
The convolutional machine learning model 30 is a machine learning model based on a CNN (Convolutional Neural Network), and is learned using learning data including material images and physical property values of materials indicated by the material images. Then, when a given material image is input, the convolutional machine learning model 30 outputs physical property values of the material indicated by the material image. Here, "physical property value" may be a physical quantity such as elastic modulus, a class defined within the range of physical quantity, or a class representing the type and ratio of materials (for example, filler or polymer) blended in the target substance. and so on. Therefore, the convolutional machine learning model 30 may be a so-called classification model or a regression model.

As shown in FIG. 2, the convolution machine learning model 30 functionally includes a parameter storage unit 31, a learning unit 32, a material image acquisition unit 33, a window setting unit 34, a window image generation unit 35, Calculation unit 36 and prediction unit 37 are included. The parameter storage unit 31 may be implemented mainly by the storage unit 12 of the physical property value prediction device 10, or may be implemented by another storage device or the like. The learning unit 32 , the material 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 implemented mainly by the processor 11 . Note that the convolutional machine learning model 30 does not have to implement all the functions shown in FIG. 2, and may implement functions other than the functions shown in FIG.

[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 material image is input to the convolutional machine learning model 30 . More specifically, the number of convolutional layers, the number of nodes used in each convolutional layer, the weighting of each node, and the like are stored as parameters of the convolutional machine learning model 30 .

[2-2-2. Learning department]
The learning unit 32 performs machine learning using a material image 100 representing a material and a plurality of learning data including physical property values of the material, so that the parameters of the convolutional machine learning model 30 stored in the parameter storage unit 31 to update. That is, the parameters of the convolutional machine learning model 30 stored in the parameter storage unit 31 are learned (updated) using a plurality of learning data.

[2-2-3. Substance Image Acquisition Unit]
The substance image acquisition unit 33 acquires a substance image 100 representing the target substance. In this embodiment, the material image acquisition unit 33 acquires the material image 100 generated by the material image generation unit 20 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 .

FIG. 5 is a diagram showing an example of processing performed by the convolutional machine learning model 30. As shown in FIG. 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. FIG. The convolution process by the computing unit 36 in the convolution layers C1-C3 may produce an image that represents the characteristics of the pixels of the material image 100 as a result of the convolution operation. The substance image obtaining unit 33 may also obtain an image representing the result of the convolution calculation performed by the calculation unit 36 as the substance image 100 indicating the target substance.

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

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 area of the material image 100 as the initial target area 110 and sets a rectangular window 200 there. The size of window 200 is preferably smaller than the size of material image 100 .

As shown in FIG. 6A , in the initial position of window 200 before moving window 200 , window 200 includes the upper left corner area of material image 100 and a non-image area 300 located outside material image 100 . contains. The calculation unit 36 uses the window 200 at this position to extract the features of pixels in the target region 110 of the material image 100 .

Next, as shown in FIG. 6B, the window setting unit 34 moves the window 200 in the material image 100 rightward by a predetermined distance from the current position. The area of the material image 100 that overlaps the window 200 after movement becomes the target area 110 . The calculation unit 36 extracts the features of the pixels in the target region 110 of the material image 100 using the window 200 after movement. This process is repeated until window 200 reaches a position that includes the right edge of material image 100 .

Next, the window setting unit 34 moves the window 200 downward by a predetermined distance from the initial position of the window 200 in the material image 100, and the calculation unit 36 uses the window 200 after movement to extract the features of pixels. Next, the window setting unit 34 moves the window 200 rightward from the current position by a predetermined distance, and the calculation unit 36 uses the window 200 after movement to extract the features of the pixels. By repeating this process, when the window 200 reaches the position including the right end of the material image 100, the window setting unit 34 sets the window 200 to the position before continuously moving the window 200 to the right (where the window 200 is the material The position including the left end of the image 100) is moved downward by a predetermined distance, and the calculation unit 36 uses the window 200 after movement to extract the features of the pixels.

As shown in FIG. 6C , the window setting unit 34 moves the window 200 rightward or downward until the window 200 finally reaches a position including the lower right corner area of the material image 100 . Unit 36 repeats the convolution process for extracting features of the pixels of material image 100 . Information on the size and initial position of the window 200 set by the window setting unit 34, the rightward or downward movement distance in one movement, and the movement range with respect to the material image 100 are parameters of the convolutional machine learning model 30. , the parameter storage unit 31 may store.

However, in the conventional convolution operation performed in this manner, for example, pixels close to the corners (sides and corners) of the material image 100 are subjected to the convolution operation less frequently. In this regard, in the present embodiment, the calculation unit 36 performs a convolution calculation on a window image 400 (see FIG. 6C), which will be described later. By doing so, information on pixels near the corners of the material image 100 can also be appropriately calculated.

[2-2-5. Window image generator]
For example, as shown in FIG. 5A, when the window 200 includes a non-image area 300 positioned outside the material image 100, the window image generation unit 35 generates the target area 110 in which the window 200 is set in the non-image area 300. A window image is generated by placing (copying) an image of a given edge region 120 of the material image 100 that is different from the .

7A to 7C are diagrams showing an example of a method for generating the window image 400. FIG. First, as shown in FIG. 7A, the window image generation unit 35 creates a rectangular non-image region 300U that contacts the upper side 100U of the material image 100 with a rectangular edge portion that contacts the lower side 100D of the material image 100. Place the image in area 120D. In addition, the window image generating unit 35 arranges the image of the rectangular edge region 120U of the material image 100, which is in contact with the upper side 100U of the material image 100, in the rectangular non-image area 300D, which is in contact with the lower side 100D of the material image 100. do.

It is preferable that the non-image area 300U and the edge area 120D have the same shape and size. Similarly, it is preferable that the shape and size of the non-image area 300D and the edge area 120U also match. In this embodiment, the non-image areas 300U and 300D and the edge areas 120U and 120D have the same shape and size. Further, in this embodiment, the vertical widths of the non-image areas 300U and 300D and the edge areas 120U and 120D are set to half the vertical width of the window 200 .

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 to the straight line passing through the center of the material image 100 and parallel to the upper side 100U and the lower side 100D). there is That is, edge region 120D would be located on the opposite side of target region 110 when window 200 is positioned to include at least a portion of edge region 120U. Similarly, edge region 120U would be located opposite region of interest 110 when window 200 is positioned to include at least a portion of edge region 120D.

Next, as shown in FIG. 7B, the window image generator 35 creates a rectangular non-image region 300L in contact with the left side 100L of the material image 100 with a rectangular edge in contact with the right side 100R of the material image 100. An image of the partial area 120R is arranged. In addition, the window image generation unit 35 arranges the image of the rectangular edge region 120L of the material image 100, which is in contact with the left side 100L of the material image 100, in the rectangular non-image area 300R, which is in contact with the right side 100R of the material image 100. do.

The non-image area 300L and the edge area 120R preferably have the same shape and size, and the non-image area 300R and the edge area 120L preferably have the same shape and size. In this embodiment, the non-image areas 300L, 300R and the edge areas 120L, 120R all have the same shape and size. Further, in this embodiment, the widths of the non-image areas 300L and 300R and the edge areas 120L and 120R in the horizontal direction are set to half the width of the window 200 in the horizontal direction.

In the material image 100, the right edge region 120R is located on the opposite side of the left edge region 120L (opposite to the straight line passing through the center of the material image 100 and parallel to the left side 100L and the right side 100R). there is That is, edge region 120R would be located on the opposite side of target region 110 when window 200 is positioned to include at least a portion of edge region 120L. Similarly, edge region 120L would be located opposite region of interest 110 when window 200 is positioned to include at least a portion of edge region 120R.

Next, as shown in FIG. 7C, the window image generation unit 35 creates a rectangular edge region 120S including the lower right corner of the material image 100 in a rectangular non-image region 300N contacting the upper left vertex of the material image 100. Place the image. In addition, the window image generating unit 35 arranges the image of the rectangular edge region 120W including the lower left corner of the material image 100 in the rectangular non-image region 300E contacting the upper right vertex of the material image 100. The image of the rectangular edge region 120E including the upper right corner of the material image 100 is arranged in the rectangular non-image region 300W that contacts the lower left vertex of the material image 100, and the rectangular non-image region 300W that contacts the lower right vertex of the At 300N, place the image of the rectangular edge region 120S containing the upper left corner of the material image 100; Thus, an image in which the images of the edge regions 120U, 120D, 120L, 120R, 120N, 120S, 120E, and 120W are arranged in the non-image regions 300U, 300D, 300L, 300R, 300N, 300S, 300E, and 300W, respectively. , becomes the window image 400 .

As described above, in this embodiment, the upper side 100U and the lower side 100D, and the left side 100L and the right side 100R of the material image 100 are each set with periodic boundaries. For this reason, the window image generation unit 35 converts the image of the edge region 120 contacting one of the two sides (two ends) to which the periodic boundary is set to the non-contact area contacting the other of the two sides. A window image 400 is generated by placing it in the image area 130 . That is, the window image generator 35 generates the window image 400 at periodic boundaries set on each side of the material image 100 .

The shape and size of the non-image region 300S and the edge region 120N preferably match, and the shape and size of the non-image region 300N and the edge region 120S also preferably match. Similarly, the non-image area 300E and edge area 120W preferably have the same shape and size, and the non-image area 300W and edge area 120E preferably have the same shape and size. In this embodiment, the non-image areas 300S, 300N, 300E, 300W and the edge areas 120S, 120N, 120E, 120W all have the same shape and size. Further, in the present embodiment, the vertical widths of the non-image areas 300S, 300N, 300E, 300W and the edge areas 120S, 120N, 120E, 120W are set to half the vertical width of the window 200. The horizontal widths of the image areas 300S, 300N, 300E and 300W and the edge areas 120S, 120N, 120E and 120W are set to half the width of the window 200 in the horizontal direction.

In material image 100, lower right edge region 120S is located diagonally (in other words, opposite to the center point of material image 100) of upper left edge region 120N, and the lower left edge Region 120W is located diagonally from upper right edge region 120E. That is, edge region 120S is positioned diagonally to target region 110 when window 200 is positioned to include at least a portion of edge region 120N. Similarly, edge region 120N would be located diagonally of target region 110 when window 200 is positioned to include at least a portion of edge region 120S. In addition, the edge region 120W is positioned diagonally from the target region 110 when the window 200 is positioned to include at least a portion of the edge region 120E. It will be located diagonally to the region of interest 110 when placed in a position that includes at least a portion of the edge region 120W.

FIG. 8 is a diagram illustrating an example of convolution processing proposed in the present disclosure. As shown in FIG. 8, the calculation unit 36 performs a convolution calculation on the window image 400 generated by the window image generation unit 35. FIG. Since the edge region of the material image 100 placed (copied) in the non-image region 300 of the material image 100 is located on the opposite side of the target region 110 set at the corner of the material image 100, the computing unit 36 performs the convolution In the processing, it becomes possible to appropriately calculate the information of the pixels near the corners (sides and corners) of the material image 100 as well.

Further, in the convolution layers C1 to C3 shown in FIG. may generate a window image 400 based on the material image 100 showing the result of the convolution operation. By doing so, the information of the pixels near the corners of the material image 100 can be appropriately calculated not only in the first convolution layer C1 but also in the convolution calculations in the second and subsequent convolution layers C2 and C3. be able to

[2-2-6. Predictor]
The prediction unit 37 calculates (predicts) 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 value of the substance based on the calculation result in the fully connected layer F shown in FIG. 5, for example.

Note that when learning data including the material image 100 and physical property values are input to the convolutional machine learning model 30, the learning unit 32 combines the physical property values calculated by the prediction unit 37 with the physical property values of the input learning data. compare. When the physical property value calculated by the prediction unit 37 is different from the physical property value of the learning 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 calculates the physical property value by performing the convolution operation again based on the updated parameters. That is, the learning unit 32 repeats updating (learning) of the parameters of the convolutional machine learning model 30 until the physical property value calculated by the prediction unit 37 matches or approximates the physical property value of the learning data.

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

[3. flowchart]
FIG. 9 is a diagram showing an example of the flow of convolution processing performed by the convolution machine learning model 30. As shown in FIG. First, the substance image acquiring unit 33 acquires the substance image 100 representing the target substance (step S110). When the convolution process is performed in the first convolution layer C1, the substance image acquisition unit 33 acquires the substance image 100 input to the convolution machine learning model 30 in step S110. Further, when the convolution processing is performed in the second and subsequent convolution layers C2 and C3, in step S110, the substance image acquiring unit 33 converts the image generated by the convolution processing of the immediately preceding layer to the target image 100. to get as

Next, the window setting unit 34 determines whether or not the window 200 set in the convolution layer includes the non-image area 300 outside the material image 100 (step S120). If the window 200 does not include the non-image area 300 (No in step S120), the calculation unit 36 performs convolution processing on the material image 100 acquired in step S110 (step S130), and ends the processing.

If the window 200 includes the non-image area 300 (Yes in step S120), the window image generator 35 creates the edge area 120 in the non-image area 300 of the material image 100 as shown in FIGS. 7A to 7C, for example. By arranging (copying) the image of , the window image 400 is generated (step S140). Then, the calculation unit 36 performs convolution processing on the window image 400 generated in step S140 (step S150), and ends the processing.

[4. summary]
As described above, in the present embodiment, when the window 200 includes the non-image area 300 located outside the material image 100, the window image generation unit 35 sets the window 200 to the non-image area 300. A window image 400 is generated by placing an image of a given edge region 120 of the material image 100 that is different from the region 110 . In this way, the edge region 120 of the material image 100 arranged at the corner of the window image 400 is located on the opposite side of the target region 110 in the material image 100. Information on pixels near the corners of the square can also be calculated appropriately.

In addition, in the present embodiment, the material image 100 is an image cut out from the three-dimensional model 90 in which periodic boundaries are set at two end faces facing each other, so that each side of the material image 100 is also set with periodic boundaries. be. The window image 400 generated by the periodic boundaries set on each side of the material image 100 shows the continuous area in the target material. By using the window image 400 generated in this way as an input image for the convolution operation, the accuracy of the physical property values calculated (predicted) thereafter can be ensured. That is, by generating a three-dimensional model 90 in which periodic boundary conditions are set by simulation, and inputting the material image 100 extracted from the three-dimensional model 90 to the convolutional machine learning model 30 proposed in the present disclosure, simulation etc. The physical property value can be calculated in a shorter time than the conventional convolutional machine learning model, and the information of the pixels near the corner of the material image 100 is also appropriately calculated in the convolution operation. can be calculated.

[5. Modification]
The present invention is not limited to the above embodiments.

(1) In the embodiment, a two-dimensional image cut out from the three-dimensional model 90 is used as the material image 100 , but the material image 100 may be a three-dimensional image representing 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 a rectangular parallelepiped image, and periodic boundaries are set on two opposing surfaces of the material image 100 .

In this case, the window setting unit 34 sets a window 200 serving 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 determine the target region 110 of the material image 100. Extract features of pixels in . Here, the window image generation unit 35 defines a spatial region of a part of the material image 100 that is in contact with one of the two surfaces (ends) on which the periodic boundary is set as the edge region 120, and defines this as the edge region 120. A window image 400 is generated by placing it in a non-image area 130 in contact with the other of the two surfaces.

The window image generation unit 35, for example, sets the spatial region of the material image 100 including the entire area of one of the two surfaces on which the periodic boundary is set as the edge region 120, and defines it as the edge region 120 of the two surfaces. A window image 400 may be generated by arranging it in the non-image area 130 in contact with the other surface. In addition, the window image generation unit 35 selects the spatial region of the material image 100 that is in contact with one of the two sides with the longest distance among the four parallel sides of the rectangular parallelepiped material image 100 as an edge. The window image 400 may be generated by setting the area 120 and arranging it in the non-image area 130 that is in contact with the other side. In addition, the window image generation unit 35 defines the spatial region of the material image 100 that is in contact with one of the two diagonally positioned corners of the material image 100 as the edge area 120, and defines the edge area 120 as the non-spatial area that is in contact with the other corner. A window image 400 may be generated by placing it in the image area 130 .

The calculation unit 36 performs a convolution calculation on the window image 400 generated in this way. In this manner, information on pixels near the corners of the material image 100 can also be appropriately calculated in the convolution calculation.

(2) In addition, the physical property value prediction device 10 may have the function of an area identification unit that identifies the area of the material image 100 that contributes to the calculated physical property value using the learned convolutional machine learning model 30. good. The region identifying unit may identify a region included in the edge region 120 of the material image 100, and two edge regions 120 located on opposite sides of the material image 100 (for example, the leftmost edge region 120L and the rightmost edge region 120L). The regions included in each edge region 120R) may be identified. By doing so, it is possible to understand the relationship between the morphology and phase separation structure of the material and the physical property values of the material.

(3) Further, the physical property value prediction device 10 calculates a desired physical property value from among the 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). It may also have a function of an image searching unit that searches for the material image 100 to be used. In this way, it becomes possible to search for heterogeneous structures of materials that achieve desired physical properties.

(4) In addition, when learning data including the material image 100 and physical property values are input to the convolutional machine learning model 30, the learning unit 32 performs convolution processing on the window image 400 generated from the material image 100. The calculated (predicted) physical property value and the physical property value of the input learning data are compared, and the parameters of the convolutional machine learning model 30 are updated based on the comparison result. That is, the physical property value prediction device 10 also contributes as a learning device. Further, by using the window image 400, the learning unit 32 can learn the convolutional machine learning model 30 in consideration of the information of the pixels near the corners of the material image 100. FIG.

10 physical property value prediction device, 11 processor, 12 storage unit, 13 communication unit, 14 display unit, 15 operation unit, 20 material image generation unit, 30 convolution machine learning model, 31 parameter storage unit, 32 learning unit, 33 material image acquisition 34 window setting unit 35 window image generation unit 36 calculation unit 37 prediction unit 40 result output unit 90 three-dimensional model 100 substance image 110 target region 120, 120U, 120D, 120L, 120R, 120N , 120S, 120E, 120W edge regions, 200 windows, 300, 300U, 300D, 300L, 300R, 300N, 300S, 300E, 300W non-image regions, 400 window images, C1, C2, C3 convolution layers, F fully connected layers .

Claims (7)

a substance image acquiring means for acquiring a substance image showing the target substance;
window setting means for setting a window on a partial target area of the material image;
if the window includes a non-image area located outside the material image, placing an image of a predetermined edge area of the material image that is different from the target area in the non-image area; , window image generating means for generating a window image;
means for performing a convolution operation on the window image;
and predicting the physical property value of the target substance. In the physical property value prediction device according to claim 1,
The physical property value prediction device, wherein the predetermined edge area is located on the opposite side of the target area in the material image. In the physical property value prediction device according to claim 1 or 2,
the material image is a rectangular or cuboid image;
The physical property value prediction device, wherein the predetermined edge area is positioned at a diagonal of the target area in the material image. In the physical property value prediction device according to claim 2 or 3,
The material image is an image in which periodic boundaries are set at two ends, which are two sides or surfaces facing each other,
The window image generating means arranges an image of an edge region contacting one of the two ends where the periodic boundary is set, in the non-image region contacting the other of the two ends. a physical property value prediction device that generates the window image by In the physical property value prediction device according to any one of claims 1 to 4,
The substance image acquiring means acquires, as the substance image, an image representing a result of a convolution operation with respect to the image representing the target substance,
The window image generating means generates the window image based on the material image representing at least part of the result of the convolution operation. A physical property value prediction device. A physical property value prediction method for predicting the physical property value of a target substance,
A substance image acquisition means of the convolutional machine learning model acquires a substance image showing the target substance,
window setting means of the convolutional machine learning model setting a window on a partial region of interest of the material image;
When the window includes a non-image area located outside the material image, the window image generating means of the convolutional machine learning model may include a predetermined area of the material image, different from the target area, in the non-image area. generating a window image by arranging the image of the edge region,
The physical property value prediction method, wherein the calculation means of the convolutional machine learning model performs a convolution calculation on the window image. A computer that predicts the physical property values of target substances
a procedure for acquiring a substance image showing the target substance;
setting a window over a partial region of interest of said material image;
if the window includes a non-image area located outside the material image, placing an image of a predetermined edge area of the material image that is different from the target area in the non-image area; , a procedure for generating a window image, and
performing a convolution operation on the window image;
program to run.

Images