Three-dimensional model generating device and three-dimensional model evaluation system

Abstract

To provide a three-dimensional model generation device and a three-dimensional model evaluation system for creating a three-dimensional model of a catalyst layer of a fuel battery in a simple way.SOLUTION: A three-dimensional model creation device 10 includes a learning model storage unit 20 that stores a learning model Lm for estimating, on the basis of one cross-section image of a catalyst layer C of a fuel battery, another cross-section image, an estimated cross-section image creation unit 32 that creates one estimated cross-section image Ei by using one actual cross-section image Ri of the catalyst layer C and the learning model Lm and creates another estimated cross-section image Ei by using the one estimated cross-section image Ei created and the learning model Lm, and a three-dimensional model creation unit 34 that creates a three-dimensional model M of the catalyst layer C on the basis of a plurality of estimated cross-section images Ei sequentially created by the estimated cross-section image creation unit 32.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to a three-dimensional model generation device and a three-dimensional model evaluation system. [Background technology]

[0002] In recent years, research and development into fuel cells has been conducted to contribute to energy efficiency in order to ensure that more people have access to affordable, reliable, sustainable and advanced energy.

[0003] The catalyst layer that constitutes the electrode of a fuel cell is a structure having pores therein. When evaluating the characteristics of a structure having pores therein, a three-dimensional model of the structure is used. JP 2019-118857 A discloses that a three-dimensional model is constructed from slice images of an exhaust gas purification catalyst. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Publication No. 2019-118857 Summary of the Invention [Problem to be solved by the invention]

[0005] There is a strong demand for technology that can easily generate three-dimensional models.

[0006] The present disclosure aims to solve the above-mentioned problems, and ultimately contributes to energy efficiency. [Means for solving the problem]

[0007] A first aspect of the present disclosure is a three-dimensional model generation device comprising: a learning model memory unit that stores a learning model for estimating an image of another cross section based on an image of one cross section of a catalyst layer of a fuel cell; an estimated cross-sectional image generation unit that generates an estimated cross-sectional image using an actual cross-sectional image of the catalyst layer and the learning model, and generates another estimated cross-sectional image using the generated estimated cross-sectional image and the learning model; and a three-dimensional model generation unit that generates a three-dimensional model of the catalyst layer based on multiple estimated cross-sectional images sequentially generated by the estimated cross-sectional image generation unit.

[0008] A second aspect of the present disclosure is a three-dimensional model evaluation system including the three-dimensional model generation device according to the first aspect.

[0009] A third aspect of the present disclosure is a three-dimensional model generation device comprising: a learning model memory unit that stores a learning model for estimating an image of another cross section based on an image of one cross section of a structure having pores therein; an estimated cross-sectional image generation unit that generates an estimated cross-sectional image using one actual cross-sectional image of the structure and the learning model, and generates another estimated cross-sectional image using the generated estimated cross-sectional image and the learning model; and a three-dimensional model generation unit that generates a three-dimensional model of the structure based on multiple estimated cross-sectional images sequentially generated by the estimated cross-sectional image generation unit. [Effects of the Invention]

[0010] According to the present disclosure, a three-dimensional model can be easily generated. [Brief explanation of the drawings]

[0011] [Figure 1] FIG. 1 is a block diagram of a three-dimensional model generating device. [Figure 2] Fig. 2A is a diagram illustrating an example of an actual cross-sectional image of a catalyst layer of a fuel cell. Fig. 2B is a diagram for explaining that an estimated cross-sectional image is generated using a learning model. Fig. 2C is a diagram showing the arrangement of an actual cross-sectional image of a catalyst layer and multiple estimated cross-sectional images. Fig. 2D is a diagram showing a three-dimensional model of a catalyst layer. [Figure 3] Fig. 3A is a diagram for explaining training data, and Fig. 3B is a diagram for explaining that a learning model is generated using the training data. [Figure 4] FIG. 4 is a flowchart showing an example of the operation of the three-dimensional model generating device. [Figure 5] FIG. 5 is a diagram showing the pore size distribution according to the intervals between portions corresponding to the cross section of the catalyst layer. [Figure 6] FIG. 6 is a block diagram of a three-dimensional model generating device. [Figure 7] Fig. 7A is a diagram illustrating pixel values ​​of pixels included in an actual cross-sectional image, Fig. 7B is a diagram for explaining correction of pixel values, and Fig. 7C is a diagram illustrating pixels whose pixel values ​​have been corrected. [Figure 8] FIG. 8 is a block diagram of a three-dimensional model generating device. [Figure 9] Fig. 9A is a diagram illustrating pixel values ​​of pixels located on the periphery of an estimated cross-sectional image, and Fig. 9B is a diagram illustrating pixels whose pixel values ​​have been interpolated. [Figure 10] FIG. 10 is a block diagram of a three-dimensional model evaluation system. DETAILED DESCRIPTION OF THE INVENTION

[0012] Conventionally, a three-dimensional model of a catalyst layer constituting an electrode of a fuel cell is generated based on multiple actual cross-sectional images of the catalyst layer. Obtaining multiple actual cross-sectional images requires a significant amount of work. In the three-dimensional model generation device disclosed herein, multiple estimated cross-sectional images are sequentially generated using a learning model for estimating an image of one cross-section of a catalyst layer of a fuel cell based on an image of another cross-section, and an actual cross-sectional image of any catalyst layer. A three-dimensional model is generated based on these sequentially generated multiple estimated cross-sectional images. Therefore, the three-dimensional model can be easily generated without requiring a significant amount of work. A three-dimensional model generation device according to one embodiment will be described with reference to the drawings.

[0013] 1 is a block diagram of a three-dimensional model generation device 10. The three-dimensional model generation device 10 is, for example, a personal computer. The three-dimensional model generation device 10 has a calculation unit 12 and a storage unit 14. The calculation unit 12 includes a processor such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit). In other words, the calculation unit 12 includes processing circuitry.

[0014] The storage unit 14 is a computer-readable recording medium. The storage unit 14 includes a volatile memory such as a random access memory (RAM) and a non-volatile memory such as a read-only memory (ROM) or flash memory. The volatile memory is used as a working memory for the processor. The non-volatile memory stores programs executed by the processor and other necessary data.

[0015] The storage unit 14 includes a learning model storage unit 20. The learning model storage unit 20 stores a learning model generated using training data, which will be described later.

[0016] The calculation unit 12 has an image input receiving unit 30, an estimated cross-sectional image generating unit 32, a three-dimensional model generating unit 34, a teacher data acquiring unit 36, and a learning unit 38. When the calculation unit 12 executes a program stored in the storage unit 14, the image input receiving unit 30, the estimated cross-sectional image generating unit 32, the three-dimensional model generating unit 34, the teacher data acquiring unit 36, and the learning unit 38 are realized.

[0017] At least some of the image input receiving unit 30, estimated cross-sectional image generating unit 32, three-dimensional model generating unit 34, teacher data acquiring unit 36, and learning unit 38 may be realized by an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array), or an electronic circuit including a discrete device.

[0018] The image input receiving unit 30 receives input of an actual cross-sectional image Ri of a catalyst layer of a fuel cell. The actual cross-sectional image Ri is obtained, for example, by using an electron microscope to irradiate an electron beam onto a cross section of a portion of the catalyst layer. The catalyst layer is a structure having pores (voids) therein. The structure is made up of various materials such as a catalyst, a catalyst support, and a promoter. In the actual cross-sectional image Ri, pixels corresponding to the pores are displayed darker than pixels corresponding to the structure. In other words, the pixel values ​​(brightness values) of the pixels corresponding to the pores are relatively small. The pixel values ​​of the pixels corresponding to the structure are relatively large, but vary depending on the materials that make up the structure.

[0019] It should be noted that pixels exhibiting relatively small pixel values ​​do not necessarily correspond to pores. Pixels exhibiting relatively small pixel values ​​may correspond, for example, to shadows of structures or noise. Similarly, pixels exhibiting relatively large pixel values ​​do not necessarily correspond to structures. Pixels exhibiting relatively large pixel values ​​may correspond, for example, to noise.

[0020] The estimated cross-sectional image generation unit 32 generates an estimated cross-sectional image using a learning model stored in the learning model storage unit 20. The learning model stored in the learning model storage unit 20 is a learning model for estimating an image of one cross-section of a catalyst layer of a fuel cell based on an image of another cross-section. The generation of the learning model will be described later with reference to FIGS. 3A and 3B. The estimated cross-sectional image generation unit 32 first generates an estimated cross-sectional image using one actual cross-sectional image Ri and the learning model.

[0021] Next, the estimated cross-sectional image generating unit 32 generates another estimated cross-sectional image using the one estimated cross-sectional image generated by the estimated cross-sectional image generating unit 32 and the learning model. Furthermore, the estimated cross-sectional image generating unit 32 generates yet another estimated cross-sectional image using the other estimated cross-sectional image generated by the estimated cross-sectional image generating unit 32 and the learning model. By repeating such processing, a plurality of estimated cross-sectional images are sequentially generated by the estimated cross-sectional image generating unit 32.

[0022] The three-dimensional model generation unit 34 generates a three-dimensional model M of the catalyst layer of the fuel cell based on the above-mentioned one actual cross-sectional image Ri and the plurality of estimated cross-sectional images sequentially generated by the estimated cross-sectional image generation unit 32. The generation of the estimated cross-sectional image and the three-dimensional model M will be described later with reference to Figures 2A, 2B, 2C, and 2D.

[0023] The training data acquisition unit 36 ​​acquires training data used to generate the learning model described above. The training data includes a set of actual cross-sectional images of adjacent portions of the catalyst layer of the fuel cell. One of the set of actual cross-sectional images is an actual cross-sectional image Rt(j) of one portion of the catalyst layer. The other of the set of actual cross-sectional images is an actual cross-sectional image Rt(j+1) of another portion adjacent to the one portion.

[0024] The learning unit 38 uses training data to generate a learning model in which an actual cross-sectional image Rt(j) of one region is input and an estimated cross-sectional image estimated as an image of a cross section of another region adjacent to the one region is output. The value of the variable j can take on values ​​from the integer 1 to the integer k-1, so that learning is performed using multiple sets of actual cross-sectional images as training data.

[0025] By performing this type of learning, the difference between the estimated cross-sectional image that is the output of the learning model and the actual cross-sectional image used as training data becomes smaller. By performing learning so that the difference becomes smaller, the estimated cross-sectional image becomes closer to the actual cross-sectional image. The learning model generated in this manner is used to generate an estimated cross-sectional image by the above-mentioned estimated cross-sectional image generation unit 32. The generation of the learning model will be described later with reference to Figures 3A and 3B. Note that after learning is performed using the actual cross-sectional image of a catalyst layer, further learning may be performed using an actual cross-sectional image of another catalyst layer that is generated in a similar manner.

[0026] FIG. 2A is a diagram illustrating an example of an actual cross-sectional image Ri of a catalyst layer C of a fuel cell. The actual cross-sectional image Ri is an electron microscope image of a cross-section of one portion of the catalyst layer C. FIG. 2B is a diagram for explaining that an estimated cross-sectional image Ei is generated using a learning model Lm. The learning model Lm is a learning model for estimating an image of a cross-section of another portion adjacent to one portion based on an image of the cross-section of the one portion of the catalyst layer C of a fuel cell. A fixed value determined when the learning model Lm is generated is used as the value of the gap G between the one portion and the other portion. The value of the gap G is, for example, 10 nm.

[0027] First, the estimated cross-sectional image generating unit 32 generates a first estimated cross-sectional image Ei(1) using the actual cross-sectional image Ri and the learning model Lm as shown in Fig. 2A. Next, the estimated cross-sectional image generating unit 32 generates a second estimated cross-sectional image Ei(2) using the generated first estimated cross-sectional image Ei(1) and the learning model Lm.

[0028] Similarly, the estimated cross-sectional image generation unit 32 generates a third estimated cross-sectional image Ei(3) using the generated second estimated cross-sectional image Ei(2) and the learning model Lm. The estimated cross-sectional image generation unit 32 generates an nth estimated cross-sectional image Ei(n) using the (n-1)th estimated cross-sectional image Ei(n-1) and the learning model Lm. In this way, the estimated cross-sectional image generation unit 32 sequentially generates a plurality of estimated cross-sectional images Ei by repeating the generation of the estimated cross-sectional image Ei. In the example shown in FIG. 2B, the number of the plurality of estimated cross-sectional images Ei is n. The number n of the estimated cross-sectional images Ei is determined in advance.

[0029] Of the n estimated cross-sectional images Ei, the n-1 estimated cross-sectional images Ei(1), ..., Ei(n-1) other than the estimated cross-sectional image Ei(n) are used with the learning model Lm to sequentially generate estimated cross-sectional images Ei(2), ..., Ei(n) of adjacent regions. In this case, the actual cross-sectional image Ri is not required. Therefore, the estimated cross-sectional image generating unit 32 can quickly generate multiple estimated cross-sectional images Ei even when the number n of estimated cross-sectional images Ei is set to a relatively large number.

[0030] FIG. 2C is a diagram showing the arrangement of an actual cross-sectional image Ri and a plurality of estimated cross-sectional images Ei of a catalyst layer C. The three-dimensional model generation unit 34 arranges the actual cross-sectional image Ri and a plurality of estimated cross-sectional images Ei(1), ..., Ei(n) at a constant interval G. FIG. 2D is a diagram showing a three-dimensional model M of the catalyst layer C. The three-dimensional model generation unit 34 generates the three-dimensional model M of the catalyst layer C based on the actual cross-sectional image Ri and the plurality of estimated cross-sectional images Ei arranged at a constant interval G.

[0031] As described above, in this embodiment, the estimated cross-sectional image generating unit 32 uses one actual cross-sectional image Ri to generate a plurality of estimated cross-sectional images Ei. Moreover, in this embodiment, a learning model Lm is used to estimate, based on an image of a cross section of one portion of the catalyst layer C, an image of a cross section of another portion adjacent to the one portion. Therefore, according to this embodiment, the three-dimensional model M can be generated easily without requiring a large number of steps.

[0032] FIG. 3A is a diagram illustrating training data. The training data is used to generate a learning model Lm. As described above, the training data includes an actual cross-sectional image Rt(j) of one portion of a catalyst layer C of a fuel cell and an actual cross-sectional image Rt(j+1) of another portion adjacent to the one portion. The variable j takes an integer value from 1 to k-1 (j=1, . . . , k-1). A set of actual cross-sectional images Rt of adjacent portions is used as training data.

[0033] When the learning model Lm is generated, the value of the distance G between adjacent portions is set to a constant value, such as 10 nm, as described above. Assume that actual cross-sectional images Rt(1), . . . , Rt(k), . . . , Rt(m) of the portions of the catalyst layer C at each distance G are obtained. The actual cross-sectional images Rt are electron microscope images of the cross section of each portion of the catalyst layer C. Of the m actual cross-sectional images Rt obtained, k actual cross-sectional images Rt(1), . . . , Rt(k) are used as a training dataset when the learning model Lm is generated.

[0034] In this embodiment, this training data is used as teacher data. The remaining mk actual cross-sectional images Rt(k+1), ..., Rt(m) are used as a validation data set or a test data set when generating a learning model Lm. Note that when cross-validation of the learning model Lm is performed, the combination of actual cross-sectional images Rt included in each of the above-mentioned training data set, validation data set, and test data set may change. Figure 3A shows an example of each data set.

[0035] 3B is a diagram illustrating the generation of a learning model Lm using training data. The learning model Lm is generated, for example, using the pix2pix method. In this case, a pair of images consisting of actual cross-sectional images Rt(j) and Rt(j+1), which are images of regions adjacent to each other with an interval G, is used as training data.

[0036] For example, a pair of images consisting of a first actual cross-sectional image Rt(1) and a second actual cross-sectional image Rt(2) is used as training data. The first actual cross-sectional image Rt(1) is an actual cross-sectional image Rt(1) of one portion of the catalyst layer C. The second actual cross-sectional image Rt(2) is an actual cross-sectional image Rt(2) of another portion adjacent to the first portion at a distance G.

[0037] The learning unit 38 generates a learning model Lm using, as training data, a pair of images consisting of an actual cross-sectional image Rt(1) of one region and an actual cross-sectional image Rt(2) of another region adjacent to the first region at an interval G. The learning model Lm generated in this manner is a learning model that inputs the actual cross-sectional image Rt(1) of one region and outputs an estimated cross-sectional image Fi(2) estimated as an image of a cross section of the other region adjacent to the first region at an interval G. By performing learning using the training data and the learning model Lm, the estimated cross-sectional image Fi(2) gradually approximates the actual cross-sectional image Rt(2) of the other region used as training data.

[0038] When a pair of images consisting of the second actual cross-sectional image Rt(2) and the third actual cross-sectional image Rt(3) is used as training data, the second actual cross-sectional image Rt(2) is the actual cross-sectional image Rt(2) of one portion of the catalyst layer C. The third actual cross-sectional image Rt(3) is the actual cross-sectional image Rt(3) of another portion adjacent to the one portion at a distance G.

[0039] The learning unit 38 generates a learning model Lm using, as training data, a pair of images consisting of an actual cross-sectional image Rt(2) of one region and an actual cross-sectional image Rt(3) of another region adjacent to the first region at an interval G. The learning model Lm generated in this manner is a learning model that inputs the actual cross-sectional image Rt(2) of one region and outputs an estimated cross-sectional image Fi(3) estimated as an image of a cross section of the other region adjacent to the first region at an interval G. By performing learning using the training data and the learning model Lm, the estimated cross-sectional image Fi(3) gradually approximates the actual cross-sectional image Rt(3) of the other region used as training data.

[0040] Similarly, when a pair of images consisting of the jth actual cross-sectional image Rt(j) and the j+1th actual cross-sectional image Rt(j+1) is used as training data, the jth actual cross-sectional image Rt(j) is the actual cross-sectional image Rt(j) of one portion of the catalyst layer C. The j+1th actual cross-sectional image Rt(j+1) is the actual cross-sectional image Rt(j+1) of another portion adjacent to the one portion with a gap G therebetween.

[0041] The learning unit 38 generates a learning model Lm using, as training data, a pair of images consisting of an actual cross-sectional image Rt(j) of one region and an actual cross-sectional image Rt(j+1) of another region adjacent to the first region at an interval G. The learning model Lm generated in this manner is a learning model that inputs the actual cross-sectional image Rt(j) of one region and outputs an estimated cross-sectional image Fi(j+1) estimated as an image of a cross section of the other region adjacent to the first region at an interval G. By performing learning using the training data and the learning model Lm, the estimated cross-sectional image Fi(j+1) gradually approximates the actual cross-sectional image Rt(j+1) of the other region used as training data.

[0042] Similarly, when a pair of images consisting of the k-1th actual cross-sectional image Rt(k-1) and the kth actual cross-sectional image Rt(k) are used as training data, the k-1th actual cross-sectional image Rt(k-1) is the actual cross-sectional image Rt(k-1) of one portion of the catalyst layer C. The kth actual cross-sectional image Rt(k) is the actual cross-sectional image Rt(k) of another portion adjacent to the one portion at an interval G.

[0043] The learning unit 38 generates a learning model Lm using, as training data, a pair of images consisting of an actual cross-sectional image Rt(k-1) of one region and an actual cross-sectional image Rt(k) of another region adjacent to the first region at an interval G. The learning model Lm generated in this manner is a learning model that inputs the actual cross-sectional image Rt(k-1) of one region and outputs an estimated cross-sectional image Fi(k) estimated as an image of a cross section of the other region adjacent to the first region at an interval G. By performing learning using the training data and the learning model Lm, the estimated cross-sectional image Fi(k) gradually approximates the actual cross-sectional image Rt(k) of the other region used as training data.

[0044] The performance of the learning model Lm can be further improved by the above-mentioned cross-validation. Furthermore, learning may be performed using an actual cross-sectional image of another catalyst layer C generated in a similar manner. The learning model Lm generated in this manner is stored in the learning model storage unit 20. The learning model Lm stored in the learning model storage unit 20 is used to generate multiple estimated cross-sectional images Ei by the estimated cross-sectional image generation unit 32, as described above with reference to FIG. 2B. This allows multiple estimated cross-sectional images Ei to be generated with high accuracy.

[0045] Fig. 4 is a flowchart showing an example of the operation of the three-dimensional model generation device 10. Fig. 4 shows the processing procedure for generating a three-dimensional model performed by the three-dimensional model generation device 10 using the generated learning model Lm. This processing procedure is performed by the calculation unit 12 of the three-dimensional model generation device 10 by executing a program stored in the storage unit 14.

[0046] When this processing procedure starts, in step S1, the image input receiving unit 30 determines whether or not one actual cross-sectional image Ri of the catalyst layer C of the fuel cell has been input. If the answer is YES in step S1, this processing procedure proceeds to step S2. If the answer is NO in step S1, this processing procedure returns to step S1.

[0047] In step S2, the image input receiving unit 30 receives input of the actual cross-sectional image Ri. In step S3, the estimated cross-sectional image generating unit 32 generates an estimated cross-sectional image Ei using the learning model Lm stored in the learning model storage unit 20. A first estimated cross-sectional image Ei(1) is generated using the actual cross-sectional image Ri and the learning model Lm. A second estimated cross-sectional image Ei(2) is generated using the generated first estimated cross-sectional image Ei(1). Thereafter, a process of generating a new estimated cross-sectional image Ei from the generated estimated cross-sectional image Ei is repeated, thereby sequentially generating multiple estimated cross-sectional images Ei.

[0048] In step S4, the estimated cross-sectional image generation unit 32 determines whether n estimated cross-sectional images Ei have been generated. If the result is YES in step S4, the process proceeds to step S5. If the result is NO in step S4, the process returns to step S3. In step S5, the three-dimensional model generation unit 34 generates a three-dimensional model M of the catalyst layer C based on the actual cross-sectional image Ri received in step S2 and the n estimated cross-sectional images Ei generated in step S3. When the processing of step S5 is completed, the process ends.

[0049] The accuracy of the three-dimensional model M depends on the interval G between the portions corresponding to the cross section of the catalyst layer C. If the interval G is too large, there is a possibility that the amount of change between the actual cross-sectional image Rt(j) and the actual cross-sectional image Rt(j+1) of two adjacent portions of the catalyst layer C will be too large. This may result in the sequential generation of a plurality of unrealistic estimated cross-sectional images Ei from the actual cross-sectional image Ri. This may result in a decrease in the accuracy of the three-dimensional model M.

[0050] If the interval G is too small, the amount of change between the actual cross-sectional image Rt(j) and the actual cross-sectional image Rt(j+1) of two adjacent portions of the catalyst layer C may be too small. This may result in the sequential generation of multiple estimated cross-sectional images Ei that do not vary significantly from the actual cross-sectional image Ri. This may result in a decrease in the accuracy of the three-dimensional model M.

[0051] As described above, in the actual cross-sectional image Ri of the catalyst layer C, the pixel values ​​of the pixels corresponding to the pores are relatively small. The same is true for the multiple estimated cross-sectional images Ei generated using the actual cross-sectional image Ri and the learning model Lm. The accuracy of the three-dimensional model M generated based on the multiple estimated cross-sectional images Ei can be evaluated using a pore size distribution that indicates the relationship between pore size and pore volume. The pore size distribution of the three-dimensional model M can be calculated using, for example, the software product GeoDict from Math2Market GmbH. Figure 5 shows the pore size distribution according to the interval G between portions corresponding to the cross section of the catalyst layer C.

[0052] In Fig. 5, graph Vb shows the pore size distribution of the three-dimensional model M generated based only on a plurality of actual cross-sectional images of the catalyst layer C as in the conventional method. In graph Vb, the pore volume peaks at pore diameter Db. In Fig. 5, graphs V15, V10, and V5 all show the pore size distribution of the three-dimensional model M generated based on the actual cross-sectional image Ri and a plurality of estimated cross-sectional images Ei by the three-dimensional model generation device 10 of this embodiment.

[0053] Graph V15 shows the pore size distribution of the three-dimensional model M generated when the value of the distance G between one portion and another portion corresponding to the cross section of the catalyst layer C is 15 nm. In graph V15, the pore volume peaks at pore diameter D15. The peak value of the pore volume in graph V15 is significantly smaller than the peak value of the pore volume in graph Vb. In addition, graph V15 is shifted overall in the direction of larger pore diameters than graph Vb.

[0054] Graph V10 shows the pore size distribution of the three-dimensional model M generated when the value of the distance G between one portion and another portion corresponding to the cross section of the catalyst layer C is 10 nm. In graph V10, the pore volume peaks at pore diameter D10. There is no significant difference between the peak pore volume value in graph V10 and the peak pore volume value in graph Vb. Furthermore, graph V10 is not shifted in the overall direction of larger pore diameters than graph Vb, as is the case with graph V15. Furthermore, the pore diameter D10 at which the pore volume peaks in graph V10 is close to the pore diameter Db at which the pore volume peaks in graph Vb.

[0055] Graph V5 shows the pore size distribution of the three-dimensional model M generated when the value of the distance G between one portion and another portion corresponding to the cross section of the catalyst layer C is 5 nm. In graph V5, the pore volume peaks at pore diameters D5p and D5q. The distance between the two peaks in graph V5 is large. The pore diameters D5p and D5q in graph V5 are relatively far from the pore diameter Db in graph Vb. The peak value of the pore volume in graph V5 is significantly smaller than the peak value of the pore volume in graph Vb. In addition, graph V5 is shifted overall in the direction of larger pore diameters than graph Vb.

[0056] Based on this analysis, it is preferable that the distance G between one portion and another portion corresponding to the cross section of the catalyst layer C is greater than 5 nm and less than 15 nm. Furthermore, it is more preferable that the distance G is 10 nm. This can prevent a decrease in the accuracy of the three-dimensional model M.

[0057] The above-described embodiment may be modified as follows: In the following modifications, explanations that overlap with the above-described embodiment will be omitted.

[0058] (Variation 1) As described above, in the actual cross-sectional image Ri used to generate the estimated cross-sectional image Ei, pixels showing pixel values ​​that are relatively small may contain noise. The same is true for the actual cross-sectional image Rt used as training data. Furthermore, pixels showing pixel values ​​that are relatively large may also contain noise. Pixels corresponding to structures may be estimated as pixels corresponding to pores due to the inclusion of noise. Pixels corresponding to pores may be estimated as pixels corresponding to structures due to the inclusion of noise.

[0059] That is, noise may affect the accuracy of the three-dimensional model M. In this first modification, in order to reduce such noise, pixel values ​​outside a predetermined gradation range of pixels included in the actual cross-sectional images Ri and Rt are corrected to pixel values ​​within the predetermined gradation range.

[0060] FIG. 6 is a block diagram of the three-dimensional model generating device 10. In FIG. 6, the same reference numerals are used for components common to FIG. 1. A description of these components will be omitted. The calculation unit 12 shown in FIG. 6 further includes a correction unit 50. The correction unit 50 can be realized by the calculation unit 12 executing a program stored in the storage unit 14.

[0061] The correction unit 50 corrects pixel values ​​outside a predetermined gradation range of pixels included in one actual cross-sectional image Ri of the catalyst layer C to pixel values ​​within the predetermined gradation range. The correction unit 50 corrects pixel values ​​outside a predetermined gradation range of pixels included in an actual cross-sectional image Rt of one portion of the catalyst layer C to pixel values ​​within the predetermined gradation range. The correction unit 50 corrects pixel values ​​outside the predetermined gradation range of pixels included in an actual cross-sectional image Rt of another portion adjacent to the one portion to pixel values ​​within the predetermined gradation range. This reduces noise, thereby improving the accuracy of the generated three-dimensional model M.

[0062] Fig. 7A is a diagram illustrating pixel values ​​of pixels included in actual cross-sectional image Ri or actual cross-sectional image Rt. In Fig. 7A, a horizontal coordinate axis and a vertical coordinate axis with the upper left corner of the image as the origin are defined as coordinate axes for expressing pixel positions by coordinate values. The pixel value of each pixel is expressed in 256 grayscale levels. In Fig. 7A, the pixel value of pixel P(h,v) located at horizontal coordinate value h and vertical coordinate value v is 192. The pixel values ​​of the eight pixels located one pixel away from pixel P(h,v) are all 82 or more and 86 or less.

[0063] FIG. 7B is a diagram for explaining pixel value correction. FIG. 7B shows the frequency distribution of pixel values ​​for a total of nine pixels, consisting of pixel P(h,v) in FIG. 7A and eight surrounding pixels located one pixel away from pixel P(h,v). Of these nine pixels, one pixel exhibits a pixel value of 82, one pixel exhibits a pixel value of 83, three pixels exhibit a pixel value of 84, two pixels exhibit a pixel value of 85, one pixel exhibits a pixel value of 86, and one pixel exhibits a pixel value of 192. The pixel exhibiting the pixel value of 192 is pixel P(h,v). The pixel values ​​of the eight surrounding pixels of pixel P(h,v) are included within a predetermined gradation range Pr of 82 to 86 inclusive. The pixel value of 192 for pixel P(h,v) is not included within the predetermined gradation range Pr that includes the pixel values ​​of the eight surrounding pixels.

[0064] When pixel P(h,v) contains noise, the pixel value of pixel P(h,v) may be outside the predetermined gradation range Pr. In the example shown in Fig. 7A, pixel value 192 of pixel P(h,v) is outside the predetermined gradation range Pr, between 82 and 86. Therefore, noise can be reduced by correcting pixel value 192 of pixel P(h,v) to a pixel value within the predetermined gradation range Pr.

[0065] FIG. 7C is a diagram illustrating a pixel P(h,v) whose pixel value has been corrected. In FIG. 7C, the same horizontal and vertical coordinate axes as in FIG. 7A are defined. As described above, of the eight pixels adjacent to pixel P(h,v) that have pixel values ​​within the predetermined gradation range Pr, the three most common pixels have a pixel value of 84. That is, the most common pixel value is 84. Therefore, the correction unit 50 corrects the pixel value 192 of pixel P(h,v) included in the actual cross-sectional image Ri or Rt, which is outside the predetermined gradation range Pr, to the pixel value 84 within the predetermined gradation range Pr. In this way, noise can be removed by correcting the pixel value using a mode filter.

[0066] In this first modification, the correction unit 50 corrects the pixel value 192 of the pixel P(h,v) to the pixel value 84, which is the most frequent value within the predetermined gradation range Pr, but is not limited to this. The correction unit 50 may correct the pixel value 192 of the pixel P(h,v) to a pixel value other than 84 that is within the predetermined gradation range Pr, which is equal to or greater than 82 and equal to or less than 86. In other words, noise may be removed by correcting the pixel value using another filter, such as a mean filter or a median filter.

[0067] Furthermore, the pixel values ​​of the pixels included in the entire actual cross-sectional image Ri or the actual cross-sectional image Rt may be binarized to a pixel value of 0 or a pixel value of 255 based on a threshold value. In this case, the predetermined gradation range Pr consists of the pixel value 0 and the pixel value 255 (discrete value). When the threshold value is, for example, the pixel value 200, the correction unit 50 corrects the pixel value 192 of the pixel P(h,v) to a pixel value of 0 within the predetermined gradation range Pr.

[0068] Furthermore, the pixel values ​​of the pixels included in the entire cross-sectional image Ri or the entire cross-sectional image Rt may be classified into pixel values ​​corresponding to multiple clusters by clustering such as k-means. In this case, the predetermined gradation range Pr is made up of pixel values ​​(discrete values) corresponding to the multiple clusters. The correction unit 50 corrects the pixel value of a pixel P(h,v) classified into a certain cluster to a pixel value within the predetermined gradation range Pr corresponding to the cluster.

[0069] In this first modification, the pixel value of each pixel is expressed in a grayscale of 256 levels, but this is not limiting. The pixel value of each pixel may be expressed in a grayscale of a number of levels other than 256. Furthermore, the pixel value of each pixel may be expressed using multiple components that make up a color space used for color images. Color spaces used for color images include, for example, RGB, HSV, HSL, CMY, and CMYK.

[0070] (Variation 2) There are eight surrounding pixels located one pixel away from a pixel included in the portion other than the peripheral portion of the estimated cross-sectional image Ei. However, the number of surrounding pixels located one pixel away from a pixel included in the peripheral portion of the estimated cross-sectional image Ei is less than eight. The amount of information that can be used to estimate the pixel values ​​of pixels included in the peripheral portion of the estimated cross-sectional image Ei is lower than that of pixels included in the portion other than the peripheral portion. This information is, for example, information regarding the continuity of pixel values ​​between adjacent pixels.

[0071] Therefore, the pixel values ​​of pixels included in the peripheral portions of the multiple estimated cross-sectional images Ei sequentially generated by the estimated cross-sectional image generating unit 32 may change more monotonically than the pixel values ​​of pixels included in portions other than the peripheral portions. In other words, a decrease in the amount of information that can be used to estimate the pixel values ​​of pixels included in the peripheral portions of the estimated cross-sectional images Ei may affect the accuracy of the three-dimensional model M.

[0072] In this second variant, in order to reduce the impact on the accuracy of the three-dimensional model M of the reduction in the amount of information that can be used to estimate the pixel values ​​of pixels included in the peripheral parts of the estimated cross-sectional image Ei, the pixel values ​​of pixels included in the peripheral parts of the estimated cross-sectional image Ei are interpolated using the pixel values ​​of pixels included in parts other than the peripheral parts.

[0073] FIG. 8 is a block diagram of the three-dimensional model generating device 10. In FIG. 8, the same reference numerals are used for components common to FIG. 1. A description of these components will be omitted. The calculation unit 12 shown in FIG. 8 further includes an interpolation unit 52. The interpolation unit 52 can be realized by the calculation unit 12 executing a program stored in the storage unit 14.

[0074] The interpolation unit 52 interpolates the pixel values ​​of pixels located in the peripheral portion of one estimated cross-sectional image Ei using pixel values ​​of pixels included in the portion other than the peripheral portion of one estimated cross-sectional image Ei. This reduces the impact on the accuracy of the three-dimensional model M caused by a decrease in the amount of information that can be used to estimate the pixel values ​​of pixels included in the peripheral portion of the estimated cross-sectional image Ei. This makes it possible to improve the accuracy of the generated three-dimensional model M.

[0075] FIG. 9A is a diagram illustrating pixel values ​​of pixels P(0,0), P(1,0), P(2,0), P(0,1), and P(0,2), which are part of the pixels located in the peripheral portion Sp of the estimated cross-sectional image Ei. In FIG. 9A, a horizontal coordinate axis and a vertical coordinate axis, each having the origin at the upper left corner of the image, are defined as coordinate axes for expressing the pixel positions in coordinate values. In FIG. 9A, at least five pixels, namely, pixel P(0,0), pixel P(1,0), pixel P(2,0), pixel P(0,1), and pixel P(0,2), are located in the peripheral portion Sp. For these five pixels, the number of surrounding pixels located one pixel away from each pixel is less than eight.

[0076] For example, the number of surrounding pixels located one pixel away from pixel P(0,0) is 3. The number of surrounding pixels located one pixel away from each of pixel P(1,0), pixel P(2,0), pixel P(0,1), and pixel P(0,2) is 5. The estimated cross-sectional image generating unit 32 generates the plurality of estimated cross-sectional images Ei before the estimated cross-sectional image Ei shown in FIG. 9A, the estimated cross-sectional image Ei shown in FIG. 9A, and the plurality of estimated cross-sectional images Ei after the estimated cross-sectional image Ei shown in FIG. 9A, in which the pixel values ​​of the above-mentioned five pixels located in the peripheral portion Sp change more monotonically than the pixel values ​​of the pixels located inside the peripheral portion Sp.

[0077] 9A, there is a possibility that the pixel value 0 of pixel P(0,0), the pixel value 0 of pixel P(1,0), the pixel value 2 of pixel P(2,0), the pixel value 2 of pixel P(0,1), and the pixel value 4 of pixel P(0,2) all change monotonically across the multiple estimated cross-sectional images Ei. Therefore, by having the interpolation unit 52 interpolate the pixel values ​​of these five pixels, the effect on the accuracy of the three-dimensional model M caused by a decrease in the amount of information that can be used to estimate the pixel values ​​of these five pixels can be reduced.

[0078] FIG. 9B is a diagram illustrating pixel P(0,0), pixel P(1,0), pixel P(2,0), pixel P(0,1), and pixel P(0,2) whose pixel values ​​have been interpolated. Of the three surrounding pixels located one pixel away from pixel P(0,0), pixel P(1,1) is the only pixel included in the portion other than the peripheral portion Sp. Interpolation unit 52 interpolates the pixel value of pixel P(0,0) located in the peripheral portion Sp using pixel value 20 of pixel P(1,1) included in the portion other than the peripheral portion Sp. In the example shown in FIG. 9B, the pixel value 20 of pixel P(0,0) after interpolation is equal to the pixel value 20 of pixel P(1,1).

[0079] Of the five surrounding pixels located one pixel away from pixel P(1,0), the two pixels included in the portion other than the peripheral portion Sp are pixel P(1,1) and pixel P(2,1). The interpolation unit 52 interpolates the pixel value of pixel P(1,0) located in the peripheral portion Sp using the pixel value of at least one of pixels P(1,1) and P(2,1) included in the portion other than the peripheral portion Sp. In the example shown in FIG. 9B, the pixel value 20 of pixel P(1,0) after interpolation is equal to the pixel value 20 of pixel P(1,1).

[0080] The interpolation unit 52 interpolates the pixel values ​​of pixels P(2,0), P(0,1), and P(0,2) located in the peripheral region Sp using the pixel values ​​of the five surrounding pixels located one pixel away from each pixel and included in the portion outside the peripheral region Sp, as with pixel P(1,0). In the example shown in FIG. 9B, the pixel value 48 of pixel P(2,0) after interpolation is equal to the pixel value 48 of pixel P(2,1). The pixel value 20 of pixel P(0,1) after interpolation is equal to the pixel value 20 of pixel P(1,1). The pixel value 51 of pixel P(0,2) after interpolation is equal to the pixel value 51 of pixel P(1,2).

[0081] 9A shows five pixels that are part of the pixels located in the peripheral portion Sp of the estimated cross-sectional image Ei, and the explanation so far has been given of the correction of the pixel values ​​of these five pixels. The interpolation unit 52 corrects the pixel values ​​of the remaining pixels located in the peripheral portion Sp of the estimated cross-sectional image Ei in the same way as these five pixels. This allows the pixel values ​​of all pixels located in the peripheral portion Sp of the estimated cross-sectional image Ei to be corrected.

[0082] In the present modified example 2, the interpolation unit 52 interpolates the pixel value of a pixel located in the peripheral portion Sp using the pixel value of a pixel located one pixel away from the pixel in question and included in the portion other than the peripheral portion Sp, but is not limited to this. The interpolation unit 52 may also interpolate the pixel value of a pixel located in the peripheral portion Sp using the pixel values ​​of multiple pixels included in the portion other than the peripheral portion Sp. The interpolation unit 52 may interpolate the pixel value of a pixel located in the peripheral portion Sp of the estimated cross-sectional image Ei by replacing it with the pixel value of a pixel located in the peripheral portion of an image obtained by enlarging the portion other than the peripheral portion Sp to the same number of pixels as the estimated cross-sectional image Ei.

[0083] (Variation 3) A three-dimensional model evaluation system including the above-described three-dimensional model generation device 10 may generate the three-dimensional model M and evaluate the generated three-dimensional model M. The three-dimensional model evaluation system according to the present modified example 3 makes it possible to easily obtain an evaluation result of the three-dimensional model M.

[0084] FIG. 10 is a block diagram of a three-dimensional model evaluation system 10A. In FIG. 10, the same reference numerals are used for components common to FIG. 1. A description of these components will be omitted. The three-dimensional model evaluation system 10A shown in FIG. 10 includes the three-dimensional model generation device 10, in which the calculation unit 12 of the three-dimensional model generation device 10 shown in FIG. 1 further includes an evaluation unit 60. The evaluation unit 60 can be realized by the calculation unit 12 executing a program stored in the memory unit 14.

[0085] The evaluation unit 60 evaluates the three-dimensional model M generated by the three-dimensional model generation unit 34 and outputs the evaluation result Ar of the three-dimensional model M to the display device 80. In addition to the pore size distribution described above, the evaluation unit 60 can measure, for example, the tortuosity of the pores and the concentration of oxygen that can flow through the pores. The tortuosity is also called the curvature ratio. The evaluation result Ar may include evaluation results that include these measurement results. By outputting the evaluation result Ar of the three-dimensional model M to the display device 80, the operator can easily determine the characteristics of the fuel cell based on the displayed evaluation result Ar.

[0086] (Variation 4) The three-dimensional model generation device 10 according to the above-described embodiment is used to generate a three-dimensional model M of a catalyst layer C that constitutes an electrode of a fuel cell, but is not limited to this. The three-dimensional model generation device 10 may also be used to generate a three-dimensional model of a structure having pores therein other than the catalyst layer C of a fuel cell. The structure having pores therein is, for example, an electrode or separator of a lithium-ion battery.

[0087] (Variation 5) The above-described modifications 1 to 4 may be combined as appropriate within a range that does not cause inconsistency.

[0088] The following additional notes are provided regarding the above-described embodiment and modifications.

[0089] (Appendix 1) The three-dimensional model generating device (10) of the present disclosure includes a learning model storage unit (20) storing a learning model (Lm) for estimating an image of another cross section based on an image of one cross section of a catalyst layer (C) of a fuel cell, an estimated cross-sectional image generating unit (32) generating an estimated cross-sectional image (Ei) using an actual cross-sectional image (Ri) of the catalyst layer and the learning model, and generating another estimated cross-sectional image using the generated estimated cross-sectional image and the learning model, and a three-dimensional model generating unit (34) generating a three-dimensional model (M) of the catalyst layer based on the multiple estimated cross-sectional images sequentially generated by the estimated cross-sectional image generating unit. This configuration allows a three-dimensional model to be generated easily without requiring a large number of steps.

[0090] (Appendix 2) In the three-dimensional model generating device described in Supplementary Note 1, the learning model may be a learning model for estimating, based on a cross-sectional image of one portion of a catalyst layer, an image of a cross-section of another portion adjacent to the one portion. With this configuration, multiple estimated cross-sectional images can be generated quickly.

[0091] (Appendix 3) In the three-dimensional model generating device described in Supplementary Note 2, the distance (G) between the one portion and the other portion may be greater than 5 nm and less than 15 nm. With this configuration, a decrease in accuracy of the three-dimensional model can be suppressed.

[0092] (Appendix 4) The three-dimensional model generating device described in Supplementary Note 1 may further include a training data acquiring unit (36) that acquires training data including an actual cross-sectional image (Rt) of one portion of the catalyst layer and an actual cross-sectional image of another portion adjacent to the one portion, and a learning unit (38) that uses the training data to generate the learning model, inputting the actual cross-sectional image of the one portion and outputting the estimated cross-sectional image estimated as an image of the cross section of the other portion. With this configuration, multiple estimated cross-sectional images can be generated with high accuracy.

[0093] (Appendix 5) In the three-dimensional model generating device described in Supplementary Note 1, the estimated cross-sectional image generating unit may generate the estimated cross-sectional image of another region adjacent to the one region by using the estimated cross-sectional image of the one region and the learning model. With this configuration, it is possible to quickly generate a plurality of estimated cross-sectional images.

[0094] (Appendix 6) The three-dimensional model generating device described in Supplementary Note 1 may further include a correction unit (50) that corrects pixel values ​​of pixels included in one of the actual cross-sectional images that are outside a predetermined gradation range (Pr) to pixel values ​​within the predetermined gradation range. With this configuration, noise is reduced, thereby improving the accuracy of the generated three-dimensional model.

[0095] (Appendix 7) The three-dimensional model generating device described in Supplementary Note 4 may further include a correction unit that corrects pixel values ​​outside a predetermined gradation range of pixels included in the actual cross-sectional image of the one region to pixel values ​​within the predetermined gradation range, and corrects pixel values ​​outside the predetermined gradation range of pixels included in the actual cross-sectional image of the other region to pixel values ​​within the predetermined gradation range. With this configuration, noise is reduced, thereby improving the accuracy of the generated three-dimensional model.

[0096] (Appendix 8) The three-dimensional model generating device described in Supplementary Note 1 may further include an interpolation unit (52) that interpolates pixel values ​​of pixels located in a peripheral portion (Sp) of one of the estimated cross-sectional images using pixel values ​​of pixels included in a portion other than the peripheral portion of one of the estimated cross-sectional images. With this configuration, it is possible to improve the accuracy of the generated three-dimensional model.

[0097] (Appendix 9) A three-dimensional model evaluation system (10A) of the present disclosure includes a three-dimensional model generation device according to any one of Supplementary Notes 1 to 8. With this configuration, the evaluation result of the three-dimensional model can be easily obtained.

[0098] (Appendix 10) The three-dimensional model generation device of the present disclosure includes a learning model storage unit storing a learning model for estimating an image of another cross section based on an image of a cross section of a structure having pores therein, an estimated cross-sectional image generation unit generating an estimated cross-sectional image using an actual cross-sectional image of the structure and the learning model, and generating another estimated cross-sectional image using the generated estimated cross-sectional image and the learning model, and a three-dimensional model generation unit generating a three-dimensional model of the structure based on the multiple estimated cross-sectional images sequentially generated by the estimated cross-sectional image generation unit. With this configuration, it is possible to generate a three-dimensional model at low cost.

[0099] Although the present disclosure has been described in detail, the present disclosure is not limited to the individual embodiments described above. Various additions, substitutions, modifications, partial deletions, etc. are possible in these embodiments without departing from the gist of the present disclosure or the spirit of the present disclosure derived from the content of the claims and their equivalents. These embodiments can also be implemented in combination. For example, in the above-described embodiments, the order of each operation and the order of each process are shown as examples and are not limited to these. The same applies when numerical values ​​or mathematical expressions are used in the description of the above-described embodiments. [Explanation of symbols]

[0100] 10... Three-dimensional model generating device 12... Calculation unit 14...Memory section 20...Learning model memory section 30... Image input receiving unit 32... Estimated cross-sectional image generating unit 34...3D model generation unit 36...Teacher data acquisition unit 38...Learning section 50...Correction section 52...Interpolation unit 60...Evaluation unit 80...Display device

Claims

1. a learning model storage unit that stores a learning model for estimating an image of another cross section based on an image of one cross section of a catalyst layer of a fuel cell; an estimated cross-sectional image generating unit that generates one estimated cross-sectional image using one actual cross-sectional image of the catalyst layer and the learning model, and generates another estimated cross-sectional image using the generated one estimated cross-sectional image and the learning model; a three-dimensional model generation unit that generates a three-dimensional model of the catalyst layer based on a plurality of estimated cross-sectional images sequentially generated by the estimated cross-sectional image generation unit; A three-dimensional model generating device comprising:

2. 2. The three-dimensional model generating device according to claim 1, A three-dimensional model generating device, wherein the learning model is a learning model for estimating a cross-sectional image of another portion adjacent to one portion of a catalyst layer based on a cross-sectional image of the one portion.

3. 3. The three-dimensional model generating device according to claim 2, A three-dimensional model generating device, wherein the distance between the one portion and the other portion is greater than 5 nm and smaller than 15 nm.

4. 2. The three-dimensional model generating device according to claim 1, a training data acquisition unit that acquires training data including an actual cross-sectional image of one portion of the catalyst layer and an actual cross-sectional image of another portion adjacent to the one portion; a learning unit that uses the training data to generate the learning model, inputting an actual cross-sectional image of one part and outputting an estimated cross-sectional image estimated as an image of a cross-section of another part.

5. 2. The three-dimensional model generating device according to claim 1, The estimated cross-sectional image generation unit generates an estimated cross-sectional image of another region adjacent to the one region using the estimated cross-sectional image of the one region and the learning model.

6. 2. The three-dimensional model generating device according to claim 1, The three-dimensional model generating device further comprises a correction unit that corrects pixel values ​​of pixels included in the one actual cross-sectional image that are outside a predetermined gradation range to pixel values ​​within the predetermined gradation range.

7. 5. The three-dimensional model generating device according to claim 4, a correction unit that corrects pixel values ​​outside a predetermined gradation range of pixels included in an actual cross-sectional image of one region to pixel values ​​within the predetermined gradation range, and corrects pixel values ​​outside the predetermined gradation range of pixels included in an actual cross-sectional image of another region to pixel values ​​within the predetermined gradation range.

8. 2. The three-dimensional model generating device according to claim 1, a three-dimensional model generating device further comprising an interpolation unit that interpolates pixel values ​​of pixels located in a peripheral portion of the one estimated cross-sectional image using pixel values ​​of pixels included in a portion other than the peripheral portion of the one estimated cross-sectional image.

9. A three-dimensional model evaluation system comprising the three-dimensional model generating device according to any one of claims 1 to 8.

10. a learning model storage unit that stores a learning model for estimating an image of another cross section based on an image of one cross section of a structure having pores therein; an estimated cross-sectional image generating unit that generates one estimated cross-sectional image using one actual cross-sectional image of the structure and the learning model, and generates another estimated cross-sectional image using the generated one estimated cross-sectional image and the learning model; a three-dimensional model generation unit that generates a three-dimensional model of the structure based on a plurality of estimated cross-sectional images sequentially generated by the estimated cross-sectional image generation unit; A three-dimensional model generating device comprising:

Images