Manufacturing process prediction device, method, program and system

Abstract

To provide manufacturing process prediction device, method, program, and system that can reduce the time required to search for a manufacturing process that achieves desired strength characteristics.SOLUTION: A manufacturing process prediction device of an embodiment has an extraction unit, an image generation unit, an image acquisition unit, and an optimization unit. The extraction unit extracts features of an image based on topological data analysis on the image. The image generation unit generates a pseudo image related to a prediction target of the manufacturing process by applying it to a generation model generated from image generation parameters corresponding to the manufacturing process parameters. The image acquisition unit acquires a first image regarding the prediction target. The optimization unit optimizes the input values applied to the generation model so that the first feature value obtained by applying the first image to the extraction unit matches the second feature value obtained by applying the pseudo second image generated by inputting the input values of the image generation parameters to the generation model to the extraction unit.SELECTED DRAWING: Figure 8

Description

[Technical field]

[0001] An embodiment of the present invention relates to a manufacturing process prediction apparatus, method, program, and system. [Background technology]

[0002] Casting materials are mixed with inorganic fillers such as silica (SiO2) according to the specifications of each product. The mechanical strength of the casting material, such as bending strength and tensile strength, varies depending on the shape and particle size of the inorganic filler, and the heat resistance (glass transition temperature) varies depending on the chemical structure of the resin. Therefore, it is very important to optimize the manufacturing process so that the strength characteristics required for the product are met.

[0003] The composition of materials and the selection of manufacturing processes are experimentally determined according to the specifications of each product. However, the experiments often involve multiple intertwined factors, and optimization is performed without knowing the true parameters to be controlled. This results in a huge number of combinations, requiring a lot of development time and costs.

[0004] In recent years, attempts have been made to reduce development costs and shorten development times through trial and error by introducing data science methods and reusing existing accumulated data or effectively utilizing new data. For example, in the case of image data, there have been many attempts to link it to material properties using deep learning and other methods. There are also methods for extracting topological information contained in images and investigating the correlation with material properties. However, these existing methods are limited to analyzing the correlation between the features of the microstructure contained in the images and the material properties, and do not provide a manufacturing process for realizing a microstructure that achieves the desired strength properties. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] JP 2019-209271 A [Non-patent literature]

[0006] [Non-Patent Document 1] Agrawal, A., & Choudhary, A. (2019). Deep materials informatics: Applications of deep learning in materials science. Mrs Communications, 9(3), 779-792. [Non-Patent Document 2] Sun, Y., Hanhan, I., Sangid, MD, & Lin, G. (2020). Predicting mechanical properties from microstructure images in fiber-reinforced polymers using convolutional neural networks. arXiv preprint arXiv:2010.03675. [Non-Patent Document 3] Gupta, S., Mukhopadhyay, T., & Kushvaha, V. (2022). Microstructural image based convolutional neural networks for efficient prediction of full-field stress maps in short fiber polymer composites. Defense Technology. [Non-Patent Document 4] Edelsbrunner, H., Letscher, D., and Zomorodian, A. 2000 “Topological persistence and simplification,” in Proceedings 41st annual symposium on foundations of computer science, 454-463, IEEE. [Non-Patent Document 5] Adams, H., Emerson, T., Kirby, M., Neville, R., Peterson, C., Shipman, P., Chepushtanova, S., Hanson, E., Motta, F., and Ziegelmeier, L. 2017 “Persistence images: A stable vector representation of persistent homology,” Journal of Machine Learning Research 18. Summary of the Invention [Problem to be solved by the invention]

[0007] An object of the present invention is to provide a manufacturing process prediction device, method, program, and system that can reduce the time required to search for a manufacturing process that achieves desired strength characteristics. [Means for solving the problem]

[0008] The manufacturing process prediction device according to the embodiment includes an extraction unit, an image generation unit, an image acquisition unit, and an optimization unit. The extraction unit extracts a feature value of the image based on a topological data analysis of the image. The image generation unit generates a pseudo image related to a prediction target of the manufacturing process by applying it to a generative model generated from image generation parameters corresponding to manufacturing process parameters. The image acquisition unit acquires a first image related to the prediction target. The optimization unit optimizes the input value applied to the generative model so that a first feature value obtained by applying the first image to the extraction unit coincides with a second feature value obtained by applying a pseudo second image generated by inputting an input value of the image generation parameter to the generative model. [Brief description of the drawings]

[0009] [Figure 1] FIG. 1 is a diagram illustrating an example of a hardware configuration of a manufacturing process prediction device. [Diagram 2] A diagram showing an example of the functional configuration of a processor [Diagram 3] An example of an SEM image [Figure 4] Schematic diagram of the process of generating pseudo-SEM images using a generative model [Diagram 5] Schematic diagram of persistent homology analysis [Figure 6] A specific example of a persistence diagram [Figure 7] Diagram showing the correlation between persistent image and strength characteristics (bending strength) [Figure 8] A diagram showing the procedure for manufacturing process prediction processing [Figure 9] FIG. 9 is a diagram showing a schematic diagram of the processing procedure from steps S1 to S5 in FIG. 8; [Figure 10] FIG. 9 is a diagram showing the transition of data generated in the processing procedure from steps S1 to S5 in FIG. 8. [Figure 11] FIG. 9 is a diagram showing a schematic process of step S6 in FIG. 8; [Figure 12] FIG. 12 is a diagram showing the transition of data generated in the processing procedure of step S6 in FIG. 11. [Figure 13] FIG. 1 shows an actual SEM image and a pseudo SEM image of sample A in Example 1. [Figure 14] FIG. 1 shows a zero-order persistence diagram of an SEM image of sample A and a zero-order persistence diagram of a pseudo SEM image of sample A in Example 1. [Figure 15] <Example 1> Pseudo SEM images and persistence diagrams when the porosity x[3] is changed [Figure 16] <Example 1> A diagram showing the correspondence between the three-dimensional porosity, which is a manufacturing process parameter, and the porosity, which is an image generation parameter. [Figure 17] <Example 2> Pseudo SEM images and persistence diagrams when the porosity x[3] is changed DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0010] Hereinafter, a manufacturing process prediction apparatus, method, program, and system according to the present embodiment will be described with reference to the drawings.

[0011] Fig. 1 is a diagram showing an example of a hardware configuration of a manufacturing process prediction apparatus 100. As shown in Fig. 1, the manufacturing process prediction apparatus 100 is a computer having a processor 1, a read only memory (ROM) 2, a random access memory (RAM) 3, an auxiliary storage device 4, an input device 5, a display device 6, and a communication device 7. Transmission and reception of data and various signals between the processor 1, the ROM 2, the RAM 3, the auxiliary storage device 4, the input device 5, the display device 6, and the communication device 7 is performed via a bus.

[0012] The processor 1 is an integrated circuit that controls the overall operation of the manufacturing process prediction apparatus 100. For example, the processor 1 has a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), and / or a floating-point unit (FPU). The processor 1 may also have an internal memory and an I / O interface. The processor 1 executes various processes by interpreting and calculating programs stored in advance in the ROM 2 or the auxiliary storage device 4. The processor 1 may be realized in part or in whole by hardware such as an application specific integrated circuit (ASIC) or a field programmable gate array (FPGA).

[0013] The ROM 2 is a non-volatile memory that stores various data. For example, the ROM 2 stores data and setting values ​​used when the processor 1 executes various processes. The ROM 2 may include a non-transitory computer-readable storage medium that stores a program executed by the processor 1.

[0014] The RAM 3 is a volatile memory used for reading and writing data. The RAM 3 temporarily stores data used when the processor 1 executes various processes. The RAM 3 provides a work area for the processor 1.

[0015] The auxiliary storage device 4 is a non-volatile memory that stores various data. For example, the auxiliary storage device 4 stores data and setting values ​​used when the processor 1 executes various processes, data generated by various processes in the processor 1, etc. The auxiliary storage device 4 is configured with an HDD (Hard Disk Drive), an SSD (Solid State Drive), an integrated circuit storage device, etc. The auxiliary storage device 4 may have a non-transitory computer-readable storage medium that stores a program executed by the processor 1.

[0016] The input device 5 accepts various operation inputs from an operator. Examples of the input device 5 that can be used include a keyboard, a mouse, various switches, a touch pad, and a touch panel display. An electrical signal corresponding to the accepted operation input (hereinafter, an operation signal) is supplied to the processor 1.

[0017] The display device 6 displays various data under the control of the processor 1. A CRT (Cathode-Ray Tube) display, a liquid crystal display, an organic EL (Electro Luminescence) display, an LED (Light-Emitting Diode) display, a plasma display, or any other display may be appropriately used as the display device 6. The display device 6 may be a projector.

[0018] The communication device 7 includes a communication interface such as a network interface card (NIC) for performing data communication between various devices connected to the manufacturing process prediction apparatus 100 via a network. An operation signal may be supplied from a computer connected via the communication device 7 or an input device included in the computer, and various data may be displayed on a display device or the like included in the computer connected via the communication device 7. However, for the sake of simplicity of the following description, unless otherwise specified, it is assumed that the source of the operation signal is the input device 5 and the display destination of the various data is the display device 6. The input device 5 can be replaced by a computer connected via the communication device 7 or an input device included in the computer, and the display device 6 can be replaced by a display device or the like included in the computer connected via the communication device 7.

[0019] The manufacturing process prediction apparatus 100 does not need to include all of the processor 1, ROM 2, RAM 3, auxiliary storage device 4, input device 5, display device 6, and communication device 7. If necessary, some of the processor 1, ROM 2, RAM 3, auxiliary storage device 4, input device 5, display device 6, and communication device 7 may not be provided. The manufacturing process prediction apparatus 100 may be provided with any additional hardware device useful for executing the processing according to this embodiment. The manufacturing process prediction apparatus 100 does not need to be physically composed of one computer, and may be composed of a computer system having multiple computers connected to each other so as to be able to communicate via wires or a network line. The allocation of the series of processing according to this embodiment to the multiple processors 1 implemented in each of the multiple computers can be arbitrarily set. All the processors 1 may execute all the processing in parallel, or a specific processing may be assigned to one or some of the processors 1, and the series of processing according to this embodiment may be executed as a whole computer system.

[0020] Fig. 2 is a diagram illustrating an example of a functional configuration of the processor 1. As illustrated in Fig. 2, the processor 1 has a functional configuration including an image acquisition unit 11, an image generation unit 12, an extraction unit 13, an optimization unit 14, a search unit 15, and a display control unit 16.

[0021] The image acquisition unit 11 acquires a first image, which is an image of a prediction target of a manufacturing process. The prediction target means a material of a product, or an object that constitutes a part or the whole of the product, in which the material is deformed, altered, or processed. The image acquired by the image acquisition unit 11 is image data obtained by photographing the material or object. The material is assumed to be a material that contains two or more types of materials. The material may be any material that can generate a pseudo image by the image generation unit 12. In the following embodiment, the prediction target is assumed to be a casting material containing an inorganic filler. Specifically, an epoxy cured material is assumed to be a matrix (base material) containing an epoxy base agent and an acid anhydride hardener, and a filler, silica (SiO2). The epoxy cured material can be used for a wide variety of products and / or applications ranging from daily necessities to factory products. For example, the epoxy cured material can be used for semiconductor resins, adhesives, switch gears, etc.

[0022] The image according to the present embodiment means image data generated by photographing the internal structure of the prediction target with an electron microscope, and is a gray level image expressing black and white shading with a predetermined gradation or a color scale image expressing RGB elements of color with a predetermined bit. As the electron microscope according to the present embodiment, a scanning electron microscope (SEM), an energy dispersive X-ray spectroscopy (EDX), a transmission electron microscope (TEM), a scanning ion microscope (SIM), etc. can be appropriately used. Hereinafter, the image according to the present embodiment is assumed to be a gray level image obtained by a scanning electron microscope. An image obtained by a scanning electron microscope is called an SEM image. FIG. 3 is a diagram showing an example of an SEM image. The observation magnification of the scanning electron microscope is preferably a magnification that makes it easy to observe the shape and particle size of silica in the epoxy cured product, and is preferably about 100 times to 500 times as an example.

[0023] The image generating unit 12 generates a pseudo image representing a predicted target of the manufacturing process by applying it to a generation model generated from image generation parameters corresponding to the manufacturing process parameters. The pseudo image is an image that simulates the photographing principle of the first image by an electron microscope. For example, the pseudo image simulates an SEM image. Hereinafter, the second image is also referred to as a pseudo image or a pseudo SEM image. The generation model is a mathematical model constructed to input input values ​​of the image generation parameters and output a pseudo image.

[0024] The manufacturing process parameters are parameters that characterize the manufacturing process to be predicted. Examples of the manufacturing process parameters include silica filling rate, particle shape (spherical or crushed), a combination blending ratio of different particle shapes, particle size and / or particle size distribution, and other blending conditions. The values ​​of the manufacturing process parameters are called manufacturing process values. The image generation parameters are parameters that are input to the generation model for generating a pseudo image, and contribute to the morphological characteristics of the pseudo image. Examples of the image generation parameters include porosity, particle size distribution, number of particle size samples, minimum particle size and / or particle arrangement. The image generation parameters include parameters that correspond to the manufacturing process parameters, and may also include parameters that do not correspond to the manufacturing process parameters. Each of the manufacturing process parameters and the image generation parameters has one or more parameters. Details of the image generation parameters will be described later.

[0025] The extraction unit 13 extracts a feature amount of an image based on a topological data analysis of the image. The feature amount quantifies a morphological feature of a prediction target depicted in the image. Specifically, the extraction unit 13 extracts a feature amount of a first image (hereinafter, a first feature amount) based on a topological data analysis of the first image. The extraction unit 13 extracts a feature amount of a second image (hereinafter, a second feature amount) based on a topological data analysis of the second image.

[0026] As shown in Fig. 2, the extraction unit 13 has a preprocessing unit 131, an analysis unit 132, and a conversion unit 133. The preprocessing unit 131 performs preprocessing on an image to generate a preprocessed image (hereinafter, a preprocessed image). The analysis unit 132 performs topological data analysis on the preprocessed image to calculate an analysis result. The conversion unit 133 converts the analysis result into numerical data. The numerical data is used as the above-mentioned feature amount.

[0027] The optimization unit 14 optimizes the input value applied to the generative model so that a first feature obtained by applying a first image to the extraction unit 13 matches a second feature obtained by applying a second image generated by inputting the input value of the image generation parameter to the generative model to the extraction unit 13. The second image means a pseudo image generated by inputting the input value of the image generation parameter to the generative model. The optimization unit 14 optimizes the input value by minimizing the distance between the first feature and the second feature. As the distance, any definition such as inner product distance (cosine similarity), Euclidean distance, Manhattan distance, or Chebyshev distance can be used. The optimized input value is called the optimal input value.

[0028] The search unit 15 uses the optimal input value to search for an optimal manufacturing process value of the manufacturing process parameter of the prediction target that satisfies the desired intensity characteristic. More specifically, the search unit 15 modulates the optimal input value to calculate multiple modulated input values, and searches for the optimal value of the manufacturing process parameter based on multiple third feature amounts obtained by applying multiple pseudo third images generated by applying the multiple modulated input values ​​to the generative model to the extraction unit 13. The optimal manufacturing process value of the manufacturing process parameter is called the optimal process value.

[0029] The display control unit 16 displays various information on the display device 6. For example, the display control unit 16 displays the first image, the second image, the third image, analysis results, numerical data, image generation parameters, manufacturing process parameters, and the like.

[0030] An example of the operation of the manufacturing process prediction apparatus 100 according to this embodiment will be described below. In the following embodiment, it is assumed that the image is an SEM image.

[0031] First, the image generating unit 12 will be described in detail with reference to Fig. 4. Fig. 4 is a diagram that illustrates a generation process of a pseudo SEM image using a generative model.

[0032] The generation model generates a pseudo SEM image by filling grains in a two-dimensional plane called a calculation domain based on image generation parameters. The pseudo SEM image is generated as a two-dimensional black and white image. The generation model fills grains while allowing overlapping. For the following description, the image generation parameters are assumed to be grain size distribution, porosity, number of grain size samples, minimum grain size, and grain arrangement. The image generation unit 12 applies the input values ​​of the image generation parameters, porosity, grain size distribution, number of grain size samples, minimum grain size, and grain arrangement, to the generation model to generate a pseudo SEM image. However, the image generation parameters according to this embodiment are not limited to this. The input order of the image generation parameters, grain size distribution x[0] and x[1], porosity x[3], number of grain size samples x[2], minimum grain size x[4], and grain arrangement x[5], is not limited to this order and may be any order.

[0033] Two image generation parameters, x[0] and x[1], are set for the particle size distribution. x[0] and x[1] are the statistical distribution function of particle size (diameter) or parameters that characterize the statistical distribution function. As the statistical distribution function, an analytical function such as a normal distribution or an exponential function is given. An exponential analytical function is given by the following formula (1). Specifically, in the analytical function of formula (1), the parameter α(x[0]) and the parameter β(x[1]) of formula (1) are used as image generation parameters.

[0034]

number

[0035] The porosity x[3] is the ratio of voids to the specified image size, in other words, it controls how many particles are filled into the specified image size. The particles are given as spheres or disks. However, in the algorithm according to this embodiment, the porosity calculated from the determined number of spheres does not completely match the specified porosity, so in order to accurately calculate the porosity, it is necessary to calculate the porosity for the obtained image.

[0036] The particle size sample number x[2] is the number of particle samples in the particle size distribution approximated by a discrete histogram. It is the number of bins when a particle size distribution is given and discretized by a histogram. The larger the sample number, the more particle sizes will be sampled from the statistical distribution.

[0037] The minimum particle size x[4] is the lower limit of the particle size. A value greater than 0 is given as a cutoff to avoid zero particle size when sampling particles from a particle size distribution.

[0038] The particle arrangement x[5] is a random number seed for generating particles. A random number is generated by substituting the random number seed into a predetermined pseudorandom function. The random number corresponds to the arrangement position of the particle in the calculation domain.

[0039] Next, the details of the processing of the extraction unit 13 will be described. As described above, the extraction unit 13 has a preprocessing unit 131 that performs preprocessing on an image to generate a preprocessed image, an analysis unit 132 that performs topological data analysis on the preprocessed image to calculate an analysis result, and a conversion unit 133 that converts the analysis result into numerical data that is a feature amount. Here, the preprocessing unit 131 performs binarization processing as preprocessing. The analysis unit 132 performs persistent homology analysis as topological data analysis to calculate a zeroth-order or first-order persistence diagram from the preprocessed image, and the conversion unit 133 converts the zeroth-order or first-order persistence diagram into numerical data such as a one-dimensional feature vector. Hereinafter, the preprocessing unit 131, the analysis unit 132, and the conversion unit 133 will be specifically described.

[0040] The preprocessing unit 131 performs preprocessing on the SEM image to generate a preprocessed image. Image data, such as SEM images, are generally grayscale images. The preprocessing is applied to the grayscale image. However, if the image has overexposed areas or blackout areas due to the imaging conditions, there is a possibility that an artificial bias is added to the data. If the grayscale image has overexposed areas or blackout areas, the preprocessing unit 131 performs preprocessing to convert the SEM image to black and white to avoid data bias due to these areas. As the preprocessing, binarization is performed. Specifically, the preprocessing unit 131 determines an appropriate threshold value from the pixel distribution expressed by 256 gradation values ​​from 0 to 255 in the grayscale image, and generates binary data (binarized image) that expresses the image data in Boolean format based on the threshold value. This makes it possible to generate a binary image that is not dependent on the imaging conditions of the scanning electron microscope. Before the binarization process, the image may be blurred with Gaussian blur to remove noise.

[0041] The analysis unit 132 performs persistent homology analysis on the preprocessed image. Persistent homology (Non-Patent Document 4) is a type of topological data analysis method (TDA: Topological Data Analysis) that applies mathematical homology. The predecessor of persistent homology is computational homology, which quantitatively calculates information about holes in each dimension (connected structure in 0 dimension, ring structure in 1 dimension, and void in 2 dimension). Persistent homology is an extension of computational homology that can analyze not only the information about the number of holes but also the scale information of the holes.

[0042] Here, persistent homology analysis will be explained by taking the data structure of a point cloud, which is a typical analysis target, as an example. Figure 5 is a diagram showing a schematic diagram of persistent homology analysis. The definition of a point cloud is R L A finite set of points on P={u i} N i=1In reality, a two-dimensional point set structure corresponding to L=2 or a three-dimensional point set structure corresponding to L=3 is used, and persistent homology is applied to obtain a quantitative index that characterizes the data structure P. In a specific analysis, we consider a disk or sphere with a radius centered on each point in the point set P, and calculate the homology of the geometric figure obtained when the radius of these disks is continuously changed. In the following, we will assume a disk.

[0043] The structure created from the point set P shown in Figure 5 is called a check complex. In a check complex, a pore structure may occur depending on the degree of overlap of the disks, and the size and shape of the pores change depending on the radius of the disks to be pasted. For example, if the radius r of the disk is continuously changed, the pore structure will change continuously. This is equivalent to changing the resolution of an image when capturing a cloud of points. In addition, the pore structure changes when the resolution is changed, reflecting the original point set. Therefore, through the continuous change in resolution of the geometric model in Figure 5 (this operation is called filtration), the characteristics of the original data structure can be captured. Filtration can construct geometric models of various data structures depending on the simplicial complex corresponding to the resolution. Examples of simplicial complexes include check complexes, alpha complexes, Lips complexes, and square complexes.

[0044] A more accurate mathematical definition of persistent homology is given as a homology linear mapping sequence obtained by adding a filtration operation to conventional homology calculations. In this embodiment, this technique is used to extract pore and scale information of a geometric model and apply it to classification and prediction of material properties.

[0045] Persistent homology can extract a global data structure rather than the correlation between individual data contained in the data. In addition, persistent homology does not require a large number of images to extract individual features, unlike other representative image analysis techniques such as Convolution Neural Network (Non-Patent Documents 2 and 3), and can extract a large amount of topological information from a single image, so it has the advantage of low cost for acquiring a data set.

[0046] The technology on which persistent homology is based is homology. Homology is a mathematical method for algebraically calculating holes that exist in a simplicial complex (= polyhedron). Specifically, it is possible to mechanically calculate homology groups of any dimension by calculating boundary operators, and obtain a representation of the hole structure in a vector space. Furthermore, an important property of homology is linear mapping, and based on this property, the persistent homology analysis described above is equivalent to "investigating the birth and death of the kth homology group of the homology linear mapping sequence induced by filtration." This means that the set of points in the left diagram of Figure 4 can be uniquely characterized by the generator pairs of the homology groups when the geometric model and filtration in the right diagram of the same figure are given. The generator ([b i ,d i ],b i is the i-th generation time, d i represents the annihilation time of the i-th homology group), the multiset D(χ)={(b i ,d i )|i=1,···,l k A persistence diagram is a two-dimensional plane defined by the birth time axis and the death time axis. A persistence diagram is also called a persistence diagram. The results of persistent homology analysis can be visualized using this persistence diagram. A persistence diagram is an example of the results of persistent homology analysis. The birth time and death time are also called the birth radius and death radius, respectively.

[0047] Figure 6 shows a specific example of a persistence diagram. The plots in Figure 6 represent each generator. The generators can be in the situations of b>d and d>b. For example, in the case of Figure 6, d>b is always the case, and generators always appear only on the upper left triangle of the first quadrant of the diagonal of the bd axis. Generators can also appear near the diagonal, which represents a noisy hole structure where holes appear and disappear immediately. On the other hand, generators away from the diagonal represent a robust hole structure where holes do not disappear easily. In this way, the persistence diagram allows you to see at a glance the data structure analyzed by persistent homology. A different persistence diagram is obtained for each image.

[0048] As mentioned above, persistence diagrams can be used as image features, and correlations between data structures can be examined using machine learning, etc. In this case, persistence diagrams need to be converted into quantitative numerical data in order to be input as explanatory variables for machine learning.

[0049] The conversion unit 133 converts the persistence diagram into numerical data. The numerical data has a format that can be input as an explanatory variable for machine learning. As an example, a vector is used as the numerical data. The persistent image method (Non-Patent Document 5) is known as a representative example of vectorization of a persistence diagram. The persistent image method regards the persistence diagram as a histogram with a specific number of bins (step width), and converts it into a one-dimensional numerical vector having a value for each bin width. When vectorizing, noisy generators near the diagonal can be suppressed to some extent by applying weights. Vectorization using a weighting function is expressed by the following equations (2) and (3).

[0050]

number

[0051] Here, D k(χ) is the k-th homology linear mapping sequence of the filtration χ, (b i ,d i ) represents each decomposed generator. σ, C, and p are parameters, and appropriate values ​​are set according to the data. In equation (2), D k A weight w is applied to each point of (χ) and the Gaussian distribution is added. In this case, the weight in equation (3) is determined according to the distance from the diagonal of the persistence diagram, and the closer to the diagonal, the smaller the value of w. In this way, the larger the value of p and the smaller the value of C, the shorter the lifetime (||d i -b i It is possible to reduce the small contribution of || (noisy generators near the diagonal in the persistence diagram). The persistent image transformed using equations (2) and (3) in this way can be used as an input for machine learning as explanatory variables that have the characteristics of the image.

[0052] FIG. 7 is a diagram showing the correlation between persistent images and strength characteristics (bending strength). The persistent image shown in FIG. 7 is a diagram in which values ​​based on the frequency of generators for each bin are plotted on a two-dimensional plane with the vertical axis representing the disappearance time and the horizontal axis representing the generation time. In other words, the persistent image is an image in which a one-dimensional numerical vector is visualized on a two-dimensional plane. The area indicated by label 0 in the persistent image has a bending strength lower than 145 MPa, meaning a component with weak bending strength. The area indicated by label 1 has a bending strength higher than 145 MPa, meaning a component with high bending strength. By comparing the persistent image based on the SEM image to be processed with the strength characteristic distribution in the persistent image shown in FIG. 7, it is possible to predict the superiority or inferiority of the strength characteristics of the SEM image to be processed. The area indicated by label 1 is called the reference area. It is possible to generate persistent images for other strength characteristics such as tensile strength by the same method as for bending strength.

[0053] Next, a procedure for generating a pseudo-SEM image by the image generating unit 12 using a generative model will be described. First, the image generating unit 12 inputs the image size of the calculation area. The image size is specified manually or automatically. For example, the vertical length and horizontal length are specified as the image size. When the image size is specified, the image generating unit 12 outputs a two-dimensional calculation area according to the specified image size in pixel values. Next, the image generating unit 12 specifies image generation parameters, which are porosity, particle size distribution, number of particle size samples, minimum particle size, and particle arrangement. The input values ​​of these image generation parameters are specified manually or automatically. Next, the image generating unit 12 applies the input values ​​of the image generation parameters, porosity, particle size distribution, number of particle size samples, minimum particle size, and particle arrangement, to the generative model to generate a pseudo-SEM image.

[0054] Next, the manufacturing process prediction processing according to this embodiment will be described with reference to Figs. 8, 9, and 10. Fig. 8 is a diagram showing the processing procedure of the manufacturing process prediction processing. Fig. 9 is a diagram showing a schematic diagram of the processing procedure from steps S1 to S5 in Fig. 8. Fig. 10 is a diagram showing the transition of data generated in the processing procedure from steps S1 to S5 in Fig. 8. Fig. 11 is a diagram showing a schematic diagram of the processing procedure of step S6 in Fig. 8. Fig. 12 is a diagram showing the transition of data generated in the processing procedure of step S6 in Fig. 11.

[0055] First, as shown in Fig. 8, Fig. 9 and Fig. 10, the image acquisition unit 11 acquires a first image (step S1). The first image is an SEM image obtained by photographing an actual casting material such as an epoxy cured product with a scanning electron microscope. Since the first image is an image of an actual epoxy cured product, the first image is also called an actual image. The manufacturing process values ​​of the manufacturing process parameters are associated with the first image.

[0056] After step S1 is performed, the extraction unit 13 extracts a first feature from the first image (step S2). Specifically, the preprocessing unit 131 performs binarization processing on the first image to generate a first binarized image, and the analysis unit 132 performs persistent homology analysis on the first binarized image to generate a first persistence map. As the persistent homology analysis, a first persistent homology analysis is performed. The conversion unit 133 applies the persistent image method to the first persistence map to calculate a first vector, which is a one-dimensional numerical vector. The first vector is used as a first feature.

[0057] After step S2, the image generator 12 uses the generative model to generate a second image from the input values ​​of the image generation parameters (step S3). The second image generated in step S3 only needs to look similar to an SEM image generated by an electron scanning microscope. The input values ​​used in step S3 do not need to be related to the manufacturing process values, and may be set by any suitable method. The pseudo image generated by the generative model is also called a pseudo SEM image.

[0058] After step S3, the extraction unit 13 extracts a second feature from the second image (step S4). Specifically, the preprocessing unit 131 performs binarization processing on the second image to generate a second binarized image, and the analysis unit 132 performs persistent homology analysis on the second binarized image to generate a second persistence map. As the persistent homology analysis, a first-order persistent homology analysis is performed. The conversion unit 133 applies the persistent image method to the second persistence map to calculate a second vector, which is a one-dimensional numerical vector. The second vector is used as the second feature.

[0059] After step S4 is performed, the optimization unit 14 optimizes the input values ​​of the image generation parameters based on the first feature extracted in step S2 and the second feature extracted in step S4 (step S5). The optimization unit 14 minimizes the inner product distance between the one-dimensional numerical vector that is the first feature and the one-dimensional numerical vector that is the second feature. As the optimization method, a mathematical optimization method represented by a gradient method or the like may be used, or a Bayesian optimization method may be used. As the optimization method of mathematical optimization, the Nelder-Mead method, the Powell method, the CG method, the BFGS method, the COBYLA method, or the like may be used.

[0060] Specifically, the optimization unit 14 calculates the dot product distance between the first feature based on the first image and the second feature based on the second image, and updates the input value based on the dot product distance. The updated input value is called an updated input value. Thereafter, the optimization unit 14 controls the image generation unit 12 and the extraction unit 13 to generate another second image based on the updated input value, and extract a second feature based on the second image. Then, the optimization unit 14 calculates the dot product distance between the first feature and the second feature, and updates the input value based on the dot product distance. The optimization unit 14 repeats the update until the number of updates reaches a predetermined number. The updated input value when the number of updates reaches the predetermined number is used as the optimal input value of the image generation parameter.

[0061] After step S5 is performed, searching unit 15 searches for optimal process values ​​of manufacturing process parameters using the optimal input values ​​of the image generation parameters obtained in step S5 (step S6).

[0062] As shown in FIG. 11 and FIG. 12, the search unit 15 first modulates the optimal input value to calculate a plurality of modulation input values. The modulation width, the number of modulation input values, and the like can be set to any value. The image generation parameters to be modulated may be all or a part of the plurality of parameters constituting the image generation parameters. The search unit 15 controls the image generation unit 12 and the extraction unit 13 to generate a third image, which is a pseudo image, for each of the plurality of modulation input values ​​based on the modulation input value, and extract a third feature amount based on the third image. Specifically, the image generation unit 12 applies the modulation input value to a generative model for each of the plurality of modulation input values ​​to generate a third image. The preprocessing unit 131 performs binarization processing on the third image to generate a third binarized image, the analysis unit 132 performs persistent homology analysis on the third binarized image to generate a third persistence diagram, and the conversion unit 133 applies the persistent image method to the third persistence diagram to convert it into a third vector. The third vector is used as a third feature.

[0063] Next, the search unit 15 searches for an optimal process value of the manufacturing process parameter based on a plurality of third feature quantities corresponding to the plurality of modulation input values, respectively. More specifically, when a specific third feature quantity that satisfies a criterion is present among the plurality of third feature quantities, the search unit 15 selects the manufacturing process value corresponding to the specific third feature quantity as the optimal process value. The criterion is that the manufacturing process value has a value less than a reference value in a reference region (label 0) of the persistent image as shown in FIG. 7. Note that the reference region and / or the reference value can be set arbitrarily.

[0064] The left diagram of FIG. 12 is a first persistent image based on persistent homology analysis of the real SEM image, which is the first image. The search unit 15 extracts a component that contributes to improving the desired strength characteristic from the correlation between the first persistent image and the strength characteristic for the real SEM image. The contributing component means a region that contributes to improving the desired strength characteristic among the regions defined by the generation time and the disappearance time in the persistent image. The contributing component can be specified and extracted based on reverse analysis, empirical knowledge, etc. The search unit 15 sets a criterion based on the contributing component. In the example of FIG. 12, as shown in the left diagram of FIG. 12, the region R0 corresponding to the generation time=0 to 2.5 and the disappearance time=2 to 4 of the persistent image is set as the reference region, and −3 is set as the reference value. If the reference region R0 has a value below the reference value, it is determined that the criterion is satisfied. The contributing component, the criterion, the reference region, and the reference value may be set by the user via the input device 5, etc.

[0065] As shown in the right diagram of FIG. 12, the image generation parameter (porosity) x[3] is modulated to obtain four modulation input values ​​in increments of 0.05, such as x[3]=0.6, x[3]=0.65, x[3]=0.7, and x[3]=0.75. For each of the four modulation input values, a third persistent image based on the modulation input value is calculated. The search unit 15 identifies a reference region R1 corresponding to the reference region R0 in the third persistent image, and calculates a representative value included in the reference region R1. The definition of the representative value can be set arbitrarily, but for example, it is preferable to set it to a statistical value of multiple values ​​corresponding to multiple bins included in the reference region R1. The statistical value can be set to a minimum value, a maximum value, an average value, or any other arbitrary statistical amount. The search unit 15 then compares the representative value with the reference value. If the representative value is below the reference value, it is determined that the criterion is met. If the criterion is met, the manufacturing process value corresponding to this representative value is selected as the optimal process value. In the example of Figure 12, since the representative values ​​for x[3] = 0.65 and x[3] = 0.7 are lower than the reference value in the reference region R0, the manufacturing process values ​​corresponding to x[3] = 0.65 and x[3] = 0.7 are selected as the optimal process values.

[0066] In the search process (step S6), the display control unit 16 may display on the display device 6 a first persistent image with a frame representing the reference region R0 superimposed thereon and / or a third persistent image with a frame representing the reference region R1 superimposed thereon, as shown in Fig. 12. At this time, modulation input values ​​that satisfy the criteria and a third pseudo SEM image and persistent image corresponding to the modulation input values ​​may be displayed. This allows the user to confirm the reliability of the selection result of the optimal process value by the search unit 15.

[0067] The modulation of the optimal input value and the search for the optimal process value by the search unit 15 are not limited to one time, but may be repeated multiple times. That is, when there are multiple modulated input values ​​that satisfy the criteria, the search unit 15 can calculate multiple new modulated input values ​​limited to the range of upper and lower limit values ​​of the multiple modulated input values, and select the optimal process value based on these multiple new modulated input values. This makes it possible to select the optimal process value in detail.

[0068] This completes the manufacturing process prediction process.

[0069] The above manufacturing process prediction process is an example, and various elements can be added, deleted and / or modified without departing from the spirit of the invention according to this embodiment.

[0070] As an example, in the above embodiment, the feature amount is a one-dimensional numerical vector obtained by vectorizing the persistence map. However, the feature amount may be a persistence map as long as the distance can be calculated in the optimization by the optimization unit 14.

[0071] As another example, in the above embodiment, the image acquired by the image acquisition unit 11 and the image generated by the image generation unit 12 are gray level images or color scale images, and a binary image is generated by performing preprocessing on the images. However, the present embodiment is not limited to this. For example, the image acquired by the image acquisition unit 11 and the image generated by the image generation unit 12 may be binary images. In this case, the preprocessing unit 131 can be omitted.

[0072] As another example, in the above embodiment, the analysis unit 132 performs persistent homology analysis on the preprocessed image. However, this embodiment is not limited to this. For example, the analysis unit 132 may perform persistent homology analysis on the image acquired by the image acquisition unit 11 and the image generated by the image generation unit 12. In this case, the preprocessing unit 131 can be omitted.

[0073] As another example, in the above embodiment, the search unit 15 searches for the optimal process value. However, the present embodiment is not limited to this. For example, the manufacturing process prediction process may be terminated when the optimization unit 14 obtains the optimal input value.

[0074] As another example, the search unit 15 may estimate the strength characteristic from a pseudo image generated by the generative model. Specifically, the search unit 15 can estimate the strength characteristic from a pseudo image generated by the generative model based on a given input value by using a neural network that inputs an image-based feature amount (e.g., a one-dimensional numerical vector) and outputs a strength characteristic value. The neural network can be generated using supervised machine learning based on a data set including an image-based feature amount related to a certain prediction target and a strength characteristic related to the prediction target. According to the above embodiment, since a given input value can be linked to a manufacturing process value, it is possible to predict a manufacturing process value of a material having a desired strength characteristic without experimentally generating a material specimen. Specifically, the search unit 15 controls the image generation unit 12 to input a modulation input value for an optimal process value to the generative model to generate a pseudo image, controls the extraction unit 13 to extract a one-dimensional numerical vector from the generated pseudo image, and applies the extracted one-dimensional numerical vector to the neural network to calculate a strength characteristic value. The search unit 15 determines whether the calculated strength characteristic value is a desired strength characteristic value. If it is determined that the strength characteristic value is not the desired strength characteristic value, the manufacturing process value is changed and the strength characteristic value is calculated in the same manner, and it is determined whether the calculated strength characteristic value is the desired strength characteristic value. If it is determined that the strength characteristic value is the desired strength characteristic value, the optimal process value corresponding to the calculated strength characteristic value is the manufacturing process value of the prediction target having the desired strength characteristic. According to this embodiment, it is possible to shorten the development time for manufacturing the material specimen, etc., in searching for the manufacturing process value having the desired strength characteristic.

[0075] According to the above embodiment, the manufacturing process prediction device 100 according to the embodiment includes an extraction unit 13, an image generation unit 12, an image acquisition unit 11, and an optimization unit 14. The extraction unit 13 extracts image features based on topological data analysis of the image. The image generation unit 12 generates a pseudo image representing a predicted target of the manufacturing process by applying it to a generative model generated from image generation parameters corresponding to the manufacturing process parameters. The image acquisition unit 11 acquires a first image representing the predicted target. The optimization unit 14 optimizes input values ​​applied to the generative model so that a first feature obtained by applying the first image to the extraction unit 13 coincides with a second feature obtained by applying a pseudo second image generated by inputting input values ​​of the image generation parameters to the generative model.

[0076] Casting materials are made by a process of mixing and kneading inorganic fillers such as silica according to the specifications of each product. The strength properties of casting materials vary depending on the shape and particle size of the inorganic filler or the chemical structure of the resin, and so optimizing the manufacturing process is important. Previously, test pieces with different compounding ratios were experimentally made to find the optimal combination. As there were a huge number of combinations, this required a lot of time and cost.

[0077] The input value of the image generation parameter is optimized so that the first feature amount and the second feature amount coincide with each other. Since the image generation parameter corresponds to the manufacturing process parameter, the optimal input value directly corresponds to the manufacturing process value of the prediction target, which is the subject of the first image. By obtaining such an optimal input value, it becomes possible to link the input value of the image generation parameter with the manufacturing process value. In addition, the characteristic value of a functional material generally depends on the material structure, and it is difficult to construct a model that quantitatively predicts the characteristic value of the prediction target directly from the manufacturing process value of the prediction target. According to the method of this embodiment, since the manufacturing process value of the prediction target can be converted into multidimensional image data, even with a small amount of input data, a generation model that is a regression model with higher accuracy can be constructed. In addition, it becomes possible to shorten the time required for searching for a manufacturing process that realizes the desired strength characteristics.

[0078] According to the above embodiment, the manufacturing process prediction device 100 further includes a search unit 15. The search unit 15 can use the optimal input value to search for the optimal process value of the manufacturing process parameter of the prediction target that satisfies the desired strength characteristic. According to the above configuration, it is possible to virtually search for the manufacturing process value that satisfies the desired strength characteristic by combining the analysis based on the topological data analysis method and the generation model that generates the pseudo SEM image. For example, it is possible to perform screening in which the manufacturing process parameter is indirectly modulated by modulating the image generation parameter corresponding to the manufacturing process parameter. In other words, since it is possible to search for the manufacturing process value without actually producing a material specimen, it is possible to shorten the time required for developing the specimen, etc.

[0079] <Example 1> Using the above manufacturing processor prediction process, virtual screening was performed using the casting material mixing process as an example. A generation model was prepared for sample A, which has a high bending strength. Sample A has a bending strength, which is one of its strength characteristics, that is greater than 160 MPa. The input values ​​of the image generation parameters of the generation model are the parameters x[0] = α and x[1] = β of the exponential distribution defined by equation (1), the number of particle size samples for the particle size distribution x[2] = nbins, the porosity x[3] = η, and the minimum particle size x[4] = r min Then, a random number seed x[5] for generating grains is set, and the optimization unit 14 optimizes six types of image generation parameters x[0] to x[5].

[0080] FIG. 13 shows the real SEM image and the pseudo SEM image of sample A. The pseudo SEM image shown in FIG. 13 is a pseudo image based on the input values ​​(optimal input values) optimized using the first-order persistent image for sample A. The optimal input values ​​obtained at this time were x[0]=0.30164573, x[1]=11.25089249, x[2]=19, x[3]=0.71789715, x[4]=1.06938176, and x[5]=220. However, the calculation range is [a,b]=[252,384], with the vertical length being a and the horizontal length being b, and the SEM image of sample A is also rescaled according to the calculation size of the generative model. As shown in FIG. 13, even though it was optimized using the results of the first-order persistent homology analysis, the distribution of the connected components of the black grain region appears to be almost reproduced.

[0081] Fig. 14 shows the zeroth-order persistence map of the SEM image and the zeroth-order persistence map of the pseudo SEM image of sample A. As shown in Figs. 13 and 14, it can be seen that both the zeroth-order and first-order persistent images reproduce the original images with high accuracy.

[0082] Figure 15 shows the pseudo-SEM images and persistence diagrams when the porosity x[3] is changed. As shown in Figure 15, the porosity x[3] is modulated to [0.6, 0.65, 0.7, 0.75] based on the optimal input values ​​x[0] to x[5]. Since the optimized x[3] is 0.71789···, x[3]=0.6, 0.65, 0.7 corresponds to the case where the porosity is reduced compared to the actual sample A, and x[3]=0.75 corresponds to the case where the porosity is increased compared to the sample A.

[0083] FIG. 16 is a diagram showing the correspondence between the three-dimensional porosity, which is a manufacturing process parameter, and the porosity, which is an image generation parameter. As shown in FIG. 16, there is a linear correspondence between the three-dimensional porosity, which is a manufacturing process parameter, and the porosity, which is an image generation parameter. This allows quantitative evaluation as a specific manufacturing process value. As described above, in this material system, using actual SEM images, it has been found that if many first-order persistent images are distributed near the generation time axis [0:2] and the annihilation time axis [2:4], the sample is classified as having strong bending strength. Therefore, since persistent images are distributed near the generation time axis [0:2] and the annihilation time axis [2:4] for x[3]=0.65, 0.7, and 0.75, it is suggested that these cases are classified as samples with strong bending strength. On the other hand, when x[3] = 0.6, the first persistent image of the pseudo-SEM image does not have a distribution in the characteristic region as described above, suggesting that in an actual sample, if the porosity is reduced to a level equivalent to x[3] = 0.6, the bending strength will be weakened. In this way, by modulating various image generation parameters based on the optimal input values, it is possible to investigate which process conditions contribute to improving bending strength.

[0084] <Example 2> Analysis was also performed on sample B, which has a small bending strength, in the same manner as in Example 1. The optimal input values ​​for Example 2 were x[0] = 0.29160089, x[1] = 12.6051394, x[2] = 17, x[3] = 0.77621263, x[4] = 1.01396426, and x[5] = 200. The calculation range for Example 2 is the same as in Example 1, with the vertical length being a and the horizontal length being b, and [a, b] = [252, 384].

[0085] FIG. 17 shows the pseudo-SEM images and persistence diagrams when the porosity x[3] is changed. As shown in FIG. 17, the porosity x[3] is modulated to [0.6, 0.65, 0.8, 0.85] based on the optimal input values ​​x[0] to x[5]. The optimized value of x[3] is 0.776, so x[3]=0.6, 0.65 corresponds to the case where the porosity is reduced compared to the actual sample B, and x[3]=0.8, 0.85 corresponds to the case where the porosity is increased compared to the sample B. In the case of 0.8, which is close to the condition of sample B, the persistent image has a weak distribution in the region that contributes to the improvement of bending strength (the generation time axis [0:2] and the disappearance time axis [2:4]), suggesting that the sample has a small bending strength. In addition, when the porosity is further increased to x[3]=0.85, it can be seen that it is clearly outside the region that contributes to bending strength. On the other hand, when x[3]=0.65, the persistent image is strongly distributed in the area that contributes to improving bending strength. However, when x[3]=0.6, the persistent image is again distributed in an area far removed from the characteristic area. From this, it can be seen that for sample B, which has low bending strength, it is effective to appropriately reduce the porosity from the current conditions in order to improve the bending strength. In this way, for samples with poor characteristics, it is possible to search for manufacturing process values ​​that move the persistent image in the direction of improving the characteristics.

[0086] Thus, according to the present embodiment, it is possible to provide a manufacturing process prediction apparatus, method, program, and system that can reduce the time required to search for a manufacturing process that achieves desired strength characteristics.

[0087] Although some embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These novel embodiments can be implemented in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their modifications are included in the scope and spirit of the invention, and are included in the scope of the invention and its equivalents described in the claims. [Explanation of symbols]

[0088] 1...processor, 2...ROM, 3...RAM, 4...auxiliary storage device, 5...input device, 6...display device, 7...communication device, 11...image acquisition unit, 12...image generation unit, 13...extraction unit, 14...optimization unit, 15...search unit, 16...display control unit, 100...manufacturing process prediction device, 131...preprocessing unit, 132...analysis unit, 133...conversion unit.

Claims

1. an extraction unit that extracts a feature amount of the image based on a topological data analysis of the image; an image generation unit that generates a pseudo image of a prediction target of a manufacturing process by applying the pseudo image to a generation model that is generated from image generation parameters corresponding to the manufacturing process parameters; An acquisition unit that acquires a first image related to the prediction target; an optimization unit that optimizes the input values ​​applied to the generative model so that a first feature value obtained by applying the first image to the extraction unit coincides with a second feature value obtained by applying a second image generated by inputting input values ​​of the image generation parameters to the generative model to the extraction unit; A manufacturing process prediction apparatus comprising:

2. The manufacturing process prediction apparatus according to claim 1 , further comprising a search unit that uses the optimized input values ​​to search for optimal values ​​of the manufacturing process parameters of the prediction target that satisfy desired strength characteristics.

3. The search unit is modulating the optimized input values ​​to produce a plurality of modulated input values; searching for the optimal value of the manufacturing process parameter based on a plurality of third feature amounts obtained by applying a plurality of third images generated by applying the plurality of modulation input values ​​to the generative model to the extraction unit; The manufacturing process prediction apparatus according to claim 2.

4. 4. The manufacturing process prediction device according to claim 3, wherein when a specific third feature quantity that satisfies a criterion is present among the plurality of third feature quantities, the search unit determines a process value corresponding to the specific third feature quantity as the optimal value.

5. The extraction unit is a pre-processing unit that performs pre-processing on the image to generate a pre-processed image; an analysis unit that performs a topological data analysis on the preprocessed image to calculate an analysis result; A conversion unit that converts the analysis result into numerical data that is the feature amount. The manufacturing process prediction apparatus according to claim 1 .

6. The manufacturing process prediction apparatus according to claim 5 , wherein the preprocessing unit performs a binarization process as the preprocessing.

7. The analysis unit performs persistent homology analysis as the topological data analysis, and calculates a persistence diagram from the preprocessed image; The conversion unit converts the persistence diagram into a one-dimensional numerical vector. The manufacturing process prediction apparatus according to claim 5.

8. The manufacturing process prediction apparatus according to claim 1 , wherein the image generation parameters include one or more parameters.

9. The manufacturing process prediction apparatus according to claim 8 , wherein the one or more parameters include porosity, grain size distribution, number of grain size samples, minimum grain size, and / or grain arrangement.

10. The manufacturing process prediction apparatus according to claim 1 , wherein the optimization unit performs optimization by minimizing a distance between the first feature amount and the second feature amount.

11. 10. The manufacturing process prediction apparatus according to claim 9, wherein the optimization unit uses a Nelder-Mead method, a Powell method, a CG method, a BFGS method, a COBYLA method or a Bayesian optimization method for the optimization.

12. 2. The manufacturing process prediction apparatus according to claim 1, wherein the image is image data generated by photographing an internal structure of the prediction target with an electron microscope.

13. acquiring a first image of a predicted target of the manufacturing process; a first extraction step of extracting a first feature amount of the first image based on a topological data analysis of the first image; an image generation step of generating a pseudo second image related to the prediction target by applying input values ​​of image generation parameters corresponding to manufacturing process parameters to a generation model; a second extraction step of extracting a second feature amount of the second image based on a topological data analysis of the second image; an optimization process for optimizing the input value applied to the generative model such that the first feature amount and the second feature amount coincide with each other; A manufacturing process prediction method comprising:

14. On the computer, an acquisition function for acquiring a first image of a predicted target of a manufacturing process; A first extraction function for extracting a first feature amount of the first image based on a topological data analysis of the first image; an image generation function that generates a pseudo second image related to the prediction target by applying input values ​​of image generation parameters corresponding to manufacturing process parameters to a generation model; A second extraction function for extracting a second feature amount of the second image based on a topological data analysis of the second image; an optimization function that optimizes the input value applied to the generative model so that the first feature amount and the second feature amount match; A manufacturing process prediction program that makes this possible.

15. an extraction unit that extracts a feature amount of the image based on a topological data analysis of the image; an image generation unit that generates a pseudo image of a prediction target of a manufacturing process by applying the pseudo image to a generation model that is generated from image generation parameters corresponding to the manufacturing process parameters; an acquisition unit that acquires a first image representing an internal structure of the prediction target; an optimization unit that optimizes the input values ​​applied to the generative model so that a first feature value obtained by applying the first image to the extraction unit coincides with a second feature value obtained by applying a second image generated by inputting input values ​​of the image generation parameters to the generative model to the extraction unit; A manufacturing process prediction system comprising:

Images