Method for predicting product strain and system thereof

Abstract

Provided are a method for predicting a product strain and a system thereof. A method for predicting a product strain according to one embodiment of the present disclosure is a method performed by a computing device, the method comprising the steps of: acquiring a three-dimensional shape model of a first product; preprocessing the three-dimensional shape model and generating graph data including nodes and edges; inputting an image dataset to a measurement model for predicting shape characteristics and physical property characteristics of an object, the image dataset being obtained by imaging, from multiple angles, the object manufactured on the basis of the three-dimensional shape model, and generating manufacturing data of the first product; comparing graph data and the manufacturing data for each node unit, and generating label data including strain data for each node unit; and fine-tuning a graph data-based graph node classification model by using a strain data set composed of pairs of the graph data and the label data. The step of generating the graph data may include the steps of: generating predictive precision reference information for each unit mesh included in the three-dimensional shape model by using the three-dimensional shape model of the first product and the manufacturing data of the first product; and according to the predictive precision reference information for each unit mesh, applying a first resolution to a first unit mesh that requires first predictive precision, and a second resolution to a second unit mesh that requires second predictive precision.

Description

Method for predicting strain of a product and its system

[0001] The present disclosure relates to a method and system for generating digital twin data of an object. More specifically, the present disclosure relates to a method and system for generating digital twin data for predicting the strain of ceramics using neural radiance fields.

[0002] In addition, the present disclosure relates to a method and system for predicting strain of a product.

[0003] In addition, the present disclosure relates to a method for generating mother mold design data and a system therefor.

[0004] In addition, the present disclosure relates to a modeling method and system using product parameters.

[0005] Digital twin technology, which recreates physical products in a virtual space, is being actively utilized across various industries to facilitate the digital transformation of products. In particular, there is a growing demand for technical solutions to acquire digital twin data that precisely replicates a product's three-dimensional shape and visual characteristics.

[0006] Neural Radiance Fields (NeRF) are attracting attention as a means of generating digital twin data. NeRF is a deep learning-based technology that reconstructs a high-resolution 3D model, including the lighting response of real-world environments, by learning the color and density distribution at the ray level from multi-angle 2D images as input.

[0007] However, for highly glossy, complexly curved objects like ceramics, accurate shape reproduction is difficult using conventional image-based modeling techniques alone, and light reflection can distort physical property information. In particular, the highly reflective surface of glazed ceramics presents technical limitations that hinder visual information alignment.

[0008] A technical problem to be solved through some embodiments of the present disclosure is to provide a method and system for generating digital twin data suitable for an object having highly reflective and curved surface characteristics, such as ceramics.

[0009] Another technical problem to be solved through some embodiments of the present disclosure is to provide a method and system for generating high-precision digital twin data by complementing a modeling structure based on NeRF (Neural Radiance Fields), which is a means for generating digital twin data of an object, and matching shape and material property information of the object.

[0010] Another technical problem to be solved through some embodiments of the present disclosure is to provide a method and system for predicting strain of a product that accurately predicts shrinkage rate and defect rate of the product.

[0011] Another technical problem to be solved through some embodiments of the present disclosure is to provide a method and system for generating mother mold design data, which predicts the shrinkage rate and defect rate of a product, predicts the initial shape of the product based on the shrinkage rate and defect rate, and generates mother mold design data necessary to implement the product based on the predicted initial shape.

[0012] Another technical problem to be solved through some embodiments of the present disclosure is to provide a modeling method and system using product parameters that can dynamically generate a shape model according to a user's design input and analyze and correct the possibility of manufacturing errors in advance through an artificial intelligence-based prediction and optimization process.

[0013] The technical problems of the present disclosure are not limited to the technical problems mentioned above, and other technical problems not mentioned will be clearly understood by those skilled in the art of the present disclosure from the description below.

[0014] According to some embodiments of the present disclosure for solving the above-described technical problem, a method for generating digital twin data of an object, the method being performed by a computing device, comprises: obtaining point cloud data, which is a ground truth value for the appearance of an object detected using a lidar sensor; obtaining an image data set obtained by photographing the object from various angles using a camera device; training a shape property prediction model that models shape properties of the object based on the point cloud data and the image data set; and training a material property prediction model that models material properties of the object based on three-dimensional shape information of the object and the image data set, wherein the three-dimensional shape information of the object may be data generated as an output value of the shape property prediction model.

[0015] In some embodiments, the shape characteristic prediction model may be a 3D Gaussian Splatting (3DGS)-based model, and the material characteristic prediction model may be a Neural Radiance Field (NeRF)-based model.

[0016] In some embodiments, the shape characteristic prediction model may be trained to estimate three-dimensional shape information of the object from the image data set by using the point cloud data as an accurate value for the appearance of the object.

[0017] In some embodiments, the step of training the physical property prediction model may include the step of obtaining visual information about a first point of the object from image data including a viewpoint corresponding to three-dimensional coordinates located at the first point based on three-dimensional shape information about the first point of the object; and the step of projecting the visual information to correspond to the three-dimensional coordinates of the first point, thereby performing spatial alignment between the visual information and the three-dimensional shape information.

[0018] According to some embodiments of the present disclosure for solving the above-described technical problem, a method for generating digital twin data of an object may include a step of obtaining a target image data set by photographing a first object from multiple angles using a camera device, a step of inputting the target image data set into a learned shape characteristic prediction model, and outputting first scanning data regarding three-dimensional shape information of the first object, a step of inputting the target image data set into a learned material property prediction model, and outputting second scanning data including material property information corresponding to each position of the first object, and a step of mapping the second scanning data corresponding to each unit element of the first scanning data, and generating digital twin data of the first object.

[0019] In some embodiments, each unit element of the first scanning data is a point unit on a surface of a mesh or a point cloud, and the step of generating digital twin data of the first object may include a step of storing physical property information of the second scanning data as metadata based on the spatial coordinates of each unit element.

[0020] In some embodiments, the step of generating digital twin data of the first object may include the step of integrating the mapped digital twin data of the first object into a three-dimensional virtual object model.

[0021] According to some embodiments of the present disclosure for solving the above-described technical problem, a system for generating digital twin data of an object includes: a communication interface; a memory in which a computer program is loaded; and one or more processors in which the computer program is executed, wherein the computer program includes instructions for performing the following operations: acquiring point cloud data, which is a ground truth value for the appearance of an object detected using a lidar sensor; acquiring an image data set in which the object is photographed from various angles using a camera device; training a shape property prediction model that models shape properties of the object based on the point cloud data and the image data set; and training a material property prediction model that models material properties of the object based on three-dimensional shape information of the object and the image data set, wherein the three-dimensional shape information of the object may be data generated as an output value of the shape property prediction model.

[0022] According to some embodiments of the present disclosure for solving the above-described technical problem, a system for generating digital twin data of an object includes: a communication interface; a memory in which a computer program is loaded; and one or more processors in which the computer program is executed, wherein the computer program may include instructions for performing an operation of acquiring a target image data set obtained by photographing a first object from multiple angles using a camera device; an operation of inputting the target image data set into a learned shape characteristic prediction model and outputting first scanning data regarding three-dimensional shape information of the first object; an operation of inputting the target image data set into a learned material property prediction model and outputting second scanning data including material property information corresponding to each position of the first object; and an operation of mapping the second scanning data corresponding to each unit element of the first scanning data to the unit element, and generating digital twin data of the first object.

[0023] According to some embodiments of the present disclosure for solving the above-described technical problem, a method for predicting a strain of a product, the method being performed by a computing device, comprises: acquiring a three-dimensional shape model of a first product; preprocessing the three-dimensional shape model and generating graph data including nodes and edges; inputting an image data set obtained by photographing an object manufactured based on the three-dimensional shape model from various angles into a measurement model for predicting shape characteristics and material properties of the object, and generating manufacturing data of the first product; comparing the graph data and the manufacturing data on a node-by-node basis, and generating correct data (Label) including deformation data for each node; and using a strain data set composed of pairs of the graph data and the correct data, the step of generating the graph data comprises: generating prediction accuracy reference information for each mesh included in the three-dimensional shape model using the three-dimensional shape model of the first product and the manufacturing data of the first product; And, according to the prediction accuracy reference information for each unit mesh, a step of applying a first resolution to a first unit mesh requiring a first prediction accuracy and a step of applying a second resolution to a second unit mesh requiring a second prediction accuracy may be included.

[0024] In some embodiments, the step of generating the graph data may include the step of converting the first unit mesh to which the first resolution is applied and the second unit mesh to which the second resolution is applied into a single graph.

[0025] In some embodiments, the step of generating prediction accuracy reference information for each unit mesh may include: inputting the graph data into the graph node classification model and obtaining a predicted strain for the first product; comparing the predicted strain with an actual strain included in manufacturing data for the first product; and correcting the prediction accuracy for a first unit mesh in which an error between the predicted strain and the actual strain is greater than or equal to a reference value as a result of the comparison.

[0026] In some embodiments, the applying step may include: receiving user-defined information distinct from the unit mesh-specific prediction accuracy reference information from the user terminal; and applying a third resolution to the first unit mesh according to the user-defined information.

[0027] In some embodiments, the step of generating the correct answer data may include: a step of aligning the scale of the three-dimensional shape model based on the manufacturing data and generating alignment data for the three-dimensional shape model; a step of projecting the alignment data onto a node corresponding to the three-dimensional shape model and generating projection data for the alignment data; and a step of restoring the projection data and the alignment data to an existing scale and calculating a distance between corresponding nodes between the restored projection data and the restored alignment data.

[0028] In some embodiments, the step of calculating the distance may include the step of displaying the three-dimensional shape model in a visually distinct manner according to the distance between the nodes.

[0029] In some embodiments, the method further comprises inputting a three-dimensional shape model of a second product and a first set of manufacturing variables for the second product into the graph node classification model, and obtaining first strain data for the second product, wherein the three-dimensional shape model of the second product and the first set of manufacturing variables are generated based on design data for the second product, and the first strain data may include values ​​that predict shrinkage and defect rates for the second product.

[0030] In some embodiments, the step of obtaining the first strain data may include the step of overlaying the first strain data onto a three-dimensional shape model of the second product and displaying the same.

[0031] In some embodiments, the step of obtaining the first strain data may include the step of inputting the first strain data into an optimal variable derivation model that has been previously learned using a plurality of manufacturing variable sets when the first strain data exceeds a preset threshold, and calculating a contribution score for each manufacturing variable included in the first manufacturing variable set; and the step of obtaining information about a manufacturing variable whose contribution score is greater than or equal to a reference value.

[0032] In some embodiments, the step of obtaining the first strain data may include the step of generating information regarding a second set of manufacturing variables that minimize shrinkage or defect rate for the second product.

[0033] According to some embodiments of the present disclosure for solving the above-described technical problem, a system for predicting a strain of a product comprises: a communication interface; a memory in which a computer program is loaded; and one or more processors in which the computer program is executed, wherein the computer program comprises: an operation for obtaining a three-dimensional shape model of a first product; an operation for preprocessing the three-dimensional shape model and generating graph data including nodes and edges; an operation for inputting an image data set obtained by photographing an object manufactured based on the three-dimensional shape from various angles into a measurement model for predicting shape characteristics and material properties of the object, and generating manufacturing data of the first product; an operation for comparing the graph data and manufacturing data on a node-by-node basis and generating correct answer data (Label) including deformation data on a node-by-node basis; And instructions for performing an operation of fine-tuning a graph node classification model based on graph data using a strain data set composed of pairs of the graph data and the correct answer data, wherein the operation of generating the graph data may include an operation of generating prediction accuracy reference information for each unit mesh included in the three-dimensional shape model using a three-dimensional shape model of the first product and manufacturing data of the first product; and an operation of applying a first resolution to a first unit mesh for which a first prediction accuracy is required, and applying a second resolution to a second unit mesh for which a second prediction accuracy is required, according to the prediction accuracy reference information for each unit mesh.

[0034] According to some embodiments of the present disclosure for solving the above-described technical problem, a method for generating mother mold design data, the method being performed by a computing device, comprises: obtaining final shape data including information on a three-dimensional shape model of a product generated based on design data for the product; predicting strain data for the product using the final shape data and a set of manufacturing variables for the product; generating initial shape data to which the strain data is reflected; and defining an intaglio shape of a mother mold for the product based on the initial shape data and generating mother mold design data for the product, wherein the generating mother mold design data may include: calculating an error between first mold design data generated based on the initial shape data and the final shape data; and updating the initial shape data such that the absolute value of the error becomes less than or equal to the reference value when the absolute value of the error is greater than or equal to a reference value.

[0035] In some embodiments, the step of generating the mother mold design data may include the step of detecting an undercut generation area in which an undercut occurs based on a demolding direction of the first mold design data, with respect to the first mold design data generated based on the initial shape data; and the step of readjusting the inclination of a mold surface included in a first area included in the first mold design data corresponding to the undercut generation area.

[0036] In some embodiments, the step of generating the mother mold design data may include the step of detecting an undercut generation area in which an undercut occurs based on a demolding direction of the first mold design data, with respect to the first mold design data generated based on the initial shape data; and the step of readjusting a segmentation structure of a mold shape including a first area included in the first mold design data corresponding to the undercut generation area.

[0037] In some embodiments, the method further comprises a step of generating the mother mold by performing 3D printing based on the mother mold design data, wherein the step of generating the mother mold may include a step of detecting an overhang structure based on a gravity direction for an output shape of the mother mold; and a step of automatically generating a support structure for preventing sagging during the 3D printing for the overhang structure.

[0038] In some embodiments, the product includes a first product and a second product, and the step of predicting the strain data may include: obtaining a first manufacturing location variable including location information on a manufacturing space where the first product is placed and a second manufacturing location variable where the second product is placed on the manufacturing space; calculating first strain data for the first product according to the first manufacturing location variable; and calculating second strain data for the second product according to the second manufacturing location variable.

[0039] In some embodiments, the step of generating the mother mold design data may include: generating mold design data for the first product based on first initial shape data reflecting the first strain data; and generating mold design data for the second product based on second initial shape data reflecting the second strain data.

[0040] According to some embodiments of the present disclosure for solving the above-described technical problem, a mother mold design data generation system comprises: a communication interface; a memory in which a computer program is loaded; and one or more processors in which the computer program is executed, wherein the computer program comprises instructions for performing an operation of obtaining final shape data including information on a three-dimensional shape model of a product generated based on design data for the product; an operation of predicting strain data for the product using the final shape data and a set of manufacturing variables for the product; an operation of generating initial shape data in which the strain data is reflected; and an operation of defining an intaglio shape of a mother mold for the product based on the initial shape data and generating mother mold design data for the product, wherein the operation of generating the mother mold design data may include an operation of calculating an error between first mold design data generated based on the initial shape data and the final shape data; and an operation of updating the initial shape data so that the absolute value of the error becomes less than or equal to the reference value when the absolute value of the error is greater than or equal to a reference value.

[0041] According to some embodiments of the present disclosure for solving the above-described technical problem, a method for modeling a product using parameters may include a step of obtaining a first parameter set for modeling a three-dimensional shape model of a product from a user terminal; a step of generating a corrected three-dimensional shape model to which corrections are applied to the three-dimensional shape model of the product based on the first parameter set; a step of inputting the corrected three-dimensional shape model into a pre-trained artificial intelligence model for predicting a defect rate and shrinkage rate of the product, and obtaining strain data for the product, wherein the strain data includes values ​​for predicting the shrinkage rate and the defect rate of the product; and a step of searching for a second parameter set at which the strain data is minimized.

[0042] In some embodiments, the step of generating the corrected three-dimensional shape model may include the step of performing calculations only on parameters among the first parameter set that are different from parameters already applied in the three-dimensional shape model of the product.

[0043] In some embodiments, the step of searching for the second parameter set may include the step of iteratively searching for a parameter combination that minimizes the shrinkage and failure rate for the first parameter set using the strain data and reliability data for the strain data.

[0044] In some embodiments, the method may further include the steps of: extracting a first parameter whose value has been changed from among the first parameter set and the second parameter set; and displaying the change data of the first parameter by overlaying it on the corrected three-dimensional shape model.

[0045] In some embodiments, the step of searching for the second parameter set may include, when a request for selecting the second parameter set is received from the user terminal, generating a recalibrated three-dimensional shape model to which correction has been applied to the corrected three-dimensional shape model based on the second parameter set; inputting the recalibrated three-dimensional shape model into the artificial intelligence model and obtaining first strain data for the product; and searching for a third parameter set in which the first strain data is minimized.

[0046] According to some embodiments of the present disclosure for solving the above-described technical problem, a modeling system using product parameters may include: a communication interface; a memory in which a computer program is loaded; and one or more processors in which the computer program is executed, wherein the computer program may include instructions for performing the following operations: obtaining a first parameter set for modeling a three-dimensional shape model of a product from a user terminal; generating a corrected three-dimensional shape model to which corrections are applied to the three-dimensional shape model of the product based on the first parameter set; inputting the corrected three-dimensional shape model into a pre-trained artificial intelligence model for predicting a defect rate and shrinkage rate of the product, and obtaining strain data for the product, wherein the strain data includes values ​​for predicting the shrinkage rate and the defect rate of the product; and searching for a second parameter set in which the strain data is minimized.

[0047] FIG. 1 is a system configuration diagram for explaining the configuration and operation of a ceramic manufacturing system according to some embodiments of the present disclosure.

[0048] FIG. 2 is a flowchart illustrating the operation of a method for generating digital twin data of an object according to some embodiments of the present disclosure.

[0049] FIG. 3 is a diagram illustrating the operation of a method for learning a measurement model for measuring an object according to some embodiments of the present disclosure.

[0050] FIG. 4 is a detailed flowchart for explaining the detailed operation of a method for generating digital twin data of an object according to some embodiments of the present disclosure, described with reference to FIG. 2.

[0051] FIG. 5 is a flowchart illustrating the operation of a method for generating digital twin data of an object according to some embodiments of the present disclosure.

[0052] FIG. 6 is a diagram illustrating a method for inferring digital twin data of a target object using a measurement model for measuring an object according to some embodiments of the present disclosure.

[0053] FIG. 7 is a flowchart illustrating the operation of a method for predicting strain of a product according to some embodiments of the present disclosure.

[0054] FIG. 8 is a diagram illustrating the operation of a method for generating graph data according to some embodiments of the present disclosure.

[0055] FIG. 9 is a detailed flowchart illustrating the operation of a method for generating graph data according to some embodiments of the present disclosure, described with reference to FIG. 7.

[0056] FIG. 10 is a detailed flowchart illustrating the operation of a method for generating unit mesh-specific prediction accuracy reference information according to some embodiments of the present disclosure, described with reference to FIG. 9.

[0057] FIG. 11 is a detailed flowchart illustrating the operation of a method for generating graph data according to some embodiments of the present disclosure, as described with reference to FIG. 9.

[0058] FIG. 12 is a detailed flowchart illustrating the operation of a method for generating correct answer data according to some embodiments of the present disclosure, as described with reference to FIG. 7.

[0059] FIG. 13 is a diagram illustrating a method for generating correct answer data according to some embodiments of the present disclosure.

[0060] FIG. 14 is a diagram illustrating a method for generating correct answer data according to some embodiments of the present disclosure.

[0061] FIG. 15 is a diagram illustrating a method for generating correct answer data according to some embodiments of the present disclosure.

[0062] FIG. 16 is a flowchart illustrating the operation of a method for predicting strain of a product according to some embodiments of the present disclosure.

[0063] FIG. 17 is a drawing illustrating a method for visually displaying the strain of a product according to some embodiments of the present disclosure.

[0064] FIG. 18 is a detailed flowchart illustrating the operation of a method for obtaining strain data for a product according to some embodiments of the present disclosure, as described with reference to FIG. 16.

[0065] FIG. 19 is an exemplary diagram illustrating a method for obtaining strain data for a product according to some embodiments of the present disclosure.

[0066] FIG. 20 is a flowchart illustrating the operation of a method for generating mother mold design data according to some embodiments of the present disclosure.

[0067] FIG. 21 is a detailed flowchart for explaining the operation of a method for generating mother mold design data according to some embodiments of the present disclosure, described with reference to FIG. 20.

[0068] FIG. 22 is a detailed flowchart illustrating the operation of a method for generating mother mold design data according to some embodiments of the present disclosure, described with reference to FIG. 20.

[0069] FIG. 23 is a detailed flowchart for explaining the operation of a method for generating mother mold design data according to some embodiments of the present disclosure, described with reference to FIG. 20.

[0070] FIG. 24 is a detailed flowchart illustrating the operation of a method for generating mother mold design data according to a manufacturing location of a product according to some embodiments of the present disclosure, as described with reference to FIG. 20.

[0071] FIG. 25 is a flowchart illustrating the operation of a modeling method using product parameters according to some embodiments of the present disclosure.

[0072] FIG. 26 is a flowchart illustrating the operation of a modeling method using product parameters according to some embodiments of the present disclosure.

[0073] FIG. 27 is a detailed flowchart for explaining the operation of a modeling method using product parameters according to some embodiments of the present disclosure, described with reference to FIG. 26.

[0074] FIG. 28 illustrates an exemplary computing device that may implement systems according to some embodiments of the present disclosure.

[0075] Hereinafter, various embodiments of the present disclosure will be described in detail with reference to the attached drawings. The advantages and features of the present disclosure, and methods for achieving them, will become clear with reference to the embodiments described in detail below together with the attached drawings. However, the technical idea of ​​the present disclosure is not limited to the following embodiments and may be implemented in various different forms. The following embodiments are provided only to complete the technical idea of ​​the present disclosure and to fully inform those skilled in the art of the present disclosure of the scope of the present disclosure, and the technical idea of ​​the present disclosure is defined only by the scope of the claims.

[0076] In describing various embodiments of the present disclosure, if it is determined that a detailed description of a related known configuration or function may obscure the gist of the present disclosure, the detailed description will be omitted.

[0077] Unless otherwise defined, the terms (including technical and scientific terms) used in the following examples may be used with meanings commonly understood by those of ordinary skill in the art to which this disclosure pertains; however, this may vary depending on the intentions of engineers working in the relevant field, precedents, the emergence of new technologies, etc. The terminology used in this disclosure is for the purpose of describing the embodiments and is not intended to limit the scope of this disclosure.

[0078] In the following examples, singular expressions include plural concepts unless the context clearly specifies that they are singular. Furthermore, plural expressions include singular concepts unless the context clearly specifies that they are plural.

[0079] In addition, terms such as first, second, A, B, (a), (b), etc. used in the following embodiments are only used to distinguish certain components from other components, and the nature, order, or sequence of the components are not limited by the terms.

[0080] Hereinafter, various embodiments of the present disclosure will be described in detail with reference to the attached drawings.

[0081] Hereinafter, with reference to FIG. 1, the configuration and operation of a ceramic manufacturing system according to some embodiments of the present disclosure will be described. FIG. 1 is a system configuration diagram for explaining the configuration and operation of a ceramic manufacturing system according to some embodiments of the present disclosure.

[0082] Referring to FIG. 1, the ceramic manufacturing system (10) may be configured to include a service server (11) and an internal database (12). However, the scope of the present disclosure is not limited thereto. In some cases, the ceramic manufacturing system (10) may be configured to further include modules / devices / systems not illustrated in FIG. 1. Alternatively, the ceramic manufacturing system (10) may be configured to exclude at least some of the components (11 and 12) illustrated in FIG. 1.

[0083] The user terminal (20) may be a device for the user to input various data required for producing ceramics.

[0084] The service server (20) can obtain point cloud data on the appearance of a ceramic product detected using a lidar sensor from a user terminal (20) and an image data set in which the appearance of the ceramic product is captured from multiple angles using a camera device. The service server (20) can train a measurement model for obtaining digital twin data for the ceramic product using the point cloud data and the image data set. The measurement model can include a shape characteristic prediction model that models the shape characteristics of the ceramic product and a material characteristic prediction model that models the material characteristics of the ceramic product. This will be described in detail with reference to FIGS. 2 to 4.

[0085] The service server (11) can obtain a target image data set obtained by photographing a first object from multiple angles using a camera device from a user terminal (20). The service server (11) can input the target image data set into the learned shape characteristic prediction model and the learned material characteristic prediction model, respectively, and generate digital twin data expressing shape characteristic information and material characteristic information for each unit element of the first object. This will be described in detail with reference to FIGS. 5 and 6.

[0086] According to this embodiment, even if there is no point cloud data on the appearance of a ceramic product detected using a lidar sensor, an artificial intelligence-based model can be created that generates digital twin data that precisely expresses shape characteristic information and material characteristic information of a ceramic product using only images of the appearance of the ceramic product captured from various angles using a camera device.

[0087] Meanwhile, the service server (11) can obtain a three-dimensional shape model of a ceramic product from a user terminal (20). The service server (11) can generate graph data by performing a predetermined preprocessing procedure on the three-dimensional shape model. The service server (11) can input an image data set of a ceramic product manufactured based on the three-dimensional shape model from the user terminal (20) from various angles into a measurement model that predicts shape characteristics and physical properties of the ceramic product, and generate manufacturing data of the ceramic product.

[0088] The service server (11) can compare the graph data and the manufacturing data by node unit and generate correct answer data including deformation data by each node unit. The service server (11) can fine-tune a graph node classification model using a strain data set composed of pairs of the graph data and the correct answer data. The graph node classification model can be an artificial intelligence-based model that predicts the deformation rate (e.g., shrinkage rate, defect rate) of a ceramic product when the graph data is input.

[0089] The service server (11) uses a three-dimensional shape model for a ceramic product and manufacturing data of the ceramic product to generate prediction accuracy reference information for each unit mesh included in the three-dimensional shape model, and can apply different resolutions to all unit meshes according to the prediction accuracy reference information for each unit mesh. The prediction accuracy reference information for each unit mesh may be reference data for determining the level of accuracy required for the strain (or shrinkage, etc.) to be predicted for each mesh unit after dividing the three-dimensional shape model of the ceramic product into mesh units of a certain size.

[0090] This will be described in detail with reference to FIGS. 10 to 15.

[0091] The service server (11) can input a three-dimensional shape model of the second product and a first manufacturing variable set for the second product into the learned graph node classification model, and obtain strain data (e.g., shrinkage rate and defect rate) for the second product. The service server (11) can perform preprocessing on the three-dimensional shape model of the second product, and input graph data for the second product, which is a result of the preprocessing, into the graph node classification model. The first manufacturing variable set may include manufacturing environment variables (e.g., kiln temperature, firing time, humidity, pressure, etc.), material property variables (e.g., clay density, moisture content, etc.), process variables (e.g., molding pressure, drying speed, demolding method, kiln stacking method, etc.) and design variables (e.g., radius of curvature, wall thickness, dimension, angle, etc.) of the second product. However, the scope of the present disclosure is not limited thereto, and the first manufacturing variable set may include various variables that affect the shrinkage rate and defect rate of the second product.

[0092] This will be described in detail with reference to FIGS. 16 to 19.

[0093] According to this embodiment, the strain of a ceramic product can be precisely predicted using a three-dimensional shape model of the ceramic product and images taken from various angles of an actually manufactured ceramic product.

[0094] Meanwhile, the service server (11) can obtain final shape data including information on a three-dimensional shape model of a ceramic product generated based on design data for the ceramic product from the user terminal (20). The service server (11) can input the final shape data and a set of manufacturing variables for the ceramic product into the graph node classification model and predict strain data for the ceramic product. The service server (11) can generate initial shape data reflecting the strain data and, based on the initial shape data, generate mother mold design data for the ceramic product.

[0095] The service server (11) calculates an error between the first mold design data generated based on the initial shape data and the final shape data, and if the absolute value of the error is greater than or equal to a reference value, the initial shape data can be updated so that the absolute value of the error becomes less than or equal to the reference value.

[0096] This is explained with reference to FIGS. 20 to 25.

[0097] According to this embodiment, the strain of a ceramic product can be predicted from the final shape data from which the ceramic product is to be produced, and mother mold design data for manufacturing the ceramic product can be generated by reflecting the predicted strain.

[0098] Meanwhile, the service server (11) receives a first parameter set for modeling the three-dimensional shape of a ceramic product from a user terminal (20), and based on the first parameter set, can search for a second parameter set that minimizes the shrinkage rate and strain rate of the ceramic product.

[0099] This will be explained with reference to FIGS. 26 to 28.

[0100] Each of the components (10 and 11) of the above-described ceramic manufacturing system (10) may be implemented by at least one computing device. For example, all functions of the ceramic manufacturing system (10) may be implemented by a single computing device, or a first function of the ceramic manufacturing system (10) may be implemented by a first computing device and a second function may be implemented by a second computing device. Alternatively, specific functions of the ceramic manufacturing system (10) may be implemented by multiple computing devices.

[0101] A computing device may include any device equipped with computing capabilities; for an example of such a device, see FIG. 29. Since a computing device is a collection of interacting components (e.g., memory, processor, etc.), it may sometimes be referred to as a "computing system." Of course, the term "computing system" can also encompass the concept of a collection of interacting computing devices.

[0102] Meanwhile, in some embodiments, the ceramic manufacturing system (10) and the user terminal (20) may communicate via a network. Here, the network may be implemented as any type of wired / wireless network, such as a local area network (LAN), a wide area network (WAN), a mobile radio communication network, or Wibro (Wireless Broadband Internet).

[0103] For ease of understanding, the following description assumes that all steps / operations of the methods described below are performed on the aforementioned service server (11). Therefore, if the subject of a specific step / operation is omitted, it can be understood that it is performed on the service server (11). However, in an actual environment, some steps / operations of the methods described below may be performed on other computing devices.

[0104] Hereinafter, with reference to FIG. 2, the operation of a method for generating digital twin data of an object according to some embodiments of the present disclosure will be described. FIG. 2 is a flowchart illustrating the operation of a method for generating digital twin data of an object according to some embodiments of the present disclosure.

[0105] Referring to FIG. 2, the service server (11) can obtain point cloud data, which is a ground truth for the appearance of an object detected using a lidar sensor from a user terminal (20) (S1000). In addition, the service server (11) can obtain an image data set obtained by photographing the object from multiple angles using a camera device from the user terminal (20) (S1100). The service server (11) can store the point cloud data and the image data set in an internal database (12).

[0106] The above point cloud data is 3D point cloud data obtained by detecting the external shape of an object using a LiDAR sensor, and each point includes spatial position coordinates (x, y, z) and reflection intensity information for the corresponding position. The service server (11) can express the external shape characteristics of the object, such as size, shape, and curvature, with high precision using the above point cloud data.

[0107] The above image data set is a collection of multiple two-dimensional images obtained by photographing an object from multiple angles using a camera device. Each image contains visual characteristics of the object, such as color, reflected light, and texture, from various viewpoints. Multi-angle photography allows for the acquisition of information about the entire surface of the object.

[0108] The service server (11) can train a shape characteristic prediction model for predicting the shape characteristic of an object based on the acquired point cloud data and image data set (S1200). The shape characteristic prediction model may be a 3D Gaussian Splatting (3DGS)-based model, and can be trained to accurately output 3D shape information (size, curvature, surface slope, etc.) of an object by interpreting visual information included in an image data set using point cloud data collected by a lidar as a ground truth. At this time, the 3DGS-based model can perform a smoothing or reconstruction process by applying a Gaussian (normal distribution) kernel around the point cloud data to visually express the 3D data, that is, the point cloud data, thereby converting it into a continuous 3D shape.

[0109] In addition, the service server (11) can additionally train a material property prediction model that predicts the material property of the object using the 3D shape information and image data set of the object output from the shape property prediction model (S1300). The material property prediction model may be a Neural Radiance Field (NeRF)-based model, and can be trained to predict material property values ​​such as gloss, reflectance, color, and texture for each location of the object based on the shape information and the image data set. At this time, the material property prediction model can precisely learn matching visual characteristic values ​​for each location by aligning (mapping or projecting) the image data set based on the 3D shape information provided from the shape property prediction model.

[0110] The 3D shape information, which is the output value of the above shape characteristic prediction model, may be data including geometric characteristics such as the size, curvature, and normal vector in 3D space regarding the external appearance of the object. The 3D shape information is subsequently used as a reference for performing alignment with visual information in the learning process of the above material property prediction model.

[0111] According to this embodiment, the service server (11) can generate digital twin data of an object that reflects both high-precision three-dimensional shape and visual characteristics by learning a shape characteristic prediction model and a material characteristic prediction model using point cloud data and an image data set collected from a user terminal (20), respectively.

[0112] That is, according to the present embodiment, in order to overcome the technical limitations that make it difficult to scan the exact shape of a porcelain product, such as its size and curvature, the service server (11) can improve the shape scanning performance of a porcelain product by additionally learning a 3DGS-based shape characteristic prediction model included in the surveying model based on the correct answer value data based on lidar scanning.

[0113] Hereinafter, with reference to FIG. 3, the operation of a method for learning a survey model for measuring an object according to some embodiments of the present disclosure will be described. FIG. 3 is a diagram illustrating the operation of a method for learning a survey model for measuring an object according to some embodiments of the present disclosure. FIG. 3 is a diagram illustrating the operation of a method for learning a survey model for measuring an object according to some embodiments of the present disclosure.

[0114] Referring to FIG. 3, the service server (11) can receive point cloud data (40) and an image data set (41) from a user terminal (20). The service server (11) can input the point cloud data (40) and the image data set (41) into a shape characteristic prediction model (30) so that the model can be trained to accurately predict three-dimensional shape information about the appearance of an object. The service server (11) inputs the point cloud data (40) into the shape characteristic prediction model (30), so that the shape characteristic prediction model (30) is trained to accurately predict three-dimensional shape information about the appearance of an object. In addition, the service server (11) can input the image data set (41) into the shape characteristic prediction model (30), thereby improving the performance of three-dimensional shape prediction based on point cloud data (40).

[0115] The shape characteristic prediction model (30) can output three-dimensional shape information (42) of an object as a learning result. The three-dimensional shape information (42) is configured in a standardized data format that reproduces the external appearance of the object, and can be implemented in the form of, for example, a mesh, a point cloud, or a voxel grid. The service server (11) can render a three-dimensional shape model of the object using the three-dimensional shape information (42).

[0116] Meanwhile, the service server (11) can input the 3D shape information (42) and image data set (41) output from the shape characteristic prediction model (30) into the material property prediction model (31). The material property prediction model (31) can extract visual characteristics (gloss, reflectance, color, etc.) from the image data set (41) by matching them to correspond to each spatial coordinate on the 3D shape information (42), and output material property information (43) through this. The material property information (43) can be added to each coordinate or unit element (mesh surface, point, etc.) of the 3D shape model, and through this, digital twin data reflecting the visual characteristics and material properties of the object can be generated.

[0117] According to this embodiment, by integrating point cloud data and image data sets to simultaneously predict shape and material properties, digital twin data of an object with many curves and prominent gloss characteristics (e.g., ceramics) can be generated with high precision.

[0118] Meanwhile, in one embodiment, the shape characteristic prediction model (30) can be trained to estimate the three-dimensional shape information (42) of the object from the image data set (41) using the point cloud data (40) as the correct value for the appearance of the object. The point cloud data (40) includes coordinate (x, y, z) information corresponding to the actual three-dimensional shape of the object, and the shape characteristic prediction model (30) can be trained to estimate the three-dimensional shape information (42) of the object by inputting the image data set (41) based on the point cloud data (40). That is, training of the shape characteristic prediction model (30) can be performed by restoring the shape structure of the object from a plurality of two-dimensional images included in the image data set (41) based on the accurate appearance standard provided by the point cloud data (40).

[0119] According to this embodiment, the shape characteristic prediction model (30) can be trained to estimate the outer shape of an object with high precision using only image data captured from multiple angles.

[0120] Hereinafter, with reference to FIGS. 3 and 4, a method for training a material property prediction model according to some embodiments of the present disclosure will be described. FIG. 4 is a detailed flowchart illustrating detailed operations of a method for generating digital twin data of an object according to some embodiments of the present disclosure, as described with reference to FIG. 2.

[0121] Referring to FIG. 4, the service server (11) can obtain visual information about the first point of the object from image data including a viewpoint corresponding to the three-dimensional coordinates located at the first point based on three-dimensional shape information about the first point of the object (S1310). The service server (11) can set three-dimensional coordinates corresponding to the first point of the object based on the three-dimensional shape information (42) of the object output from the shape characteristic prediction model (30). Thereafter, the service server (11) can obtain visual information about the point from image data (41) including a viewpoint corresponding to the first point. For example, visual characteristics such as color, reflectance, and glossiness of the area where the first point is located are extracted.

[0122] Thereafter, the service server (11) can perform spatial alignment between the visual information and the 3D shape information by projecting the visual information for the first point to correspond to the 3D coordinates of the first point (S1320). By performing step S1320, the service server (11) can enable the physical property prediction model (31) to learn visual characteristics that are precisely aligned for each point in the 3D space, and can generate a measurement model that can highly accurately predict the physical property corresponding to each 3D coordinate.

[0123] According to this embodiment, the shape characteristic prediction model generates three-dimensional shape information that accurately reflects the actual appearance, and the material characteristic prediction model predicts material information based on aligned visual characteristics, thereby constructing a measurement model that can generate high-precision digital twin data in which both shape and material properties are consistent.

[0124] Hereinafter, the operation of a method for generating digital twin data of an object according to some embodiments of the present disclosure will be described with reference to FIG. 5. FIG. 5 is a flowchart illustrating the operation of a method for generating digital twin data of an object according to some embodiments of the present disclosure.

[0125] Referring to FIG. 5, the service server (11) can acquire a target image data set obtained by photographing a first object from various angles from a user terminal (20) (S2000). The target image data set is composed of a plurality of two-dimensional images obtained by photographing the first object from various viewpoints, and each image can include visual characteristics such as color, gloss, and reflectivity of the surface of the first object.

[0126] The service server (11) can input the target image data set into a learned shape characteristic prediction model and output first scanning data regarding the three-dimensional shape information of the first object (S2100). The first scanning data is data generated through the shape characteristic prediction model and can include information expressing the appearance of the first object in the form of coordinates, size, curvature, surface slope, etc. in three-dimensional space. The shape characteristic prediction model can be a model learned in advance based on point cloud data and an image data set so as to be able to precisely estimate the appearance of an object from an image.

[0127] The service server (11) can input the target image data set into a learned material property prediction model and output second scanning data including material property information corresponding to each location of the first object (S2200). The material property prediction model receives the 3D shape information obtained from the first scanning data and the target image data set, and predicts visual or material property values ​​such as gloss, reflectivity, and color corresponding to each 3D coordinate. In this process, the target image data is projected (mapped) to match the 3D shape information, so that material properties can be extracted based on accurate visual information corresponding to each location.

[0128] Thereafter, the service server (11) can integrate the first scanning data and the second scanning data to generate digital twin data of the first object (S2300). More specifically, the service server (11) can map the second scanning data corresponding to each unit element of the first scanning data to generate digital twin data of the first object. The digital twin data is stored in an internal database (12) in a form that reflects both the three-dimensional shape and visual material properties, and can be utilized for various purposes such as virtual environment simulation, manufacturing process optimization, and quality inspection.

[0129] In one embodiment, the first scanning data is data including three-dimensional shape information of the first object, and each unit element may be configured as a surface unit of a mesh or a point unit on a point cloud. In the case of mesh-based, the first scanning data is expressed as a set of polygonal surfaces such as triangles and squares, and each surface includes position coordinates and normal vector information in three-dimensional space. In the case of point cloud-based, the first scanning data is configured as a plurality of points distributed in three-dimensional space, and each point includes position coordinate (x, y, z) information.

[0130] The service server (11) extracts corresponding material property information from the second scanning data based on the spatial coordinates of each unit element (mesh surface or point) of the first scanning data, and stores this as metadata for each unit element, thereby generating digital twin data of the first object. The second scanning data includes material property values ​​such as gloss, reflectance, and color corresponding to each position of the first object, and the service server (11) can add material property information corresponding to the corresponding position by referring to the position coordinates corresponding to each unit element of the first scanning data.

[0131] According to this embodiment, the digital twin data of the first object may be structured data in which each 3D shape unit (mesh surface or point) contains unique material property information as metadata. This data structure enables high-precision analysis and processing by comprehensively reflecting shape and material properties when the digital twin data is utilized in various applications such as rendering, simulation, and quality inspection.

[0132] Therefore, according to this embodiment, it is possible to effectively create integrated digital twin data that accurately reflects the visual and physical characteristics of an object, beyond simple three-dimensional shape reproduction.

[0133] Hereinafter, with reference to FIG. 6, a method for generating digital twin data of an object according to some embodiments of the present disclosure will be described. FIG. 6 is a diagram illustrating a method for inferring digital twin data of a target object using a measurement model for measuring the object according to some embodiments of the present disclosure.

[0134] Referring to FIG. 6, the target image data set (60) is composed of a set of multiple two-dimensional images obtained by photographing the target object from various angles. The target image data set (60) includes the external characteristics and visual characteristics of the target object, and the service server (11) can receive the target image data set (60) from the user terminal (20).

[0135] The service server (11) can input the target image data set (60) into each of the shape characteristic prediction model (30) and the material characteristic prediction model (31).

[0136] The shape characteristic prediction model (30) can output three-dimensional shape information (61) of a target object based on a target image data set (60). The three-dimensional shape information (61) is data that expresses the external appearance of the target object in the form of spatial coordinates, size, curvature, surface slope, etc. The shape characteristic prediction model (30) is a pre-learned model, and may be a model learned to precisely restore the external structure of the object from input image data. The learning method of the shape characteristic prediction model (30) is described in detail with reference to FIGS. 2 and 3.

[0137] Meanwhile, the service server (11) can input the target image data set (60) into the material property prediction model (31), predict visual and material property information corresponding to each location of the target object, and output material property information (62). The material property information (62) can include characteristics such as specularity, reflectance, and color for each location. The material property prediction model (31) can be an artificial intelligence-based model designed to estimate accurate material property corresponding to each location by matching image data to match the three-dimensional shape information (61). The learning method of the material property prediction model (31) has been described in detail with reference to FIGS. 2 to 4.

[0138] The service server (11) can generate digital twin data (63) for a target object by combining the three-dimensional shape information (61) output from the shape characteristic prediction model (30) and the material characteristic information (62) output from the material characteristic prediction model (31). The digital twin data (63) can be configured as an integrated three-dimensional virtual object model that includes not only the three-dimensional shape of the target object but also the material characteristic for each location.

[0139] The digital twin data (63) generated through the above process can be utilized in various industrial fields such as rendering, simulation, manufacturing process optimization, and quality inspection.

[0140] Meanwhile, in one embodiment, the service server (11) can integrate digital twin data generated by mapping the material property information of the second scanning data to each unit element of the first scanning data into a three-dimensional virtual object model. The service server (11) can generate a single integrated three-dimensional virtual object model by combining the three-dimensional shape information of the object expressed by the first scanning data and the material property information mapped from the second scanning data. The three-dimensional virtual object model is configured in a form in which each unit element (mesh surface or point) includes both its own spatial coordinates and material property information.

[0141] For example, for a single mesh face or point, not only the position coordinates (x, y, z) but also visual and physical properties such as specularity, reflectance, and color can be stored.

[0142] The 3D virtual object model integrated in this manner can be utilized not only for visual rendering but also for physical simulations (e.g., thermal analysis, reflectance simulation, lighting response simulation, etc.). The service server (11) outputs the integrated 3D virtual object model, which can then be applied to various application systems such as virtual environments (VR / AR), manufacturing process simulation, and quality analysis.

[0143] According to this embodiment, 3D shape information and material property information are combined and managed in a single object model, allowing digital twin data to be directly applied to various industrial applications without separate post-processing. Furthermore, this embodiment improves data consistency and integration, significantly enhancing processing efficiency and system performance.

[0144] Hereinafter, the operation of a method for predicting strain of a product according to some embodiments of the present disclosure will be described with reference to FIG. 7. FIG. 7 is a flowchart for explaining the operation of a method for predicting strain of a product according to some embodiments of the present disclosure.

[0145] Referring to FIG. 7, the service server (11) can obtain a three-dimensional shape model of a first product from a user terminal (20) (S3000). The three-dimensional shape model is generated based on design data of the first product and may include 3D shape data in a design CAD file or a neutral format (e.g., STL, OBJ, etc.) generated based on the design CAD file. The three-dimensional shape model can serve as a basis for conversion into a graph structure in a subsequent step.

[0146] The service server (11) can preprocess the three-dimensional shape model and generate graph data including nodes and edges (S3100). The graph data refers to graph structure data generated by defining the center point or vertex of each mesh as a node based on a polygonal mesh constituting the surface of the three-dimensional shape model, and defining the connection relationship between adjacent nodes as an edge. The graph data is a conversion result for expressing the shape structure of a product appropriately for an artificial intelligence-based model, and each node can include shape information such as location coordinates, curvature, thickness, etc.

[0147] The service server (11) can input a set of image data obtained by photographing an object manufactured based on the three-dimensional shape model from multiple angles into a surveying model that predicts shape characteristics and material properties of the object, and generate manufacturing data of the first product. The surveying model may include a shape characteristic prediction model and a material characteristic prediction model, as described above with reference to FIGS. 2 and 3. The manufacturing data refers to digital twin data acquired through a 3D scanning device such as a lidar or a structured light scanner for the first product actually manufactured. The digital twin data is as described above with reference to FIG. 2. The service server (11) can infer manufacturing data by inputting image data obtained by photographing the manufactured first product from multiple angles into an AI-based surveying model that predicts shape and material properties.

[0148] Thereafter, the service server (11) can compare the graph data and the manufacturing data by node unit, calculate values ​​such as shrinkage rate or strain rate that occurred at a location corresponding to each node, and generate correct data (Label) including the deformation data by node unit (S3300). For example, the service server (11) can align point cloud data on a digital twin of a manufactured actual first product with nodes on the graph data, and then calculate the distance difference between corresponding points to calculate a deformation vector or scalar shrinkage rate. This value can be stored in the internal database (12) as a label for supervised learning used in subsequent learning.

[0149] The service server (11) can fine-tune a graph node classification model based on graph data using a strain data set composed of pairs of the graph data and the correct answer data (S3400). The graph node classification model is a machine learning or deep learning-based model trained to predict whether shrinkage has occurred, the size of the deformation, the possibility of defects, etc. for each node, and in particular, may include a graph neural network structure such as a graph convolutional network (GCN), a graph attention network (GAT), etc. The service server (11) can perform fine-tuning for domain adaptation based on a pre-trained graph node classification model.

[0150] According to this embodiment, by using a three-dimensional shape model for a product and digital twin data for a manufactured product, the degree of deformation of an actually manufactured product can be predicted from the three-dimensional shape model.

[0151] Hereinafter, a method for generating graph data according to some embodiments of the present disclosure will be described with reference to FIG. 8. FIG. 8 is a diagram illustrating the operation of a method for generating graph data according to some embodiments of the present disclosure.

[0152] Referring to FIG. 8, the service server (11) can obtain a three-dimensional shape model (80) of the first product from the user terminal (20).

[0153] Thereafter, the service server (11) can preprocess the three-dimensional shape model (80) and generate graph data (81) including nodes and edges. The graph data (81) may be graph structure data generated by defining the center point or vertex of each mesh as a node based on a polygonal mesh forming the surface of the three-dimensional shape model (80) and defining the connection relationship between adjacent nodes as an edge.

[0154] Thereafter, the service server (11) can input the graph data and the manufacturing variable set (82) for the product into the fine-tuned graph node classification model, and obtain strain data (83) for the first product. The strain data (83) may be a visual representation of the expected failure rate, as illustrated in FIG. 8. However, the scope of the present disclosure is not limited thereto, and the shrinkage rate and failure rate may be numerically expressed as data for each node.

[0155] Hereinafter, a method for generating graph data according to some embodiments of the present disclosure will be described with reference to FIG. 9. FIG. 9 is a detailed flowchart illustrating the operation of a method for generating graph data according to some embodiments of the present disclosure, as described with reference to FIG. 7.

[0156] Referring to FIG. 9, the service server (11) can generate prediction accuracy reference information for each unit mesh included in the 3D shape model of the first product and manufacturing data of the first product (S3310). The prediction accuracy reference information refers to data indicating the level of prediction accuracy required for each area by determining the relative importance or sensitivity required for deformation prediction for each mesh area. For example, an area where deformation is severe or defects frequently occur can be classified as a first unit mesh that requires high-precision prediction, and an area that is not can be classified as a second unit mesh that allows relatively low precision.

[0157] Accordingly, the service server (11) can configure graph data in such a way that, based on the prediction accuracy reference information for each unit mesh, high-resolution mesh processing (e.g., increasing node density) is performed on the first unit mesh requiring the first prediction accuracy, and low-resolution mesh processing is performed on the second unit mesh requiring the second prediction accuracy. This heterogeneous resolution application method can contribute to improving the prediction performance of key areas while maintaining the computational efficiency of the entire learning model.

[0158] According to this embodiment, by comparing and learning a design-based 3D shape model with actual manufacturing results, the shrinkage or strain rate at each location of the product can be precisely predicted, which can be used to prevent defects in advance or reverse engineer optimal manufacturing conditions.

[0159] Meanwhile, in one embodiment, the service server (11) can convert the first unit mesh applied with a high-resolution first resolution and the second unit mesh applied with a low-resolution second resolution into a single graph. The first resolution has a relatively high node density (high-resolution), and the second resolution has a relatively low node density (low-resolution). This is to reduce unnecessary calculations resulting from meshing the entire product with the same resolution, and to simultaneously secure prediction efficiency and accuracy by concentrating precision on important areas.

[0160] As described above, after defining multiple unit meshes with different resolutions, the service server (11) can integrate the heterogeneous density meshes into a single graph structure. The service server (11) can include multiple nodes constituting the high-resolution mesh and a relatively small number of nodes constituting the low-resolution mesh in a single connected graph, while generating a consistent connection relationship (edge ​​list or adjacency matrix) by reflecting the adjacency between each node.

[0161] For example, the service server (11) may configure edges based on a distance standard set within the same resolution for nodes within the first unit mesh, and may form edges based on an interpolation method or an adjacency standard between nodes located at the boundary between the first unit mesh and the second unit mesh so as to maintain connectivity between cross-resolutions. At this time, in order to maintain structural consistency of the entire graph, overlapping area interface processing may be performed considering the difference in node density between the high-resolution area and the low-resolution area. Through this transformation, a non-uniformly distributed node structure generated from a mesh-based 3D model can be integrated into a single computable graph model.

[0162] Furthermore, the single graph can be directly utilized as input for classification or regression models based on a Graph Neural Network (GNN). This can be provided with attribute information for each node (e.g., location, curvature, thickness, etc.) and distance information or joint characteristics for edges (e.g., presence of a boundary, joint strength, etc.). This allows for the effective construction of predictive input data with an irregular structure that reflects the regional complexity or vulnerability of product shapes.

[0163] According to this embodiment, the single graph, even if generated from heterogeneous resolution meshes, is configured to process the entire product structure as a single, connected computational unit, thereby enabling consistent learning and inference in future graph-based machine learning models. This allows for focused precision in areas requiring high-precision predictions, while simultaneously reducing the overall prediction model's weight.

[0164] Hereinafter, with reference to FIG. 10, a method for generating unit mesh-specific prediction accuracy reference information according to some embodiments of the present disclosure will be described. FIG. 10 is a detailed flowchart illustrating the operation of a method for generating unit mesh-specific prediction accuracy reference information according to some embodiments of the present disclosure, as described with reference to FIG. 9.

[0165] Referring to FIG. 10, the service server (11) inputs graph data into a graph node classification model and can obtain a predicted strain for the first product (S3311). The predicted strain is an output value of the graph-based prediction model and may include strain data (e.g., shrinkage rate or twist direction) for each node.

[0166] Thereafter, the service server (11) can compare the predicted strain with the actual strain included in the manufacturing data of the first product (S3312). The actual strain may refer to ground-truth strain information calculated after alignment from digital twin or 3D scan data. The service server (11) can quantify the difference between the predicted strain and the actual strain for each unit mesh or node to derive an error.

[0167] Thereafter, the service server (11) can correct the prediction accuracy for the first unit mesh in which the error between the predicted strain and the actual strain is greater than or equal to a reference value (S3313). The reference value may be set differently depending on the situation. For example, if the predicted strain of the first unit mesh shows a difference greater than or equal to the reference value from the actual strain, the service server (11) can increase the resolution of the corresponding mesh region, re-segment the mesh, and then design the graph node classification model to selectively relearn only the corresponding region. By having the service server (11) automatically identify only the region in which the error is greater than or equal to the reference value and reflect it back into the model learning structure, an iterative correction structure that gradually improves the overall prediction accuracy can be implemented.

[0168] This correction procedure can be applied continuously, not just in a single epoch, but also within a set number of iterations or iterative learning loops, ultimately leading to convergence in predictive performance. Furthermore, this embodiment allows for localized prediction performance control without increasing the computational burden of the overall model, thereby achieving a balance between efficiency and accuracy.

[0169] Hereinafter, a method for generating graph data according to some embodiments of the present disclosure will be described with reference to FIG. 11. FIG. 11 is a detailed flowchart illustrating the operation of a method for generating graph data according to some embodiments of the present disclosure, as described with reference to FIG. 9.

[0170] Referring to FIG. 11, the service server (11) can receive user-defined information distinct from the prediction accuracy reference information for each unit mesh from the user terminal (20) (S3121). That is, the user can receive prediction accuracy information for the first unit mesh through the user interface provided on the user terminal (20). For example, the user can click on a specific area or draw a selection box on a screen visualizing a 3D shape model displayed on the user terminal (20) to designate the corresponding unit mesh.

[0171] Thereafter, the service server (11) can apply a fixed user-defined resolution (third resolution) to the unit mesh separately from the existing automatic resolution (e.g., first resolution or second resolution) according to the received user-defined information (S3122).

[0172] The user-driven control described above can be applied at the mesh generation stage and can influence the overall predictive model learning and inference performed thereafter. According to this embodiment, by allowing the user to reflect critical design elements that the service server (11) could not automatically determine based on their own judgment, both practical applicability and scalability can be secured. Furthermore, if a user repeatedly selects a specific area, the pattern can be learned and reflected in future automation.

[0173] According to this embodiment, a hybrid mesh precision control system in which automatic control and manual control coexist can be implemented by providing a user selection-based feedback interface in a graph-based prediction precision control technology.

[0174] Hereinafter, a method for generating correct answer data according to some embodiments of the present disclosure will be described with reference to FIG. 12. FIG. 12 is a detailed flowchart illustrating the operation of a method for generating correct answer data according to some embodiments of the present disclosure, as described with reference to FIG. 7.

[0175] Referring to FIG. 12, the service server (11) can align the scale of the three-dimensional shape model of the first product based on the manufacturing data of the first product and generate matching data for the three-dimensional shape model (S3310). The service server (11) can align the geometric center of the three-dimensional shape model with the manufacturing data of the first product and adjust the rotation, translation, and scaling factors in the three-axis directions (X, Y, Z). The service server (11) can use an Iterative Closest Point (ICP) algorithm or an IoU (Maximization of Intersection over Union)-based optimization algorithm for optimal matching. Based on the scale information determined through step S3310, the service server (11) can reduce the three-dimensional shape model or enlarge the manufacturing data to match the sizes of the three-dimensional shape model and the manufacturing data.

[0176] Hereinafter, step S3310 will be described in more detail with reference to FIG. 13. FIG. 13 is a diagram illustrating a method for generating correct answer data according to some embodiments of the present disclosure.

[0177] Referring to FIG. 13, a first screen (130) before the manufacturing data and the 3D shape model are aligned and a second screen (131) after the alignment are shown.

[0178] Referring to the second screen (131), the service server (11) aligns the manufacturing data of the first product shown in the first screen (130) with the geometric center of the three-dimensional shape model of the first product, and adjusts the rotation, translation, and scaling factors in the three-axis direction, thereby aligning the scale of the manufacturing data and the three-dimensional shape model, and can generate matching data for the three-dimensional shape model.

[0179] Again, this is explained with reference to Fig. 12.

[0180] After step S3310, the service server (11) can project the alignment data onto a node corresponding to the three-dimensional shape model and generate projection data for the alignment data (S3320). The service server (11) can generate projection data by projecting each point of the manufacturing data onto the nearest curved surface or segmented surface of the three-dimensional shape model. The projection data may be a point on a curved surface existing in an actual shape, and may be a coordinate value of a corresponding manufacturing result for each node on the graph data.

[0181] Referring to FIG. 14, step S3320 is further described. FIG. 14 is a diagram illustrating a method for generating correct answer data according to some embodiments of the present disclosure.

[0182] Referring to FIG. 14, the service server (11) can generate projection data by projecting the matching data for a three-dimensional shape model, which is reduced based on manufacturing data, onto the nearest curved surface or segmented surface of the three-dimensional shape model.

[0183] Again, this is explained with reference to Fig. 12.

[0184] After step S3320, the service server (11) can restore the projection data and the alignment data to their original scales, and calculate the distance between the corresponding nodes between the restored projection data and the restored alignment data (S3320). At this time, the calculated distance vector is used as a value indicating how much each node has been deformed during the manufacturing process, and based on this, the deformation amount (vector and scalar size) per node can be estimated.

[0185] The processing results are stored in the internal database (12) in the form of correct answer data (Label), and have a structure that includes strain information for each node. Thereafter, the correct answer data can be directly used for graph-based prediction model learning (Fine-tuning), thereby forming a learning dataset.

[0186] In this case, in one embodiment, the service server (11) can display the three-dimensional shape model in a visually distinct manner according to the distance between the nodes. This will be further explained with reference to FIG. 15. FIG. 15 is a diagram illustrating a method for generating correct answer data according to some embodiments of the present disclosure.

[0187] Referring to FIG. 15, the service server (11) can restore the matching data for the 3D shape model and the projection data for the matching data to the original scale, and calculate the distance between the corresponding nodes between the two restored data. Thereafter, the service server (11) can display the distance between the corresponding nodes between the two restored data in a visually distinct manner on the 3D shape model. Referring to FIG. 15, the magnitude of the displacement intensity for each area for all areas of the 3D shape model can be expressed through a user interface such as a user terminal (20).

[0188] According to this embodiment, unlike the Iterative Closet Point (ICP) algorithm that assumes only a simple rigid body transformation, the actual deformation occurring in the ceramics process can be accurately reflected by using an alignment that takes into account the change in scale (shrinkage).

[0189] Furthermore, according to this embodiment, the processes of scale matching, projection, and reverse scale restoration are automated through an optimization algorithm, effectively reducing manual matching time and increasing accuracy. Therefore, this embodiment has the advantage of efficiently automating the alignment process between a 3D shape model and manufacturing data.

[0190] In addition, according to this embodiment, there is an advantage in that the amount of deformation of the product can be displayed as a color map or arrow on a three-dimensional shape model, making it easy to determine how much deformation has occurred in a specific area.

[0191] In addition, according to this embodiment, the quantitative deformation data obtained according to this embodiment can be used as a label for learning of an artificial intelligence-based shrinkage and deformation prediction model or a process optimization model in the future, thereby having the advantage of strengthening the link between digital twin data and the production process.

[0192] Hereinafter, with reference to FIG. 16, a method for predicting strain of a product according to some embodiments of the present disclosure will be described. FIG. 16 is a flowchart illustrating the operation of a method for predicting strain of a product according to some embodiments of the present disclosure.

[0193] Referring to FIG. 16, the service server (11) can input a three-dimensional shape model of a second product and a first manufacturing variable set for the second product into a graph node classification model and obtain first strain data for the second product (S3500). The graph node classification model is as described above with reference to FIG. 7. The three-dimensional shape model of the second product and the first manufacturing variable set are generated based on design data for the second product, and the first strain data may include values ​​that predict shrinkage and defect rates for the second product.

[0194] The three-dimensional shape model of the second product is shape information generated based on design data for the second product, such as a CAD-based model or a mesh-based STL shape file, and is converted into graph data through a preprocessing process. The preprocessed three-dimensional shape is normalized into a graph structure including nodes and edges, and each node represents a characteristic point of the product surface (e.g., center of surface, center of curvature, etc.), and each node can be assigned shape characteristics such as curvature, thickness, and surface texture as features.

[0195] The first set of manufacturing variables may include physical or environmental conditions used in the product manufacturing process. For example, the first set of manufacturing variables may consist of process-specific settings, such as the maximum kiln temperature, firing time, molding pressure, humidity, and drying method. The first manufacturing variables may be encoded in numerical or categorical form and input into a graph node classification model.

[0196] The above graph node classification model is an AI model based on supervised learning that has been pre-trained based on the shape model and manufacturing data of the first product. It can receive the shape and manufacturing conditions of a new product (second product) as input and provide the shrinkage rate and defect rate for each node or unit area of ​​the second product as predicted output values. At this time, the generated predicted value is composed of first strain data, and the first strain data can include the predicted shrinkage rate (%) or probability (score) of the defect rate for each node or unit mesh.

[0197] This allows users to proactively predict areas prone to shrinkage or warpage without requiring separate manufacturing experiments for new products. If necessary, they can improve quality by modifying designs or changing process conditions in these areas. This process flow can reduce the number of product development iterations and lower costs by enabling simulation and virtual tuning before actual manufacturing.

[0198] Meanwhile, in one embodiment, the service server (11) may display the first strain data by overlaying it on a three-dimensional shape model of the second product. This will be further described with reference to FIG. 17. FIG. 17 is a diagram illustrating a method for visually displaying the strain of a product according to some embodiments of the present disclosure.

[0199] Referring to FIG. 17, predicted strain data (e.g., shrinkage or defect rate) for a second product may be overlaid on a three-dimensional shape model (170) of the second product and displayed in a color distinct from the three-dimensional shape model (170) or an area with shrinkage or distortion may be indicated with an arrow. However, the scope of the present disclosure is not limited thereto, and the predicted strain data may be displayed in various ways.

[0200] Hereinafter, with reference to FIG. 18, a method for obtaining strain data for a product according to some embodiments of the present disclosure is described. FIG. 18 is a detailed flowchart illustrating the operation of a method for obtaining strain data for a product according to some embodiments of the present disclosure, as described with reference to FIG. 16.

[0201] Referring to FIG. 18, when the first strain data exceeds a preset threshold, the service server (11) inputs the first strain data into an optimal variable derivation model learned using multiple manufacturing variable sets, and can calculate a contribution score for each manufacturing variable included in the first manufacturing variable set (S3510).

[0202] Thereafter, the service server (11) can obtain information about manufacturing variables whose contribution score is greater than or equal to the reference value (S3520).

[0203] If the first strain data predicting the shrinkage or defect rate of the second product exceeds a preset threshold, the service server (11) can quantitatively analyze the cause of the result from the perspective of a specific manufacturing variable. To this end, the service server (11) can use a pre-learned optimal variable derivation model (or variable influence analysis model) based on multiple sets of manufacturing variables to calculate a contribution score for each manufacturing variable on the strain.

[0204] The above optimal variable derivation model is an artificial intelligence model constructed by learning multiple manufacturing variables including manufacturing environment variables, process variables, material properties variables, design variables, etc. and corresponding strain data, and for example, a decision tree-based regression model, a random forest model, an importance evaluation model based on SHAP analysis, or an influence calculation algorithm of the Permutation Importance method can be used.

[0205] When the service server (11) inputs the first manufacturing variable set and the first strain data for the second product into the optimal variable derivation model, the optimal variable derivation model quantifies the strain contribution of each manufacturing variable to produce a feature importance score. Thereafter, among the produced contribution scores, manufacturing variables whose scores are higher than a preset reference value are extracted, and these variables can be considered as factors that have a major influence on the prediction results of the shrinkage rate or defect rate.

[0206] For example, if the first strain data is 1.8% and exceeds the reference shrinkage (e.g., 1.2%), and the contribution scores for each manufacturing variable such as firing temperature, drying humidity, and molding pressure are calculated, if the firing temperature exceeds the reference value by 0.42, the molding pressure by 0.31, etc., the present invention can present these high-contribution variables to the user or use them as input values ​​for a subsequent process correction or design change algorithm.

[0207] According to this embodiment, it is possible to implement a prediction and optimization system that not only predicts shrinkage rate or defect rate, but also identifies the cause of a problem and derives an improvement plan.

[0208] Meanwhile, in one embodiment, the service server (11) can display information regarding manufacturing variables with contribution scores exceeding a threshold value on a 3D shape model of the second product. The service server (11) can display, using a color heatmap, which areas or shape features of the 3D shape model are related to the manufacturing variables with the highest influence. For example, if the area around a cup handle is determined to have a large influence from the curvature variable, the service server (11) can display the cup handle area in red.

[0209] Hereinafter, FIG. 18 will be described in more detail with reference to FIG. 19. FIG. 19 is an exemplary drawing for explaining a method of obtaining strain data for a product according to some embodiments of the present disclosure.

[0210] Referring to FIG. 19, the first manufacturing variable set to be introduced is assumed to be the first manufacturing variable, the second manufacturing variable, and the third manufacturing variable. The service server (11) can calculate a contribution score for each manufacturing variable. The service server (11) can obtain information regarding the second manufacturing variable (192) whose contribution score is greater than or equal to a reference value. The service server (11) can generate information regarding the second manufacturing variable set (191) in which the shrinkage rate or defect rate for the second product is minimized.

[0211] According to this embodiment, a manufacturing variable that minimizes product strain is derived from among multiple manufacturing variables, and the ceramic manufacturing process can be optimized by adjusting the derived manufacturing variables. Furthermore, according to this embodiment, strain data on shrinkage and strain of a product according to manufacturing variables can be accumulated, and the relationship between each manufacturing variable and the strain data can be learned based on the accumulated data. Thereafter, a manufacturing variable for minimizing product strain is derived based on the above relationship, and the adjustment value of the derived manufacturing variable is predicted, thereby optimizing the ceramic manufacturing process.

[0212] Hereinafter, with reference to FIG. 20, a method for generating mother mold design data according to some embodiments of the present disclosure will be described. FIG. 20 is a flowchart illustrating the operation of a method for generating mother mold design data according to some embodiments of the present disclosure.

[0213] Referring to FIG. 20, the service server (11) can obtain final shape data including information regarding a three-dimensional shape model of a product generated based on design data for the product (S4000). At this time, the final shape data represents the final target shape of the product defined by the user, and may be shape information extracted from a CAD model or intermediate design results.

[0214] Thereafter, the service server (11) can predict strain data for the product using the final shape data and the manufacturing variable set for the product (S4100). The manufacturing variable set may include elements such as temperature, material properties, molding pressure, and cooling time, and the service server (11) can calculate strain (shrinkage) through a machine learning-based model based on these variables and past accumulated data. The service server (11) can input the final shape data and the manufacturing variables for the product into the graph node classification model described above with reference to FIGS. 7 to 19, and obtain strain data for the product as an output value of the graph node classification model. This has been described in detail with reference to FIG. 16.

[0215] Thereafter, the service server (11) can generate initial shape data reflecting the above strain data (S4200). The initial shape data is a shape calculated in reverse direction so that the final shape can accurately reach the target shape when shrunk during molding, and the actual pre-molding state of the product can be digitally expressed. For example, assuming the target shape is F(final) and the initial shape is F(init), the relationship between the target shape and the initial shape can be expressed as follows.

[0216] F(init)=F(final)+strain

[0217] The above strain refers to, for example, strain data predicted by the graph node classification model.

[0218] Thereafter, the service server (11) can define the negative shape of the mother mold for the product based on the initial shape data and generate mother mold design data for the product (S4300). The mother mold design data includes the three-dimensional shape of the mold for forming, and is data that implements the outer shape of the product in negative form.

[0219] Referring to FIG. 21, step S4300 is further described. FIG. 21 is a detailed flowchart for explaining the operation of a method for generating mother mold design data according to some embodiments of the present disclosure, described with reference to FIG. 20.

[0220] Referring to FIG. 21, the service server (11) can calculate the error between the first mold design data generated by adding bubbles to the initial shape data and the final shape data (S4310). This error refers to the difference between the predicted molding result and the target shape.

[0221] If the absolute value of the error is greater than or equal to the reference value, the service server (11) can update the initial shape data so that the absolute value of the error is less than or equal to the reference value (S4311). If the absolute value of the error is greater than or equal to the reference value, the service server (11) repeatedly updates the initial shape data to correct the error so that it is less than or equal to the reference value. Through this iterative optimization process, a more precise mother mold shape can be derived. For example, the service server (11) can minimize the error by applying an iterative technique or numerical optimization (Newton's method, gradient descent method, genetic algorithm, etc.).

[0222] According to this embodiment, product-specific mother mold design data can be automatically generated based on shrinkage and manufacturing variables. That is, as described with reference to FIGS. 7 to 19, in parallel with predicting product strain, it is possible to reverse-engineer the design of the mother mold based on the predicted strain to generate an accurate three-dimensional shape model.

[0223] Hereinafter, a method for generating mother mold design data according to some embodiments of the present disclosure will be described with reference to FIG. 22. FIG. 22 is a detailed flowchart illustrating the operation of a method for generating mother mold design data according to some embodiments of the present disclosure, as described with reference to FIG. 20.

[0224] Referring to FIG. 22, the service server (11) can detect an undercut occurrence area where an undercut occurs based on the demolding direction of the first mold design data generated based on the initial shape data (S4320). The undercut phenomenon refers to a demolding difficulty problem that may occur when removing a product from a mold depending on the shape of the ceramic. The service server (11) can automatically determine a surface or engraved structure formed at an angle smaller than a reference slope by calculating the angle between the normal vector of each mold shape surface and the set demolding direction vector of the mold. The undercut occurrence area corresponds to a shape portion where the product is not easily separated from the mold after being molded or where physical interference may occur during separation.

[0225] Thereafter, the service server (11) can readjust the inclination of the mold surface included in the first area included in the first mold design data corresponding to the undercut occurrence area (S4321). The service server (11) can prevent the undercut phenomenon by identifying the first area corresponding to the detected undercut occurrence area and readjusting the inclination (draft angle) of the mold surface included in the first area.

[0226] For example, if the surface where the undercut occurred is formed with a vertical or reverse slope, the service server (11) can correct the surface so that an inclination angle of about 2 to 5 degrees is applied based on the de-molding direction. This inclination adjustment can be performed using shape editing techniques such as surface offset, surface rotation, or curved interpolation.

[0227] According to this embodiment, the service server (11) can prevent demolding defects and improve the stability of the molding process by performing pre-detection and automatic correction for shapes with a risk of undercut when generating mother mold design data.

[0228] Hereinafter, a method for generating mother mold design data according to some embodiments of the present disclosure will be described with reference to FIG. 23. FIG. 23 is a detailed flowchart illustrating the operation of a method for generating mother mold design data according to some embodiments of the present disclosure, as described with reference to FIG. 20.

[0229] Referring to FIG. 23, the service server (11) can detect an undercut occurrence area where an undercut occurs based on the demolding square of the first mold design data generated based on the initial shape data (S4330). Step S4330 is identical to step S4320 of FIG. 22, and therefore, a detailed description thereof is omitted for convenience of understanding.

[0230] Thereafter, the service server (11) can readjust the division structure of the mold shape included in the first area included in the first mold design data corresponding to the undercut occurrence area (S4331). That is, if the mold was previously divided into two (Core / Cavity) based on a single axis or plane, the mold division design can be automatically changed so that the undercut area does not affect the mold demolding direction by switching to a side (Side Core) or multi-axis (Sliding Core) division structure based on the part including the undercut area.

[0231] For example, if there is a shape that includes an undercut, such as a product handle or an internal latch, demolding is not possible with the existing mold splitting boundary, so the service server (11) can update the mold design data so that the area is split and opened laterally by adding a sliding block structure that includes the area.

[0232] The above-described segmentation structure readjustment can be implemented through a CAD-based parametric modeling function or a shape recognition-based automatic segmentation algorithm, and has the effect of preventing product damage or demoulding failure during mold manufacturing and increasing manufacturing efficiency.

[0233] Meanwhile, in one embodiment, the service server (11) can generate a mother mold for the product by performing 3D printing on the mother mold design data for the product. The service server (11) can convert the designed mother mold shape data into a 3D printing output shape to generate the mother mold.

[0234] In this case, the service server (11) can detect an overhang structure based on the gravity direction of the output shape of the mother mold. The overhang structure refers to a shape portion that floats in the air without support below during 3D printing.

[0235] Since 3D printing operates in a bottom-up layering manner, if a shape is tilted vertically or at a certain reference angle or more without a support, the area may sag or collapse in the direction of gravity during printing. Generally, a surface with an inclination of 45 degrees or more is considered an overhang structure. Accordingly, the service server (11) can calculate the angle between the normal vector of each mold shape surface and the direction of gravity (-Z axis), and if the angle is greater than a reference value, the area can be identified as an overhang area.

[0236] Thereafter, the service server (11) can automatically generate a support structure for the overhang structure to prevent sagging during the 3D printing process. The support structure is a temporary support structure that physically supports the overhang portion, and can be made of a separate material that can be removed or dissolved after printing is completed. For example, a columnar structure that supports the lower end of the protrusion can be formed as a support for a portion that protrudes from the main body, such as a cup handle.

[0237] The generation of support structures is performed automatically by CAD or slicer algorithms, and support shapes, support materials, support densities, etc. can be optimized according to the output purpose.

[0238] The service server (11) can prevent geometric problems that may occur during mother mold output in advance through output settings that include automatic support generation, and improve the precision of the mold shape and the success rate of printing.

[0239] Hereinafter, with reference to FIG. 24, a method for generating mother mold design data according to a manufacturing location of a product according to some embodiments of the present disclosure will be described. FIG. 24 is a detailed flowchart illustrating the operation of a method for generating mother mold design data according to a manufacturing location of a product according to some embodiments of the present disclosure, as described with reference to FIG. 20.

[0240] Referring to FIG. 24, the service server (11) can obtain a first manufacturing location variable including location information on a manufacturing space where a first product is placed, and a second manufacturing location variable where a second product is placed on the manufacturing space (S4110). The first manufacturing location variable is information corresponding to the location where the first product is placed, and the second manufacturing location variable is information corresponding to the location where the second product is placed. The manufacturing location variables are defined based on an absolute coordinate system or a relative placement coordinate system within the manufacturing equipment, and can indicate the locations where products are formed.

[0241] Thereafter, the service server (11) can calculate first strain data for the first product according to the first manufacturing location variable, and can calculate second strain data for the second product according to the second manufacturing location variable (S4111). The service server (11) can predict how the first product and the second product will shrink / deform during actual molding according to location-dependent environmental variables such as heat distribution, cooling speed, and pressure conditions at the location where each of the first and second products is manufactured in the manufacturing space.

[0242] Thereafter, the service server (11) can generate mold design data for the first product based on the first initial shape data reflecting the first strain data (S4340). Similarly, the service server (11) can generate mold design data for the second product based on the second initial shape data reflecting the second strain data (S4341).

[0243] According to this embodiment, even when a product is composed of multiple products (e.g., a first product and a second product), the strain rate can be independently predicted for each product, and mold design data can be generated based on the results, tailored to the location. In other words, according to this embodiment, even in a process environment where multiple products are manufactured simultaneously, shrinkage characteristics can be independently reflected for each product by considering the manufacturing location information for each product. Mold design data generation based on this can contribute to improved molding precision.

[0244] Hereinafter, with reference to FIG. 25, a modeling method using product parameters according to some embodiments of the present disclosure will be described. FIG. 25 is a flowchart illustrating the operation of a modeling method using product parameters according to some embodiments of the present disclosure.

[0245] Referring to FIG. 25, the service server (11) can obtain a first parameter set for modeling a three-dimensional shape model of a product from a user terminal (20) (S5000). The first parameter set comprises various parametric variables that influence design, such as the degree, curvature, dimensions, and number of curves of a curve. The user can directly input the parameter values ​​through the user terminal (20) or modify them based on an existing model.

[0246] Thereafter, the service server (11) can generate a corrected 3D shape model with corrections applied to the 3D shape model of the product based on the first parameter set (S5100). At this time, the corrections can be performed according to mathematical definitions such as splines, NURBS curves, or polygon meshes, and the calculations can be performed incrementally instead of recalculating the entire shape when parameters change. This allows for efficient reflection of changes even for complex shapes.

[0247] Thereafter, the service server (11) inputs the 3D shape model into a pre-trained artificial intelligence model for predicting the product's defect rate and shrinkage rate, and can obtain strain data for the product (S5200). At this time, the strain data includes values ​​that predict the shrinkage rate and defect rate for the product. The artificial intelligence model may be a graph node classification model described with reference to FIGS. 7 to 19. The artificial intelligence model can predict the possibility of deformation according to the design shape based on past manufacturing cases or simulation data.

[0248] Thereafter, the service server (11) can search for a second parameter set that minimizes the above strain data (S5300). This search can be efficiently performed using an iterative search technique (e.g., Bayesian optimization or selective sampling technique) rather than a full-scale search method, and optimal design conditions can be derived by deriving parameter combinations that satisfy highly reliable strain minimum conditions.

[0249] According to the present embodiment, a parametric design support technology can be implemented that dynamically generates a shape model according to a user's design input and analyzes and corrects the possibility of manufacturing errors in advance through an artificial intelligence-based prediction and optimization process. That is, assuming that a product is generated through parameter-based modeling in a digital 3D design environment, the shape of the product can be designed to change according to input variables (e.g., degree of curve, dimension, number, etc.). The user can directly adjust the parameters through a user interface provided on the user terminal (20), and can visualize the prediction results according to various correction directions in real time using an artificial intelligence-based defect rate and shrinkage rate prediction model, and derive an optimal correction direction.

[0250] Meanwhile, in one embodiment, the service server (11) may perform calculations only on parameters that have changed compared to the parameters of a previously applied three-dimensional shape model, rather than performing calculations on all parameters included in the first parameter set in bulk.

[0251] For example, if the first parameter set includes curve degree, curvature, thickness, length, etc., and the user changes only the thickness, the service server (11) calculates new geometry only for shape elements related to thickness, and can reuse existing calculation results for shapes according to other unchanged parameters.

[0252] This method can be implemented as an incremental update method or a cache-based reuse method, and can prevent unnecessary recalculation of the entire shape and effectively reduce the computational cost required for complex modeling operations.

[0253] Therefore, according to the present embodiment, faster and more resource-efficient parametric model generation is possible by avoiding redundant calculations for all parameters and selectively performing operations only on changed parameters.

[0254] Meanwhile, in one embodiment, the service server (11) may perform an iterative search by considering not only the strain data acquired in the preceding step S5200 but also the reliability data for the strain data when performing the search in step S5300 of FIG. 25. The reliability data is information that quantifies the error range, prediction variance, or probability-based reliability of the strain prediction value produced by the artificial intelligence model for a specific parameter combination, and is used as an indicator of how confident the artificial intelligence model is about a specific combination.

[0255] The service server (11) can repeatedly and selectively sample parameter combinations expected to have a large information gain based on the prediction results for the initial parameter combinations and the corresponding reliability data, and perform a search centered on combinations whose predicted values ​​are likely to minimize shrinkage and defect rates.

[0256] This iterative search can be implemented using Bayesian Optimization or other probabilistic iterative search methods, which can improve computational efficiency by considering both prediction accuracy and confidence rather than exhaustively searching the entire parameter space, and can converge on parameter combinations that minimize shrinkage and defect rates using fewer resources.

[0257] That is, according to the present embodiment, ceramic design optimization can be achieved more efficiently by first exploring parameter combinations with high uncertainty by using not only strain data but also reliability data thereof.

[0258] Hereinafter, with reference to FIG. 26, a modeling method using product parameters according to some embodiments of the present disclosure will be described. FIG. 26 is a flowchart illustrating the operation of a modeling method using product parameters according to some embodiments of the present disclosure.

[0259] Referring to FIG. 26, after the second parameter set is derived in step S5300 of FIG. 25, the service server (11) can extract the first parameter whose value has been changed in the first parameter set from among the first parameter set and the second parameter set (S5400). That is, the service server (11) filters only the parameters that have actually changed among all parameters, thereby enabling the visual focus on the parameters that have a key impact on the change.

[0260] Thereafter, the service server (11) can display the change data of the first parameter by overlaying it on the corrected three-dimensional shape model (S5500). The service server (11) can visualize it in the form of color, transparency, or animation, and the user can intuitively confirm the location, scope, directionality, etc. of the design change through the user interface of the user terminal (20).

[0261] For example, areas where curvature has changed may be shaded red, areas where thickness has increased may be indicated by a bold border, and this can provide visual feedback to the user about the change history, thereby supporting design decision-making.

[0262] According to this embodiment, by providing a visual overlay of parametric design changes based on a geometric model rather than simple numerical information, the effects of design changes can be intuitively understood, thereby improving the reliability and efficiency of the iterative optimization process.

[0263] Hereinafter, with reference to FIG. 27, a modeling method using product parameters according to some embodiments of the present disclosure will be described. FIG. 27 is a detailed flowchart illustrating the operation of a modeling method using product parameters according to some embodiments of the present disclosure, as described with reference to FIG. 26.

[0264] Referring to FIG. 27, after the second parameter set is derived in step S5300 of FIG. 25, the service server (11) can determine whether a selection request for the second parameter set is received from the user terminal (10) (S5310). The user can accept the second parameter set automatically derived by the system or request additional review of the corresponding conditions.

[0265] If the service server (11) does not receive a request for selecting the second parameter set from the user terminal (20), the user may assume that the existing first parameter set is the optimal parameter combination for modeling the three-dimensional shape model of the product. In this case, the service server (11) may determine that the existing first parameter set is the optimal parameter combination for modeling the three-dimensional shape model of the product, and may terminate the process without searching for any other parameter sets.

[0266] On the other hand, if a request for selection of re-parameters is received from the user terminal (20), the service server (11) can generate a re-calibrated 3D shape model with correction applied based on the second parameter set (S5320). The service server (11) can re-generate a 3D shape model reflecting the previous design state, or incrementally update the 3D shape model by reflecting only some of the changed parameters.

[0267] Thereafter, the service server (11) inputs the recalibrated 3D shape model into an artificial intelligence model (e.g., a graph node classification model) and can obtain new first strain data for the product (S5330). Unlike existing strain data, the first strain data is a shrinkage rate and defect rate prediction value based on the design results based on the second parameter.

[0268] Thereafter, the service server (11) can search for a third parameter set that minimizes the first strain data (S5340). Similar to the search method described above in step S5300, the service server (11) can newly calculate an optimal parameter combination that is most suitable for the updated conditions by utilizing an iterative optimization algorithm (e.g., Bayesian search, information gain-based sampling, etc.). Thereafter, the service server (11) can store the third parameter set and the strain data according to the third parameter set in an internal database (12) and utilize it for reinforcement learning of a graph node classification model that predicts future strain data.

[0269] According to this embodiment, by realizing a structure that dynamically repeats the redesign and prediction-optimization loop according to the user's design change or re-examination request, more precise product quality improvement and user-tailored optimization can be realized.

[0270] Hereinafter, an exemplary computing device (1000) capable of implementing a system according to some embodiments of the present disclosure will be described with reference to FIG. 28.

[0271] FIG. 28 illustrates an exemplary computing device capable of implementing systems according to some embodiments of the present disclosure. The computing device (1000) of FIG. 28 may include one or more processors (1100), a system bus (1600), a communication interface (1200), a memory (1400) for loading a computer program (1500) executed by the processor (1100), and a storage (1300) for storing the computer program (1500). For example, the computing device of FIG. 28 may be the service server (11) described with reference to FIG. 1.

[0272] The processor (1100) controls the overall operation of each component of the computing device (1000). The processor (1100) may perform operations for at least one application or program for executing methods / operations according to various embodiments of the present disclosure. The memory (1400) stores various data, commands, and / or information. The memory (1400) may load one or more computer programs (1500) from the storage (1300) to execute methods / operations according to various embodiments of the present disclosure. The storage (1300) may non-temporarily store one or more computer programs (1500). The computer program (1500) may include one or more instructions implementing methods / operations according to various embodiments of the present disclosure. When the computer program (1500) is loaded into the memory (1400), the processor (1100) can perform methods / operations according to various embodiments of the present disclosure by executing one or more instructions.

[0273] The computer program (1500) may be a program related to a digital twin data generation system of an object.

[0274] In some embodiments, the computer program (1500) includes instructions for performing the following operations: obtaining point cloud data, which is a ground truth for the appearance of an object detected using a lidar sensor; obtaining an image data set in which the object is photographed from various angles using a camera device; training a shape property prediction model that models shape properties of the object based on the point cloud data and the image data set; and training a material property prediction model that models material properties of the object based on three-dimensional shape information of the object and the image data set, wherein the three-dimensional shape information of the object may be data generated as an output value of the shape property prediction model.

[0275] In some embodiments, the computer program (1500) may include instructions for performing an operation of acquiring a target image data set by photographing a first object from multiple angles using a camera device, an operation of inputting the target image data set into a learned shape characteristic prediction model, and an operation of outputting first scanning data regarding three-dimensional shape information of the first object, an operation of inputting the target image data set into a learned material property prediction model, and an operation of outputting second scanning data including material property information corresponding to each position of the first object, and an operation of mapping the second scanning data corresponding to each unit element of the first scanning data, and generating digital twin data of the first object.

[0276] The computer program (1500) may be a program related to a strain prediction system of a product.

[0277] In some embodiments, the computer program (1500) includes instructions for performing an operation of obtaining a three-dimensional shape model of a first product, an operation of preprocessing the three-dimensional shape model and generating graph data including nodes and edges, an operation of inputting an image data set obtained by photographing an object manufactured based on the three-dimensional shape from various angles into a measurement model for predicting shape characteristics and material properties of the object, an operation of generating manufacturing data of the first product, an operation of comparing the graph data and manufacturing data for each node and generating correct data (Label) including deformation data for each node, and an operation of fine-tuning a graph node classification model based on graph data using a strain data set composed of pairs of the graph data and the correct data, wherein the operation of generating the graph data includes an operation of generating prediction accuracy reference information for each unit mesh included in the three-dimensional shape model using the three-dimensional shape model of the first product and the manufacturing data of the first product, and an operation of performing a first prediction accuracy reference information for each unit mesh based on the prediction accuracy reference information for each unit mesh for a first unit mesh. The operation may include applying a resolution and applying a second resolution to a second unit mesh for which a second prediction accuracy is required.

[0278] The computer program (1500) may be a program related to a mother mold design data generation system.

[0279] In some embodiments, the computer program (1500) includes instructions for performing an operation of obtaining final shape data including information about a three-dimensional shape model of the product generated based on design data for the product, an operation of predicting strain data for the product using the final shape data and a set of manufacturing variables for the product, an operation of generating initial shape data in which the strain data is reflected, and an operation of defining an intaglio shape of a mother mold for the product based on the initial shape data and generating mother mold design data for the product, wherein the operation of generating the mother mold design data may include an operation of calculating an error between first mold design data generated based on the initial shape data and the final shape data, and an operation of updating the initial shape data such that the absolute value of the error becomes less than or equal to the reference value when the absolute value of the error is greater than or equal to a reference value.

[0280] The computer program (1500) may be a program for a modeling system using product parameters.

[0281] In some embodiments, the computer program (1500) may include instructions for performing an operation of obtaining a first parameter set for modeling a three-dimensional shape model of a product from a user terminal, an operation of generating a corrected three-dimensional shape model to which corrections are applied to the three-dimensional shape model of the product based on the first parameter set, an operation of inputting the corrected three-dimensional shape model into a pre-trained artificial intelligence model for predicting a defect rate and a shrinkage rate of the product, an operation of obtaining strain data for the product, the strain data including values ​​for predicting the shrinkage rate and the defect rate of the product, and an operation of searching for a second parameter set at which the strain data is minimized.

[0282] In some embodiments, the computing system described with reference to FIG. 28 may be configured using one or more physical servers included in a server farm based on cloud technologies such as virtual machines. In this case, at least some of the components illustrated in FIG. 28, such as the processor (1100), memory (1400), and storage (1300), may be virtual hardware, and the communication interface (1200) may also be configured as a virtualized networking element such as a virtual switch.

[0283] Although the operations are depicted in a particular order in the drawings, this should not be construed as requiring that the operations be performed in the particular order depicted, or in any sequential order, or that all depicted operations must be performed to achieve the desired results. In certain situations, multitasking and parallel processing may be advantageous.

[0284] Although the embodiments of the present disclosure have been described with reference to the attached drawings, those skilled in the art will appreciate that the present disclosure can be implemented in other specific forms without changing the technical concepts or essential features thereof. Therefore, it should be understood that the embodiments described above are exemplary in all respects and not restrictive. The scope of protection of the present invention should be interpreted by the following claims, and all technical ideas within a scope equivalent thereto should be interpreted as being included within the scope of the technical ideas defined by the present disclosure.

Claims

1. In a method performed by a computing device, A step of obtaining a three-dimensional shape model of the first product; A step of preprocessing the above 3D shape model and generating graph data including nodes and edges; A step of inputting an image data set of an object manufactured based on the above 3D shape model, which is photographed from multiple angles, into a measurement model that predicts shape characteristics and material properties of the object, and generating manufacturing data of the first product; A step of comparing the graph data and the manufacturing data by node unit and generating correct data (Label) including the deformation data by node unit; and A step of fine-tuning a graph node classification model based on graph data using a strain data set composed of the above graph data and the above correct answer data pairs, The steps for generating the above graph data are: A step of generating prediction accuracy reference information for each unit mesh included in the three-dimensional shape model using the three-dimensional shape model of the first product and manufacturing data of the first product; and A step of applying a first resolution to a first unit mesh requiring a first prediction accuracy and applying a second resolution to a second unit mesh requiring a second prediction accuracy, according to the prediction accuracy reference information for each unit mesh, A method for predicting the strain of a product.

2. In paragraph 1, The steps for generating the above graph data are: A step of converting the first unit mesh to which the first resolution is applied and the second unit mesh to which the second resolution is applied into a single graph, A method for predicting the strain of a product.

3. In paragraph 1, The step of generating the prediction accuracy reference information for each unit mesh is as follows: A step of inputting the above graph data into the graph node classification model and obtaining a predicted strain rate for the first product; A step of comparing the predicted strain with the actual strain included in the manufacturing data of the first product; and As a result of the above comparison, a step of correcting the prediction accuracy for the first unit mesh in which the error between the predicted strain and the actual strain is greater than or equal to a reference value is included. A method for predicting the strain of a product.

4. In paragraph 1, The steps to apply the above are: A step of receiving user-defined information distinct from the prediction accuracy reference information for each unit mesh from a user terminal; and According to the above user-defined information, a step of applying a third resolution to the first unit mesh is included. A method for predicting the strain of a product.

5. In paragraph 1, The steps for generating the above correct answer data are: A step of aligning the scale of the three-dimensional shape model based on the above manufacturing data and generating alignment data for the three-dimensional shape model; A step of projecting the above alignment data onto a node corresponding to the three-dimensional shape model and generating projection data for the above alignment data; and A step of restoring the projection data and the alignment data to the original scale, and calculating the distance between corresponding nodes between the restored projection data and the restored alignment data, A method for predicting the strain of a product.

6. In paragraph 5, The steps for calculating the above distance are: Including a step of displaying in a visually distinct manner on the three-dimensional shape model according to the distance between the nodes. A method for predicting the strain of a product.

7. In paragraph 1, Further comprising a step of inputting a three-dimensional shape model of a second product and a first manufacturing variable set for the second product into the graph node classification model, and obtaining first strain data for the second product, The three-dimensional shape model of the second product and the first manufacturing variable set are, It was created based on the design data for the above second product, The above first strain data is, Including values ​​predicting shrinkage and defect rates for the second product, A method for predicting the strain of a product.

8. In paragraph 7, The step of obtaining the above first strain data is: A step of overlaying the first strain data onto a three-dimensional shape model of the second product and displaying the same, A method for predicting the strain of a product.

9. In paragraph 7, The step of obtaining the above first strain data is: When the first strain data exceeds a preset threshold, a step of inputting the first strain data into an optimal variable derivation model learned using multiple manufacturing variable sets, and calculating a contribution score for each manufacturing variable included in the first manufacturing variable set; and A step of obtaining information about a manufacturing variable whose contribution score is greater than or equal to a reference value, A method for predicting the strain of a product.

10. In paragraph 7, The step of obtaining the above first strain data is: A step of generating information about a second set of manufacturing variables that minimizes shrinkage or defect rate for the second product, A method for predicting the strain of a product.

11. Communication interface; Memory into which computer programs are loaded; and Including one or more processors on which the above computer program is executed, The above computer program, An action of obtaining a three-dimensional shape model of the first product; An operation of preprocessing the above 3D shape model and generating graph data including nodes and edges; An operation of inputting an image data set of an object manufactured based on the above three-dimensional shape, which is photographed from multiple angles, into a measurement model that predicts shape characteristics and material properties of the object, and generating manufacturing data of the first product; An operation of comparing the above graph data and manufacturing data by node unit and generating correct data (Label) including the transformation data by node unit; and Includes instructions for performing an operation of fine-tuning a graph node classification model based on graph data using a strain data set composed of the above graph data and the above correct answer data pairs, The action of generating the above graph data is: An operation of generating prediction accuracy reference information for each unit mesh included in the three-dimensional shape model using the three-dimensional shape model of the first product and manufacturing data of the first product; and An operation including applying a first resolution to a first unit mesh requiring a first prediction accuracy and applying a second resolution to a second unit mesh requiring a second prediction accuracy, according to the prediction accuracy reference information for each unit mesh. Product strain prediction system.

Images