Artificial Intelligence-based Generative Design Method, Apparatus, and Computer Program

Abstract

Embodiments of the present invention are artificial intelligence-based generative design methods executed by a computing device, including a step of generating a new design for a product, a step of predicting the performance of the new design, and a step of obtaining an optimal design among the new designs, and can provide an artificial intelligence-based generative design method.

Description

Technical Field

[0001] The present invention relates to an artificial intelligence-based generative design method, apparatus, and computer program.

Background Art

[0002] Generative design is a design technology proposed by artificial intelligence software. The generative design technology is characterized in that a computer automatically generates an optimal design plan that meets customer requirements from the initial stage of product development.

[0003] In an environment where data is insufficient, it is essential to utilize synthetic data to train an artificial intelligence model. Synthetic data refers to data artificially generated through various automation technologies such as simulations.

[0004] However, in the case of 3D engineering data, since synthetic data with physical meaning must be generated while understanding the engineering domain that requires technical know-how compared to other data, it is difficult to construct a large amount of 3D engineering data that can be used for artificial intelligence learning in actual product development sites.

[0005] In particular, it is a design technology mainly used for products where functions and performance are important. By introducing this method, various customer needs can be met, the review time of prototypes can be saved, and the delivery time can be shortened.

[0006] Conventional generative design technologies mainly utilize topology optimization technology, but topology optimization technology has problems such as being unable to utilize historical design data, being unable to create designs with excellent external aesthetics, and creating designs that are irrelevant to customer preferences.

Prior Art Documents

Patent Documents

[0007]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0008] Embodiments of the present invention aim to provide an AI-based generative design method, apparatus, and computer program that can generate product designs based on artificial intelligence, analyze the performance of the designs, and provide a design performance prediction and optimal design model learned from the learning data for the performance of the designs.

[0009] Embodiments of the present invention aim to provide an AI-based generative design method, apparatus, and computer program that can automatically generate multiple product design proposals applicable to a product based on artificial intelligence using a small number of reference designs.

[0010] Embodiments of the present invention aim to provide an AI-based generative design method, apparatus, and computer program that can model and provide an optimal design proposal for a product corresponding to the target performance of the product required by a user, and provide a solution result obtained by analyzing the performance of the optimal design proposal.

[0011] One embodiment of the present invention aims to provide an optimized generative design method, apparatus, and computer program based on reference data and implicit neural representations that can provide a Deep Generative Design service capable of generating a new design using a small amount of reference data and training the generated new design on an artificial intelligence model.

[0012] One embodiment of the present invention aims to provide an optimized generative design method, apparatus, and computer program based on reference data and implicit neural representations that can simultaneously solve the problems of data shortage and shortage of AI experts occurring at the product development site in the manufacturing industry.

[0013] One embodiment of the present invention aims to effectively explore new design proposals that were not considered in the past during the conceptual design stage and propose an optimized design considering various performances at a speed close to real-time, and provides an optimized generative design method, apparatus, and computer program based on reference data and implicit neural representations.

Means for Solving the Problems

[0014] An embodiment of the present invention is an artificial intelligence-based generative design method executed by a computing device, including a step of generating a new design for a product, a step of predicting the performance of the new design, and a step of obtaining an optimal design among the new designs, and can provide an artificial intelligence-based generative design method.

[0015] An embodiment of the present invention is that the step of generating a new design for the product includes a step of applying phase optimization to first reference data to obtain a first new design, a step of applying parametric design to the first reference data to obtain a second new design, a step of applying the parametric design to the first new design to further obtain the second new design, and a step of applying the phase optimization to the second new design to further obtain the first new design, and can provide an artificial intelligence-based generative design method.

[0016] An embodiment of the present invention is that the step of generating a new design for the product includes a step of mapping a first new design and a second new design to one latent space using a design generation model, and a step of using a third new design generated by the design generation model as the second reference data, and the first new design and the second new design can have different representation methods and formats, and can provide an artificial intelligence-based generative design method.

[0017] In an embodiment of the present invention, the process of predicting the performance of the new design includes: analyzing the new design using a performance analysis module; learning a design performance prediction model based on the analysis results of the new design; predicting the performance of the new design based on the design performance prediction model; providing the learned design performance prediction model when the predicted design performance meets a preset reference value; and when the predicted design performance does not meet the preset reference value, collecting additional analysis results for the new design through adaptive sampling and re-learning the design performance prediction model based on the collected additional analysis results. An artificial intelligence-based generative design method can be provided that includes these steps.

[0018] In an embodiment of the present invention, the process of obtaining the optimal design among the new designs includes: obtaining a first target performance regarding the performance information of the target product desired by the user; using the first target performance as the input of an inverse design model to obtain a first optimal design solution, and learning the first optimal design solution as the input of a forward design model so that the same second target performance as the first target performance is predicted; and obtaining a second optimal design solution by re-learning the design performance prediction model through additional analysis results of additional sampled design solutions selected by adaptive sampling considering the uncertainty of the first optimal design. An artificial intelligence-based generative design method can be provided that includes these steps.

[0019] An embodiment of the present invention can further provide an artificial intelligence-based generative design method that further includes: evaluating the optimal design; accumulating part of the optimal design as design data; analyzing the design data; and recommending the design data.

[0020] Embodiments of the present invention can provide an AI-based generative design method that includes a step of obtaining a first optimal design plan using the first target performance as an input to an inverse design model, and learning such that a second target performance identical to the first target performance is predicted using the first optimal design plan as an input to a forward design model. The step of obtaining the first optimal design plan includes inputting the user's predicted preference for the design predicted from the user preference model, in addition to the first target performance, into the inverse design model.

[0021] Embodiments of the present invention can provide an AI-based generative design apparatus including a processor, a network interface, a memory, and a computer program loaded into the memory and executed by the processor. The computer program includes instructions for generating a new design for a product, instructions for predicting the performance of the new design, and instructions for obtaining an optimal design among the new designs.

[0022] Embodiments of the present invention can provide an AI-based generative design computer program stored in a computer-readable recording medium for performing, in combination with a computing device, a step of generating a new design for a product, a step of predicting the performance of the new design, and a step of obtaining an optimal design among the new designs.

[0023] Embodiments of the present invention can provide an AI-based generative design method that, in an AI-based generative design method executed by a computing device, includes a step of applying generative design to reference data to obtain a plurality of new designs having similarity to the reference data.

[0024] Embodiments of the present invention can provide an artificial intelligence-based generative design method further including: a step of filtering the plurality of new designs; a step of obtaining a plurality of additional new designs generated using a design generation model based on the filtered plurality of new designs; a step of filtering the plurality of additional new designs; and a step of applying the generative design to the filtered plurality of additional new designs as the plurality of reference data to obtain the plurality of new designs.

[0025] Embodiments of the present invention can provide an artificial intelligence-based generative design method further including a step of obtaining the reference data, wherein the first reference data includes 2D reference data or 3D reference data, the 2D reference data includes at least one of shape image data, pattern data, skeleton data, graph data, SDF data (distance information), depth data, negative function data, Grammar, and multi-view data obtained from data in each form, and the 3D reference data includes at least one of CAD data, voxel data, mesh data, point cloud data, octree data, shape parameter data, pattern data, skeleton data, graph data, SDF data (distance information), negative function (function formula or deep learning model).

[0026] Embodiments of the present invention can provide an artificial intelligence-based generative design method, wherein the generative design uses at least one of (i) a topology optimization methodology using at least one of SIMP, ESO, BESO, LevelSet, MMC, and deep learning, (ii) a size and shape optimization methodology, and (iii) a parametric design methodology.

[0027] Embodiments of the present invention can provide an artificial intelligence-based generative design method that (i) constructs the objective function of the generative design from a performance function part corresponding to requirements input by a user and a similarity function part for determining similarity to the reference data, or (ii) constructs the sensitivity function of the generative design from a part obtained by differentiating a performance function corresponding to requirements input by a user and a similarity function part for determining similarity to the reference data.

[0028] Embodiments of the present invention can provide an artificial intelligence-based generative design method, wherein the step of filtering the plurality of new designs or additional new designs includes at least one of a performance filtering step and a similarity filtering step of new designs. The performance filtering step filters low-performance new designs by evaluating requirements set by the user, and the similarity filtering step evaluates and filters similarity on a high-dimensional domain where the shape of the new design exists, or maps the shape of the new design to a low-dimensional domain to evaluate and filter similarity.

[0029] Embodiments of the present invention can provide an artificial intelligence-based generative design method, wherein the design generation model includes at least one of a deep learning model, a Boolean model, a Morphing model, and an Interpolation model.

[0030] Embodiments of the present invention can provide an artificial intelligence-based generative design method, including a step of learning an implicit neural representation model using the plurality of new designs and mapping the plurality of new designs to a low-dimensional continuous and parametric space z to obtain a plurality of design mapping data, and including at least one of a design exploration step, a design interpolation step, a design performance prediction step, a design optimization step, and a design inverse design step using the plurality of design mapping data in the low-dimensional continuous and parametric space z.

[0031] Embodiments of the present invention can provide an AI-based generative design method that performs topology optimization by applying a topology optimization technique to the plurality of design mapping data in the low-dimensional continuous and parametric space z in the design optimization process.

[0032] Embodiments of the present invention include a processor, a network interface, a memory, and a computer program loaded into the memory and executed by the processor. The computer program can include instructions for applying generative design to reference data for a product to obtain a plurality of new designs having similarity to the reference data, thereby providing an AI-based generative design apparatus.

[0033] Embodiments of the present invention can provide an AI-based generative design computer program stored in a computer-readable recording medium for executing, in combination with a computing device, a process of applying generative design to reference data for a product to obtain a plurality of new designs having similarity to the reference data.

[0034] Embodiments of the present invention can provide an AI-based generative design method that, in a method executed by a computing device, includes a process of mapping a plurality of designs X for a product to a low-dimensional latent space z using an implicit neural representation model to obtain a plurality of design mapping data Z, and a process of using the plurality of design mapping data in the latent space z to include at least one of a design exploration process, a design interpolation process, a design performance prediction process, a design optimization process, and a design inverse design process.

[0035] Embodiments of the present invention can provide an AI-based generative design method in which the design exploration process explores a first design mapping data Z1 among the plurality of design mapping data Z in the latent space z and obtains a first design X1 corresponding to the first design mapping data Z1.

[0036] In an embodiment of the present invention, in the design interpolation step, N pieces of data are selected from the plurality of design mapping data Z in the latent space z and interpolated with the second design mapping data Z2 by an interpolation method, and a second design X2 corresponding to the interpolated second design mapping data Z2 is obtained, and an artificial intelligence-based generative design method can be provided.

[0037] In an embodiment of the present invention, in the design prediction step, a performance prediction model learned from the plurality of design mapping data Z in the latent space z and the performance data Y of the plurality of designs X corresponding to the plurality of design mapping data Z is used to predict predicted performance data Y3' via the third design mapping data Z3 for the third design X3, and an artificial intelligence-based generative design method can be provided.

[0038] In an embodiment of the present invention, in the design optimization step, based on the performance data Y of the plurality of designs X, the fourth design mapping data Z4 corresponding to the fourth design X4 is optimized in the latent space z, and an optimized fourth design X4' is obtained, and an artificial intelligence-based generative design method can be provided.

[0039] In an embodiment of the present invention, in the inverse design step, an inverse design model learned from the performance data Y in the latent space z and the plurality of design mapping data Z corresponding to the performance data Y is used to obtain the fifth design mapping data Z5 corresponding to the fifth performance data Y5, and the fifth design data X5 corresponding to the fifth design mapping data Z5 is obtained, and an artificial intelligence-based generative design method can be provided.

Advantages of the Invention

[0040] In an embodiment of the present invention, an artificial intelligence-based generative design method, apparatus, and computer program that provide a technology for generating 2D / 3D design data based on a 2D / 3D deep learning-based generative design technology can be provided.

[0041] Embodiments of the present invention can provide an AI-based generative design method, apparatus, and computer program that construct a system for generating a large amount of engineering 3D design data using an AI model in the design process and provide a 2D / 3D deep learning-based inverse design technology capable of real-time optimal design.

[0042] Embodiments of the present invention can provide an optimized generative design method, apparatus, and computer program based on reference data and implicit neural representation, which can reflect reference data for multi-views of a product during concept design of the product and enable new design.

[0043] Embodiments of the present invention can provide an optimized generative design method, apparatus, and computer program based on reference data and implicit neural representation, which can provide a three-dimensional new design of a product having aesthetic properties for multi-views desired by a user while satisfying the performance of the product.

[0044] Embodiments of the present invention can provide an optimized generative design method, apparatus, and computer program based on reference data and implicit neural representation, which can map all seed data in a low-dimensional latent space z by deep learning, then process the data mapped in the latent space z, and generate synthetic data for learning of an AI model.

[0045] Embodiments of the present invention can provide an optimized generative design method, apparatus, and computer program based on reference data and implicit neural representation, which can generate and search design data in real time using design mapping data in a low-dimensional space.

[0046] Embodiments of the present invention can provide an AI-based generative design method, apparatus, and computer program that construct a design recommendation system by ensuring dimensionality reduction feature extraction and clustering techniques for a 2D / 3D design data set.

[0047] One embodiment of the present invention can provide an optimized generative design method, apparatus, and computer program based on reference data and implicit neural representations that can search for various designs that satisfy design requirements in a low-dimensional space z called a design space.

[0048] One embodiment of the present invention can provide an optimized generative design method, apparatus, and computer program based on reference data and implicit neural representations that can generate designs that satisfy the engineering requirements of a product.

[0049] One embodiment of the present invention can provide an optimized generative design method, apparatus, and computer program based on reference data and implicit neural representations that can provide a method for automatically labeling a generated new design by CAE analysis automation.

[0050] Embodiments of the present invention can provide an artificial intelligence-based generative design method, apparatus, and computer program that can provide artificial intelligence-based design verification and evaluation solutions.

[0051] Embodiments of the present invention can provide an artificial intelligence-based generative design method, apparatus, and computer program that can generate a large amount of design data based on artificial intelligence using a small amount of design data.

[0052] Embodiments of the present invention can provide an artificial intelligence-based generative design method, apparatus, and computer program that can provide full-stack technology available from product design to manufacturing by providing an organically connected design generation artificial intelligence model, design evaluation artificial intelligence model, and design recommendation artificial intelligence model.

Brief Description of the Drawings

[0053]

Figure 1

Figure 2

Figures 3a - 3d

Figure 4

Figures 5a - 5d

Figures 6a - 6d

Figure 7

Figures 8a - 8b

Figure 9

Figure 10

Figure 11

Figure 12

Figure 13

Figure 14

Figure 15

Mode for Carrying Out the Invention

[0054] Embodiments of the present invention can be subjected to various modifications, so the embodiments are shown in the drawings and will be described in detail in this specification. However, this is not intended to limit the embodiments of the present invention to specific forms, and includes all modifications, equivalents, or alternatives included in the spirit and technical scope of the present invention.

[0055] The terms used in this specification are for explaining the embodiments and are not intended to limit the present invention. In this specification, the singular form also includes the plural form unless otherwise particularly mentioned in the context. "Comprises" and / or "comprising" used in the specification do not exclude the existence or addition of one or more other components in addition to the mentioned components. The same reference numerals throughout the specification refer to the same components, and "and / or" includes each of the mentioned components and all combinations of one or more of them. For example, although "first", "second", etc. are used to explain various components, it goes without saying that these components are not limited by these terms. These terms are merely used to distinguish one component from another. Therefore, it goes without saying that the first component mentioned below can be the second component within the technical idea of the present invention. In this specification, including at least one of A, B, C, and D means including all various combinations of one or two or more of A, B, C, and D.

[0056] Unless otherwise defined, all terms (including technical and scientific terms) used herein may be used in a meaning commonly understood by one of ordinary skill in the art to which this invention belongs. Also, terms defined in commonly used dictionaries are not to be interpreted ideally or overly unless specifically defined otherwise.

[0057] As used in the specification, the terms "part" or "module" mean hardware components such as software, FPGA, or ASIC, and the "part" or "module" serves some role. However, the "part" or "module" is not meant in a sense limited to software or hardware. The "part" or "module" may be configured to be located in an addressable storage medium or may be configured to reproduce one or more processors. Thus, by way of example, the "part" or "module" includes components such as software components, object-oriented software components, class components, and task components, and processes, functions, attributes, procedures, subroutines, segments of program code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays, and variables. The functions provided within the components and the "part" or "module" can be combined in fewer components and "parts" or "modules" or further separated with additional components and "parts" or "modules".

[0058] Spatially relative terms such as "below", "beneath", "lower", "above", "upper", etc. can be used to easily describe the correlation between one component and another as shown in the drawings. Spatially relative terms should be understood as terms that include different directions of components during use or operation in addition to the directions shown in the drawings. For example, when the components shown in the drawings are inverted, the component described as "below" or "beneath" another component may be arranged "above" the other component. Therefore, the exemplary term "below" can include both the downward and upward directions. The component can also be oriented in other directions, and thus spatially relative terms can be interpreted according to the orientation.

[0059] Hereinafter, with reference to the accompanying drawings, an artificial intelligence-based generative design method, apparatus, and computer program according to an embodiment of the present invention will be described in detail.

[0060] FIG. 1 is a hardware configuration diagram of an artificial intelligence-based generative design apparatus according to an embodiment of the present invention.

[0061] Referring to FIG. 1, an artificial intelligence-based generative design apparatus 100 (hereinafter, "computing apparatus 100") according to an embodiment of the present invention may include one or more processors 110, a memory 120 that loads a computer program 141 executed by the processor 110, a communication interface 130, and a storage 140 that stores the computer program 141.

[0062] Here, the artificial intelligence-based generative design apparatus 100 according to an embodiment of the present invention is not limited to the components shown in FIG. 1 and may further include other general-purpose components.

[0063] In various embodiments, the computing device 100 refers to any type of hardware device including at least one processor 110, and can be understood to also include software configurations operating on the corresponding hardware device according to the embodiments.

[0064] The computing device 100 can be understood to include, but is not limited to, servers, smartphones, tablet PCs, desktops, notebooks, and both user clients and applications driven by each device.

[0065] Each step of the artificial intelligence-based generative design method according to the embodiments of the present invention is described as being executed by the computing device 100, but the subject matter of each step is not limited thereto, and at least a part of each step may be executed by different computing devices according to the embodiments.

[0066] The processor 110 controls the overall operation of each component of the computing device 100. The processor 110 may be configured to include a CPU (Central Processing Unit), an MPU (Micro Processor Unit), an MCU (Micro Controller Unit), a GPU (Graphic Processing Unit), or any form of processor well-known in the technical field of the present invention.

[0067] In addition, the processor 110 can execute operations on at least one application or program for executing the artificial intelligence-based generative design method according to the embodiments of the present invention, and the computing device 100 can be provided with one or more processors.

[0068] In addition, the processor 110 may further include a RAM (Random Access Memory, not shown) and a ROM (Read-Only Memory, not shown) that temporarily and / or permanently store signals (or data) processed internally. Further, the processor 110 may be implemented in the form of a system on chip (SoC) including at least one of a graphic processing unit, a RAM, and a ROM.

[0069] The memory 120 stores various data, instructions, and / or information. The memory 120 can load the computer program 141 from the storage 140 to execute the artificial intelligence-based generative design method according to an embodiment of the present invention. When the computer program 141 is loaded into the memory 120, the processor 110 can execute the method by executing one or more instructions constituting the computer program 141. The memory 120 may be implemented as a volatile memory such as a RAM, but the technical scope of the present disclosure is not limited thereto.

[0070] The bus BUS provides a communication function between components of the computing device 100. The bus BUS can be implemented as various forms of buses such as an address bus, a data bus, and a control bus.

[0071] The communication interface 130 supports wired and wireless Internet communications of the computing device 100. Further, the communication interface 130 can also support various communication methods other than Internet communication. For example, the communication interface 130 can support at least one of short-range communication, mobile communication, and broadcast communication methods. For this purpose, the communication interface 130 may be configured to include a communication module well known in the technical field of the present invention. In some embodiments, the communication interface 130 may be omitted.

[0072] Storage 140 can non-temporarily store computer program 141. Storage 140 may be configured to include non-volatile memories such as ROM (Read Only Memory), EPROM (Erasable Programmable ROM), EEPROM (Electrically Erasable Programmable ROM), flash memory, hard disks, removable disks, or any form of computer-readable recording medium well-known in the technical field to which the present invention pertains.

[0073] When computer program 141 is loaded into memory 120, it can include one or more instructions for causing processor 110 to perform an artificial intelligence-based generative design method according to an embodiment of the present invention. That is, processor 110 may execute the method according to various embodiments of the present invention by executing one or more instructions.

[0074] The steps of the artificial intelligence-based generative design method according to an embodiment of the present invention may be directly implemented in hardware, may be implemented by a software module executed by the hardware, or may be implemented by a combination thereof. The software module may also reside in RAM (Random Access Memory), ROM (Read Only Memory), EPROM (Erasable Programmable ROM), EEPROM (Electrically Erasable Programmable ROM), flash memory, hard disk, removable disk, CD-ROM, or any form of computer-readable recording medium well-known in the technical field to which the present invention pertains.

[0075] The steps of the artificial intelligence-based generative design method according to the embodiments of the present invention may be implemented as a program (or application) and stored in a medium for execution in combination with a computer which is hardware. The components of the present invention may be executed as software programming or software elements. Similarly, the embodiments include various algorithms implemented in a combination of data structures, processes, routines, or other programming constructs and may be implemented in programming languages or scripting languages such as C, C++, Java, assembler, etc. The functional aspects may be realized by algorithms executed on one or more processors.

[0076] Hereinafter, the artificial intelligence-based generative design method according to the embodiments of the present invention will be described.

[0077] FIG. 2 is a flowchart of the artificial intelligence-based generative design method according to the embodiments of the present invention. FIGS. 3a to 3d are conceptual block diagrams of the artificial intelligence-based generative design method according to the embodiments of the present invention.

[0078] Referring to FIGS. 2 to 3d, the artificial intelligence-based generative design method according to the embodiments of the present invention may include at least one of a step S10 of generating a design, a step S15 of exploring the design, a step S20 of predicting design performance, a step S30 of obtaining an optimal design solution, a step S40 of evaluating the optimal design solution, a step S50 of accumulating design data, a step S60 of analyzing the design data, and a step S70 of recommending the design data.

[0079] The artificial intelligence-based generative design method according to the embodiments of the present invention can generate, evaluate, and recommend the design of a product based on artificial intelligence. Here, the product may be any structure having a shape or an outer appearance. For example, there are automobiles, automobile wheels, automobile rims, watches, smartphones, bridges, buildings, etc.

[0080] For the sake of convenience of explanation, in this specification, the description is based on the wheel of an automobile. However, the scope of the rights of the present invention is not limited to the wheel of an automobile and can be applied to all products that can be designed.

[0081] Hereinafter, the process S10 of generating a design in the artificial intelligence-based generative design method according to an embodiment of the present invention will be described.

[0082] FIG. 4 is a flowchart of a process of generating a design based on artificial intelligence, reference data, and implicit neural representation according to an embodiment of the present invention. FIGS. 5A to 5D are conceptual block diagrams of a process of generating a design based on artificial intelligence, reference data, and implicit neural representation according to an embodiment of the present invention. FIGS. 6A to 6D are diagrams showing design generation examples according to an embodiment of the present invention.

[0083] In step S10, the computing device 100 can generate second reference data using the first reference data. The computing device 100 can generate a relatively large number of second reference data using a small amount of first reference data.

[0084] In other words, the computing device 100 can generate and evaluate a new design of a product using reference data. Here, the product may be any structure having a shape or outer form. For example, there are automobiles, automobile wheels, wheel hubs of automobiles, watches, smartphones, bridges, buildings, and the like. For the sake of convenience of explanation, in this specification, the description is based on the wheel of an automobile. However, the scope of the rights of the present invention is not limited to the wheel of an automobile and can be applied to all products that can be designed.

[0085] Referring to FIGS. 4 to 5d, a design generation method (S10) according to an embodiment of the present invention can include at least one of a step S100 of obtaining first reference data, a step S200 of obtaining a new design by generative design, a step S300 of filtering the new design, a step S400 of checking the number of new designs, a step S500 of obtaining additional new designs by a design generation model, a step S600 of filtering the additional new designs and repeating the above steps using the additional new designs as reference data, a step S700 of evaluating the new designs, and a step S800 of expressing the new designs in a low-dimensional space.

[0086] As shown in FIG. 4, in step S100, the computing device 100 can obtain first reference data for a product.

[0087] The computing device 100 can receive the first reference data via the communication interface 130 or obtain the first reference data stored in the storage 140.

[0088] The reference data can include designs of products shown by designers, designed by engineers, and aesthetically and performance-verified. Alternatively, the reference data refers to a technical blueprint including essential elements of the system and may be copied by a third party. The third party can improve or modify the reference data as needed.

[0089] The reference data can include at least one of a pattern image (e.g., Pattern Img) not related to the product, an image for a 2D product (e.g., Wheel Img), a skeleton image for a 2D product (e.g., 2D Skel.Img), a skeleton image for a 3D product (e.g., 3D Skel.Img), a cross-sectional image for a 3D product (e.g., Cross Img), a view image for a 3D product (e.g., View Img), a black-and-white image for the product (e.g., Black Img), and a color image for the product (e.g., Color Img). The reference data can include 2D reference data or 3D reference data.

[0090] In this case, the 2D reference data can include at least one of shape image data, pattern data, skeleton data, graph data, SDF data (distance information), depth data, a negative function (function formula or deep learning model), Grammar, and multi-view data obtained from data of each form.

[0091] Also, the 3D reference data can include at least one of CAD data, voxel data, mesh data, point cloud data, octree data, shape parameter data, pattern data, skeleton data, graph data, SDF data (distance information), and a negative function (function formula or deep learning model). Since the reference data is used as a seed to generate a new design, it may be called seed data.

[0092] In the following specification, the term of the design or design plan for the product means data that can be represented as the input and output of the computing device 100. Also, the design or design plan for the product may be 2D or 3D design data, may be black-and-white or color design data, and may be all file formats representing the design of the product such as an image file, a CAD file, or a modeling file. Design and design can be used interchangeably with the same meaning.

[0093] FIG. 7 is a flowchart of a process of acquiring reference data in a generative design method based on reference data and implicit neural representations according to an embodiment of the present invention. FIGS. 8A and 8B show examples of reference data generated by preprocessing input data.

[0094] In step S210, the computing device 100 can receive input data via the communication interface 130 or obtain the input data stored in the storage 140. The input data can include at least one of an image of a 2D product, an image of a 3D product, and a modeling of a 3D product.

[0095] In step S220, the computing device 100 can preprocess the input data by a preset preprocessing method. For example, the preset preprocessing method can include at least one of a method of extracting a pattern image from the input data (see (1) in FIG. 8A, S221), a method of extracting a skeleton image (see (2) in FIG. 8A, S222), a method of extracting a view image (see (3) in FIG. 8A, S223), a method of extracting a cross-sectional image (see (1) in FIG. 8B, S224), and a method of extracting a black-and-white image from a color image (S225).

[0096] Here, the view image of a 3D product can mean an image when the 3D product is viewed from a specific direction (view). Further, the computing device 100 can provide information about the specific view together while extracting the view image of the 3D product (for example, [x, y, z]=[1, 1, 1]).

[0097] Here, the cross-sectional image of the three-dimensional product can mean an image of the cross-section of the three-dimensional product at a depth according to the value input by the user while viewing the three-dimensional product from a specific direction (view). Further, the computing device 100 can provide information regarding the specific view and information regarding the depth of the pixels while extracting the cross-sectional image of the three-dimensional product (for example, [x, y, z]=[0, 0, 1]).

[0098] In step S220, the computing device 100 can execute one or two or more of steps S221 to S225 in parallel with respect to the input data, and can execute two or more in sequential combination, thereby obtaining a pre-processed image.

[0099] In step S230, the computing device 100 can provide the pre-processed image, and the pre-processed image can be used as reference data. Here, the pre-processed image can include at least one of a pattern image (for example, Pattern Img), an image of a two-dimensional product (for example, Wheel Img), a skeleton image of a two-dimensional product (for example, 2D Skel.Img), a skeleton image of a three-dimensional product (for example, 3D Skel.Img), a cross-sectional image of a three-dimensional product (for example, Cross Img), a view image of a three-dimensional product (for example, View Img), a black-and-white image of a product (for example, Black Img), a color image of a product (for example, Color Img), or can include a combination of two or more.

[0100] As shown in FIG. 4, in step S200, the computing device 100 can obtain a new design by generative design. Specifically, the computing device 100 can generate a plurality of new designs by applying generative design based on the reference data.

[0101] Generative design can use at least one of (i) a topology optimization methodology using at least one of a SIMP model, an ESO model, a BESO model, a LevelSet model, an MMC model, and a deep learning model, (ii) a size / shape optimization methodology, and (iii) a parametric design methodology.

[0102] In step S120, the computing device 100 can obtain a first novel design by topology optimization (TO). Referring to FIGS. 5a to 5d, as an example, the computing device 100 can apply topology optimization to the first reference data to obtain one or more first novel designs. As another example, the computing device 100 can apply topology optimization to the second novel design obtained in step S130 described below to obtain one or more first novel designs.

[0103] Topology optimization is defined as an optimization problem for a reference design that can satisfy the user's design requirements based on physics. The optimal design can be distinguished into a shape optimization that optimizes the shape and size of an object, a material optimization that optimizes the material properties, and a topology optimization that optimizes the structure of the object.

[0104] In step S130, the computing device 100 can obtain a second new design through Parametric Design (PD). Referring to FIGS. 5a to 5d, as an example, the computing device 100 can apply parametric design to the first reference data to obtain one or more second new designs. As another example, the computing device 100 can apply parametric design to the first new design obtained in step S120 described above to obtain one or more second new designs.

[0105] Parametric design is to perform parameterization on a reference design that can meet the user's design requirements based on rules.

[0106] In various embodiments, the computing device 100 can execute step S130 on the first new design obtained in step S200 to obtain a second new design, and execute step S120 again on the second new design to obtain a first new design, and steps S120 and S130 can be repeatedly executed. Thereby, the computing device 100 can obtain a plurality of first and second new designs. That is, even if the number of first reference data obtained by step S100 is small, a relatively large number of first new designs and second new designs can be obtained by steps S120 and S130.

[0107] In step S140, the computing device 100 can learn the design generation model DGM. As an example, the computing device 100 can learn a design generation model DGM that can generate a new third new design using at least one of the first reference data, the first new design, and the second new design.

[0108] Here, the learned design generation model DGM can have a latent space z learned by at least one of the first reference data, the first new design, and the second new design. The latent space z has a reduced dimension compared to the first reference data, the first new design, and the second new design, and can be expressed as mapping the first reference design, the first new design, and the second new design to the reduced dimension (low dimension).

[0109] Since the first new design generated by phase optimization and the second new design generated by parametric design have different representation methods such as dimension and length, they cannot be integrated into one model based on physics (rules). By extracting the characteristics of the first new design and the second new design and mapping them to the latent space z with a low dimension by the design generation model DGM, both the design data of the first new design and the second new design with different representations can be used.

[0110] In various embodiments, the design generation model DGM can be composed of an artificial intelligence model that can perform a new design using the latent space z while extracting design data and mapping it to the latent space z. The design generation model DGM can be composed of all generative models using deep learning. As an example, the design generation model can be composed of at least one of neural implicit representations such as DeepSDF, DualSDF, and Deep Implicit Templates in addition to GAN models such as Vanilla GAN, DCGAN, cGAN, WGAN, EBGAN, BEGAN, CycleGAN, DiscoGAN, StarGAN, SRGAN, SEGAN, SDF-based GAN, Shape GAN, and PolyGen models. However, the design generation model is not limited to GAN models and neural implicit representations, and all generative models based on deep learning that can generate new design data can be used.

[0111] In various embodiments, the computing device 100 can perform preprocessing on the design data of the first reference data, the first new design, and the second new design before learning the design generation model. As an example, the computing device 100 can preprocess the design data in a format such as a mesh, SDF (Signed Distance Function), and then learn with a deep learning model suitable for each preprocessing format. The preprocessing method of the design data can apply voxel, point cloud, etc., but is not limited thereto, and all methods that can be borrowed by ordinary technicians can be applied. Referring to FIGS. 5a to 5d, as an example, the design generation model DGM can include a discriminator that classifies the authenticity of the input design and a generator that generates a new design using a vector mapped to the latent space z. The discriminator and the generator are models that competitively learn against each other to improve each other's performance.

[0112] The discriminator is learned to classify real data as real and fake data as fake, and the generator is learned to generate fake data that cannot be distinguished from real data by the discriminator. When the learning of the discriminator and the generator is completed, the generator can learn the probability distribution of the real data and generate fake data according to the probability distribution, so that the discriminator cannot distinguish the fake data generated by the generator from the real data.

[0113] Referring to FIGS. 5a to 5d, the Neural Implicit Representation method optimizes the mapped data Z of each data X in the latent space z through the Auto-Decoder methodology. At this time, the mapped data enters as input together with the corresponding coordinates, and the model learns the distance value. When the learning is completed, the corresponding Neural Implicit Representation model has learned the three-dimensional shape in a negative function manner, and can output the corresponding three-dimensional shape by inputting appropriate mapped data.

[0114] High-dimensional data such as the first reference data, the first new design, and the second new design may be mapped to the latent space z of the generator having a low dimension. Then, a new third new design can be generated by utilizing the characteristics of the first reference data, the characteristics of the first new design, and the characteristics of the second new design that are matched to the low dimension.

[0115] In step S150, the computing device 100 can generate a new third new design using the latent space z of the learned design generation model DGM. The characteristics of the first reference data, the characteristics of the first new design, and the characteristics of the second new design can be mapped to the latent space z.

[0116] In various embodiments, the computing device 100 applies an Interpolation method to the characteristics of the first reference data, the characteristics of the first new design, and the characteristic information of the second new design mapped to the latent space z to generate the characteristic information of the new third new design, and can generate the third new design by utilizing the design generation model DGM based on the characteristic information of the third new design.

[0117] Here, as the interpolation method, at least two or more of the characteristics of the first reference data, the characteristics of the first new design, and the characteristic information of the second new design, such as the average value, the median value, the sum, the difference, the maximum value, the minimum value, and the selection of any value between the maximum and the minimum, can be applied to determine the characteristic information of the new third new design.

[0118] A method for generating a generative design similar to reference data can be selected and used from two options. One is a method that reflects the similarity to the reference data in the objective function, and the other is a method that reflects the similarity to the reference data in the sensitivity function obtained by differentiating the objective function.

[0119] First, a method of reflecting similarity in the objective function, which is the first method, will be introduced. The objective function of the generative design may be composed of the sum of the performance objective function part PERF corresponding to matters that the user should be satisfied with during design and the similarity part SIM that determines the similarity with the reference data. JPEG2025519277000002.jpg8170

[0120] The performance objective function part PERF is an optimization problem regarding product performance and can include elements such as the engineering performance of the product, the manufacturability of the product, and the aesthetics of the product. That is, it can mean the requirements set by the user. The similarity objective function part SIM is an optimization problem regarding similarity to the reference data and can include the similarity between the new design and the reference data. The similarity objective function part can adjust the influence degree on the entire objective function of the generative design by the weight λ.

[0121] Second, a method of reflecting similarity in the sensitivity function, which is the second method, will be introduced. The sensitivity function may be composed of the sum of the part obtained by differentiating the performance objective function part PERF corresponding to matters that the user should be satisfied with during design and the similarity part with the reference data. JPEG2025519277000003.jpg10170

[0122] The sensitivity function represents the rate of change of the objective function, and this value is repeatedly calculated to obtain the optimal design value. At this time, the sensitivity can be increased as the design is more similar to the reference data (reference design), and the new design (generated design) can be induced to be similar to the reference data (reference design). At this time, the influence degree of the reference data (reference design) of the generative design can be adjusted by the weight λ.

[0123] In summary, the generative design of the present invention can generate a new design similar to the reference data by assigning the similarity SIM to the objective function or the sensitivity function, and can adjust the degree of similarity by the weight λ. The present invention includes all generative design methods that modify the objective function or the sensitivity function to create a generative design similar to the reference data.

[0124] Furthermore, referring to FIGS. 5a to 5d, the computing device 100 can execute an abnormal filtering process for excluding abnormal designs among the third new designs generated by the design generation model DGM. Also, the computing device 100 can execute a similarity filtering process for excluding similar designs in the third new design. Here, the computing device 100 can execute at least one of the abnormal filtering process and the similarity filtering process.

[0125] Also, in various embodiments, the computing device 100 can apply at least one of the abnormal filtering process and the similarity filtering process to at least one of the first reference data, the first new design, the second new design, and the third new design.

[0126] The computing device 100 can execute the abnormal filtering process using the abnormal filtering module NAM. Here, the abnormal filtering module NAM can be a binary classifier that classifies the input design data as normal (Normal, 1) or abnormal (Abnormal, 0), or a multi-classifier that classifies the degree of abnormality of the design data into a plurality of classes or represents it as a probability.

[0127] The anomaly filtering model can be a supervised learning model or an unsupervised learning model. The supervised learning-based anomaly filtering model can learn using training data of input designs labeled as normal or abnormal, and the unsupervised learning-based anomaly filtering model can extract features of unlabeled input designs and classify the normality or abnormality of the input designs based on the extracted features.

[0128] The anomaly filtering model can be composed of a convolutional neural network (CNN) consisting of a feature extraction unit that extracts features using an image of the input design and a classifier that classifies the extracted features.

[0129] In step S160 of FIG. 4, the computing device 100 can obtain at least one of the first new design, the second new design, and the third new design as the second reference data. Further, the computing device 100 can execute the subsequent steps including step S100 using the second reference data as the first reference data, and by repeating a plurality of steps, a large amount of second reference data can be generated using a relatively small amount of first reference data.

[0130] Referring to FIGS. 3a to 3d and FIGS. 5a to 5d, in step S160 of various embodiments, the computing device 100 can sample at least one of the first new design, the second new design, and the third new design and select a predetermined number of designs. In various embodiments, the computing device 100 can determine the number of designs for performing step S200 described below in consideration of the performance of the computing device 100 for design of experiment (DOE). Here, the predetermined number of designs can mean at least 1000 or more designs for deep learning. The computing device 100 can perform optimal sampling by a sampling technique such as Latin Hypercube Sampling (LHS), but is not limited thereto.

[0131] Referring to FIGS. 3a to 3d and FIGS. 5a to 5d, in step S160 of various embodiments, when the second reference data is two-dimensional design data, the computing device 100 can execute a postprocessing step of changing the second reference data into three-dimensional design data. As a postprocessing method for changing two-dimensional design data into three-dimensional design data, there are 3D mesh, B-rep, etc., and various methods that can be borrowed by those skilled in the art can be applied.

[0132] Although the above-described first reference data, first new design, second new design, third new design, and second reference data are expressed in the singular, they of course include plural concepts.

[0133] FIG. 6a is a diagram showing various examples of new designs generated using a pattern image as reference data.

[0134] Referring to FIG. 6a, in step S200 of various embodiments, the computing device 100 can utilize the pattern image as reference data to perform generative design (e.g., phase optimization, size and shape optimization, parametric design) to ensure the engineering performance of the product, and generate a new design of a product similar to the pattern image through the similar parts. That is, even without reference data (e.g., initial rim wheel image) for the product to be newly designed, the computing device 100 can use the pattern image as reference data to generate a new design for a product that is aesthetically similar.

[0135] FIG. 6b is a diagram showing various examples of new designs generated using the skeleton image as reference data.

[0136] Referring to FIG. 6b, in step S200 of various embodiments, the computing device 100 can utilize the skeleton image for a 2D product (e.g., 2D Skel.Img) or the skeleton image for a 3D product (e.g., 3D Skel.Img) as reference data to proceed with generative design (e.g., phase optimization, size and shape optimization, parametric design) to generate a new design.

[0137] In various embodiments, when applying the phase optimization methodology in generative design, by using the skeleton image for a 2D (3D) product instead of the general image for a 2D (3D) product, it is possible to prevent the influence caused by the density (thickness, volume) biased in the general image for a 2D (3D) product from affecting the new design. Also, by using the skeleton image for a 2D (3D) product, the density (thickness, volume) distribution for the product in the new design may not be biased.

[0138] In various embodiments, the computing device 100 can generate a novel design for a product having a density (thickness, volume) that increases uniformly according to a volume ratio, which is a user-specified parameter.

[0139] For example, a novel design with a large volume ratio can provide a design for a product having a relatively increased density (thickness, volume) compared to a novel design with a small volume ratio.

[0140] FIG. 6c is a diagram showing various examples of novel designs generated using a view image or a cross-sectional image as reference data.

[0141] Referring to the upper two diagrams in FIG. 6c, the computing device 100 can generate a novel design similar to a view image in a specific direction in one of an image for a 2D product, an image for a 3D product, and a modeling for a 3D product.

[0142] Here, the computing device 100 can extract a view image O1 in a specific direction using a view extraction model from one of an image for a 2D product, an image for a 3D product, and a modeling for a 3D product. Also, the computing device 100 can extract information regarding the specific direction view together. Here, the information regarding the specific direction view may be to represent the direction of viewing the product as a vector (for example, [x, y, z]=[1, 1, 1]).

[0143] Furthermore, the computing device 100 can apply phase optimization using the view image O1 in a specific direction as reference data to generate a novel design N1. In this case, the novel design can be a design similar to the same direction view as the specific direction view image.

[0144] Furthermore, the computing device 100 can extract a plurality of view images O1, O2 for a plurality of view directions, and can apply generative designs (e.g., phase optimization, size and shape optimization, parametric design) to the plurality of view images respectively to generate a plurality of new designs N1, N2.

[0145] Here, the plurality of new designs may be phase-optimized designs corresponding to the plurality of view images respectively. In this case, the objective functions of the phase optimization are formed in a plurality so as to have similarity parts corresponding to the plurality of view images, and the plurality of new designs can be obtained by obtaining the answers of the respective objective functions of the plurality of phase optimizations.

[0146] Furthermore, the computing device 100 can use information ([1,1,1], [1,0,1]) regarding a plurality of view directions to combine a plurality of new designs corresponding to the plurality of view directions to generate a new image for a 2D product or a new modeling for a 3D product.

[0147] In various embodiments, the computing device 100 can extract a plurality of view images O1, O2 for a plurality of view directions, and can apply generative designs (e.g., phase optimization, size and shape optimization, parametric design) to the plurality of view images together to generate one new design. In this case, the objective function of the phase optimization is formed by a plurality of similarity parts corresponding to the plurality of view images, and the new design can be obtained by obtaining the answer of one objective function of the phase optimization in which there are a plurality of similarity parts (L1 distance term).

[0148] On the other hand, referring to the lower figure in Fig. 6c, the computing device 100 can generate a new design similar to a cross-sectional image at a predetermined depth by user input in a view in a specific direction from among an image for a 2D product, an image for a 3D product, and a modeling for a 3D product.

[0149] Here, the computing device 100 can extract a cross-sectional image O3 of the product at a specific direction and a predetermined depth using one of an image for a two-dimensional product, an image for a three-dimensional product, and a modeling for the three-dimensional product. Here, the cross-sectional image of the product can mean an image of the cross-section of the product at a depth according to a value input by the user while viewing the two-dimensional or three-dimensional product in a specific direction (view).

[0150] Also, the computing device 100 can extract information regarding a specific direction view and information regarding depth together. Here, the information regarding the specific direction view can represent the direction of viewing the product as a vector (for example, [x, y, z] = [0, 0, 1]). Regarding the depth information, assuming that the total depth of the product based on the specific direction view is 100%, the portion of the product farthest from the user can be defined as 100%, and the portion of the product closest to the user can be defined as 0% (for example, depth = 50%).

[0151] Furthermore, the computing device 100 can generate a new design N3 by applying generative design (for example, phase optimization, size and shape optimization, parametric design) to the cross-sectional image O3 at a specific direction and a predetermined depth as reference data. In this case, the new design can be a design similar to the same direction view and the predetermined depth as the cross-sectional image of the product at a specific direction and a predetermined depth.

[0152] Furthermore, the computing device 100 can extract a plurality of cross-sectional images O3 for a specific direction and a plurality of depths, and can apply generative design (for example, phase optimization, size and shape optimization, parametric design) to each of the plurality of cross-sectional images to generate a plurality of new designs N3.

[0153] Here, the plurality of new designs may be phase-optimized designs corresponding to each of the plurality of cross-sectional images. In this case, a plurality of objective functions for phase optimization are formed so as to correspond to the plurality of cross-sectional images, and the plurality of new designs can be obtained by obtaining the answers of the respective objective functions for the plurality of phase optimizations.

[0154] Furthermore, the computing device 100 can combine a plurality of new designs corresponding to a plurality of depths using information ([0,0,1]) for a specific direction and the plurality of depths to generate a new image for a two-dimensional product or a new modeling for a three-dimensional product.

[0155] In various embodiments, the computing device 100 can extract a plurality of cross-sectional images O3 for a plurality of depth directions, and can apply phase optimization to the plurality of cross-sectional images together to generate one new design.

[0156] In this case, a plurality of similarity portions corresponding to the plurality of cross-sectional images are formed in the objective function for phase optimization, and a new design can be obtained by obtaining the answer of one objective function for phase optimization in which a plurality of similarity portions (L1 distance term) exist.

[0157] In various embodiments, the computing device 100 can apply generative design (e.g., phase optimization, size and shape optimization, parametric design) to view cross-sectional images extracted from a plurality of direction views and a plurality of depths for a product as reference data to generate a plurality of new designs.

[0158] Referring to FIGS. 4 to 5d, in step S300, the computing device 100 can filter a plurality of new designs. Here, the new design may be any one of the first reference data, the first new design, the second new design, the third new design, and the second reference data.

[0159] In operation S300, the computing device 100 can perform at least one of a performance filtering process and a similarity filtering process for a new design.

[0160] In various embodiments, in the performance filtering process, the computing device 100 can evaluate a new design based on requirements set by a user at design time, and can perform a process of excluding new designs that do not meet the requirements set by the user. In this case, a new design that fails to pass the requirements is referred to as a low-performance new design.

[0161] In various embodiments, in the similarity filtering process, the computing device 100 can perform filtering to exclude similar designs among a plurality of new designs.

[0162] In various embodiments, the computing device 100 can evaluate a similarity degree on a high-dimensional domain where the shape of a new design exists. The computing device 100 can perform an X-space filtering process. Here, the computing device 100 measures the distance between two of a plurality of new designs in the X space, and if the inter-distance is smaller than a preset first threshold value, at least one of the two can be excluded as a similar design. Such a process can be performed for all of the plurality of new designs. Here, the X space means a high-dimensional domain where the shape of the new design exists by itself. That is, the X space means the new design itself generated by generative design.

[0163] In various embodiments, the computing device 100 can map the shape of a new design to a low-dimensional domain to evaluate the similarity. The computing device 100 can perform a Z-space filtering process. Here, the computing device 100 projects a plurality of new designs into the latent space z using an encoding model to extract low-dimensional feature values, measures the distance between the low-dimensional feature values of both designs, and if the inter-distance is smaller than a preset second threshold, at least one of the two can be excluded when viewed as similar designs. Such a process can be performed for all the plurality of new designs. Here, the Z-space means a space where the low-dimensional values of new designs are mapped.

[0164] In step S300, the computing device 100 can perform one or both of a performance filtering process and a similarity filtering process. Also, the computing device 100 can perform one or both of the above-described X-space filtering process and Z-space filtering process. However, the filtering process of the present invention is not limited to this, and various filtering methods capable of classifying similar designs can be applied.

[0165] Referring to FIGS. 4 to 5d, in step S400, the computing device 100 can check whether the number of filtered new designs exceeds a preset threshold. Here, if the number of filtered new designs does not exceed the preset threshold, the computing device 100 can execute the next step S500, and if it exceeds the preset threshold, the computing device 100 can execute the next step S700.

[0166] In step S500, the computing device 100 can apply a design generation model to the new design to generate additional new designs.

[0167] The computing device 100 can learn a design generation model. As an example, the computing device 100 can learn a design generation model that can generate a new additional new design using at least one of reference data and a new design.

[0168] Here, the learned design generation model can have a latent space z learned by at least one of reference data and a new design. The latent space z has reduced dimensions compared to the reference data and the new design, and can be expressed as mapping the reference design and the new design to the reduced dimensions (low dimensions).

[0169] In various embodiments, the design generation model can be composed of an artificial intelligence model that can perform a new design using the latent space z while extracting design data and mapping it to the latent space z.

[0170] The design generation model can be composed of all Generative Models using deep learning. As an example, the generative model can be composed of at least one of Vanilla GAN, DCGAN, cGAN, WGAN, EBGAN, BEGAN, CycleGAN, DiscoGAN, StarGAN, SRGAN, SEGAN, SDF-based GAN, Shape GAN, and PolyGen model. However, the design generation model is not limited to the GAN model, and all deep learning-based Generative Models that can generate or interpolate new design data can be used. For example, the design generation model can include at least one of a deep learning model, a boolean model, a morphing model, and an interpolation model.

[0171] The computing device 100 can generate a new additional new design using the latent space z of the learned generative model.

[0172] In step S600, the computing device 100 can perform filtering to exclude similar designs from the additional new designs.

[0173] Step S600 can directly apply the filtering process described in step S300. Further, the computing device 100 can execute step S100 of using the additional new design or the filtered additional new design as reference data. Here, the computing device 100 may perform preprocessing steps such as extracting a pattern image or a skeleton image using the additional new design or the filtered additional new design.

[0174] The computing device 100 can repeat step S200 of applying generative design using the additional new design and the filtered additional new design to generate a new design.

[0175] In step S700, the computing device 100 can evaluate the new design.

[0176] The computing device 100 can determine whether the new design performance obtained by analyzing the new design meets preset requirements. The requirements may be information input by the user during design generation or preset information for each product.

[0177] Here, the new design performance can include at least one of a first analysis result corresponding to the design requirements desired by the user, a second analysis result corresponding to the physical performance of the product corresponding to the design, and a third analysis result corresponding to the manufacturability which is a matter to be considered when manufacturing the product corresponding to the design.

[0178] Correspondingly, the preset requirements can include at least one of a first requirement corresponding to a preset design requirement, a second requirement corresponding to a preset physical performance that the product needs to minimize, and a third requirement corresponding to a preset manufacturability for matters that need to be minimally considered when manufacturing the product.

[0179] In various embodiments, the computing device 100 can evaluate whether the requirements are met by comparing one of the first to third requirements corresponding to at least one of the first to third analysis results of the new design.

[0180] The computing device 100 can repeatedly execute steps S100 to S600. Here, the computing device 100 can repeatedly execute steps S100 to S600 until the number of new designs filtered in step S400 is more than a preset third threshold.

[0181] Referring to FIGS. 4 to 5d, in step S800, the computing device 100 can execute a step of representing the new design in a low-dimensional space.

[0182] Hereinafter, for convenience of explanation, the new design input to the implicit neural representation model is referred to as input design Xi, and the output is referred to as output design Xo.

[0183] In step S800, the computing device 100 can learn an implicit neural representation model using a plurality of new designs Xi.

[0184] The implicit neural representation model can represent a plurality of new designs Xi in a low-dimensional continuous parametric space z, which can be called a plurality of design mapping data Z.

[0185] The computing device 100 can represent a plurality of new designs Xi as continuous values on a low-dimensional space z using an implicit neural representation model.

[0186] The design mapping data Z has information on the new design Xi represented on the low-dimensional space z that corresponds one-to-one with the new design Xi. Since the values between the plurality of design mapping data Z are continuous, it is represented differentiably.

[0187] The computing device 100 can execute at least one of a search process for a new design, an interpolation process for a design, a performance prediction process for a design, an optimization process for a design, and an inverse design process for a design using a plurality of design mapping data Z in a low-dimensional continuous parametric space z.

[0188] The computing device 100 is represented on a continuous low-dimensional space z using an implicit neural representation model, and the design mapping data Z corresponding to the new design Xi can be stored in the storage 140.

[0189] The new design Xi can include all seed data such as phase optimization, size and shape optimization, parametric design, and user data included in the generative design described above.

[0190] After mapping all the seed data to the low-dimensional latent space z through deep learning for the implicit neural representation model, the computing device 100 can process the design mapping data Z in the latent space z to generate new design mapping data Z'.

[0191] The computing device 100 can generate an output design Xo that is distinguished from the conventional new design Xi using the new design mapping data Z' in the latent space z. Since the output design Xo is newly synthesized data, it can be called synthetic data.

[0192] Hereinafter, with reference to FIG. 2, a search process for a design, an interpolation process S15 of the design, a performance prediction process S20 of the design, an optimization process S30 of the design, and an inverse design process of the design will be described.

[0193] In step S15, the computing device 100 can execute a design search process of searching for the first design mapping data Z1 among a plurality of design mapping data Z in the low-dimensional space z and obtaining the first design X1 corresponding to the first design mapping data Z1.

[0194] Since the design mapping data Z represented on the low-dimensional space z is represented by continuous numerical values, it is easier to search than the new design X represented by two-dimensional or three-dimensional design data. Therefore, in order to find a specific first design X1, the first design X1 can be quickly found by searching for the first design mapping data Z1 corresponding to the first design X1 on the low-dimensional space z.

[0195] After the computing device 100 learns the implicit neural representation model as a new design X, it can search for a desired new design in real time using the design mapping data Z represented in the latent space z.

[0196] FIG. 6d is a diagram showing an example of design interpolation in the latent space.

[0197] Referring to FIG. 6d, in step S15, the computing device 100 can execute a design interpolation process of interpolating a plurality of second design mapping data Z1 among a plurality of design mapping data Z in the low-dimensional space z and obtaining the second design X2 corresponding to the interpolated second design mapping data Z1'.

[0198] In various embodiments, the computing device 100 can execute a design interpolation process of selecting N data among a plurality of design mapping data Z in the low-dimensional space z, interpolating with the second design mapping data Z2 by an interpolation method, and obtaining the second design X2 corresponding to the interpolated second design mapping data Z2.

[0199] In various embodiments, the computing device 100 can generate new design mapping data by applying an interpolation method (e.g., median, average, or average with other weights) to at least one design mapping data represented in the low-dimensional space z. Since such an interpolation process is performed within the low-dimensional space z, it can be executed with simple operations.

[0200] Conventionally, a plurality of new designs generated by generative design cannot be represented in a continuous design space. Therefore, a design C located between design A and design B cannot be generated. However, embodiments of the present invention can interpolate design mapping data corresponding to all designs in a low-dimensional latent space z to represent a new design.

[0201] That is, after the computing device 100 learns the implicit neural representation model as a new design X, it can generate a desired new design in real time using the design mapping data Z represented in the latent space z.

[0202] Hereinafter, a process S20 of predicting the performance of a design by an AI-based generative design method according to an embodiment of the present invention will be described.

[0203] FIG. 9 is a flowchart of a process of predicting the performance of a design by an AI-based generative design method according to an embodiment of the present invention. FIG. 10 is a conceptual block diagram of a process of predicting design performance according to an embodiment of the present invention.

[0204] In process S20, the computing device 100 can predict the performance of the design generated in process S10. The generated design can include second reference data.

[0205] Referring to FIGS. 9 and 10, step S20 can include at least one of a preprocessing step S21 of the second reference data, a step S22 of analyzing the second reference data, a step S23 of learning a design performance prediction model, a step S24 of predicting the performance of the second reference data, a step S25 of determining whether the predicted value of the design performance meets the standard, a step S26 of providing the learned design performance prediction model, and a step S27 of adaptive sampling.

[0206] Referring to FIG. 10, in step S21, the computing device 100 can perform preprocessing on the second reference data that is the design generated in step S10. Since a high computational cost is required for the computing device 100 to analyze the design data using the performance analysis module, it is not possible to perform analysis on all the second reference data. Therefore, the computing device 100 can perform preprocessing on the second reference data to select a preset number of third reference data from the second reference data. In other words, the third reference data means a part of the preprocessed second reference data.

[0207] The number of the third reference data may be less than the number of the second reference data. For example, the computing device 100 can select a preset number of third reference data from the second reference data through the preprocessing step. Here, the preprocessing step can be similarly applied to the step of sampling a preset number of designs by applying a sampling technique to a plurality of designs in the description regarding FIGS. 4 to 5d.

[0208] However, step S21 is not necessarily a step that must be executed in step S20, and it can be excluded as necessary and the remaining steps can be executed.

[0209] In step S22, the computing device 100 can analyze the third reference design using the performance analysis module. The performance analysis module can provide a plurality of analysis results by analyzing the third reference design.

[0210] Here, the plurality of analysis results can include at least one of a first analysis result corresponding to design requirements desired by the user, a second analysis result corresponding to the physical performance of the product corresponding to the design, and a third analysis result corresponding to manufacturability, which are matters to be considered when manufacturing the product corresponding to the design.

[0211] Examples of design requirements can include the length of a specific part of the product, the weight of the product, the volume of the product, and the like.

[0212] Examples of the physical performance of the product can include the rigidity of the structure, the strength of the structure, the weight of the structure, and the like.

[0213] Examples of matters to be considered during product manufacturing include moldability by injection molding, moldability by extrusion molding, manufacturing cost per unit, manufacturing period, and the like.

[0214] The analysis results can be represented as absolute numerical values (e.g., 120, 45, etc.), relative numerical values (e.g., probability, percentage, etc.), classification into multiple classes (e.g., upper / middle / lower, etc.), binary classification (e.g., 0 / 1, etc.), and the like.

[0215] In step S220 of various embodiments, the computing device 100 can automatically label at least one of the plurality of analysis results with one of the second reference design or the third reference design and store it as analysis data. As shown in the following table, the second / third reference design X can match the first to third analysis results (Y1, Y2, Y3) analyzed by the performance analysis module and store them as analysis data.

[0216]

Table 1

[0217] In step S23, the computing device 100 can learn a design performance prediction model. In various embodiments, the computing device 100 can use the analysis data as learning data to learn the design performance prediction model.

[0218] The design performance prediction model can be trained using the analysis results labeled with the second or third reference design, and the trained design performance prediction model can output the analysis results expected as the input of the second or third reference design.

[0219] The design performance prediction model can include at least one of a first design performance prediction model trained with the first analysis results labeled with the design, a second design performance prediction model trained with the second analysis results labeled with the design, and a third design performance prediction model trained with the third analysis results labeled with the design.

[0220] Here, the trained first design performance prediction model can output the design requirements desired by the user. Also, the trained second design performance prediction model can output the physical performance corresponding to the design. Also, the trained third design performance prediction model can output the manufacturability, which is a matter to be considered when manufacturing the product corresponding to the design.

[0221] The design performance prediction model can include multiple artificial intelligence models. The multiple artificial intelligence models can include at least one of a machine learning model such as a random forest model and a support vector machine (SVM), and a deep learning model such as a convolutional neural network (CNN) series model and a recurrent neural network (RNN) series model. It can be composed of one of the implicit neural representation series, Transformer series, PointNet series, 3D CNN series specialized for 3D deep learning, and graph neural network (GNN), and multiple models can also be ensembled. Also, all models can utilize both 2D data and 3D data.

[0222] In step S24, the computing device 100 can predict the performance of the second or third reference data using the learned design performance prediction model. Alternatively, the computing device 100 can obtain an analysis result by inputting the second or third reference data into the learned design performance prediction model.

[0223] The learned design performance prediction model can output and provide at least one of a first analysis result corresponding to design requirements desired by a user, a second analysis result corresponding to the physical performance of a product corresponding to the design, and a third analysis result corresponding to manufacturability which are considerations when manufacturing a product corresponding to the design.

[0224] In various embodiments, the computing device 100 can provide a first analysis result using the first design performance prediction model by inputting the third reference data, can provide a second analysis result using the second design performance prediction model, and can provide a third analysis result using the third design performance prediction model.

[0225] In step S25, the computing device 100 can determine whether the obtained analysis result meets a preset reference value. For example, the computing device 100 can determine whether the first to third analysis results meet the criteria when compared to the preset reference values corresponding to each of them. The preset reference values can include critical design requirements corresponding to the first analysis result, critical physical performance corresponding to the second analysis result, and critical manufacturability corresponding to the third analysis result.

[0226] In step S26, the computing device 100 can provide the learned design performance prediction model when the analysis result of the second or third reference data meets the preset reference value in step S250. The computing device 100 can store the learned design performance prediction model in the storage 140.

[0227] In step S27, when the computing device 100 does not meet the criteria in step S25, it can apply Adaptive Sampling to the second or third reference data to obtain fourth reference data. Adaptive Sampling is a technique that indicates the position of data where the performance of the model is expected to be highest when design data is added.

[0228] In various embodiments, the computing device 100 can add design data with high uncertainty (e.g., the standard deviation of the analysis result value) to the predicted value (analysis result) of the design performance prediction model and re-learn the design performance prediction model to reduce the uncertainty.

[0229] In various embodiments, the computing device 100 can perform additional analysis on the fourth reference data to obtain additional analysis results, and re-learn the design performance prediction model based on the additional analysis results.

[0230] In summary, when the predicted design performance does not meet the preset reference value, the computing device 100 can collect additional analysis results for the newly generated design in step S10 via Adaptive Sampling, and re-learn the design performance prediction model based on the collected additional analysis results.

[0231] In step S20, the computing device 100 can execute a design prediction process of predicting predicted performance data Y3' via third design mapping data Z3 for a third design X3 using a performance prediction model learned as a plurality of design mapping data Z in a low-dimensional space z and performance data Y of a plurality of designs X corresponding to the plurality of design mapping data Z.

[0232] In other words, by learning the performance prediction model with design mapping data Z and performance data Y corresponding to design X, the learned hierarchical prediction model can output predicted performance data with specific design mapping data as the input.

[0233] The computing device 100 can train a performance prediction model using Design X and performance data Y corresponding to each design as training data. Specifically, the computing device 100 can train a performance prediction model using design mapping data Z of Design X represented in a low dimension using a transient neural representation model and performance data Y as training data.

[0234] In this case, the trained performance prediction model can predict performance data Y' for a new Design X together with an implicit neural representation model. That is, the computing device 100 can receive an input of a new Design X using the performance prediction model and predict the performance of a product of the new design.

[0235] FIG. 11 is a flowchart of a process of obtaining an optimal design by an artificial intelligence-based generative design method according to an embodiment of the present invention. FIG. 12 is a conceptual block diagram of a process of obtaining an optimal design according to an embodiment of the present invention. FIG. 13 is a diagram showing a conceptual diagram of a deep learning-based optimal design model according to an embodiment of the present invention.

[0236] In step S30, the computing device 100 can obtain an optimal design based on target performance desired by a user.

[0237] Referring to FIGS. 11 and 12, step S30 may include at least one of a step S31 of obtaining first target performance, a step S32 of obtaining a first optimal design plan, and a step S33 of obtaining a second optimal design plan.

[0238] Referring to FIG. 12, in step S31, the computing device 100 can obtain first target performance regarding performance information of a target product desired by a user.

[0239] Here, the target performance can include at least one of first target information corresponding to the design requirements for the product desired by the user, second target information corresponding to the physical performance of the product, and third target information corresponding to the manufacturability which are matters to be considered when manufacturing the product. The first target information to the third target information can respectively correspond to the first analysis result to the third analysis result.

[0240] In step S32, the computing device 100 can obtain a first optimal design plan based on the first target performance.

[0241] In various embodiments, the computing device 100 can use the first target performance as the input of the inverse design model IDM, obtain the first optimal design plan as the output, use the first optimal design plan as the input of the forward design model FDM, and learn the inverse design model and the forward design model so that the second target performance the same as the first target performance is predicted as the output. In this step, the computing device 100 can obtain, in real time, the first optimal design plan for the first target performance as the output of the inverse design model.

[0242] Here, the forward design model may be composed of the design performance prediction model learned in step S20.

[0243] Here, since the optimal design plan for the target performance is generated using the inverse design model and the forward design model, both the inverse design model and the forward design model can be called the optimal design model.

[0244] The computing device 100 can obtain, in real time, the optimal design plan for the target performance for the purpose of global optimization using the inverse design model. The computing device 100 can obtain the target performance for the optimal design plan for the purpose of local optimization using the forward design model.

[0245] Various optimization techniques can be used in the optimal design model. As an example, Gradient-based methods (e.g., backpropagation, SQP, etc.) and heuristic optimization methods (e.g., GA, etc.) can be used.

[0246] At least one of the inverse design model IDM and the forward design model FDM can include multiple deep learning models. The multiple deep learning models can include at least one of the convolutional neural network (CNN) series models and the recurrent neural network (RNN) series models, or can include at least one of the PointNet series, 3D CNN series, and graph neural network (GNN) specialized for 3D deep learning. Alternatively, the multiple models may be an ensemble. Also, all models can utilize both 2D data and 3D data.

[0247] Here, the optimal design model can be a model learned by a deep learning-based inverse design methodology. The deep learning-based inverse design methodology is a technique that learns a forward network as a surrogate model, relearns an inverse network that predicts an optimal solution based on this, and outputs an optimal solution in real time when an input value is given to the inverse network. The deep learning-based inverse design method can show better performance than conventional optimal design techniques as the design problem is of higher dimension. Also, the deep learning-based inverse design method can selectively use supervised and unsupervised learning.

[0248] In various embodiments, after receiving the first target performance as an input, the inverse design model derives the first optimal design as an output, and the forward design model is trained such that after receiving the first optimal design derived from the inverse design model as an input, the same second target performance as the first target performance which is the input of the inverse design model is predicted as it is.

[0249] In various embodiments, at least one of the forward design model and the inverse design model is trained such that the second target performance which is the output of the forward design model is predicted as the first target performance which is the input of the inverse design model, and the computing device 100 can acquire, as the first optimal design, the design output in real time from the inverse design model trained based on the first target performance.

[0250] In step S33, the computing device 100 can sample the second optimal design from the first optimal design.

[0251] In various embodiments, the computing device 100 can acquire the second optimal design by re-training the forward design model through collecting additional analysis results by adaptive sampling considering the first optimal design and the uncertainty of the performance prediction therefor.

[0252] Here, uncertainty-considering adaptive sampling means quantifying the uncertainty of the first optimal design predicted by the inverse design model through an uncertainty quantification model, and adaptively sampling the first optimal design in which the target performance of the first optimal design is higher than the preset performance and at the same time the quantified uncertainty is higher than the preset value. Here, the uncertainty quantification model can include at least one of a Bayesian neural network, an ensemble-based model (such as Deep Ensemble), and a Gaussian process.

[0253] In various embodiments, the computing device 100 can apply uncertainty-aware adaptive sampling to the first optimal design to obtain a design case that is additionally sampled in the first optimal design case. The additionally sampled design case can also be referred to as the first-1 optimal design case.

[0254] In various embodiments, the computing device 100 can proceed with the analysis of the additionally sampled design case and re-learn the design performance prediction model in step S20 based on the analyzed data labeled as the analysis result.

[0255] In various embodiments, the computing device 100 can utilize the re-learned design prediction model to re-execute the optimal design for the first target performance, thereby deriving a second optimal design case in a performance region that is improved compared to the overall performance region of the existing first optimal design case.

[0256] In various embodiments, the computing device 100 can use the first optimal design case or the second optimal design case as the first reference data in step S10 to generate a new design.

[0257] In step S30, the computing device 100 can execute a design optimization process of optimizing the fourth design mapping data Z4 corresponding to the fourth design X4 in the low-dimensional space z based on the performance data Y of a plurality of designs X and obtaining the optimized fourth design data X4'.

[0258] When the performance data of a specific design meets the requirements desired by the user, the computing device 100 can obtain the design mapping data corresponding to the corresponding design, optimize the fourth design mapping data Z4 based on the corresponding design mapping data, and obtain the optimized fourth design X4' by obtaining the optimized fourth design X4', so as to obtain the optimized fourth design X4' that has the requirements desired by the user.

[0259] In various embodiments, the computing device 100 can perform phase optimization on a plurality of design mapping data in the low-dimensional space z.

[0260] In conventional automatic parameterization optimal design techniques, for 3D shape optimization, the designer needs to perform parameterization based on the base design and define the design variables to be optimized. As a result, the optimal design can only be found within the trivial design space. However, in the embodiments of the present invention, since the 3D shape is automatically parameterized in the latent space, the optimal design can be performed in the original design space.

[0261] In step S30, the computing device 100 can execute an inverse design process of obtaining the fifth design mapping data Z5 corresponding to the fifth performance data Y5 and obtaining the fifth design X5 corresponding to the fifth design mapping data Z5 by using the inverse design model learned with the performance data Y in the low-dimensional space z and the plurality of design mapping data Z corresponding to the performance data Y.

[0262] Here, the inverse design model can output the design X that satisfies the requirements when the user inputs the performance data Y for the desired requirements. The computing device 100 can output the design for a specific product that satisfies the input performance specifications by receiving the input of the desired performance specifications for a specific product. That is, if the user only knows the performance data of the product, the user can inverse-design the design of the product.

[0263] FIG. 14 is a conceptual block diagram of the evaluation of the optimal design plan, the accumulation of design data, the analysis of design data, and the design data recommendation process according to the embodiments of the present invention.

[0264] Referring back to FIGS. 2 and 14, in step S40, the computing device 100 can evaluate the optimal design solution. The optimal design solution can include the first optimal design solution and the second optimal design solution of step S30. In various embodiments, the computing device 100 can determine whether the performance of the optimal design solution analyzed meets a preset requirement.

[0265] Here, the optimal design solution performance can include at least one of a first analysis result corresponding to the design requirements desired by the user, a second analysis result corresponding to the physical performance of the product corresponding to the design, and a third analysis result corresponding to the manufacturability which is a matter to be considered when manufacturing the product corresponding to the design.

[0266] Correspondingly, the preset requirements can include at least one of a first requirement corresponding to the preset design requirements, a second requirement corresponding to the preset physical performance that the product should have to be minimized, and a third requirement corresponding to the preset manufacturability which is a matter to be considered minimally when manufacturing the product.

[0267] In various embodiments, the computing device 100 can evaluate whether the requirements are met by comparing at least one of the first to third analysis results of the optimal design solution with one of the first to third requirements corresponding thereto.

[0268] In step S50, the computing device 100 can store the design data that meets the preset requirements among the optimal design solutions. In various embodiments, the computing device 100 can repeatedly execute steps S10 to S30 until the stored design data meets a preset number. The preset number may be 50,000 to 150,000, and preferably may be about 100,000.

[0269] In step S60, the computing device 100 can analyze the plurality of stored design data.

[0270] In step S60 of various embodiments, the computing device 100 can classify design data according to characteristics of a plurality of design data by using a clustering model. The computing device 100 can select a representative design representing the cluster from among the plurality of design data classified into a plurality of clusters.

[0271] In step S60, in various embodiments, the computing device 100 can perform an anomaly filtering process for excluding abnormal design data from the accumulated design data. Further, the computing device 100 can perform a similarity filtering process for removing duplicate design data from the accumulated design data. The anomaly filtering process can be applied in the same manner as the anomaly filtering process in step S100 described above.

[0272] In step S60, in various embodiments, the computing device 100 can visualize corresponding design data by using the optimal design plan performance and the characteristics of the design data.

[0273] FIG. 15 is a conceptual block diagram of a method for collecting preference information for design data according to an embodiment of the present invention and predicting a preferred design by using a learned preference prediction model.

[0274] Referring to FIG. 15, in step S60 of various embodiments, the computing device 100 can predict a user preference for design data by using a preference prediction model. Further, the computing device 100 can recommend design data to the user based on the predicted user preference for the design data.

[0275] Referring to the top view in FIG. 15, in various embodiments, the computing device 100 can perform a process of collecting learning data for learning a Preference Prediction Model. Here, the computing device 100 can perform a process of obtaining user preferences for the design data itself from the user through an interrogation process. Various methods such as rank, relative evaluation, rating, etc. can be used as the user evaluation method for obtaining user preferences.

[0276] In various embodiments, the computing device 100 can perform a process of obtaining user preferences for the deformed design data after changing at least one of dimensions (2D->3D, 3D->2D), color (black and white->color), material (e.g., metal material, etc.) of the design data in order to obtain accurate user preferences for the design data applied to the product. Here, the user preferences can include information such as the style for the design, the price for the design, etc.

[0277] Referring to the bottom view of FIG. 15, in various embodiments, the computing device 100 can learn a Preference Prediction Model using user preference information (Preference Data) for the design data. In various embodiments, the computing device 100 can predict a user-preferred user preference design among a plurality of design data using the learned preference prediction model. As an example, the preference prediction model can learn the preference (y) labeled for Design X as learning data, and the learned preference prediction model can output the predicted preference (y) for the new Design X.

[0278] In various embodiments, the computing device 100 can learn a preference prediction model based on user preference evaluations for at least one of all the above-described designs or design proposals, and obtain the user predicted preference for the design predicted by the learned preference prediction model.

[0279] FIG. 13 is a diagram showing a conceptual diagram of a deep learning-based optimal design model according to an embodiment of the present invention.

[0280] Referring to FIG. 13, the computing device 100 can obtain an optimal design plan (x) based on the target performance (y) of the product and the user preference (c) for the design using the optimal design model.

[0281] In various embodiments, the inverse design model of the optimal design model receives the first target performance (y) and the user preference (c) as inputs, and then derives the first optimal design plan (x) as an output. The forward design model of the optimal design model is learned such that after receiving the first optimal design plan (x) derived from the inverse design model as an input, the same second target performance (y1, y2,..., yn) as the first target performance (y) that is the input of the inverse design model is predicted as it is.

[0282] In various embodiments, the computing device 100 can obtain an optimal design plan by using the user predicted preference for the design predicted by the preference prediction model and the target performance as inputs to the optimal design model. That is, by reflecting the output value of the preference prediction model as an input to the optimal design model, the optimal design model can provide an optimal design plan with high user preference.

[0283] In various embodiments, the computing device 100 can obtain user preference design performance for a user preference design predicted using a design performance prediction model or a performance analysis module. Here, the user preference design performance can include at least one of a first analysis result related to user design requirements, a second analysis result corresponding to the physical performance of the product, and a third analysis result corresponding to manufacturability, which are matters considered when manufacturing a product corresponding to the design.

[0284] In various embodiments, the computing device 100 can propose the user preference design performance as the first target performance of the product to the user in step S30. Thereby, in step S30, the computing device 100 can provide a first optimal design solution that not only satisfies engineering performance but also has aesthetic properties that the user may like.

[0285] In step S70, the computing device 100 can propose a user-recommended design among the design data. In various embodiments, the computing device 100 can propose a user-recommended design selected from the optimal design solutions using a design recommendation model. The user can derive a design solution for manufacturing through subsequent detailed design within the concept having engineering performance and aesthetic properties from the user-recommended design.

[0286] All of the above-described models can mean artificial intelligence models. Here, the artificial intelligence model can be composed of one or more network functions, and one or more network functions can be composed of a set of interconnected computing units that can generally be called "nodes". Such "nodes" may also be called "neurons". One or more network functions are composed of including at least one or more nodes. The nodes (or neurons) constituting one or more network functions may be interconnected by one or more "links".

[0287] Within an artificial intelligence model, one or more nodes connected via links can relatively form the relationship between input nodes and output nodes. The concepts of input nodes and output nodes are relative. Any node that has an output node relationship with one node can have an input node relationship with other nodes, and vice versa. As described above, the relationship between an output node and an input node can be generated centered around a link. One input node may be connected via links to one or more output nodes, and vice versa can also hold true.

[0288] In the relationship between an input node and an output node connected via one link, the output node can determine its value based on the data input to the input node. Here, the node interconnecting the input node and the output node can have a weight. The weight may be variable and may be varied by a user or an algorithm to execute the desired function of the artificial intelligence model. For example, when one or more input nodes are interconnected by respective links to one output node, the output node can determine the output node value based on the values input to the input nodes connected to the output node and the weights set for the respective links corresponding to the input nodes.

[0289] As described above, an artificial intelligence model is formed by one or more nodes interconnected via one or more links to form the relationship between input nodes and output nodes within the artificial intelligence model. The characteristics of the artificial intelligence model can be determined by the number of nodes and links within the artificial intelligence model, the correlation between nodes and links, and the values of the weights assigned to each link. For example, if there are two artificial intelligence models with the same number of nodes and links but different weight values between the links, the two artificial intelligence models can be recognized as different from each other.

[0290] Among the nodes that make up an artificial intelligence model, some can form one layer based on the distance from the initial input node. For example, a set of nodes at a distance of n from the initial input node can form the nth layer. The distance from the initial input node can be defined by the minimum number of links that must be traversed to reach the node from the initial input node. However, such a layer definition is arbitrary for illustrative purposes, and the layer order within the artificial intelligence model can be defined in a different way. For example, the layer of nodes can also be defined by the distance from the final output node.

[0291] The initial input node can mean one or more nodes into which data is directly input without passing through links in relation to other nodes within the artificial intelligence model. Or, within the artificial intelligence model network, in the relationship between nodes based on links, it can mean a node that does not have other input nodes connected to the link. Similarly, the final output node can mean one or more nodes that do not have output nodes in relation to other nodes within the artificial intelligence model. Also, a hidden node can mean a node that makes up the artificial intelligence model, other than the initial input node and the final output node. The artificial intelligence model according to an embodiment of the present invention can be an artificial intelligence model in which the number of nodes in the input layer may be more than the number of nodes in the hidden layer closer to the output layer, and the number of nodes decreases as it progresses from the input layer to the hidden layer.

[0292] The artificial intelligence model can include one or more hidden layers. The hidden nodes of the hidden layer can take as inputs the outputs of the previous layer and the outputs of the surrounding hidden nodes. The number of hidden nodes for each hidden layer may be the same or different. The number of nodes in the input layer may be determined based on the number of data fields of the input data, and may be the same as or different from the number of hidden nodes. The input data input to the input layer may be computed by the hidden nodes of the hidden layer, or may be output by a fully connected layer (FCL) that is the output layer.

[0293] The computing device 100 can learn an artificial intelligence model using learning data. The computing device 100 can perform learning on one or more network functions that constitute the artificial intelligence model using a learning data set.

[0294] In one embodiment, the computing device 100 can input each learning input data set to one or more network functions, extract the output data values computed by the one or more network functions as feature values, and learn the artificial intelligence model based on the extracted feature values.

[0295] In other embodiments, the computing device 100 can input each learning input data set to one or more network functions, and compare each of the output data computed by the one or more network functions with each of the learning output data sets corresponding to the labels of the learning input data sets to derive an error.

[0296] In the learning of the artificial intelligence model, the learning input data can be input to the input layer of one or more network functions, and the learning output data can be compared with the outputs of the one or more network functions. The computing device 100 can learn the artificial intelligence model based on the error between the computation results of the one or more network functions for the learning input data and the learning output data (labels).

[0297] In addition, the computing device 100 can adjust the weights of one or more network functions in a backpropagation manner based on errors. That is, the computing device 100 can adjust the weights based on the error between the calculation result of one or more network functions for the learning input data and the learning output data so that the output of one or more network functions approaches the learning output data.

[0298] When the learning of one or more network functions has been executed for a pre-determined number of epochs or more, the computing device 100 can determine whether to interrupt the learning using verification data. The pre-determined number of epochs can be part of the overall learning target number of epochs. The verification data may be composed of at least a part of the labeled learning dataset. That is, the computing device 100 executes the learning of the artificial intelligence model through the learning dataset, and after the learning of the artificial intelligence model has been repeated for a pre-determined number of epochs or more, it can determine whether the learning effect of the artificial intelligence model is at a pre-determined level or more using the verification data. For example, when the computing device 100 performs learning with a target number of repeated learning times of 10 using 100 pieces of learning data, after executing 10 repeated learning times which is the pre-determined number of epochs, it executes 3 repeated learning times using 10 pieces of verification data. If the change in the output of the artificial intelligence model during the 3 repeated learning times is below the pre-determined level, it can be determined that further learning is meaningless and the learning can be terminated. That is, the verification data can be used to determine the completion of learning based on whether the learning effect for each epoch in the repeated learning of the artificial intelligence model is above a certain level. The numbers of the above-mentioned learning data, verification data, and the number of repetitions are merely examples and are not limited thereto.

[0299] The computing device 100 can generate an artificial intelligence model by testing the performance of one or more network functions using a test data set and determining the activation status of the one or more network functions. The test data can be used to verify the performance of the artificial intelligence model and can be composed of at least a part of the learning data set. For example, 70% of the learning data set can be utilized for the learning of the artificial intelligence model (i.e., learning to adjust the weights to output result values similar to the labels), and 30% can be utilized as test data for verifying the performance of the artificial intelligence model.

[0300] The computing device 100 can input a test data set into the artificial intelligence model for which learning has been completed, measure the error, and determine the activation status of the artificial intelligence model based on whether it is above a pre-determined performance. The computing device 100 can verify the performance of the artificial intelligence model for which learning has been completed using the test data, and if the performance of the artificial intelligence model for which learning has been completed is above the pre-determined standard, the corresponding artificial intelligence model can be activated so that it can be used in other applications.

[0301] Also, the artificial intelligence model can be learned in at least one of the methods of supervised learning, unsupervised learning, and semi-supervised learning. The learning of the artificial intelligence model is performed to minimize the error of the output. In the learning of the artificial intelligence model, the learning data is repeatedly input into the artificial intelligence model, the error between the output of the artificial intelligence model for the learning data and the target is calculated, and the error of the artificial intelligence model is backpropagated from the output layer to the input layer in the direction to reduce the error to update the weights of each node of the artificial intelligence model.

[0302] In the case of supervised learning, learning data with correct answers labeled for each piece of learning data is used (i.e., labeled learning data). In the case of unsupervised learning, there may be cases where correct answers are not labeled for each piece of learning data. That is, for example, the learning data in the case of supervised learning for data classification may be data with categories labeled for each piece of learning data. An error can be calculated by inputting the learning data into an artificial intelligence model and comparing the output (category) of the artificial intelligence model with the label of the learning data.

[0303] As another example, in the case of unsupervised learning for data classification, an error can be calculated by comparing the input learning data with the output of the artificial intelligence model. The calculated error can be backpropagated in the artificial intelligence model in the reverse direction (i.e., from the output layer to the input layer), and the connection weights of each node in each layer of the artificial intelligence model can be updated according to the backpropagation.

[0304] The change amount of the connection weight of each updated node can be determined according to the learning rate. The calculation of the artificial intelligence model for the input data and the backpropagation of the error can constitute a learning cycle (epoch). The learning rate can be applied to vary according to the number of repetitions of the learning cycle of the artificial intelligence model. For example, at the initial stage of learning of the artificial intelligence model, a high learning rate can be used to enhance efficiency by enabling the artificial intelligence model to quickly secure a certain level of performance, and at the later stage of learning, a low learning rate can be used to improve accuracy.

[0305] In the training of an artificial intelligence model, generally, the training data may be a subset of the actual data (i.e., the data to be processed using the trained artificial intelligence model). Therefore, there may be a training cycle in which the error for the training data decreases, but the error for the actual data increases. Overfitting is a phenomenon in which the model over-learns the training data and the error for the actual data increases. Overfitting can cause an increase in the error of the machine learning algorithm. Various optimization methods can be used to prevent such overfitting. To prevent overfitting, the training data can be increased, or methods such as regularization, dropout (omitting some of the nodes in the network during the training process) can be applied.

[0306] In various embodiments, the artificial intelligence model can include a plurality of artificial intelligence models. The plurality of artificial intelligence models can include at least one of a random forest model, which is a machine learning model, a support vector machine (SVM), and deep learning models such as a convolutional neural network (CNN) series model and a recurrent neural network (RNN) series model. It can be composed of a neural implicit representation series, a Transformer series, a PointNet series, a 3D CNN series, and a graph neural network (GNN) specialized for 3D deep learning, and multiple models can also be ensembled. Also, all models can utilize both 2D data and 3D data.

[0307] The present invention has been described with reference to the embodiments shown in the drawings, which are merely exemplary, and those having ordinary knowledge in the technical field will understand that various modifications and equivalent other embodiments will be possible hereafter. Therefore, the true technical protection scope of the present invention should be determined by the technical idea of the appended registered claims.

Explanation of Signs

[0308] 100 Computing device

Claims

1. In an artificial intelligence-based generative design method executed by a computing device, a step of generating a new design for a product; a step of exploring the generated new design; a step of predicting the performance of the new design; a step of obtaining an optimal design among the new designs, an artificial intelligence-based generative design method.

2. The step of generating a new design for the product includes a step of applying phase optimization to first reference data to obtain a first new design, a step of applying parametric design to the first reference data to obtain a second new design, a step of applying the parametric design to the first new design to further obtain the second new design, and a step of applying the phase optimization to the second new design to further obtain the first new design, the artificial intelligence-based generative design method according to Claim 1.

3. The step of generating a new design for the product includes a step of mapping a first new design and a second new design to one latent space using a design generation model, and a step of using a third new design generated by the design generation model as second reference data, wherein the first new design and the second new design can have different representation methods and formats, the artificial intelligence-based generative design method according to Claim 1.

4. The step of predicting the performance of the new design includes a step of analyzing the new design using a performance analysis module, a step of learning a design performance prediction model based on the analysis result of the new design, a step of predicting the performance of the new design based on the design performance prediction model, a step of providing the learned design performance prediction model when the predicted design performance meets a preset reference value, and when the predicted design performance does not meet the preset reference value, collecting additional analysis results for the new design through adaptive sampling, and re-learning the design performance prediction model based on the collected additional analysis results, the artificial intelligence-based generative design method according to Claim 1.

5. The step of obtaining an optimal design among the new designs is A step of obtaining a first target performance regarding performance information of a target product desired by a user, obtaining a first optimal design plan using the first target performance as an input to an inverse design model, and learning so that a second target performance identical to the first target performance is predicted by using the first optimal design plan as an input to a forward design model; and a step of obtaining a second optimal design plan by re-learning the design performance prediction model through additional analysis results of additional sampled design plans selected by adaptive sampling considering the uncertainty of the first optimal design plan. The artificial intelligence-based generative design method according to claim 4, comprising:

6. The step of evaluating the optimal design, the step of accumulating a part of the optimal design as design data, the step of analyzing the design data, and the step of recommending the design data. The artificial intelligence-based generative design method according to claim 1, further comprising:

7. The step of obtaining a first optimal design plan by using the first target performance as an input to an inverse design model, and learning so that a second target performance identical to the first target performance is predicted by using the first optimal design plan as an input to a forward design model is The artificial intelligence-based generative design method according to claim 5, comprising a step of obtaining the first optimal design plan by additionally inputting a user's predicted preference for a design predicted from a user preference model to the inverse design model together with the first target performance.

8. In an artificial intelligence-based generative design method executed by a computing device, An artificial intelligence-based generative design method, comprising a step of applying generative design to reference data to obtain a plurality of new designs having similarity to the reference data.

9. The generative design is (i) A topology optimization methodology using at least one of SIMP, ESO, BESO, LevelSet, MMC, and deep learning, (ii) A size and shape optimization methodology, and (iii) The artificial intelligence-based generative design method according to claim 8, using at least one of parametric design methodologies.

10. (i) Composing the objective function of the generative design from a performance function part corresponding to requirements input by a user and a similarity function part for determining similarity to the reference data, or (ii) The generative design sensitivity function according to claim 8, wherein the generative design sensitivity function is composed of a part obtained by differentiating a performance function corresponding to requirements input by a user and a similarity function part for determining similarity with the reference data.

11. A step of learning an implicit neural representation model using the plurality of new designs, mapping the plurality of new designs into a low-dimensional continuous and parametric space (z) to obtain a plurality of design mapping data, and using the plurality of design mapping data in the low-dimensional continuous and parametric space (z), including at least one of a design exploration step, a design interpolation step, a design performance prediction step, a design optimization step, and a design inverse design step. The artificial intelligence-based generative design method according to claim 8.

12. In the design optimization step, the artificial intelligence-based generative design method according to claim 11, wherein a phase optimization technique for the plurality of design mapping data is applied in the low-dimensional continuous and parametric space (z) to perform phase optimization.

13. In a method executed by a computing device, a step of mapping a plurality of designs (X) for a product into a low-dimensional latent space (z) using an implicit neural representation model to obtain a plurality of design mapping data (Z), and using the plurality of design mapping data in the latent space (z), including at least one of a design exploration step, a design interpolation step, a design performance prediction step, a design optimization step, and a design inverse design step. An artificial intelligence-based generative design method.

14. The design exploration step is to explore a first design mapping data (Z1) among the plurality of design mapping data (Z) in the latent space (z) and obtain a first design (X1) corresponding to the first design mapping data (Z1). The artificial intelligence-based generative design method according to claim 13.

15. The design interpolation step is to select N data among the plurality of design mapping data (Z) in the latent space (z) and interpolate them with a second design mapping data (Z2) by an interpolation method, and obtain a second design (X2) corresponding to the interpolated second design mapping data (Z2). The artificial intelligence-based generative design method according to claim 13.

16. The design performance prediction step predicts predicted performance data (Y3') via third design mapping data (Z3) for a third design (X3) using a performance prediction model learned from the plurality of design mapping data (Z) in the latent space (z) and the performance data (Y) of the plurality of designs (X) corresponding to the plurality of design mapping data (Z). The artificial intelligence-based generative design method according to claim 13.

17. The design optimization step optimizes fourth design mapping data (Z4) corresponding to a fourth design (X4) in the latent space (z) based on the performance data (Y) of the plurality of designs (X), and obtains an optimized fourth design (X4'). The artificial intelligence-based generative design method according to claim 13.

18. The inverse design process uses an inverse design model learned from the performance data (Y) in the latent space (z) and the plurality of design mapping data (Z) corresponding to the performance data (Y) to obtain fifth performance data (Y ５ ), and obtain fifth design mapping data (Z ５ ) corresponding thereto, and obtain fifth design data (X ５ ) corresponding to the fifth design mapping data (Z ５ ). The artificial intelligence-based generative design method according to claim 13.

Images