Design support device and design support method

Abstract

To provide a design support system capable of generating high-quality drawings from a density distribution of a structure derived from structural optimization calculations in a short time.SOLUTION: A design support system includes: a clustering condition input unit that receives input of a clustering condition; a clustering unit that follows the clustering condition; a machine learning control unit that performs machine learning on each cluster obtained; a drawing generation unit that generates new drawings from machine learning results; a drawing determination unit that determines whether a basic drawing to be compared and the new drawing are the same; an optimization / analysis condition input unit that receives input of calculation conditions for structural optimization; an analysis model generation unit that generates an analysis model according to calculation conditions; a structural optimization calculation unit that calculates using the analysis model; and a drawing generation condition input unit that displays the structural optimization calculation results and clustering results, and receives input of weights for each cluster required for drawing generation. The drawing generation unit generates a proposed drawing using the calculation results and weights of structural optimization.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to a design support apparatus and a design support method. [Background technology]

[0002] Conventionally, in design support technology for predicting the optimal shape of mechanical structures, there is a method for predicting the optimal shape by using topology optimization calculations and image processing, which is a type of machine learning. Here, topology optimization is a structural optimization technique that derives the optimal density distribution of a material for a given structure based on the design variables of the structure, and is also called topology optimization.

[0003] Patent Document 1 discloses a modeling process including a learning method that enables the generation of physically realistic CAD data that can be directly used in design and / or manufacturing processes, the modeling process including providing a dataset including a functional structure by the learning method (S10), and training a generative autoencoder on the dataset by the learning method (S20), and may further include training a latent space classifier by the learning method (S30), and may further include providing one or more latent vectors by the generation method (S100), and generating a functional structure by the generation method (S200), and may further include performing topology optimization of a mechanical assembly of rigid parts represented by the generated functional structure by the generation method (S300).

[0004] In addition, there is a method for generating CAD drawings from the results of optimization calculations in accordance with the model form of the analysis, where CAD is an abbreviation for Computer Aided Design.

[0005] Patent Document 2 discloses a design optimization support system used in the design of structures, which includes a finite element method model conversion device that converts CAD data into a finite element method model, an optimization device that constructs an optimized finite element method model using an optimization method, and a CAD data conversion device that converts the optimized finite element method model into CAD data, and which also includes a means for selecting a finite element method model creation method that is suited to the type of CAD data and the model form, a means for selecting an optimization method that is suited to the optimization purpose, and a CAD data conversion means that converts the optimized finite element method model into CAD data in accordance with the model form. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] Patent Publication No. 2021-12693 [Patent Document 2] Japanese Patent Application Laid-Open No. 2006-11729 Summary of the Invention [Problem to be solved by the invention]

[0007] Conventional topology design support technology uses machine learning to learn a tree representing the functional structure of a CAD drawing using a deep neural network (DNN), generates a new CAD drawing from vector data called latent vectors, and then performs topology optimization calculations using the generated CAD drawing to predict the optimal shape. In this case, the predicted optimal shape is represented by a density distribution that varies from 0 to 1 across the entire calculation domain. In other words, the optimal shape of a mechanical structure is predicted by assuming that areas with a density of 0 do not contain any structures and areas with a density of 1 contain structures.

[0008] The optimal shape predicted in this way also includes density areas other than 0 and 1. In areas with intermediate density values ​​such as 0.3 or 0.7, it is up to the engineer to decide whether or not a structure is included. As a result, interpretations vary depending on the engineer, which can lead to variations in the final CAD drawing, making it difficult to obtain high-quality CAD drawings.

[0009] The method described in Patent Document 1 does not sufficiently consider the point that high-quality CAD drawings can be obtained regardless of whether they are done by an engineer or not.

[0010] In addition, in the technology for generating CAD drawings from the results of optimization calculations, the shape of the CAD drawing is determined by smoothly connecting the density distribution of the calculation domain in the predicted optimal shape. Attributes of CAD drawings include characteristic information such as manufacturing cost and efficiency.

[0011] In the design optimization support system described in Patent Document 2, the designer is responsible for selecting the optimization objective. Therefore, selecting an appropriate optimization objective from multiple characteristics and the like is left to the designer's skill, experience, and so on, and there are cases where the designer needs to repeatedly review the optimization objective. Therefore, there is room for improvement in that more effort and time are required for design.

[0012] The present disclosure aims to generate high-quality drawings in a short time from the density distribution of a structure derived by structural optimization calculations. [Means for solving the problem]

[0013] a structural optimization calculation unit that performs structural optimization calculations using the analytical model; a drawing generation condition input unit that displays the structural optimization calculation results and the clustering results and accepts input of weights for each cluster required for drawing generation; and a display unit that displays information including at least one of the clustering conditions, the calculation results, and the generated drawing. The drawing generation unit generates a proposed drawing using the structural optimization calculation results and the weights, and the display unit displays the proposed drawing. [Effects of the Invention]

[0014] According to the present disclosure, high-quality drawings can be generated in a short time from the density distribution of a structure derived by structural optimization calculations. [Brief explanation of the drawings]

[0015] [Figure 1] 1 is a schematic configuration diagram showing a design support device according to an embodiment; [Figure 2A] 10 is a flowchart showing a process of inputting clustering conditions in Phase 1. [Figure 2B] 10 is a flowchart showing the clustering process in Phase 1. [Figure 3] 1 is a flowchart showing the machine learning process in Phase 1. [Figure 4A] 10 is a flowchart showing the topology optimization calculation process in Phase 2. [Figure 4B] 10 is a flowchart showing a drawing generation process in Phase 2. [Figure 5] FIG. 2B is a diagram showing an example of data acquired in step S101 of FIG. 2A. [Figure 6] FIG. 2B is a diagram showing an example of an input screen for clustering conditions used in step S102 of FIG. 2A. [Figure 7] FIG. 2C is a diagram showing an example of a screen showing the clustering results displayed in step S204 of FIG. 2B. [Figure 8] FIG. 1 is a conceptual diagram illustrating an example of CycleGAN. [Figure 9] FIG. 10 is a diagram showing an example of an input screen for topology optimization calculation. [Figure 10] FIG. 4B is a diagram showing an example of the result of the topology optimization calculation obtained in step S403 of FIG. 4A. [Figure 11] FIG. 4C is a diagram showing an example of an input screen for conditions for generating a new drawing in step S502 of FIG. 4B. [Figure 12] FIG. 4C is a diagram showing an example of a screen displayed in step S504 of FIG. 4B. DETAILED DESCRIPTION OF THE INVENTION

[0016] The present disclosure relates to a design support technology that generates a specific drawing based on an input condition vector (weight) from the results of topology optimization having density shading obtained by numerical analysis.

[0017] Hereinafter, the embodiments will be described with reference to the drawings. [Example]

[0018] FIG. 1 is a schematic diagram showing the configuration of a design support device according to an embodiment.

[0019] As shown in this figure, the design support device includes a clustering condition input unit 101, a clustering unit 102, a clustering result display unit 103, a machine learning control unit 104, a drawing generation unit 105, a drawing discrimination unit 106, an optimization / analysis condition input unit 107, an analysis model generation unit 108, a topology optimization calculation unit 109 (structural optimization calculation unit), a drawing generation condition input unit 110, a drawing display unit 111, a database 112 (storage device), and a computer 113.

[0020] In the clustering condition input unit 101, an operator inputs information (clustering conditions) required for clustering, such as CAD drawings, topology optimization calculation results, and characteristic information.

[0021] The clustering unit 102 performs clustering using a self-organizing map according to the input clustering conditions. Here, clustering is a type of machine learning, and refers to a technique for grouping data based on the similarity between data. Clustering is also called "cluster analysis" or "data clustering."

[0022] The clustering result display unit 103 displays the results of clustering using the self-organizing map, along with the average value of the characteristic information of each cluster.

[0023] The machine learning control unit 104 controls the drawing generation unit 105 and the drawing determination unit 106, and performs machine learning using the topology optimization calculation results, CAD drawings, and condition vectors as inputs using a method that applies CycleGAN. Here, CycleGAN is an abbreviation for Cycle Generative Adversarial Networks.

[0024] The drawing generation unit 105 generates a new drawing using the topology optimization calculation result, the CAD drawing, and the condition vector as input information.

[0025] Drawing discrimination unit 106 receives two drawings as input information and discriminates whether or not the two drawings are the same.

[0026] In the optimization / analysis condition input section 107, analysis conditions, objective functions, and volume constraint conditions required for topology optimization calculation are input.

[0027] The analytical model generation unit 108 generates an analytical model based on analytical conditions, objective functions, and volume constraints required for topology optimization calculation.

[0028] The topology optimization calculation unit 109 performs topology optimization calculations using the finite element method for the structural analysis portion of the generated analysis model and the density method for the topology optimization portion.

[0029] The drawing generation condition input unit 110 displays the results of topology optimization calculation and clustering, and inputs the weight of each cluster required for drawing generation.

[0030] The drawing display unit 111 displays the drawing generated by the drawing generation unit 105 and the drawing generation conditions.

[0031] The database 112 stores the data obtained from each section.

[0032] Next, an example of a processing procedure in the design support device, that is, a design support method will be described.

[0033] The design support method of this embodiment can be broadly divided into two processes. The first process (Phase 1) is a process of clustering data such as CAD drawings and applying machine learning to the data. The second process (Phase 2) is a process of performing topology optimization calculations and generating drawings from the results.

[0034] The following description will be given with reference to a mechanical structure.

[0035] FIG. 2A is a flowchart showing the input process of the clustering conditions in Phase 1.

[0036] In steps S101 to S104 shown in the figure, the clustering conditions are inputted from the clustering condition input unit 101 and stored in the database 112.

[0037] First, in step S101, past CAD drawings, topology optimization calculation results, characteristic information such as manufacturing cost and efficiency, strength resistance performance, and heat resistance performance are acquired from the database 112.

[0038] FIG. 5 is a diagram showing an example of the data acquired in step S101.

[0039] As shown in this figure, the CAD drawing of "AA001" is associated with the result of the topology optimization calculation, "AA001TOP," and is managed by linking it to characteristic information such as manufacturing cost, efficiency, strength performance, and heat resistance as its attributes. It is assumed that this past data has been stored in the database 112 in advance. In this case, this past data is the data that serves as the basis for supervised learning. The past CAD drawing may be one obtained as a result of the topology optimization calculation, or it may be a drawing that a designer actually created in the past.

[0040] In step S102, the operator inputs the conditions necessary for clustering using the clustering condition input unit 101.

[0041] FIG. 6 is a diagram showing an example of a screen for inputting clustering conditions.

[0042] As shown in this figure, "Machine Structure 1" has been entered as the learning model name. The data attributes shown in Figure 5 are displayed in a corresponding format as variables to be clustered. Figure 6 displays the CAD drawing, topology optimization calculation results, manufacturing cost, efficiency, strength performance, and heat resistance performance, with manufacturing cost, efficiency, strength performance, and heat resistance performance selected as the clustering conditions. It is also possible to enter the number of clusters to divide the acquired data into. Here, "10" has been entered as the number of clusters.

[0043] In step S103, information such as the learning model name, variables that are clustering conditions, and the number of clusters that were input in step S102 is acquired.

[0044] In step S104, the information acquired in step S103 is registered in the database 112.

[0045] FIG. 2B is a flowchart showing the clustering process in Phase 1.

[0046] In steps S201 to S205 shown in the figure, the clustering unit 102 classifies the data into clusters, and the clustering result display unit 103 displays the clustering results.

[0047] In step S201, the clustering unit 102 acquires from the database 112 past CAD drawings, topology optimization calculation results, and characteristic information such as manufacturing cost, efficiency, strength resistance, and heat resistance.

[0048] In step S202, the information acquired in steps S101 to S104 is acquired by the clustering unit 102. Here, the name of the learning model, the variables to be clustered, and the number of clusters are acquired.

[0049] In step S203, the clustering unit 102 performs clustering. Various clustering methods have been proposed. Here, clustering is performed using a method called a self-organizing map. A self-organizing map is a type of neural network that models the visual area of ​​the cerebral cortex. In a self-organizing map, weight vectors are randomly placed on a map and one input vector is prepared. The similarity between all weight vectors on the map and the input vector is calculated. Euclidean distance is used to determine the similarity. The vector with the smallest distance is found, and the weight vectors in its vicinity are changed using the following formula (1).

[0050]

number

[0051] In the formula, W u is the weight vector, θ is the neighborhood radius, and α is the learning coefficient. U is the input vector, which contains the variables to be clustered. Also, n represents the number of iterations. In this way, a self-organizing map is used to divide the data into clusters with high similarity. Here, it is divided into 10 clusters.

[0052] In step S204, the clustering result display unit 103 displays the clustering results.

[0053] FIG. 7 is a diagram showing an example of a display screen of the clustering results.

[0054] In this figure, variables to be clustered, such as manufacturing cost, efficiency, strength resistance, and heat resistance, are displayed as clusters with high similarity, ranging from "1" to "10." The table shown in this figure also displays information such as the average values ​​of variables, such as manufacturing cost, efficiency, strength resistance, and heat resistance, for each cluster. This allows the operator to understand the similarity characteristics of each cluster, such as cluster "1" having low manufacturing cost and cluster "3" having high efficiency. If the operator is satisfied with the results, he or she presses the "Decide" button, and if he or she wants to start over, he or she presses the "Redo" button and reviews the clustering conditions from step S101.

[0055] In step S205, the results of the clustering in step S203 are registered in the database 112. Here, the clustering information divided into 10 groups and the information on the average value of each variable are registered.

[0056] FIG. 3 is a flowchart showing the machine learning process in Phase 1.

[0057] The machine learning shown in this diagram is executed by the machine learning control unit 104 controlling the drawing generation unit 105 and the drawing determination unit 106. Although various machine learning methods have been proposed, here machine learning is performed by applying an image processing technique called CycleGAN.

[0058] FIG. 8 is a conceptual diagram illustrating an example of CycleGAN.

[0059] CycleGAN generates new images from images of horses and other objects. For example, using a zebra as the real image, it performs image discrimination to determine whether the newly generated image is the same as an image of a zebra, and then uses machine learning to adjust the image generation parameters so that they are the same. At this point, a new image is generated from a real image of a zebra, and image discrimination is performed to determine whether the newly generated image is the same as an image of a horse, and then machine learning is used to adjust the image generation parameters so that they are the same. These machine learning processes are repeated until they are determined to be the same.

[0060] As such, CycleGAN is characterized by its ability to generate an image that resembles a zebra from an image of a horse or other object. In this case, by using a condition vector, it is possible to generate an image that resembles a zebra or an image that resembles a deer from an image of a horse. For example, if the real image is a zebra, machine learning is performed by inputting the condition vector [1 0 0 0 0], and if the real image is a deer, machine learning is performed by inputting the condition vector [0 1 0 0 0]. If you want to generate an image that resembles a zebra after machine learning, inputting the image of a horse and the condition vector [1 0 0 0 0] will generate an image that resembles a zebra. Furthermore, by inputting the condition vector [1 1 0 0 0], images that resemble both zebras and deer can also be generated.

[0061] A neural network is used to generate images. A neural network is a mathematical model that aims to represent the characteristics of the brain, which consists of a large number of nerve cells, through computer simulation. A neural network uses X-rays to represent each layer of artificial neurons. i Then, it is given by the recurrence formula (2) below.

[0062]

number

[0063] In the formula, Ai is the weight parameter, B i is the bias parameter. f is the activation function. Through machine learning, A i B i In the case of a three-layer network, X1 is the input layer, X2 is the middle layer, and X3 is the output layer. A network with multiple middle layers is called a deep neural network. Deep neural networks are used in CycleGAN image processing.

[0064] In step S301, the machine learning control unit 104 acquires from the database 112 past CAD drawings, the results of topology optimization calculations, the clustering conditions acquired in the step shown in FIG. 2A, and all clustering data acquired in the step shown in FIG. 2B.

[0065] In step S302, one of the clusters clustered by the machine learning control unit 104 is extracted, and the topology optimization calculation result and condition vector belonging to this cluster are input to the drawing generation unit 105 to generate a new drawing. Here, since there are 10 clusters, the topology optimization calculation result and condition vector [1 0 0 0 0 0 0 0 0 0] belonging to cluster number "1" are input to generate a new drawing.

[0066] In step S303, the new drawing generated by the machine learning control unit 104 in step S302 and a drawing related to the input result of the topology optimization calculation are input to the drawing determination unit 106, which determines whether the new drawing and the drawing related to the input result of the topology optimization calculation are the same. If it is determined that they are not the same, the parameters of the drawing generation unit 105 are adjusted. Here, the drawing related to the input result of the topology optimization calculation is used as the "real drawing" for comparison, and is also called the "base drawing."

[0067] In step S304, the machine learning control unit 104 inputs the original drawing used in step S303 and the condition vector into the drawing generation unit 105 to generate a new drawing. Here too, the condition vector [1 0 0 0 0 0 0 0 0 0] is used.

[0068] In step S305, the new drawing generated by the machine learning control unit 104 in step S304 and the result of the topology optimization calculation used in step S302 are input to the drawing determination unit 106, which determines whether the new drawing and the calculation result are the same. If it is determined that they are not the same, the parameters of the drawing generation unit 105 are adjusted. Here, the determination of whether they are the same is made by machine learning. Specifically, this is made using a neural network.

[0069] In step S306, the machine learning control unit 104 performs machine learning on all topology optimization calculation results and drawings for the data belonging to the cluster, repeating steps S302 to S305 until convergence is achieved. Furthermore, similar machine learning is performed on all remaining data in each cluster until convergence is achieved. Taking the example of machine learning data for cluster number "3," machine learning is performed using the condition vector [0 0 1 0 0 0 0 0 0 0]. The convergence criteria are typically entered arbitrarily by the user. For example, if the user inputs a value of "99%," the calculation is considered to have converged and terminated when image discrimination determines the data is "same" with a probability of 99% or higher. If the input value of "99%" is too high and convergence is not achieved, the user can force the calculation to terminate and relax the convergence criteria, for example, by re-entering "95%." An input value of 95% is generally considered appropriate.

[0070] In step S307, information resulting from machine learning by the machine learning control unit 104 is registered in the database 112. Here, the machine learning information in which the parameters of the generated drawing have been adjusted so that it is the same as the drawing to be made authentic is registered in the database.

[0071] This concludes Phase 1.

[0072] Next, phase 2 will be described.

[0073] FIG. 4A is a flowchart showing the topology optimization calculation process in Phase 2.

[0074] In the topology optimization calculation process shown in this figure, the conditions required for the topology optimization calculation are input by the optimization / analysis condition input unit 107, the analysis model generation unit 108 generates an analysis model in accordance with the conditions input by the optimization / analysis condition input unit 107, and the topology optimization calculation unit 109 performs the topology optimization calculation targeting the mechanical structure.

[0075] In step S401, the optimization and analysis condition input unit 107 receives the optimization conditions and analysis conditions input by the operator.

[0076] FIG. 9 is a diagram showing an example of the input screen.

[0077] In this figure, "Hook" has been entered as the analysis model name. An analysis model for topology optimization calculations has been created by the operator. In the calculation domain, which is a rectangular area, a boundary condition constrained by a wall has been applied to the left side of the figure, and a load condition of 1000 N has been applied to the center right side of the figure. In addition, "aluminum" has been entered as the material condition. The topology optimization calculation conditions entered are those that minimize the average compliance, which is expressed as the product of the load and displacement, and "40% or less" has been entered as the volume constraint condition. In other words, the optimization conditions have been set so that the volume of the rectangular area is 40% or less and the stiffness is maximized at that time.

[0078] In step S402, the analytical model generation unit 108 generates an analytical model based on the conditions input in step S401. The calculation domain is divided into mesh-like regions called finite elements, and an analytical model having information on the analytical model, analysis and optimization conditions, etc. is generated.

[0079] In step S403, the topology optimization calculation unit 109 inputs the analysis model generated in step S402 and performs topology optimization calculations. Here, the finite element method is used for the structural analysis portion, and the density method is used for the topology optimization portion. This calculation method assigns a density that varies from 0 to 1 to each finite element as a design variable, and performs optimization by calculating the product of the density and Young's modulus as the Young's modulus of the finite element. Therefore, a density of 0 indicates a state where no structure is present, and a density of 1 indicates a state where a structure is present.

[0080] Fig. 10 shows an example of the results of topology optimization calculation. In the figure, the shading of the divided calculation areas represents the level of density. The darker the area, the higher the density.

[0081] As shown in this figure, the structure of the mechanical structure is expressed by the shading of the calculation domain.

[0082] FIG. 4B is a flowchart showing the drawing generation process in Phase 2.

[0083] In the drawing generation process shown in this figure, the drawing generation condition input unit 110 inputs conditions for generating a new drawing from the results of the topology optimization calculation, the machine learning control unit 104 inputs the information input in the drawing generation condition input unit 110 to the drawing generation unit 105 to generate a new drawing, and the drawing display unit 111 displays the generated drawing.

[0084] In step S501, the machine learning control unit 104 acquires all the information input in the steps shown in FIGS. 2A, 2B, and 3.

[0085] In step S502, the operator inputs, via the drawing generation condition input unit 110, conditions for generating a new drawing from the results of the topology optimization calculation.

[0086] FIG. 11 is a diagram showing an example of an input screen for inputting conditions for creating a new drawing.

[0087] The operator generates a new drawing by entering the analysis model name, learning model name, and drawing generation conditions. In this figure, "Hook" is entered as the analysis model name, and "Machine Structure 1" is displayed as the learning model name.

[0088] In the display area for the topology optimization calculation results, the topology optimization calculation results shown in FIG. 4A are displayed.

[0089] The display area at the bottom of Fig. 11 displays the clustering results shown in Fig. 2B. The operator generates a new drawing by changing the weight of each cluster number between 0 and 1. Cluster number "3" has a weight of "1," cluster number "5" has a weight of "0.5," and cluster number "7" has a weight of "1." The weights of the other clusters are set to "0."

[0090] In step S503, the machine learning control unit 104 inputs the information input in step S502 into the drawing generation unit 105 to generate a new drawing. Here, as a result of the topology optimization calculation for the analysis model name "hook", a condition vector [0 0 1 0 0.5 0 1 0 0 0] in which "1", "0.5", and "1" are entered in the elements corresponding to cluster numbers "3", "5", and "7" is input to the drawing generation unit 105 to generate a new drawing. In this way, a new screen is generated in which the properties of cluster numbers "3", "5", and "7" are possessed in a ratio of 1:0.5:1.

[0091] In step S504, the drawing display unit 111 displays the drawing generated in step S503.

[0092] FIG. 12 is a diagram showing an example of the generated drawing display screen.

[0093] In this figure, the analysis model name "Hook" and the learning model name "Mechanical Structure 1" are displayed. Also displayed are the drawing (proposed drawing) generated in step S503 and the drawing generation conditions entered in step S502.

[0094] In step S505, the drawing generated as a result of the topology optimization calculation shown in Fig. 4A is registered in the database 112. After the drawing is displayed in step S504, the process may return to step S502 as necessary to repeat input by the operator.

[0095] In this way, past CAD drawings and the results of topology optimization calculations are divided into clusters with high similarity based on characteristic information such as manufacturing cost and efficiency, and new drawings are generated from the results of topology optimization calculations by machine learning for each cluster. This allows for the generation of shapes that are not dependent on the engineer's experience, as opposed to shapes that engineers previously predicted based on the density shades obtained from topology optimization calculations based on their experience. Furthermore, new drawings can be generated that take into account characteristics other than the objective function of the optimization calculation, such as manufacturing cost and efficiency. Furthermore, by applying condition vectors, drawings that take into account multiple characteristics such as manufacturing cost and efficiency can be generated in a short period of time.

[0096] As described above, according to the present disclosure, high-quality drawings can be obtained in a short time.

[0097] In this disclosure, clustering, machine learning, and topology optimization calculations are described as being performed on the same computer. However, they can be performed on different computers by using a network environment. That is, the design support method according to the present disclosure may be implemented by configuring a cloud in which multiple servers installed in different locations are linked via the Internet. In this case, the multiple servers can be said to constitute a design support system. [Explanation of symbols]

[0098] 101: clustering condition input unit, 102: clustering unit, 103: clustering result display unit, 104: machine learning control unit, 105: drawing generation unit, 106: drawing discrimination unit, 107: optimization / analysis condition input unit, 108: analysis model generation unit, 109: topology optimization calculation unit, 110: drawing generation condition input unit, 111: drawing display unit, 112: database, 113: computer.

Claims

1. A design support device that generates a proposed drawing from a calculation result of structural optimization that derives an optimized density distribution for a structure, a clustering condition input unit that accepts input of clustering conditions; a clustering unit that performs clustering according to the clustering conditions; a machine learning control unit that performs machine learning on each of the clusters obtained by the clustering; a drawing generation unit that generates a new drawing from the result of the machine learning; a drawing discrimination unit that compares the basic drawing to be compared with the new drawing and determines whether these drawings are the same; an optimization / analysis condition input unit that receives input of calculation conditions necessary for the structural optimization calculation; an analytical model generation unit that generates an analytical model in accordance with the calculation conditions; a structural optimization calculation unit that performs the calculation for the structural optimization using the analytical model; a drawing generation condition input unit that displays the calculation results of the structural optimization and the clustering results and receives input of weights for each of the clusters required for drawing generation; a display unit that displays information including at least one of the clustering conditions, the calculation results, and the generated drawing, the drawing generation unit generates a proposed drawing using the calculation result of the structural optimization and the weight; The display unit displays the proposed drawing.

2. The computer aided design system according to claim 1 , wherein the clustering conditions include at least one of a manufacturing cost and an efficiency of the structure.

3. The design support system according to claim 1 , wherein the display unit displays information including characteristic information of each of the clusters.

4. The design support system according to claim 1 , wherein the calculation of the structural optimization uses a finite element method and a density method.

5. The computer-aided design support system according to claim 1 , wherein the clustering is performed using a self-organizing map.

6. The design support system according to claim 1 , wherein the structural optimization is topology optimization.

7. The design support device according to claim 6 , wherein the drawing generation unit generates the proposed drawing through the machine learning.

8. A design support method for generating a proposed drawing from a calculation result of structural optimization that derives an optimized density distribution for a structure, comprising: a clustering condition input unit that receives input of clustering conditions; a clustering unit performing clustering in accordance with the clustering conditions; a machine learning control unit that performs machine learning on each of the clusters obtained by the clustering; a drawing generation unit that generates a new drawing from the results of the machine learning; A drawing determination unit compares the basic drawing to be compared with the new drawing and determines whether these drawings are the same; an optimization / analysis condition input unit that receives input of calculation conditions necessary for the structural optimization calculation; an analytical model generation unit generates an analytical model in accordance with the calculation conditions; a structural optimization calculation unit that performs the calculation for the structural optimization using the analytical model; a drawing generation condition input unit that displays the calculation results of the structural optimization and the clustering results, and receives input of weights for each of the clusters required for drawing generation; a display unit displays information including at least one of the clustering conditions, the calculation results, and the generated drawing; the drawing generation unit generates a proposed drawing using the calculation result of the structural optimization and the weight; The display unit displays the proposed drawing.

9. The design support method according to claim 8 , wherein the clustering conditions include at least one of a manufacturing cost and an efficiency of the structure.

10. The design support method according to claim 8 , wherein the display unit displays information including characteristic information of each of the clusters.

11. The design support method according to claim 8 , wherein the calculation of the structural optimization uses a finite element method and a density method.

12. The design support method according to claim 8 , wherein the clustering is performed using a self-organizing map.

13. The design support method according to claim 8 , wherein the structural optimization is topology optimization.

14. The design support method according to claim 13 , wherein the drawing generation unit generates the proposed drawing by the machine learning.

Images