Three-dimensional model generation device, three-dimensional object formation system, three-dimensional object formation device, three-dimensional model generation method, three-dimensional model generation program, and recording medium

Abstract

A three-dimensional model generation device 1 comprises a training unit 11 and a generation unit 12. The training unit 11 constructs an estimation model for estimating a three-dimensional object by performing machine learning using training data including structural mechanics parameters and three-dimensional objects. The generation unit 12 generates a three-dimensional object using the estimation model constructed by the training unit 11.

Description

3D model generation apparatus, 3D object formation system, 3D object formation apparatus, 3D model generation method, 3D model generation program, and recording medium

[0001] This application claims priority to Japanese Patent Application No. 2024-160422, filed on 17 September 2024, which is incorporated herein by reference in its entirety.

[0002] This disclosure relates to a three-dimensional model generation apparatus, a three-dimensional object formation system, a three-dimensional object formation apparatus, a three-dimensional model generation method, a three-dimensional model generation program, and a recording medium.

[0003] In recent years, not only image generation AI like StableDiffusion, but also 3D shape generation technologies using AI have been developed (see, for example, Non-Patent Document 1).

[0004] Alex N, Heewoo J, Prafulla D, Pamela M and Mark C:Point-e: A system for generating 3d point clouds from complex prompts, arXiv:2212.08751, 2022.JJ Park, et. al. DeepSDF: Learning Continuous Signed Distance Functions for Shape Representation, CVPR, pp.165-174, 2019William E. Lorensen and Harvey E.Cline: Marchingcubes: A high resolution 3d surface construction algorithm, ACM SIGGRAPH Computer Graphics, pp.163-169, 1987Nakahashi.K: Building-cube method for flow problems with broadband characteristic length, International Conference on Computational Fluid Dynamics, pp.77-81, 2022.

[0005] However, the technologies described in Non-Patent Document 1 are all aimed at the fields of computer graphics and computer vision, and are heavily reliant on the mapping between language and images. For this reason, it is difficult to apply them to cases where there are no existing large-scale datasets or multimodal datasets, such as structural mechanical information. Currently, there are no 3D object estimation models that take structural mechanical information into account. In this specification, a 3D object estimation model refers to a model (parameter-to-3D model) that takes structural mechanical parameters as input and estimates a 3D object that satisfies those structural mechanical parameters.

[0006] The technology described herein was developed in view of these circumstances, and its purpose is to provide a technology for constructing a three-dimensional object estimation model that satisfies structural mechanical parameters by specifying those structural mechanical parameters.

[0007] To solve the above problems, a three-dimensional model generation apparatus according to one aspect of the present invention includes a learning unit that constructs an estimation model for estimating the shape of a three-dimensional object by performing machine learning using training data that includes structural mechanical parameters and three-dimensional object data.

[0008] Another aspect of this disclosure is a three-dimensional object forming system. This system comprises a learning unit that constructs an estimation model for estimating the shape of a three-dimensional object by performing machine learning using training data including structural mechanical parameters and three-dimensional object data; a generation unit that generates a three-dimensional object using the estimation model; and a forming unit that forms a three-dimensional object using the three-dimensional object generated by the three-dimensional model generation device.

[0009] A further aspect of the present invention is a three-dimensional object forming apparatus. This apparatus comprises a learning unit that constructs an estimation model for estimating a three-dimensional object by performing machine learning using training data including structural mechanical parameters and three-dimensional object data; a generation unit that generates a three-dimensional object using the estimation model; and a forming unit that forms a three-dimensional object using the three-dimensional object.

[0010] Yet another aspect of the present invention is a method for generating a three-dimensional model. This method includes a learning step of constructing an estimation model for estimating the shape of a three-dimensional object by performing machine learning using training data which includes structural mechanical parameters and three-dimensional object data.

[0011] Furthermore, any combination of the above components, as well as conversions of the expressions of this disclosure between methods, apparatus, systems, recording media, computer programs, etc., are also valid forms of this disclosure.

[0012] According to this disclosure, by specifying structural mechanical parameters, a three-dimensional object estimation model that satisfies those structural mechanical parameters can be constructed.

[0013] This is a functional block diagram of a 3D model generation device according to an embodiment. This is a flowchart showing the processing procedure of a 3D model generation method according to an embodiment. This is a schematic diagram showing the principle of shape representation using DeepSDF. This is a schematic diagram showing the learning of SDF values ​​using DeepSDF. This is a graph showing an example of the objective function history. This is a diagram showing load conditions and boundary conditions for topology optimization. This is a diagram showing the object at the end of optimization when arbitrary load conditions are set. This is a schematic diagram showing a network that allows control of the output shape using structural mechanical parameters based on the DeepSDF network. This is a diagram showing the loss progression of training data and validation data. This is a diagram showing the reconstruction results of 3D objects with and without positional encoding. This is a diagram showing the results of comparing the distribution of the error in reconstructing test data relative to training data using a histogram with Chamfer-L1 Distance. This is a diagram showing an example of a 3D object reconstructed using test data. This is a diagram showing the specific flow of validation 2. This is a diagram showing the strain energy before input to the network and the strain energy output from the network. This figure shows an example of a 3D object reconstructed based on data containing outliers.

[0014] Preferred embodiments will be described below with reference to the drawings. The same or equivalent components, members, and processes shown in each drawing will be denoted by the same reference numerals, and redundant descriptions will be omitted as appropriate. Furthermore, the embodiments are illustrative and not limiting to the invention, and not all features or combinations thereof described in the embodiments are necessarily essential to the invention.

[0015] Furthermore, the dimensions (thickness, length, width, etc.) of each component shown in the drawings may be enlarged or reduced as appropriate for ease of understanding. Moreover, the dimensions of multiple components do not necessarily represent their relative sizes; even if component A is depicted as thicker than component B in the drawing, component A may actually be thinner than component B.

[0016] The terms used in this specification will be described below. In the following, the terms will be described using specific examples, but it goes without saying that the terms are not limited to these specific examples. In fact, note that the terms used in this specification widely include concepts within the range commonly used in the technical field to which this disclosure pertains, even if they do not apply to these specific examples.

[0017] In this specification, "statistical features" mainly refer to distributive and summary quantities derived from three-dimensional object data, the shape of a three-dimensional object (including surface / volume representation), or structural mechanics parameters (strain energy, each component of the load vector, volume, height, etc.). Statistical features include, but are not limited to, mean, variance, skewness, kurtosis, percentile, histogram frequency, cumulative distribution, distribution of distances and errors (e.g., statistical quantities of Chamfer-L1 Distance), distribution statistics of geometric quantities such as surface curvature and thickness, and spatial statistics of voxel density (including the density ρ of SIMP) and SDF values. These statistical features can be used for learning, verification, and evaluation of machine learning, and can be global feature quantities obtained by aggregating local quantities such as individual points, surfaces, voxels, etc., or local statistics calculated for each region.

[0018] In this specification, "outer shape" mainly refers to the geometric shape forming the outer boundary of a three-dimensional object. For example, in the SDF representation, it refers to the object surface specified by the isosurface of SDF(·)=0 (including the polygon mesh extracted by the marching cubes method, etc.), and includes surface geometries such as contour, silhouette, surface roughness, surface normal, and local curvature. The outer shape is conceptually distinguished from the internal structure such as internal voids and ribs, but the measurement of the outer shape may include derived quantities such as bounding box dimensions, volume, surface area, circumscribed sphere / inscribed sphere radius, aspect ratio, etc.

[0019] In this specification, the "internal structure" mainly refers to the arrangement, connectivity, thickness, internal patterns such as ribs / webs / honeycombs / lattices / channels / cavities, etc., and their geometric and topological relationships within the 3D object. In the SDF representation, it includes the composition of the regions where SDF(·) < 0 (material side) and SDF(·) > 0 (void side), the material distribution based on SIMP or density field ρ(x), and further includes the presence or absence of connected components and through channels in terms of elements / voxels. The internal structure is defined or generated according to structural mechanics parameters such as load conditions, boundary conditions, and volume fraction, and is determined by optimization or the output of the generation model. The evaluation indicators of the internal structure may include minimum thickness, connectivity, porosity, stress concentration indicators, and flow path continuity, etc.

[0020] In this specification, the "structural mechanics parameters" broadly include physical quantities that define or describe the mechanical behavior and characteristics of 3D objects, as well as design conditions. That is, the structural mechanics parameters include, for example, physical property values such as strain energy, stress, stiffness, natural frequency, buckling load, and heat conduction characteristics, and load conditions, boundary conditions, material properties (Young's modulus, etc.), design constraint conditions (volume, weight, etc.).

[0021] The structural mechanics parameters are not limited to those obtained from linear static analysis. That is, the structural mechanics parameters may include physical quantities and conditions obtained from dynamic analysis (e.g., impact energy absorption in collision analysis, maximum reaction force, natural frequency in vibration analysis, etc.), non-linear analysis (e.g., stress considering material non-linearity or geometric non-linearity), or coupled analysis (e.g., heat conduction characteristics and fluid resistance obtained from structural-fluid-thermal coupled analysis).

[0022] The following describes a 3D model generation device and a 3D model generation method for generating 3D objects. Here, "3D object" broadly includes not only 3D structures with physical existence, but also the data itself in CAD software such as STL data, or 3D shapes generated using such data by CAD or 3D printers. "3D object data" refers to data about 3D objects that serve as training data for machine learning; for example, the volume and height of a 3D object fall into this category. "3D object shape" is the target of estimation in the estimation model constructed by machine learning, and is an abstract shape that characterizes the 3D object.

[0023] [3D Model Generation Apparatus] Figure 1 is a functional block diagram of a 3D model generation apparatus 1 according to an embodiment of the present disclosure. The 3D model generation apparatus 1 comprises a learning unit 11 and a generation unit 12.

[0024] The learning unit 11 constructs an estimation model for estimating the shape of a 3D object by performing machine learning using training data that includes structural mechanical parameters and 3D object data.

[0025] The generation unit 12 generates a 3D object using the estimation model constructed by the learning unit 11. For example, the generation unit 12 can generate a 3D object by providing a trained decoder with latent vectors randomly obtained from a Gaussian distribution. In this case, for example, the latent vectors may be sampled from a multivariate normal distribution with mean 0 and a predetermined covariance.

[0026] Furthermore, the 3D model generation apparatus of this embodiment can solve the problems of this disclosure if it can construct an estimation model for estimating the shape of a 3D object, so it is not essential to include the generation unit 12.

[0027] [3D Model Generation Method] Figure 2 is a flowchart showing the processing procedure of a 3D model generation method according to an embodiment. This 3D generation method includes a learning step S1 and a generation step S2.

[0028] In this method, in learning step S1, an estimation model for estimating a 3D object is constructed by performing machine learning using training data that includes structural mechanical parameters and a 3D object.

[0029] In this method, a 3D object is generated in generation step S2 using the estimation model constructed in learning step S1. For example, in generation step S2, a 3D object can be generated by providing a trained decoder with latent vectors randomly obtained from a Gaussian distribution. In this case, for example, the latent vectors may be sampled from a multivariate normal distribution with mean 0 and a predetermined covariance.

[0030] Furthermore, the 3D model generation method of the embodiment can solve the problems of this disclosure if an estimation model for estimating 3D objects can be constructed, so it is not essential to include the generation step S2.

[0031] Below, we will take DeepSDF as an example of machine learning in the embodiment and explain in detail the operation of the 3D model generation device and the processing of the 3D model generation method.

[0032] [DeepSDF] DeepSDF is a decoder-type neural network for modeling three-dimensional objects (see, for example, Non-Patent Document 2). DeepSDF describes the shape as a decision boundary of a trained feedforward network in order to represent the SDF (Signed Distance Function). Figure 3 schematically shows the principle of shape representation by DeepSDF. As shown in Figure 3, the SDF is a continuous function that outputs the distance from a given point (hereinafter referred to as "spatial sample point") to the nearest surface for given spatial coordinates in space. However, its sign is encoded as negative when the spatial sample point is inside the target object, positive when it is outside, and 0 when it is on the boundary. The SDF is expressed as follows, for spatial coordinate x∈R 3 Let SDF(x) = s: x ∈ R be a scalar. 3 , s∈R

[0033] According to this embodiment, by specifying structural mechanical parameters, it is possible to realize a device that generates a three-dimensional object that satisfies those structural mechanical parameters.

[0034] Here, the object surface is represented by isosurfaces with SDF(•)=0. A polygon mesh of this object interface can be obtained using the marching cube method (see, for example, Non-Patent Document 3). DeepSDF uses a deep neural network to predict SDF values ​​from spatial sample points. That is, a trained deep neural network can predict the SDF value of any spatial sample point, and thus can generate isosurfaces with an SDF value of 0.

[0035] The right-hand figure in Figure 3 shows a 3D object reconstructed by DeepSDF. As shown here, DeepSDF learns a continuous field in space that corresponds to the distance from the object's surface.

[0036] Figure 4 schematically illustrates the learning of SDF values ​​using DeepSDF. Figure 4(a) shows the training of a single neural network for a specific target shape. Here, when the coordinates (x, y, z) of spatial sample points are input to the neural network, SDF values ​​are output. Specifically, given a target shape, a set of pairs X consisting of the coordinates (x, y, z) of spatial sample points and their SDF values ​​is prepared. X:={(x,s):SDF(x)=s} In this embodiment, in the training set S, a number of fully connected neural networks f are trained to appropriately approximate the SDF given within the target domain Ω. θ Train the parameter θ.

[0037] Training is performed by minimizing the sum of losses between the predicted SDF values ​​and the actual SDF values ​​of spatial sample points in X, according to the following L1 loss function: L(f θ (x),s)=|clamp(f θ(x), δ)-clamp(s, δ) | Here, clamp(x, δ) = min(δ, max(-δ, x)), and the stability of learning is achieved by controlling the maximum and minimum values of the SDF input to the neural network using the parameter δ. In this embodiment, δ = 0.1.

[0038] [Learning in the Latent Vector Space] Fig. 4(a) assumes learning a single neural network for a specific one target shape. That is, in this method, for multiple target shapes, it is necessary to train neural networks individually. Although this is feasible, it is not practically beneficial. Therefore, a model that can represent various target shapes, discover common characteristics, and embed them in a low-dimensional latent space is desirable. Thus, to be able to represent various object surfaces with one model, a latent vector L that distinguishes individual objects is used as an input. Fig. 4(b) shows the state of learning using the latent vector. Here, in addition to the coordinates of the spatial sample points as the first input, the latent vector as the second input is input to the neural network. The latent vector is stored together with query information at the spatial sample points of each target shape. Formally, for a certain target shape i, f θ is the latent vector L ｉ and is a function of the query position x in 3D space, and outputs the SDF of the target shape. f θ (L i , x) ~ SDF i (x)

[0039] By conditioning the output of the network by this latent vector, this formulation enables modeling of multiple objects with a single neural network. When the decoder model f θ is given, the continuous surface related to the latent vector L is similarly represented by the zero等值面 of f θ (L, x). The polygon mesh of the object is obtained by the aforementioned marching cube method.

[0040] [Formulation of an Autodecoder-based DeepSDF] In formulating the autodecoder-based DeepSDF, we perform regularization of the latent vector, which was also used in the original DeepSDF. This is shown to approximate a multivariate Gaussian distribution with a mean of 0, based on the interpretation of the decoder in DeepSDF as a probabilistic model. There is a dataset consisting of N target shapes, each representing a signed distance function SDF. i i=1 N Let it be represented as follows. Prepare K spatial sample points and their signed distance values. X i ={(x j ,s j ):s j =SDF i (x j )}

[0041] In the case of an autodecoder, there is no encoder, so each latent vector z ｉ This is the training shape X ｉ This pairs with the set of spatial sample points X representing the target shape. ｉ latent vector z ｉ The posterior probability of is decomposed as follows:

[0042] Here, θ is parameterized as the likelihood of the SDF. In the latent vector space, the prior distribution p(z) of the latent vectors ｉ ) with mean 0 and variance σ ２ We assume a multivariate Gaussian distribution. This prior distribution is based on the idea that, ideally, the latent vectors should follow an unbiased and well-behaved distribution. This assumption is adopted in this specification because it has been empirically shown to be necessary for convergence to a good solution. In the autodecoder-based DeepSDF formulation, the likelihood of the SDF is expressed via a deep FFNN (Feed Forward Neural Network), and to the extent that generality is not lost, we assume that the likelihood takes the following form: p θ (s j |z i :x j )=exp(-L(f θ (zi ,x j ),s j ))

[0043] SDF predicted value s ~ j =f θ (z i ,x j ),s j ) is represented using a network. L(s ～ ｊ s ｊ ) shows the network's predicted and actual SDF values ​​s ｊ This is a loss function to bring the two values ​​closer together. In this specification, a loss function is used that assumes a Gaussian distribution for the SDF values. During training, the individual latent vectors {z ｉ} ｉ＝１ Ｎ And with respect to the network parameter θ, we maximize the common posterior log probability across all training shapes. From the above, the loss function in this specification is expressed by the following equation. The first term of this equation is the loss of the SDF value, and the second term is the loss related to the distribution of the latent vector, ||z|| ２ ２ Now we are calculating the L2 norm of the latent vector z.

[0044] During inference, θ is fixed after training. Target shape X ｉ Shape code L ｉ This can be expressed using Maximum A Posteriori (MAP) estimation as follows:

[0045] This formulation is valid for spatial sample points of arbitrary size and distribution, and the gradient of z with respect to the loss can be calculated individually for each spatial sample point. This allows DeepSDF to handle the shape of partial observations, such as depth maps.

[0046] [Dataset Generation] In this specification, to create a 3D object dataset with a structural mechanical basis, topology optimization targeting small deformation problems of linear elastic bodies using the Building-Cube method (see, for example, Non-Patent Document 4) is employed. Topology optimization is one of the optimization methods that determines the arrangement of members and materials, including voids, to maximize or minimize the objective function within the allowed design space. By using the Building-Cube method, topology optimization that achieves the large area, high resolution, and low volume constraints required in applications such as automotive structural design becomes possible. The creation of a dataset by topology optimization will be explained in detail below.

[0047] [Linear topology optimization method based on the Building-Cube method] Topology optimization is performed with the aim of maximizing stiffness, that is, minimizing compliance. Under the conditions, ρ minimizes Ω = u・f. ｊ We find the solution where u is the displacement solution, ρ is the density, and k ｊ is the stiffness matrix of element j. Here, we use the SIMP method (solid isotropic material with penalization), which allows for an intermediate density state between voids (ρ=0) and the base material within the element, and determines the final density state by density redistribution.

[0048] Figure 5 shows an example of the objective function history. In this example, the calculation is terminated after 50 iterations because it is determined that the objective function has sufficiently converged. When the objective function decreases monotonically and then converges to a constant value, it can be determined that normal optimization has been performed on the objective function.

[0049] [Calculation of 3D Shape using Linear Topology Optimization] Figure 6 shows the load and boundary conditions for topology optimization in one example. In this example, the design domain is a rectangular prism, the dimensions are fixed, and the topology that emerges is changed by varying the load direction to create a dataset.

[0050] The load direction was represented by a three-component vector. The x-component was given in the range of 1 to 0.5 in increments of 0.1, the y-component in the range of 0 to 1 in increments of 0.1, and the z-component in the range of -1 to 0 in increments of 0.01. By removing overlaps of identical shapes due to the same load direction, 6667 shapes were created. Of the total data, 300 were used as validation data, 300 as test data, and the remainder as training data. The structural mechanical parameters obtained by topology optimization and forward analysis consist of six components: strain energy in the small deformation region, each component of the load vector (3D), volume, and height (1D). During training with DeepSDF, 3D shape training data including structural mechanical parameters was created by pairing it with the shape data. The conditions used for topology optimization at this time are summarized below.

[0051] Material: Aluminum Density: 2.7 g / cm³ ３ Young's modulus: 70 GPa Poisson's ratio: 0.3 Material model: Linear elastic material Volume fraction: 5% Minimum size: 1.5 mm Number of cells: 1,048,576 Number of nodes: 8 Optimal number of steps: 50 Computation time: 20 minutes Number of datasets: 6667 Number of training data: 6067 Number of validation data: 300 Number of test data: 300

[0052] Figure 7 shows the shape at the end of optimization when arbitrary load conditions are set. Isosurfaces were created with a threshold of ρ=0.5, which is common for density distributions. It can be seen that the shape changes significantly with respect to the load direction, indicating that a variety of shapes for training are ensured.

[0053] In the example above, 3D shape training data was created by providing numerical parameters for load and boundary conditions for topology optimization. However, this is not the only way; the training data may also be data obtained from actual measurements. For example, in the case of the external shape of an object, the training data may be based on data obtained from a 3D scanner, photogrammetry, CAD, etc. Alternatively, in the case of the internal structure of an object, the training data may be based on data obtained from X-ray CT, MRI, ultrasonic testing, etc.

[0054] Alternatively, the training data may include, for example, a set of three-dimensional object data generated by using a parametric CAD model and varying its design parameters (e.g., rib thickness, hole diameter, etc.), and the corresponding set of structural mechanical parameters (e.g., simulation results).

[0055] When using data obtained from actual measurements as training data, it may be advantageous to use data outside the actual measurement range. In such cases, data from unobserved regions can be created through extrapolation and applied to the training data. In this case, the validity of the extrapolation can be further enhanced by explicitly defining the training space boundary or by using it in conjunction with physical constraints.

[0056] The 3D objects generated using the method disclosed herein can be exported in STL / OBJ format, used for CAE (Computer-Aided Engineering) analysis, and integrated with external systems via on-premise or cloud APIs, allowing for widespread deployment and dissemination to users and related systems. Thus, the method disclosed herein can be easily and quickly deployed in various industrial fields.

[0057] Furthermore, structural mechanical parameters are not limited to those obtained from linear static analysis. In other words, structural mechanical parameters may include physical quantities and conditions obtained from dynamic analysis (e.g., impact energy absorption in collision analysis, maximum reaction force, natural frequency in vibration analysis), nonlinear analysis (e.g., stress considering material nonlinearity and geometric nonlinearity), or coupled analysis (e.g., thermal conductivity characteristics and fluid resistance obtained from structural-fluid-thermal coupled analysis).

[0058] [Lipschitz Regularization] The learning unit 11 may perform Lipschitz regularization when learning the estimation model. For example, the learning unit 11 may constrain the Lipschitz constant of the network by applying spectral normalization to the weight matrix of the neural network, or by adding a gradient penalty term to the loss function. This makes the change in the output (SDF value) in response to input changes in the latent space smoother, enabling geometrically consistent and stable shape generation in shape interpolation and extrapolation.

[0059] [Neural Network Based on DeepSDF Considering Structural Mechanical Parameters] Figure 8 schematically shows a network based on the DeepSDF network that allows control of the output shape using structural mechanical parameters. The values ​​input to the decoder network are a latent vector (313 dimensions) randomly obtained from a Gaussian distribution, the coordinates of spatial sample points (60 dimensions), and structural mechanical parameters (6 dimensions) associated with the shape. Here, the coordinates of the sample points are expanded from 3 dimensions to 60 dimensions using the Positional encoding method. Details of Positional encoding will be described later. The values ​​of the structural mechanical parameters are normalized to between 0 and 1.

[0060] The number of spatial sample points was set to 30,000 for both training and reconstruction. To prevent overfitting, a feedforward neural network (FFNN) consisting of eight fully connected layers with 20% dropout applied to each layer was used. All internal layers were 512-dimensional, and ReLU was used as the activation function.

[0061] By adding information about the coordinates of spatial sample points and structural mechanical parameters to the input layer and inserting it into each layer, the network can be made more efficient at learning the relationship between shape and structural mechanical parameters. The output layer of the network outputs the SDF value at the spatial sample point. The error between the output SDF value and the SDF value of the 3D object in the training data is calculated, and the network parameters are updated using the Adam optimization algorithm (learning rate η=0.001).

[0062] Figure 9 shows the loss progression for both the training and validation data. Here, training is stopped at 80 epochs, when no further decrease is observed from the loss value in the validation data.

[0063] [Positional encoding] Positional encoding is a 3D vector of spatial sample points x=(x,y,z) T This is converted into a 60-dimensional Fourier feature. If the positional encoding is the function φ:R3→R60, the specific conversion formula is expressed as follows.

[0064] By using positional encoding, it is expected that the accuracy of reconstructing geometrically complex object interfaces can be improved.

[0065] Figure 10 shows the reconstruction results of a 3D object with and without positional encoding. Figure 10 evaluates the spatial resolution of the model in this embodiment by how accurately the original training shape can be reconstructed after training on one training shape. As shown in Figure 10, it can be seen that by using positional encoding, even the fine details of the object can be reconstructed, and the spatial resolution can be improved.

[0066] The reliability of the estimation results for three-dimensional objects generated using the method disclosed herein can be verified by performing simulations or by materializing the generated three-dimensional objects using a 3D printer or the like.

[0067] Furthermore, the generation unit 12 (or a separately provided uncertainty evaluation unit) may, along with generating 3D objects, quantify and present the uncertainty of the generation results. Specifically, the uncertainty may be quantified using methods such as prediction variance estimation, ensemble methods, or MC dropout (Monte Carlo Dropout), and presented as a numerical confidence level, or regions with low confidence levels may be visualized.

[0068] In another embodiment of this disclosure, the 3D model generation device 1 may further include an exploration control unit (AI agent). The exploration control unit autonomously adjusts the structural mechanical parameters input to the generation unit 12 (or a trained estimation model) based on a predetermined design objective and explores the design space.

[0069] The exploration control unit may receive feedback from the performance evaluation of the generated 3D object (e.g., simulation results from cooperating CAE software, or output from a separately established surrogate model) and iteratively adjust parameters to approach the design objective. The exploration control unit may be implemented using, for example, reinforcement learning, Bayesian optimization, or evolutionary algorithms. Alternatively, the exploration control unit may be configured as a multi-agent system in which multiple exploration control units cooperate or compete.

[0070] The exploration control unit (or generation unit 12) may, when given multiple conflicting design objectives (e.g., maximizing stiffness and minimizing weight), perform multi-objective optimization to search for a set of optimal design options with trade-off relationships (a Pareto optimal solution group). In particular, in this case, the solution group may be efficiently derived by intensively searching the vicinity of a Pareto surface (or Pareto front) in the latent space. The derived Pareto optimal solution group may be presented visually so that the designer can compare and consider them.

[0071] The 3D model generation device 1 may further include a parameter conversion unit. The parameter conversion unit may interpret design instructions in natural language input from the user (for example, instructions such as "I want the impact absorption amount to be 10 kJ and the maximum reaction force to be 120 kN") using a large-scale language model (LLM) or the like, automatically extract or set the corresponding structural mechanical parameters, and input them to the generation unit 12 (or exploration control unit).

[0072] The 3D model generation device 1 may further include an explanation generation unit that analyzes the characteristics and performance of the generated 3D object and generates an explanation in natural language. In this case, the user can interactively iterate on the shape by providing feedback in natural language on the generation results and explanations, and the parameter conversion unit interpreting this feedback and readjusting the parameters.

[0073] The learning unit 11 or the generation unit 12 may also have a function to visualize the learned latent space. For example, the learning unit 11 or the generation unit 12 may map the latent space to two or three dimensions using a dimensionality reduction method such as t-SNE (t-distributed Stochastic Neighbor Embedding) or UMAP, and present the overall picture of the design space and the data distribution to the user.

[0074] In the embodiments described above, the system included a "learning unit" or "learning step" for training and building an AI model. However, it is not limited to this, and in other embodiments, the shape of a 3D object may be generated using a pre-built, trained model.

[0075] Specifically, a three-dimensional model generation apparatus of one embodiment includes a storage unit that stores a trained estimation model for estimating the shape of a three-dimensional object, which is constructed by performing machine learning using training data that includes structural mechanical parameters and three-dimensional object data, and a generation unit that generates the shape of a three-dimensional object by inputting structural mechanical parameters into the trained estimation model.

[0076] In this embodiment, the generation unit may calculate the confidence level of the estimation model by predictive variance, ensemble, or MC dropout.

[0077] Alternatively, in this embodiment, the generation unit may sample latent vectors from a multivariate normal distribution and generate a three-dimensional object based on these latent vectors and structural dynamic parameters.

[0078] Furthermore, in this embodiment, the generation unit may export the generated 3D object to STL / OBJ and link with an external system via CAE analysis or a cloud API.

[0079] One embodiment of a 3D model generation method includes a generation step of generating the shape of a 3D object from input structural mechanical parameters using a trained estimation model that estimates the shape of a 3D object by performing machine learning on training data including structural mechanical parameters and 3D object data.

[0080] In one embodiment, a 3D model generation program causes a computer to perform a generation step in which it generates the shape of a 3D object from input structural mechanical parameters, using a trained estimation model that estimates the shape of a 3D object by performing machine learning on training data including structural mechanical parameters and 3D object data.

[0081] One embodiment of the recording medium records a 3D model generation program characterized by causing a computer to execute a generation step in which it generates the shape of a 3D object from input structural mechanical parameters, using a trained estimation model that estimates the shape of a 3D object by performing machine learning using training data including structural mechanical parameters and 3D object data.

[0082] According to this embodiment, even if the system has not performed training itself using a learning unit or training steps, it is possible to generate the shape of a 3D object using a pre-built, trained model.

[0083] [Verification] To confirm the effectiveness of this disclosure, the inventors performed two types of verification: Verification 1 and Verification 2. Verification 1 verifies whether the trained model of the embodiment can accurately reconstruct the SDF values ​​of the input training shape and test shape, and the optimized latent vector and structural mechanical parameters based on those values. This is a verification to confirm the generalization performance of the model of the embodiment. Verification 2 verifies the structural mechanical performance of an object generated from structural mechanical parameters and a random latent vector with a normal distribution. This is a verification to confirm whether the model of the embodiment can generate an object that satisfies predetermined structural mechanical performance. Strain energy at small deformations was used for performance comparison.

[0084] (Verification 1: Verification of the accuracy of reconstructing the training and test shapes) First, we evaluate the model's ability to represent the training shape from a latent vector of a predetermined number of dimensions and structural mechanical parameters. This verifies the model's expressive power. The input values ​​to the model are structural mechanical parameters and latent vectors optimized based on the corresponding SDF values ​​of the shapes. The number of dimensions of the latent vectors is 313, and the Chamfer-L1 Distance formula below was used for a quantitative comparison of reconstruction accuracy. This method quantitatively evaluates shape error by calculating the difference in coordinates of each vertex in both directions for two sets of triangular meshes and summing the values ​​of the minimums. The average value of the 6067 training data points was 0.0187 m.

[0085] Next, we evaluated the ability to make predictions for unknown shapes, i.e., test data. As with the above, for quantitative comparison, using Chamfer-L1 Distance, the average value for 300 test data points was 0.0210m.

[0086] Figure 11 shows the results of comparing the distribution of errors in reconstructing the test data relative to the training data using a histogram with Chamfer-L1 Distance. From these results, it can be confirmed that the reconstruction of the test data was achieved with a similar level of accuracy to that of the training data.

[0087] Figure 12 shows an example of a 3D object reconstructed using the test data described above. Figure 13 shows four examples, in which the original object is on the left and the reconstructed object is on the right. From these results, it can be seen that while the overall object is reproduced in the reconstructed object compared to the original object, the surface irregularities are not reproduced sufficiently, which is thought to have led to a decrease in reconstruction accuracy. However, since there is an advantage in that stress concentration is less likely to occur due to the reduced surface irregularities, this result can be said to be rather desirable from an engineering standpoint.

[0088] (Verification 2: Verification of generated objects and accuracy using mechanical parameters) Here, we verify whether it is possible to generate objects that reflect the specified mechanical parameters using test data. The difference from the reconstruction of the test data shape in Verification 1 is that instead of using latent vectors optimized from SDF values, we use latent vectors randomly obtained from a Gaussian distribution. The reason for this will be explained at the end of this chapter. As above, strain energy in the small deformation region is used as a quantitative indicator.

[0089] Figure 13 shows the specific flow of Verification 2. The strain energy obtained by linear strain analysis was verified to match the input strain energy of an object generated by feeding structural mechanical parameters, spatial sample points, and latent vectors randomly obtained from a Gaussian distribution within the test data to a trained decoder. The direction of the load used in the linear strain analysis was the same as the load direction given as the input value for the structural mechanical parameters.

[0090] The verification results are shown below for various accuracy levels. Number of test data: 300 Average accuracy: 88.8% Median accuracy: 91.2% Highest accuracy: 99.8% Lowest accuracy: 30.2%

[0091] As shown in these results, the average accuracy of the 300 test data points was 88.8%, and the median accuracy for each value was 91.2%, demonstrating that the model in this embodiment has high accuracy in reflecting structural mechanical parameters.

[0092] One possible reason why the median accuracy is higher than the mean accuracy is that the influence of outliers is less. Figure 14 shows the strain energy before input to the network and the strain energy output from the network. Below are the required accuracy and the number and percentage of data that satisfy it. Required Accuracy Number Percentage (%) 95% ≥ 71 23.7 90% ≥ 181 60.5 85% ≥ 233 77.7 80% ≥ 252 84.0 75% ≥ 275 91.7 70% ≥ 283 24.3 As shown here, the reflection accuracy is generally above 70%, while some data has a reflection accuracy of around 30%. This trend was often observed in data where the generation of the lower part of the object was unstable, as shown in Figure 15. From this, it can be seen that more training shapes need to be learned in order to reduce outliers.

[0093] As a specific example, the technology disclosed herein can be suitably applied to the design of Gigacast (large-scale aluminum die-cast integral molding) structures in the automotive sector. In this case, there is the advantage of being able to generate an optimal shape that achieves both strength and weight reduction while integrating multiple parts and taking into account manufacturing constraints (e.g., draft angles).

[0094] Furthermore, the technology disclosed herein can also be applied to the design of impact-absorbing members such as crash boxes (example of application to dynamic nonlinear problems). For example, by training the system with approximately 22,000 cases of dynamic nonlinear analysis data, a shape was generated by specifying the target impact absorption energy and maximum reaction force, resulting in a shape that satisfies the target performance with an accuracy of over 95%. Moreover, optimization considering manufacturing constraints is also possible in the design of thin plate structures.

[0095] [Each aspect of this disclosure] The aspects of this disclosure are summarized below. A three-dimensional model generation device in one aspect of this disclosure includes a learning unit that constructs an estimation model for estimating the shape of a three-dimensional object by performing machine learning using training data that includes structural mechanical parameters and three-dimensional object data.

[0096] The 3D model generation device may further include a generation unit that generates 3D objects using an estimated model.

[0097] The 3D model generation apparatus may further include a forming unit that forms a 3D object using 3D objects.

[0098] Structural mechanical parameters may include, at a minimum, strain energy.

[0099] The structural mechanical parameters may further include each component of the load vector, as well as the volume and height of the three-dimensional object.

[0100] The learning unit may perform machine learning using DeepSDF, and in addition to the structural mechanical parameters, the coordinates of each spatial sample point may also be input to the learning unit.

[0101] The 3D model generation device may further include a preprocessing unit that expands the dimension of the spatial sample points using positional encoding.

[0102] The learning unit may perform machine learning using latent vectors in addition to constant parameters of structural mechanics.

[0103] The learning unit may perform Lipschitz regularization during the training of the estimation model.

[0104] Another aspect of the three-dimensional object display system of this disclosure comprises a three-dimensional model generation device having the aforementioned learning unit and generation unit, and a display unit that images and displays the three-dimensional objects generated by the three-dimensional model generation device. Here, the display unit that images and displays the three-dimensional objects is used to surface model the three-dimensional objects generated by the generation unit using a method such as polygons, and to display them on the display unit. The display unit may be any suitable display device, such as a liquid crystal display, a head-mounted display, a tablet, or a smartphone.

[0105] A 3D object forming system in another aspect of the present disclosure comprises a 3D model generation device having the aforementioned learning unit and generation unit, and a forming unit that forms a 3D object using the 3D objects generated by the 3D model generation device. Here, the forming unit that forms the 3D object is a unit that forms a tangible 3D object based on the 3D objects generated by the generation unit, and examples of this include 3D printers and CAD (the same applies hereinafter).

[0106] A three-dimensional object forming apparatus in another aspect of the present disclosure comprises: a learning unit that constructs an estimation model for estimating a three-dimensional object by performing machine learning using training data including structural mechanical parameters and three-dimensional object data; a generation unit that generates a three-dimensional object using the estimation model; and a forming unit that forms a three-dimensional object using the three-dimensional object.

[0107] In another aspect of the present disclosure, a three-dimensional model generating apparatus comprises an autodecoder-type neural network that outputs a signed distance function (SDF), and includes: (i) positional encoding for spatial sample point coordinates; (ii) an input / output configuration for injecting the structural dynamic parameters into each layer; and (iii) an L1 loss using L2 regularization (multivariate normal prior with mean 0) and clamp(δ) for latent vectors.

[0108] In a 3D model generation apparatus of another aspect of the present disclosure, the generation unit calculates the confidence level of the estimated model estimated by the learning unit by predictive variance, ensemble, or MC dropout.

[0109] In another embodiment of the present disclosure, a three-dimensional model generation apparatus samples latent vectors from a multivariate normal distribution and generates a three-dimensional object based on these latent vectors and structural mechanical parameters.

[0110] In a 3D model generation apparatus of another aspect of this disclosure, the learning unit creates data from unobserved regions by extrapolation and applies it as training data.

[0111] In a 3D model generation apparatus of another aspect of this disclosure, the generation unit exports the generated 3D object to STL / OBJ and connects with an external system via CAE analysis or a cloud API.

[0112] A method for generating a three-dimensional model in another aspect of the present disclosure includes a learning step of constructing an estimation model for estimating the shape of a three-dimensional object by performing machine learning using training data including structural mechanical parameters and three-dimensional object data, and a generation step of generating a three-dimensional object using the estimation model.

[0113] The 3D model generation method may further include a generation step of generating a 3D object using an estimated model.

[0114] According to this embodiment, by specifying structural mechanical parameters, a three-dimensional object that satisfies those structural mechanical parameters can be generated.

[0115] A three-dimensional model generation program in one aspect of the present disclosure causes a computer to perform a learning step in which it constructs an estimation model for estimating the shape of a three-dimensional object by performing machine learning using training data that includes structural mechanical parameters and three-dimensional object data.

[0116] According to this embodiment, by specifying structural mechanical parameters, a program can be implemented in the software to generate a three-dimensional object that satisfies those structural mechanical parameters.

[0117] A recording medium in one aspect of the present disclosure records a 3D model generation program characterized by causing a computer to perform a learning step of constructing an estimation model for estimating the shape of a 3D object by performing machine learning using training data including structural mechanical parameters and 3D object data.

[0118] According to this embodiment, by specifying structural mechanical parameters, a program that generates a three-dimensional object satisfying those structural mechanical parameters can be recorded on a recording medium.

[0119] A 3D model generation apparatus in another aspect of the present disclosure includes a storage unit that stores a trained estimation model for estimating the shape of a 3D object constructed by performing machine learning using training data including structural mechanical parameters and 3D object data, and a generation unit that generates the shape of a 3D object by inputting structural mechanical parameters into the trained estimation model.

[0120] According to this embodiment, the shape of a three-dimensional object can be generated using a pre-trained estimation model.

[0121] The generation unit may calculate the confidence level of the estimation model using predictive variance, ensemble, or MC dropout.

[0122] The generation unit may sample latent vectors from a multivariate normal distribution and generate a three-dimensional object based on these latent vectors and structural dynamic parameters.

[0123] The generation unit may export the generated 3D object to STL / OBJ and link with an external system via CAE analysis or a cloud API.

[0124] A method for generating a three-dimensional model in another aspect of the present disclosure includes a generation step of generating the shape of a three-dimensional object from input structural mechanical parameters using a trained estimation model that estimates the shape of a three-dimensional object by performing machine learning on training data including structural mechanical parameters and three-dimensional object data.

[0125] According to this embodiment, the shape of a three-dimensional object can be generated using a pre-trained estimation model.

[0126] A 3D model generation program in another aspect of the present disclosure causes a computer to perform a generation step of generating the shape of a 3D object from input structural mechanical parameters, using a trained estimation model that estimates the shape of a 3D object by performing machine learning on training data including structural mechanical parameters and 3D object data.

[0127] According to this embodiment, a program that generates the shape of a 3D object using a trained estimation model can be implemented in software.

[0128] A recording medium in another aspect of the present disclosure records a 3D model generation program characterized by causing a computer to perform a generation step of generating the shape of a 3D object from input structural mechanical parameters, using a trained estimation model that estimates the shape of a 3D object by performing machine learning using training data including structural mechanical parameters and 3D object data.

[0129] According to this embodiment, a program that generates the shape of a 3D object using a trained estimation model can be recorded on a recording medium.

[0130] The large number of 3D objects created using the method disclosed herein can be used for quality control, risk assessment, manufacturing simulations, and more. They can also be widely applied to the analysis of internal structures that are not visible from the outside, such as the statistical generation of casting defects inside die castings.

[0131] The method disclosed herein embeds the relationship between structural mechanical parameters and three-dimensional shape into a low-dimensional latent space. This latent space is continuous and possesses excellent interpolation and extrapolation capabilities, making it possible to learn a wide range of design spaces and generate diverse shapes even from a relatively small amount of training data (a few-shot approach).

[0132] In conventional design methods, exploring high-dimensional design spaces and understanding the complex nonlinear relationships between performance and shape have been difficult due to the limitations of human cognitive abilities. The technology disclosed herein overcomes these limitations by enabling AI to perform multidimensional spatial exploration, making it possible to discover innovative structures that go beyond human preconceptions and rules of thumb.

[0133] The present invention has been described above based on embodiments. These embodiments are illustrative, and it will be understood by those skilled in the art that various modifications are possible in combinations of their components and processing processes, and that such modifications also fall within the scope of the present invention.

[0134] Although the embodiment mainly described an example using DeepSDF, the machine learning model used by the learning unit 11 is not limited to this. For example, point cloud-based, voxel-based, and mesh-based generation models, as well as diffusion models and transformer-based models, can also be applied by incorporating structural mechanical parameters as conditions.

[0135] The technology of this disclosure is further extensible. For example, the technology of this disclosure may include parameters that indicate material anisotropy as structural mechanical parameters. The technology of this disclosure may also be applied to multiphysics (coupled analysis) problems and the design of multimaterials (composite materials). Furthermore, in learning, the concept of Physics-Informed Neural Networks (PINNs) may be introduced, and learning in accordance with physical laws may be promoted by adding the residuals of the governing equations of the relevant physical phenomena (e.g., equations of motion, heat conduction equations, etc.) to the loss function.

[0136] The technology disclosed herein is not limited to a parameter-to-3D configuration. In another embodiment, a 3D-to-3D configuration may be used, which takes existing 3D object data (and modification instructions or target parameters) as input and outputs modified 3D object data. For example, an existing design proposal may be input and automatically modified to a shape that satisfies specific manufacturing or performance requirements.

[0137] Any combination of the embodiments and modifications described above is also useful as an embodiment of the present invention. The new embodiments resulting from these combinations possess the combined effects of each of the embodiments and modifications that are combined.

[0138] In understanding the technical concept abstracted from the embodiments and modifications, that technical concept should not be interpreted restrictively to the content of the embodiments and modifications. The embodiments and modifications described above are merely examples, and many design changes, such as changes, additions, and deletions of components, are possible. In the embodiments, the content in which such design changes are possible is emphasized with the notation "embodiment." However, design changes are also permitted in content without such notation.

[0139] The 3D model generation apparatus and 3D model generation method disclosed herein can be widely used in the design and manufacture of vehicles such as automobiles, aircraft, and ships, various industrial parts, buildings such as office buildings and houses, and structures such as bridges and dams.

[0140] 1...3D model generation device, 11...Learning unit, 12...Generation unit, S1...Learning step, S2...Generation step.

Claims

1. A 3D model generation device characterized by comprising a learning unit that constructs an estimation model for estimating the shape of a 3D object by performing machine learning using training data that includes structural mechanical parameters and 3D object data.

2. The three-dimensional model generation apparatus according to claim 1, further comprising a generation unit that generates a three-dimensional object using the estimation model.

3. The three-dimensional model generation apparatus according to claim 2, further comprising a display unit that converts the three-dimensional object into an image and displays it.

4. The three-dimensional model generation apparatus according to claim 2, further comprising a forming unit that forms a three-dimensional object using the three-dimensional object.

5. The three-dimensional model generation apparatus according to claim 1, characterized in that the structural mechanical parameters include at least strain energy.

6. The three-dimensional model generation apparatus according to claim 5, characterized in that the structural mechanical parameters further include each component of the load vector, and the volume and height of the three-dimensional object.

7. The three-dimensional model generation apparatus according to claim 6, characterized in that the learning unit performs machine learning using DeepSDF, and in addition to the structural mechanical parameters, the coordinates of each spatial sample point are further input to the learning unit.

8. The 3D model generation apparatus according to claim 7, further comprising a preprocessing unit that expands the dimension of the spatial sample points using positional encoding.

9. The three-dimensional model generation apparatus according to claim 7, characterized in that the learning unit performs machine learning using latent vectors in addition to structural mechanics constant parameters.

10. The three-dimensional model generation apparatus according to claim 1, characterized in that the learning unit performs Lipschitz regularization in the learning of the estimation model.

11. A three-dimensional object forming system comprising: a three-dimensional model generation apparatus according to claim 2; and a forming unit for forming a three-dimensional object using three-dimensional objects generated by the three-dimensional model generation apparatus.

12. A three-dimensional object forming apparatus comprising: a learning unit that constructs an estimation model for estimating a three-dimensional object by performing machine learning using training data including structural mechanical parameters and three-dimensional object data; a generation unit that generates a three-dimensional object using the estimation model; and a forming unit that forms a three-dimensional object using the three-dimensional object.

13. The three-dimensional model generation apparatus according to claim 1, wherein the estimation model is an autodecoder-type neural network that outputs a signed distance function (SDF), and includes (i) positional encoding for spatial sample point coordinates, (ii) an input / output configuration for injecting the structural dynamic parameters into each layer, and (iii) an L1 loss using L2 regularization (multivariate normal prior with mean 0) and clamp(δ) for latent vectors.

14. The three-dimensional model generation apparatus according to claim 2, characterized in that the generation unit calculates the reliability of the estimated model estimated by the learning unit by predictive variance, ensemble, or MC dropout.

15. The three-dimensional model generation apparatus according to claim 2, characterized in that the generation unit samples latent vectors from a multivariate normal distribution and generates a three-dimensional object based on the latent vectors and structural mechanical parameters.

16. The three-dimensional model generation apparatus according to claim 1, characterized in that the learning unit creates data from an unobserved region by extrapolation and applies it as training data.

17. The three-dimensional model generation apparatus according to claim 2, characterized in that the generation unit exports the generated three-dimensional object to STL / OBJ and cooperates with an external system via CAE analysis or cloud API.

18. A method for generating a three-dimensional model, characterized by including a learning step of constructing an estimation model for estimating the shape of a three-dimensional object by performing machine learning using training data that includes structural mechanical parameters and three-dimensional object data.

19. A method for generating a three-dimensional model according to claim 18, comprising a generation step of generating a three-dimensional object using the estimation model.

20. A 3D model generation program characterized by causing a computer to perform a learning step to construct an estimation model for estimating the shape of a 3D object by performing machine learning using training data that includes structural mechanical parameters and 3D object data.

21. A recording medium that stores a 3D model generation program, characterized by causing a computer to perform a learning step to construct an estimation model for estimating the shape of a 3D object by performing machine learning using training data that includes structural mechanical parameters and 3D object data.

22. A three-dimensional model generation device comprising: a storage unit that stores a trained estimation model for estimating the shape of a three-dimensional object constructed by performing machine learning using training data including structural mechanical parameters and three-dimensional object data; and a generation unit that generates the shape of a three-dimensional object by inputting structural mechanical parameters into the trained estimation model.

23. The three-dimensional model generation apparatus according to 22, characterized in that the generation unit calculates the reliability of the estimation model by predictive variance, ensemble, or MC dropout.

24. The three-dimensional model generation apparatus according to 22, characterized in that the generation unit samples latent vectors from a multivariate normal distribution and generates a three-dimensional object based on the latent vectors and structural mechanical parameters.

25. The three-dimensional model generation apparatus according to 22, characterized in that the generation unit exports the generated three-dimensional object to STL / OBJ and cooperates with an external system via CAE analysis or cloud API.

26. A method for generating a three-dimensional model, characterized by including a generation step of generating the shape of a three-dimensional object from input structural mechanical parameters, using a trained estimation model that estimates the shape of a three-dimensional object by performing machine learning using training data including structural mechanical parameters and three-dimensional object data.

27. A 3D model generation program characterized by causing a computer to perform a generation step of generating the shape of a 3D object from input structural mechanical parameters, using a trained estimation model that estimates the shape of a 3D object by performing machine learning using training data including structural mechanical parameters and 3D object data.

28. A recording medium that stores a 3D model generation program, characterized in that it causes a computer to execute a generation step of generating the shape of a 3D object from input structural mechanical parameters, using a trained estimation model that estimates the shape of a 3D object by performing machine learning using training data including structural mechanical parameters and 3D object data.

Images